Large Language Models (LLMs) have emerged as a new information channel. Meanwhile, one critical but under-explored question is: Is it possible to bypass the safety alignment and inject harmful information into LLMs stealthily? In this paper, we propose to reformulate knowledge editing as a new type of safety threat for LLMs, namely Editing Attack, and conduct a systematic investigation with a newly constructed dataset EditAttack. Specifically, we focus on two typical safety risks of Editing Attack including Misinformation Injection and Bias Injection. For the first risk, we find that editing attacks can inject both commonsense and long-tail misinformation into LLMs, and the effectiveness for the former one is particularly high. For the second risk, we discover that not only can biased sentences be injected into LLMs with high effectiveness, but also one single biased sentence injection can degrade the overall fairness. Then, we further illustrate the high stealthiness of editing attacks. Our discoveries demonstrate the emerging misuse risks of knowledge editing techniques on compromising the safety alignment of LLMs and the feasibility of disseminating misinformation or bias with LLMs as new channels.

Background: Traditional supervised learning (SL) assumes data points are independently and identically distributed (i.i.d.), which overlooks dependencies in real-world data. Reinforcement learning (RL), in contrast, models dependencies through state transitions. Objectives: This study aims to bridge SL and RL by reformulating SL problems as RL tasks, enabling the application of RL techniques to a wider range of SL scenarios. We aim to model SL data as interconnected, and develop novel temporal difference (TD) algorithms that can accommodate diverse data types. Our objectives are to (1) establish conditions where TD outperforms ordinary least squares (OLS), (2) provide convergence guarantees for the generalized TD algorithm, and (3) validate the approach empirically using synthetic and real-world datasets. Methods: We reformulate traditional SL as a RL problem by modeling data points as a Markov Reward Process (MRP). We then introduce a concept analogous to the inverse link function in generalized linear models, allowing our TD algorithm to handle various data types. Our analysis, grounded in variance estimation, identifies conditions where TD outperforms OLS. We establish a convergence guarantee by conceptualizing the TD update rule as a generalized Bellman operator. Empirical validation begins with synthetic data progressively matching theoretical assumptions to verify our analysis, followed by evaluations on real-world datasets to demonstrate practical utility. Results: Our theoretical analysis shows that TD can outperform OLS in estimation accuracy when data noise is correlated. Our approach generalizes across various loss functions and SL datasets. We prove that the Bellman operator in our TD framework is a contraction, ensuring convergence for both expected and stochastic TD updates. Empirically, TD outperforms SL baselines when data aligns with its assumptions, remains competitive across diverse datasets, and is robust to hyperparameter choices. Conclusions: This study demonstrates that SL can be reformulated as a problem of interconnected data modeled by an MRP, effectively solved using TD learning. Our generalized TD is theoretically sound, with convergence guarantees, and practically effective. It generalizes OLS, offering superior performance on correlated data. This work enables RL techniques to benefit SL tasks, offering a pathway for future advancements.

Deep research agents have attracted growing attention for their potential to orchestrate multi-stage research workflows, spanning literature synthesis, methodological design, and empirical verification. Despite these strides, evaluating their research capability faithfully is rather challenging due to the difficulty of collecting frontier research questions that genuinely capture researchers’ attention and intellectual curiosity. To address this gap, we introduce DeepResearch Arena, a benchmark grounded in academic seminars that capture rich expert discourse and interaction, better reflecting real-world research environments and reducing the risk of data leakage. To automatically construct DeepResearch Arena, we propose a Multi-Agent Hierarchical Task Generation (MAHTG) system that extracts research-worthy inspirations from seminar transcripts. The MAHTG system further translates research-worthy inspirations into high-quality research tasks, ensuring the traceability of research task formulation while filtering noise. With the MAHTG system, we curate DeepResearch Arena with over 10,000 high-quality research tasks from over 200 academic seminars, spanning 12 disciplines, such as literature, history, and science. Our extensive evaluation shows that DeepResearch Arena presents substantial challenges for current state-of-the-art agents, with clear performance gaps observed across different models.

Zelin Tan, Hejia Geng, Xiaohang Yu, Mulei Zhang, Guancheng Wan, Yifan Zhou, Qiang He, Xiangyuan Xue, Heng Zhou, Yutao Fan, Zhong-Zhi Li, Zaibin Zhang, Guibin Zhang, Chen Zhang, Zhenfei Yin, Philip Torr, Lei Bai. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Composed image retrieval aims to retrieve images based on a query that consists of a reference image and text describing desired modifications to that image. It has recently attracted attention for its ability to tailor image retrieval to user intentions by combining information-rich reference images with concise natural language instructions. Despite its current success, the robustness of composed image retrieval methods to either (1) common corruptions or (2) variations of the textual descriptions have never been systematically evaluated. In this paper, we perform the first robustness study of composed image retrieval, establishing three new benchmarks for a systematic evaluation of robustness to common corruption (in both the textual and visual domains) and robustness in text understanding. For analysis of natural image corruption, we introduce two new large-scale benchmark datasets, CIRR-C and FashionIQ-C, for the open domains and fashion domains respectively–both of which feature 75 visual corruptions and 35 textual corruptions. To facilitate robust evaluation of text understanding, we introduce a new diagnostic dataset CIRR-D by expanding the CIRR dataset with synthetic data, specifically probing text understanding across variations in: numerical, attribute, object removal, and background. We introduce BenchCIR, a testbed for evaluating composed image retrieval model robustness with standardized evaluation protocols. Through benchmarking ten published models in the testbed, we reveal insights into how the composition of visual and textual modalities affects model robustness. The code is in https://suntongtongtong.github.io/BenchCIR/

We present DreamBeast, a novel method based on score distillation sampling (SDS) for generating fantastical 3D animal assets composed of distinct parts. Existing SDS methods often struggle with this generation task due to a limited understanding of part-level semantics in text-to-image diffusion models. While recent diffusion models, such as Stable Diffusion 3, demonstrate a better part-level understanding, they are prohibitively slow and exhibit other common problems associated with single-view diffusion models. DreamBeast overcomes this limitation through a novel part-aware knowledge transfer mechanism. For each generated asset, we efficiently extract part-level knowledge from the Stable Diffusion 3 model into a 3D Part-Affinity implicit representation. This enables us to instantly generate Part-Affinity maps from arbitrary camera views, which we then use to modulate the guidance of a multi-view diffusion model during SDS to create 3D assets of fantastical animals. DreamBeast significantly enhances the quality of generated 3D creatures with user-specified part compositions while reducing computational overhead, as demonstrated by extensive quantitative and qualitative evaluations.

Tong Liu, Zhixin Lai, Jiawen Wang, Gengyuan Zhang, Shuo Chen, Philip Torr, Vera Demberg, Volker Tresp, Jindong Gu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Large Audio-Language Models (LALMs), such as GPT-4o, have recently unlocked audio dialogue capabilities, enabling direct spoken exchanges with humans.The potential of LALMs broadens their applicability across a wide range of practical scenarios supported by audio dialogues.However, given these advancements, a comprehensive benchmark to evaluate the performance of LALMs in the open-ended audio dialogue understanding remains absent currently.To address this gap, we propose an Audio Dialogue Understanding Benchmark (ADU-Bench), which consists of 4 benchmark datasets.They assess the open-ended audio dialogue ability for LALMs in 3 general scenarios, 12 skills, 9 multilingual languages, and 4 categories of ambiguity handling.Notably, we firstly propose the evaluation of ambiguity handling in audio dialogues that expresses different intentions beyond the same literal meaning of sentences, e.g., "Really!?" with different intonations.In summary, ADU-Bench includes over 20,000 open-ended audio dialogues for the assessment of LALMs.Through extensive experiments on 16 LALMs, our analysis reveals that existing LALMs struggle with mathematical symbols and formulas, understanding human behavior such as roleplay, comprehending multiple languages, and handling audio dialogue ambiguities from different phonetic elements, such as intonations, pause positions, and homophones.The benchmark is available at https://adu-bench.github.io/.

Haoyang Su, Renqi Chen, Shixiang Tang, Zhenfei Yin, Xinzhe Zheng, Jinzhe Li, Biqing Qi, Qi Wu, Hui Li, Wanli Ouyang, Philip Torr, Bowen Zhou, Nanqing Dong. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The challenge in open-world object detection, similarly to few- and zero-shot learning, is to generalize beyond the class distribution of the training data. In this paper, we propose a general class-agnostic objectness measure to limit bias toward labeled samples. One issue in open-world detection is that previously unseen objects are often misclassified as known categories or filtered as background by classifiers. To prevent this, we explicitly model the joint distribution of objectness and category labels using variational approximation. However, without sufficient labeled data, minimizing the KL divergence between the estimated posterior and a static normal prior fails to converge. Our theoretical analysis identifies the root cause of this failure and motivates adopting a Gaussian prior with variance dynamically adapted to the estimated posterior as a surrogate. To further reduce misclassification, we introduce an energy-based margin loss that encourages unknown objects to move toward high-density regions of the distribution, thus reducing the uncertainty of unknown detections. Our Open-World OBJectness modeling (OWOBJ) boosts novel object detection, especially in low-data regimes. OWOBJ is a flexible plugin that outperforms baselines in Open-World, Few-Shot, and zero-shot Open-Vocabulary Object Detection.

Many practical vision-language applications require models that understand negation, e.g., when using natural language to retrieve images which contain certain objects but not others. Despite advancements in vision-language models (VLMs) through large-scale training, their ability to comprehend negation remains underexplored. This study addresses the question: how well do current VLMs understand negation? We introduce NegBench, a new benchmark designed to evaluate negation understanding across 18 task variations and 79k examples spanning image, video, and medical datasets. The benchmark consists of two core tasks designed to evaluate negation understanding in diverse multimodal settings: Retrieval with Negation and Multiple Choice Questions with Negated Captions. Our evaluation reveals that modern VLMs struggle significantly with negation, often performing at chance level. To address these shortcomings, we explore a data-centric approach wherein we finetune CLIP models on large-scale synthetic datasets containing millions of negated captions. We show that this approach can result in a 10% increase in recall on negated queries and a 28% boost in accuracy on multiple-choice questions with negated captions.

One-Shot Federated Learning (OSFL), a special decentralized machine learning paradigm, has recently gained significant attention. OSFL requires only a single round of client data or model upload, which reduces communication costs and mitigates privacy threats compared to traditional FL. Despite these promising prospects, existing methods face challenges due to client data heterogeneity and limited data quantity when applied to real-world OSFL systems. Recently, Latent Diffusion Models (LDM) have shown remarkable advancements in synthesizing high-quality images through pretraining on large-scale datasets, thereby presenting a potential solution to overcome these issues. However, directly applying pretrained LDM to heterogeneous OSFL results in significant distribution shifts in synthetic data, leading to performance degradation in classification models trained on such data. This issue is particularly pronounced in rare domains, such as medical imaging, which are underrepresented in LDM’s pretraining data. To address this challenge, we propose Federated Bi-Level Personalization (FedBiP), which personalizes the pretrained LDM at both instance-level and concept-level. Hereby, FedBiP synthesizes images following the client’s local data distribution without compromising the privacy regulations. FedBiP is also the first approach to simultaneously address feature space heterogeneity and client data scarcity in OSFL. Our method is validated through extensive experiments on three OSFL benchmarks with feature space heterogeneity, as well as on challenging medical and satellite image datasets with label heterogeneity. The results demonstrate the effectiveness of FedBiP, which substantially outperforms other OSFL methods. Our code is available at https://github.com/HaokunChen245/FedBiP.

We introduce Olympus, a new approach that transforms Multimodal Large Language Models (MLLMs) into a unified framework capable of handling a wide array of computer vision tasks. Utilizing a controller MLLM, Olympus delegates over 20 specialized tasks across images, videos, and 3D objects to dedicated modules. This instruction-based routing enables complex workflows through chained actions without the need for training heavy generative models. Olympus easily integrates with existing MLLMs, expanding their capabilities with comparable performance. Experimental results demonstrate that Olympus achieves an average routing accuracy of 94.75% across 20 tasks and precision of 91.82% in chained action scenarios, showcasing its effectiveness as a universal task router that can solve a diverse range of computer vision tasks.

We propose DiTFlow, a method for transferring the motion of a reference video to a newly synthesized one, designed specifically for Diffusion Transformers (DiT). We first process the reference video with a pre-trained DiT to analyze cross-frame attention maps and extract a patch-wise motion signal called the Attention Motion Flow (AMF). We guide the latent denoising process in an optimization-based, training-free, manner by optimizing latents with our AMF loss to generate videos reproducing the motion of the reference one. We also apply our optimization strategy to transformer positional embeddings, granting us a boost in zero-shot motion transfer capabilities. We evaluate DiTFlow against recently published methods, outperforming all across multiple metrics and human evaluation.

Localizing events in videos based on semantic queries is a pivotal task in video understanding research and user-oriented applications like video search. Yet, current research predominantly relies on natural language queries (NLQs), overlooking the potential of using multimodal queries (MQs) that incorporate images to flexibly represent semantic queries, particularly when it is difficult to express non-verbal or unfamiliar concepts in words. To bridge this gap, we introduce ICQ, a new benchmark designed for localizing events in videos with MQs, alongside an evaluation dataset ICQ-Highlight. To adapt and reevaluate existing video localization models for this new task, we propose 3 Multimodal Query Adaptation methods and a novel Surrogate Fine-Tuning strategy, serving as strong baseline methods. ICQ systematically benchmarks 12 state-of-the-art backbone models, spanning from specialized video localization models to Video Large Language Models. Our extensive experiments highlight the high potential of using MQs in real-world applications. We believe this is a first step toward video event localization with MQs1.

Large language models (LLMs) frequently generate confident yet inaccurate responses, introducing significant risks for deployment in safety-critical domains.We present a novel, test-time approach to detecting model hallucination through systematic analysis of information flow across model layers.We target cases when LLMs process inputs with ambiguous or insufficient context.Our investigation reveals that hallucination manifests as usable information deficiencies in inter-layer transmissions.While existing approaches primarily focus on final-layer output analysis, we demonstrate that tracking cross-layer information dynamics (LI) provides robust indicators of model reliability, accounting for both information gain and loss during computation.LI integrates easily with pretrained LLMs without requiring additional training or architectural modifications.

Individual Large Language Models (LLMs) have demonstrated significant capabilities across various domains, such as healthcare and law.Recent studies also show that coordinated multi-agent systems exhibit enhanced decisionmaking and reasoning abilities through collaboration.However, due to the vulnerabilities of individual LLMs and the difficulty of accessing all agents in a multi-agent system, a key question arises: If attackers only know one agent, could they still generate adversarial samples capable of misleading the collective decision?To explore this question, we formulate it as a game with incomplete information, where attackers know only one target agent and lack knowledge of the other agents in the system.With this formulation, we propose M-Spoiler, a framework that simulates agent interactions within a multi-agent system to generate adversarial samples.These samples are then used to manipulate the target agent in the target system, misleading the system's collaborative decisionmaking process.More specifically, M-Spoiler introduces a stubborn agent that actively aids in optimizing adversarial samples by simulating potential stubborn responses from agents in the target system.This enhances the effectiveness of the generated adversarial samples in misleading the system.Through extensive experiments across various tasks, our findings confirm the risks posed by the knowledge of an individual agent in multi-agent systems and demonstrate the effectiveness of our framework.We also explore several defense mechanisms, showing that our proposed attack framework remains more potent than baselines, underscoring the need for further research into defensive strategies.Our source code is available at here.

Large language models (LLMs) excel in diverse applications but face dual challenges: generating harmful content under jailbreak attacks and over-refusing benign queries due to rigid safety mechanisms.These issues severely affect the application of LLMs.Existing approaches can be divided into three types: contrastive decoding, activation manipulation, and prompting strategies.However, all these approaches face challenges like inefficiency, fragility, or architectural constraints, ultimately failing to strike a balance between safety and usability.These problems are more obvious in multimodal large language models (MLLMs), especially in terms of heightened over-refusal in cross-modal tasks and new security risks arising from expanded attack surfaces.We propose Magic Image 1 , an optimization-driven visual prompt framework that enhances security and reduces over-refusal at the same time.The Magic Image is optimized using gradients derived from harmful/benign training samples.Using the magic image can modify the model's original safety alignment, maintaining robust safety while reducing unnecessary denials.Experiments demonstrate its effectiveness in preserving model performance and improving safety-responsiveness balance across datasets, including unseen data, offering a practical solution for reliable MLLM deployment.

Match-cuts are powerful cinematic tools that create seamless transitions between scenes, delivering strong visual and metaphorical connections. However, crafting match-cuts is a challenging, resource-intensive process requiring deliberate artistic planning. In MatchDiffusion, we present the first training-free method for match-cut generation using text-to-video diffusion models. MatchDiffusion leverages a key property of diffusion models: early denoising steps define the scene's broad structure, while later steps add details. Guided by this insight, MatchDiffusion employs "Joint Diffusion" to initialize generation for two prompts from shared noise, aligning structure and motion. It then applies "Disjoint Diffusion", allowing the videos to diverge and introduce unique details. This approach produces visually coherent videos suited for match-cuts. User studies and metrics demonstrate MatchDiffusion's effectiveness and potential to democratize match-cut creation.

We present Stable Virtual Camera (Seva), a generalist diffusion model that creates novel views of a scene, given any number of input views and target cameras. Existing works struggle to generate either large viewpoint changes or temporally smooth samples, while relying on specific task configurations. Our approach overcomes these limitations through simple model design, optimized training recipe, and flexible sampling strategy that generalize across view synthesis tasks at test time. As a result, our samples maintain high consistency without requiring additional 3D representation-based distillation, thus streamlining view synthesis in the wild. Furthermore, we show that our method can generate high-quality videos lasting up to half a minute with seamless loop closure. Extensive benchmarking demonstrates that Seva outperforms existing methods across different datasets and settings. Project page with code and model: https://stable-virtual-camera.github.io/.

Fine-tuning large language models (LLMs) on human preferences, typically through reinforcement learning from human feedback (RLHF), has proven successful in enhancing their capabilities. However, ensuring the safety of LLMs during fine-tuning remains a critical concern, and mitigating the potential conflicts in safety and helpfulness is costly in RLHF. To address this issue, we propose a supervised learning framework called Bi-Factorial Preference Optimization (BFPO), which re-parameterizes a joint RLHF objective of both safety and helpfulness into a single supervised learning objective. In the supervised optimization, a labeling function is used to capture global preferences ranking to balance both safety and helpfulness. To evaluate BFPO, we develop a benchmark including comprehensive discriminative and generative tasks for helpfulness and harmlessness. The results indicate that our method significantly outperforms existing approaches in both safety and helpfulness. Moreover, BFPO eliminates the need for human prompting and annotation in LLM fine-tuning while achieving the same level of safety as methods that heavily rely on human labor, with less than 10\% of the computational resources. The training recipes and models will be released.

Recent research shows that fine-tuning on benign instruction-following data can inadvertently undo the safety alignment process and increase a model's propensity to comply with harmful queries. While instruction-following fine-tuning is important, task-specific fine-tuning-where models are trained on datasets with clear ground truth answers (e.g., multiple choice questions)-can enhance model performance on specialized downstream tasks. Understanding and mitigating safety risks in the task-specific setting remains distinct from the instruction-following context due to structural differences in the data. Our work demonstrates how malicious actors can subtly manipulate the structure of almost *any* task-specific dataset to foster significantly more dangerous model behaviors, while maintaining an appearance of innocuity and reasonable downstream task performance. To address this issue, we propose a novel mitigation strategy that mixes in safety data which *mimics* the task format and prompting style of the user data, showing this is significantly more effective and efficient than existing baselines at re-establishing safety alignment while maintaining similar task performance.

Large language models (LLMs) are often deployed to do constrained tasks, with narrow domains. For example, customer support bots can be built on top of LLMs, relying on their broad language understanding and capabilities to enhance performance. However, these LLMs are adversarially susceptible, potentially generating outputs outside the intended domain. To formalize, assess and mitigate this risk, we introduce domain certification; a guarantee that accurately characterizes the out-of-domain behavior of language models. We then propose a simple yet effective approach dubbed VALID that provides adversarial bounds as a certificate. Finally, we evaluate our method across a diverse set of datasets, demonstrating that it yields meaningful certificates.

Since neural classifiers are known to be sensitive to adversarial perturbations that alter their accuracy, certification methods have been developed to provide provable guarantees on the insensitivity of their predictions to such perturbations. On the other hand, in safety-critical applications, the frequentist interpretation of the confidence of a classifier (also known as model calibration) can be of utmost importance. This property can be measured via the Brier score or the expected calibration error. We show that attacks can significantly harm calibration, and thus propose certified calibration providing worst-case bounds on calibration under adversarial perturbations. Specifically, we produce analytic bounds for the Brier score and approximate bounds via the solution of a mixed-integer program on the expected calibration error. Finally, we propose novel calibration attacks and demonstrate how they can improve model calibration through adversarial calibration training. The code will be publicly released upon acceptance.

Vision-Language Models (VLMs) are powerful tools for processing and understanding text and images. We study the processing of visual tokens in the language model component of LLaVA, a prominent VLM. Our approach focuses on analyzing the localization of object information, the evolution of visual token representations across layers, and the mechanism of integrating visual information for predictions. Through ablation studies, we demonstrated that object identification accuracy drops by over 70\% when object-specific tokens are removed. We observed that visual token representations become increasingly interpretable in the vocabulary space across layers, suggesting an alignment with textual tokens corresponding to image content. Finally, we found that the model extracts object information from these refined representations at the last token position for prediction, mirroring the process in text-only language models for factual association tasks. These findings provide crucial insights into how VLMs process and integrate visual information, bridging the gap between our understanding of language and vision models, and paving the way for more interpretable and controllable multimodal systems.

Preference learning is a central component for aligning current LLMs, but this process can be vulnerable to data poisoning attacks. To address this concern, we introduce PoisonBench, a benchmark for evaluating large language models' susceptibility to data poisoning during preference learning. Data poisoning attacks can manipulate large language model responses to include hidden malicious content or biases, potentially causing the model to generate harmful or unintended outputs while appearing to function normally. We deploy two distinct attack types across eight realistic scenarios, assessing 22 widely-used models. Our findings reveal concerning trends: (1) Scaling up parameter size does not always enhance resilience against poisoning attacks and the influence on model resilience varies among different model suites. (2) There exists a log-linear relationship between the effects of the attack and the data poison ratio; (3) The effect of data poisoning can generalize to extrapolated triggers that are not included in the poisoned data. These results expose weaknesses in current preference learning techniques, highlighting the urgent need for more robust defenses against malicious models and data manipulation.

Generating high-quality 3D content from text, single images, or sparse view images remains a challenging task with broad applications. Existing methods typically employ multi-view diffusion models to synthesize multi-view images, followed by a feed-forward process for 3D reconstruction. However, these approaches are often constrained by a small and fixed number of input views, limiting their ability to capture diverse viewpoints and, even worse, leading to suboptimal generation results if the synthesized views are of poor quality. To address these limitations, we propose Flex3D, a novel two-stage framework capable of leveraging an arbitrary number of high-quality input views. The first stage consists of a candidate view generation and curation pipeline. In the second stage, the curated views are fed into a Flexible Reconstruction Model (FlexRM), built upon a transformer architecture that can effectively process an arbitrary number of inputs. Through extensive exploration of design and training strategies, we optimize FlexRM to achieve superior performance in both reconstruction and generation tasks. Our results demonstrate that Flex3D achieves state-of-the-art performance, with a user study winning rate of over 92% in 3D generation tasks when compared to several of the latest feed-forward 3D generative models.

Despite the success of Instruction Tuning (IT) in training large language models (LLMs), such models often leverage spurious or biased features learnt from their training data and can become misaligned, leading to undesired behaviours. While existing techniques can steer model behaviour at inference-time, they are often post-hoc and do not embed steering as an intrinsic model feature. In this work, we introduce Focus Instruction Tuning (FIT), which trains LLMs to condition their responses by focusing on specific features whilst ignoring others, leading to different behaviours based on what features are specified. Across diverse benchmarks, we demonstrate that FIT: (i) successfully steers behaviour at inference time; (ii) increases robustness by amplifying core task signals and down-weighting spurious cues; (iii) mitigates social bias by suppressing demographic attributes; and (iv) generalises under distribution shifts and to previously unseen focus features. FIT therefore offers a lightweight, intrinsic mechanism for building more robust, fair, and easily controllable LLMs.

Neurons in large language models often exhibit \emph{polysemanticity}, simultaneously encoding multiple unrelated concepts and obscuring interpretability. Instead of relying on post-hoc methods, we present \textbf{MoE-X}, a mixture-of-experts (MoE) language model designed to be \emph{intrinsically} interpretable. Our approach is motivated by the observation that, in language models, wider networks with sparse activations are more likely to capture interpretable factors. however, directly training such large sparse networks is computationally prohibitive. MoE architectures offer a scalable alternative by activating only a subset of experts for any given input, inherently aligning with interpretability objectives. In MoE-X, we establish this connection by rewriting the MoE layer as an equivalent sparse, large MLP. This approach enables efficient scaling of the hidden size while maintaining sparsity. To further enhance interpretability, we enforce sparse activation within each expert and redesign the routing mechanism to prioritize experts with the highest activation sparsity. These designs ensure that only the most salient features are routed and processed by the experts. We evaluate MoE-X on chess and natural language tasks, showing that it achieves performance comparable to dense models while significantly improving interpretability. MoE-X achieves a perplexity better than GPT-2, with interpretability surpassing even sparse autoencoder (SAE)-based approaches.

Recent advances in generative models have demonstrated remarkable capabilities in producing high-quality images, but their reliance on large-scale unlabeled data has raised significant safety and copyright concerns. Efforts to address these issues by erasing unwanted concepts have shown promise. However, many existing erasure methods involve excessive modifications that compromise the overall utility of the model. In this work, we address these issues by formulating a novel minimalist concept erasure objective based *only* on the distributional distance of final generation outputs. Building on our formulation, we derive a tractable loss for differentiable optimization that leverages backpropagation through all generation steps in an end-to-end manner. We also conduct extensive analysis to show theoretical connections with other models and methods. To improve the robustness of the erasure, we incorporate neuron masking as an alternative to model fine-tuning. Empirical evaluations on state-of-the-art flow-matching models demonstrate that our method robustly erases concepts without degrading overall model performance, paving the way for safer and more responsible generative models.

Recent advances in operating system (OS) agents have enabled vision-language models (VLMs) to directly control a user’s computer. Unlike conventional VLMs that passively output text, OS agents autonomously perform computer-based tasks in response to a single user prompt. OS agents do so by capturing, parsing, and analysing screenshots and executing low-level actions via application programming interfaces (APIs), such as mouse clicks and keyboard inputs. This direct interaction with the OS significantly raises the stakes, as failures or manipulations can have immediate and tangible consequences. In this work, we uncover a novel attack vector against these OS agents: Malicious Image Patches (MIPs), adversarially perturbed screen regions that, when captured by an OS agent, induce it to perform harmful actions by exploiting specific APIs. For instance, a MIP can be embedded in a desktop wallpaper or shared on social media to cause an OS agent to exfiltrate sensitive user data. We show that MIPs generalise across user prompts and screen configurations, and that they can hijack multiple OS agents even during the execution of benign instructions. These findings expose critical security vulnerabilities in OS agents that have to be carefully addressed before their widespread deployment.

Evaluating large language models (LLMs) is crucial for both assessing their capabilities and identifying safety or robustness issues prior to deployment. Reliably measuring abstract and complex phenomena such as safety' androbustness' requires strong construct validity, that is, having measures that represent what matters to the phenomenon. With a team of 29 expert reviewers, we conduct a systematic review of 445 LLM benchmarks from leading conferences in natural language processing and machine learning. Across the reviewed articles, we find patterns related to the measured phenomena, tasks, and scoring metrics which undermine the validity of the resulting claims. To address these shortcomings, we provide eight key recommendations and detailed actionable guidance to researchers and practitioners in developing LLM benchmarks.

Large Language Models (LLMs) demonstrate impressive zero-shot performance across a wide range of natural language processing tasks. Integrating various modality encoders further expands their capabilities, giving rise to Multimodal Large Language Models (MLLMs) that process not only text but also visual and auditory modality inputs. However, these advanced capabilities may also pose significant safety problems, as models can be exploited to generate harmful or inappropriate content through jailbreak attack. While prior work has extensively explored how manipulating textual or visual modality inputs can circumvent safeguards in LLMs and MLLMs, the vulnerability of audio-specific Jailbreak on Large Audio-Language Models (LALMs) remains largely underexplored. To address this gap, we introduce \textbf{Jailbreak-AudioBench}, which consists of the Toolbox, curated Dataset, and comprehensive Benchmark. The Toolbox supports not only text-to-audio conversion but also various editing techniques for injecting audio hidden semantics. The curated Dataset provides diverse explicit and implicit jailbreak audio examples in both original and edited forms. Utilizing this dataset, we evaluate multiple state-of-the-art LALMs and establish the most comprehensive Jailbreak benchmark to date for audio modality. Finally, Jailbreak-AudioBench establishes a foundation for advancing future research on LALMs safety alignment by enabling the in-depth exposure of more powerful jailbreak threats, such as query-based audio editing, and by facilitating the development of effective defense mechanisms.

Changing facial expressions, gestures, or background details may dramatically alter the meaning conveyed by an image. Notably, recent advances in diffusion models greatly improve the quality of image manipulation while also opening the door to misuse. Identifying changes made to authentic images, thus, becomes an important task, constantly challenged by new diffusion-based editing tools. To this end, we propose a novel approach for ReliAble iDentification of inpainted AReas (RADAR). RADAR builds on existing foundation models and combines features from different image modalities. It also incorporates an auxiliary contrastive loss that helps to isolate manipulated image patches. We demonstrate these techniques to significantly improve both the accuracy of our method and its generalisation to a large number of diffusion models. To support realistic evaluation, we further introduce BBC-PAIR, a new comprehensive benchmark, with images tampered by 28 diffusion models. Our experiments show that RADAR achieves excellent results, outperforming the state-of-the-art in detecting and localising image edits made by both seen and unseen diffusion models. Further information about our code, data and models, including separate licensing terms, will be publicly available at https://alex-costanzino.github.io/radar/.

Coordinating multiple embodied agents in dynamic environments remains a core challenge in artificial intelligence, requiring both perception-driven reasoning and scalable cooperation strategies. While recent works have leveraged large language models (LLMs) for multi-agent planning, a few have begun to explore vision-language models (VLMs) for visual reasoning. However, these VLM-based approaches remain limited in their support for diverse embodiment types. In this work, we introduce VIKI-Bench, the first hierarchical benchmark tailored for embodied multi-agent cooperation, featuring three structured levels: agent activation, task planning, and trajectory perception. VIKI-Bench includes diverse robot embodiments, multi-view visual observations, and structured supervision signals to evaluate reasoning grounded in visual inputs. To demonstrate the utility of VIKI-Bench, we propose VIKI-R, a two-stage framework that fine-tunes a pretrained vision-language model (VLM) using Chain-of-Thought annotated demonstrations, followed by reinforcement learning under multi-level reward signals. Our extensive experiments show that VIKI-R significantly outperforms baselines method across all task levels. Furthermore, we show that reinforcement learning enables the emergence of compositional cooperation patterns among heterogeneous agents. Together, VIKI-Bench and VIKI-R offer a unified testbed and method for advancing multi-agent, visual-driven cooperation in embodied AI systems.

Recent interest in Multi-Agent Systems of Large Language Models (MAS LLMs) has led to an increase in frameworks leveraging multiple LLMs to tackle complex tasks. However, much of this literature appropriates the terminology of MAS without engaging with its foundational principles. In this position paper, we highlight critical discrepancies between MAS theory and current MAS LLMs implementations, focusing on four key areas: the social aspect of agency, environment design, coordination and communication protocols, and measuring emergent behaviours. Our position is that many MAS LLMs lack multi-agent characteristics such as autonomy, social interaction, and structured environments, and often rely on oversimplified, LLM-centric architectures. The field may slow down and lose traction by revisiting problems the MAS literature has already addressed. Therefore, we systematically analyse this issue and outline associated research opportunities; we advocate for better integrating established MAS concepts and more precise terminology to avoid mischaracterisation and missed opportunities.

Academic poster generation is a crucial yet challenging task in scientific communication, requiring the compression of long-context interleaved documents into a single, visually coherent page. To address this challenge, we introduce Paper2Poster, the first benchmark and metric suite for poster generation, which pairs recent conference papers with author-designed posters and evaluates outputs on (i) Visual Quality—semantic alignment with human posters, (ii) Textual Coherence—language fluency, (iii) Holistic Assessment—six fine-grained aesthetic and informational criteria scored by a VLM-as-judge, and notably (iv) PaperQuiz—the poster’s ability to convey core paper content as measured by VLMs answering generated quizzes. Building on this benchmark, we propose PosterAgent, a top‐down, visual‐in‐the‐loop multi‐agent pipeline: the (a) Parser distills the paper into a structured asset library; the (b) Planner aligns text–visual pairs into a binary‐tree layout that preserves reading order and spatial balance; and the (c) Painter–Commenter loop refines each panel by executing rendering code and using VLM feedback to eliminate overflow and ensure alignment.In our comprehensive evaluation, we find that GPT‐4o outputs—though visually appealing at first glance—often exhibit noisy text and poor PaperQuiz scores; We find that reader engagement is the primary aesthetic bottleneck, as human‐designed posters rely largely on visual semantics to convey meaning.Our fully open‐source Paper2Poster pipeline outperforms GPT‐4o–based systems across nearly all metrics while consuming 87 \% fewer tokens. These findings chart clear directions for the next generation of fully automated poster‐generation models.

Watermarking, the practice of embedding imperceptible information into media such as images, videos, audio, and text, is essential for intellectual property protection, content provenance and attribution. The growing complexity of digital ecosystems necessitates watermarks for different uses to be embedded in the same media. However, to detect and decode all watermarks, they need to coexist well with one another. We perform the first study of coexistence of deep image watermarking methods and, contrary to intuition, we find that various open-source watermarks can coexist with only minor impacts on image quality and decoding robustness. The coexistence of watermarks also opens the avenue for ensembling watermarking methods. We show how ensembling can increase the overall message capacity and enable new trade-offs between capacity, accuracy, robustness and image quality, without needing to retrain the base models.

Training vision language models (VLMs) aims to align visual representations from a vision encoder with the textual representations of a pretrained large language model (LLM). However, many VLMs exhibit reduced factual recall performance compared to their LLM backbones, raising the question of how effective multimodal fine-tuning is at extending existing mechanisms within the LLM to visual inputs. We argue that factual recall based on visual inputs requires VLMs to solve a two-hop problem: (1) forming entity representations from visual inputs, and (2) recalling associated factual knowledge based on these entity representations. By benchmarking 14 VLMs with various architectures (LLaVA, Native, Cross-Attention), sizes (7B-124B parameters), and training setups on factual recall tasks against their original LLM backbone models, we find that 11 of 14 models exhibit factual recall degradation. We select three models exhibiting high- and two models with low performance degradation, and use attribution patching, activation patching, and probing to show that degraded VLMs struggle to use the existing factual recall circuit of their LLM backbone, because they resolve the first hop too late in the computation. In contrast, high-performing VLMs resolve entity representations early enough to reuse the existing factual recall mechanism. Finally, we demonstrate two methods to recover performance: patching entity representations from the LLM backbone into the VLM, and prompting with chain-of-thought reasoning. Our results highlight that the speed of early entity resolution critically determines how effective VLMs are in using preexisting LLM mechanisms. More broadly, our work illustrates how mechanistic analysis can explain and unveil systematic failures in multimodal alignment.

Generating high-resolution 3D shapes using volumetric representations such as Signed Distance Functions (SDFs) presents substantial computational and memory challenges. We introduce Direct3D-S2, a scalable 3D generation framework based on sparse volumes that achieves superior output quality with dramatically reduced training costs. Our key innovation is the Spatial Sparse Attention (SSA) mechanism, which greatly enhances the efficiency of Diffusion Transformer (DiT) computations on sparse volumetric data. SSA allows the model to effectively process large token sets within sparse volumes, significantly reducing computational overhead and achieving a 3.9$\times$ speedup in the forward pass and a 9.6$\times$ speedup in the backward pass. Our framework also includes a variational autoencoder (VAE) that maintains a consistent sparse volumetric format across input, latent, and output stages. Compared to previous methods with heterogeneous representations in 3D VAE, this unified design significantly improves training efficiency and stability. Our model is trained on public datasets, and experiments demonstrate that Direct3D-S2 not only surpasses state-of-the-art methods in generation quality and efficiency, but also enables training at 1024³ resolution using only 8 GPUs—a task typically requiring at least 32 GPUs for volumetric representations at $256^3$ resolution, thus making gigascale 3D generation both practical and accessible. Project page: https://www.neural4d.com/research-page/direct3d-s2.

Retrieval-Augmented Generation (RAG) based on knowledge graphs (KGs) enhances large language models (LLMs) by providing structured and interpretable external knowledge. However, existing KG-based RAG methods struggle to retrieve accurate and diverse information from text-rich KGs for complex real-world queries. Process Reward Models (PRMs) offer a way to align the retrieval process of KG-based RAG with query-specific knowledge requirements, but they heavily rely on process-level supervision signals that are expensive and hard to obtain on KGs. To address this challenge, we propose GraphFlow, a framework that efficiently retrieves accurate and diverse knowledge required for real-world queries from text-rich KGs. GraphFlow employs a transition-based flow matching objective to jointly optimize a retrieval policy and a flow estimator. The flow estimator factorizes the reward of the retrieval outcome into the intermediate retrieval states. Such reward factorization guides the retrieval policy to retrieve candidates from KGs in proportion to their reward. This allows GraphFlow to explore high-quality regions of KGs that yield diverse and relevant results. We evaluate GraphFlow on the STaRK benchmark, which includes real-world queries from multiple domains over text-rich KGs. GraphFlow outperforms strong KG-RAG baselines, including GPT-4o, by 10\% on average in hit rate and recall. It also shows strong generalization to unseen KGs, demonstrating its effectiveness and robustness.

Large Language Models (LLMs) with in-context learning (ICL) ability can quickly adapt to a specific context given a few demonstrations (demos). Recently, Multimodal Large Language Models (MLLMs) built upon LLMs have also shown multimodal ICL ability, i.e., responding to queries given a few multimodal demos, including images, queries, and answers. While ICL has been extensively studied on LLMs, its research on MLLMs remains limited. One essential question is whether these MLLMs can truly conduct multimodal ICL, or if only the textual modality is necessary. We investigate this question by examining two primary factors that influence ICL: 1) Demo content, i.e., understanding the influences of demo content in different modalities. 2) Demo selection strategy, i.e., how to select better multimodal demos for improved performance. Experiments revealed that multimodal ICL is predominantly driven by the textual content whereas the visual information in the demos has little influence. Interestingly, visual content is still necessary and useful for selecting demos to increase performance. Motivated by our analysis, we propose a simple yet effective approach, termed Mixed Modality In-Context Example Selection (MMICES), which considers both visual and language modalities when selecting demos. Extensive experiments are conducted to support our findings and verify the improvement brought by our method.

Background: Traditional supervised learning (SL) assumes data points are independently and identically distributed (i.i.d.), which overlooks dependencies in real-world data. Reinforcement learning (RL), in contrast, models dependencies through state transitions. Objectives: This study aims to bridge SL and RL by reformulating SL problems as RL tasks, enabling the application of RL techniques to a wider range of SL scenarios. We aim to model SL data as interconnected, and develop novel temporal difference (TD) algorithms that can accommodate diverse data types. Our objectives are to (1) establish conditions where TD outperforms ordinary least squares (OLS), (2) provide convergence guarantees for the generalized TD algorithm, and (3) validate the approach empirically using synthetic and real-world datasets. Methods: We reformulate traditional SL as a RL problem by modeling data points as a Markov Reward Process (MRP). We then introduce a concept analogous to the inverse link function in generalized linear models, allowing our TD algorithm to handle various data types. Our analysis, grounded in variance estimation, identifies conditions where TD outperforms OLS. We establish a convergence guarantee by conceptualizing the TD update rule as a generalized Bellman operator. Empirical validation begins with synthetic data progressively matching theoretical assumptions to verify our analysis, followed by evaluations on real-world datasets to demonstrate practical utility. Results: Our theoretical analysis shows that TD can outperform OLS in estimation accuracy when data noise is correlated. Our approach generalizes across various loss functions and SL datasets. We prove that the Bellman operator in our TD framework is a contraction, ensuring convergence for both expected and stochastic TD updates. Empirically, TD outperforms SL baselines when data aligns with its assumptions, remains competitive across diverse datasets, and is robust to hyperparameter choices. Conclusions: This study demonstrates that SL can be reformulated as a problem of interconnected data modeled by an MRP, effectively solved using TD learning. Our generalized TD is theoretically sound, with convergence guarantees, and practically effective. It generalizes OLS, offering superior performance on correlated data. This work enables RL techniques to benefit SL tasks, offering a pathway for future advancements.

While convolution and self-attention mechanisms have dominated architectural design in deep learning, this survey examines a fundamental yet understudied primitive: the Hadamard product. Despite its widespread implementation across various applications, the Hadamard product has not been systematically analyzed as a core architectural primitive. We present the first comprehensive taxonomy of its applications in deep learning, identifying four principal domains: higher-order correlation, multimodal data fusion, dynamic representation modulation, and efficient pairwise operations. The Hadamard product's ability to model nonlinear interactions with linear computational complexity makes it particularly valuable for resource-constrained deployments and edge computing scenarios. We demonstrate its natural applicability in multimodal fusion tasks, such as visual question answering, and its effectiveness in representation masking for applications including image inpainting and pruning. This systematic review not only consolidates existing knowledge about the Hadamard product's role in deep learning architectures but also establishes a foundation for future architectural innovations. Our analysis reveals the Hadamard product as a versatile primitive that offers compelling trade-offs between computational efficiency and representational power, positioning it as a crucial component in the deep learning toolkit.

Existing top-performance autonomous driving systems typically rely on the multi-modal fusion strategy for reliable scene understanding. This design is however fundamentally restricted due to overlooking the modality-specific strengths and finally hampering the model performance. To address this limitation, in this work, we introduce a novel modality interaction strategy that allows individual per-modality representations to be learned and maintained throughout, enabling their unique characteristics to be exploited during the whole perception pipeline. To demonstrate the effectiveness of the proposed strategy, we design DeepInteraction++, a multi-modal interaction framework characterized by a multi-modal representational interaction encoder and a multi-modal predictive interaction decoder. Specifically, the encoder is implemented as a dual-stream Transformer with specialized attention operation for information exchange and integration between separate modality-specific representations. Our multi-modal representational learning incorporates both object-centric, precise sampling-based feature alignment and global dense information spreading, essential for the more challenging planning task. The decoder is designed to iteratively refine the predictions by alternately aggregating information from separate representations in a unified modality-agnostic manner, realizing multi-modal predictive interaction. Extensive experiments demonstrate the superior performance of the proposed framework on both 3D object detection and end-to-end autonomous driving tasks.

Referring segmentation is a fundamental vision-language task that aims to segment out an object from an image or video in accordance with a natural language description. One of the key challenges behind this task is leveraging the referring expression for highlighting relevant positions in the image or video frames. A paradigm for tackling this problem in both the image and the video domains is to leverage a powerful vision-language ("cross-modal") decoder to fuse features independently extracted from a vision encoder and a language encoder. Recent methods have made remarkable advances in this paradigm by exploiting Transformers as cross-modal decoders, concurrent to the Transformer's overwhelming success in many other vision-language tasks. Adopting a different approach in this work, we show that significantly better cross-modal alignments can be achieved through the early fusion of linguistic and visual features in intermediate layers of a vision Transformer encoder network. Based on the idea of conducting cross-modal feature fusion in the visual feature encoding stage, we propose a unified framework named Language-Aware Vision Transformer (LAVT), which leverages the well-proven correlation modeling power of a Transformer encoder for excavating helpful multi-modal context. This way, accurate segmentation results can be harvested with a light-weight mask predictor. One of the key components in the proposed system is a dense attention mechanism for collecting pixel-specific linguistic cues. When dealing with video inputs, we present the video LAVT framework and design a 3D version of this component by introducing multi-scale convolutional operators arranged in a parallel fashion, which can exploit spatio-temporal dependencies at different granularity levels. We further introduce unified LAVT as a unified framework that could handle both image and video inputs with enhanced segmentation capability on unified referring segmentation task. Our methods surpass previous state-of-the-art methods on seven benchmarks for referring image segmentation and referring video segmentation. The code to reproduce our experiments is available at LAVT-RS.

Multimodal Large Language Models (MLLMs) have gained significant attention due to their rich internal implicit knowledge for cross-modal learning. Although advances in bringing audio-visuals into LLMs have resulted in boosts for a variety of Audio-Visual Question Answering (AVQA) tasks, they still face two crucial challenges: 1) audio-visual ambiguity, and 2) audio-visual hallucination. Existing MLLMs can respond to audio-visual content, yet sometimes fail to describe specific objects due to the ambiguity or hallucination of responses. To overcome the two aforementioned issues, we introduce the CAT+, which enhances MLLM to ensure more robust multimodal understanding. We first propose the Sequential Question-guided Module (SQM), which combines tiny transformer layers and cascades Q-Formers to realize a solid audio-visual grounding. After feature alignment and high-quality instruction tuning, we introduce Ambiguity Scoring Direct Preference Optimization (AS-DPO) to correct the problem of CAT+ bias toward ambiguous descriptions. To explore the hallucinatory deficits of MLLMs in dynamic audio-visual scenes, we build a new Audio-visual Hallucination Benchmark, named AVHbench. This benchmark detects the extent of MLLM's hallucinations across three different protocols in the perceptual object, counting, and holistic description tasks. Extensive experiments across video-based understanding, open-ended, and close-ended AVQA demonstrate the superior performance of our method. The AVHbench is released at https://github.com/rikeilong/Bay-CAT.

We present a novel data-driven Parametric Linear Blend Skinning (PLBS) model meticulously crafted for generalized 3D garment dressing and animation. Previous data-driven methods are impeded by certain challenges including overreliance on human body modeling and limited adaptability across different garment shapes. Our method resolves these challenges via two goals: 1) Develop a model based on garment modeling rather than human body modeling. 2) Separately construct low-dimensional sub-spaces for modeling in-plane deformation (such as variation in garment shape and size) and out-of-plane deformation (such as deformation due to varied body size and motion). Therefore, we formulate garment deformation as a PLBS model controlled by canonical 3D garment mesh, vertex-based skinning weights and associated local patch transformation. Unlike traditional LBS models specialized for individual objects, PLBS model is capable of uniformly expressing varied garments and bodies, the in-plane deformation is encoded on the canonical 3D garment and the out-of-plane deformation is controlled by the local patch transformation. Besides, we propose novel 3D garment registration and skinning weight decomposition strategies to obtain adequate data to build PLBS model under different garment categories. Furthermore, we employ dynamic fine-tuning to complement high-frequency signals missing from LBS for unseen testing data. Experiments illustrate that our method is capable of modeling dynamics for loose-fitting garments, outperforming previous data-driven modeling methods using different sub-space modeling strategies. We showcase that our method can factorize and be generalized for varied body sizes, garment shapes, garment sizes and human motions under different garment categories.

Neural Radiance Fields (NeRF) have garnered remarkable success in novel view synthesis. Nonetheless, the task of generating high-quality images for novel views persists as a critical challenge. While the existing efforts have exhibited commendable progress, capturing intricate details, enhancing textures, and achieving superior Peak Signal-to-Noise Ratio (PSNR) metrics warrant further focused attention and advancement. In this work, we propose NeRF-VPT, an innovative method for novel view synthesis to address these challenges. Our proposed NeRF-VPT employs a cascading view prompt tuning paradigm, wherein RGB information gained from preceding rendering outcomes serves as instructive visual prompts for subsequent rendering stages, with the aspiration that the prior knowledge embedded in the prompts can facilitate the gradual enhancement of rendered image quality. NeRF-VPT only requires sampling RGB data from previous stage renderings as priors at each training stage, without relying on extra guidance or complex techniques. Thus, our NeRF-VPT is plug-and-play and can be readily integrated into existing methods. By conducting comparative analyses of our NeRF-VPT against several NeRF-based approaches on demanding real-scene benchmarks, such as Realistic Synthetic 360, Real Forward-Facing, Replica dataset, and a user-captured dataset, we substantiate that our NeRF-VPT significantly elevates baseline performance and proficiently generates more high-quality novel view images than all the compared state-of-the-art methods. Furthermore, the cascading learning of NeRF-VPT introduces adaptability to scenarios with sparse inputs, resulting in a significant enhancement of accuracy for sparse-view novel view synthesis. The source code and dataset are available at https://github.com/Freedomcls/NeRF-VPT.

Diffusion-based models have gained significant popularity for text-to-image generation due to their exceptional image-generation capabilities. A risk with these models is the potential generation of inappropriate content, such as biased or harmful images. However, the underlying reasons for generating such undesired content from the perspective of the diffusion model's internal representation remain unclear. Previous work interprets vectors in an interpretable latent space of diffusion models as seman-tic concepts. However, existing approaches cannot discover directions for arbitrary concepts, such as those related to inappropriate concepts. In this work, we propose a novel self-supervised approach to find interpretable latent directions for a given concept. With the discovered vectors, we further propose a simple approach to mitigate inappropriate generation. Extensive experiments have been conducted to verify the effectiveness of our mitigation approach, namely, for fair generation, safe generation, and responsible text-enhancing generation. Project page: https://interpretdiffusion.github.io.

Before developing a Document Layout Analysis (DLA) model in real-world applications, conducting comprehensive robustness testing is essential. However, the robustness of DLA models remains underexplored in the literature. To address this, we are the first to introduce a robustness benchmark for DLA models, which includes 450K document images of three datasets. To cover realistic corruptions, we propose a perturbation taxonomy with 12 common document perturbations with 3 severity levels inspired by realworld document processing. Additionally, to better understand document perturbation impacts, we propose two metrics, Mean Perturbation Effect (mPE) for perturbation assessment and Mean Robustness Degradation (mRD) for robustness evaluation. Furthermore, we introduce a self-titled model, i.e., Robust Document Layout Analyzer (RoDLA), which improves attention mechanisms to boost extraction of robust features. Experiments on the proposed benchmarks (PubLayNet-P, DocLayNet-P, andM6Doc-P) demonstrate that RoDLA obtains state-of-the-art mRD scores of 115.7, 135.4, and 150.4, respectively. Compared to previous methods, RoDLA achieves notable improvements in mAP of +3.8%, +7.1% and +12.1%, respectively.

Existing open-vocabulary image segmentation methods re-quire a fine-tuning step on mask labels and/or image-text datasets. Mask labels are labor-intensive, which limits the number of categories in segmentation datasets. Con-sequently, the vocabulary capacity of pre-trained VLMs is severely reduced after fine-tuning. However, without fine-tuning, VLMs trained under weak image-text supervision tend to make suboptimal mask predictions. To alleviate these issues, we introduce a novel recurrent framework that progressively filters out irrelevant texts and enhances mask quality without training efforts. The recurrent unit is a two-stage segmenter built upon a frozen VLM. Thus, our model retains the VLM's broad vocabulary space and equips it with segmentation ability. Experiments show that our method outperforms not only the training-free counter-parts, but also those fine-tuned with millions of data sam-ples, and sets the new state-of-the-art records for both zero-shot semantic and referring segmentation. Concretely, we improve the current record by 28.8, 16.0, and 6.9 mloU on Pascal VOC, COCO Object, and Pascal Context.

While transformer models exhibit strong capabilities on linguistic tasks, their complex architectures make them difficult to interpret.Recent work has aimed to reverse engineer transformer models into human-readable representations called circuits that implement algorithmic functions.We extend this research by analyzing and comparing circuits for similar sequence continuation tasks, which include increasing sequences of Arabic numerals, number words, and months.By applying circuit interpretability analysis, we identify a key sub-circuit in both GPT-2 Small and Llama-2-7B responsible for detecting sequence members and for predicting the next member in a sequence.Our analysis reveals that semantically related sequences rely on shared circuit subgraphs with analogous roles.Additionally, we show that this sub-circuit has effects on various math-related prompts, such as on intervaled circuits, Spanish number word and months continuation, and natural language word problems.This mechanistic understanding of transformers is a critical step towards building more robust, aligned, and interpretable language models.

Youssef Mohamed, Runjia Li, Ibrahim Said Ahmad, Kilichbek Haydarov, Philip Torr, Kenneth Church, Mohamed Elhoseiny. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Neural surface reconstruction is sensitive to the camera pose noise, even when state-of-the-art pose estimators like COLMAP or ARKit are used. Existing Pose-NeRF joint optimisation methods have struggled to improve pose accuracy in challenging real-world scenarios. To overcome the challenges, we introduce the pose residual field (PoRF), a novel implicit representation that uses an MLP for regressing pose updates. Compared with the conventional per-frame pose parameter optimisation, this new representation is more robust due to parameter sharing that leverages global information over the entire sequence. Furthermore, we propose an epipolar geometry loss to enhance the supervision that leverages the correspondences exported from COLMAP results without the extra computational overhead. Our method yields promising results. On the DTU dataset, we reduce the rotation error of COLMAP poses by 78\%, leading to the reduced reconstruction Chamfer distance from 3.48mm to 0.85mm. On the MobileBrick dataset that contains casually captured unbounded 360-degree videos, our method refines ARKit poses and improves the reconstruction F1 score from 69.18 to 75.67, outperforming that with the provided ground-truth pose (75.14). These achievements demonstrate the efficacy of our approach in refining camera poses and improving the accuracy of neural surface reconstruction in real-world scenarios.

AI agents are commonly trained with large datasets of demonstrations of human behavior. However, not all behaviors are equally safe or desirable. Desired characteristics for an AI agent can be expressed by assigning desirability scores, which we assume are not assigned to individual behaviors but to collective trajectories. For example, in a dataset of vehicle interactions, these scores might relate to the number of incidents that occurred. We first assess the effect of each individual agent's behavior on the collective desirability score, e.g., assessing how likely an agent is to cause incidents. This allows us to selectively imitate agents with a positive effect, e.g., only imitating agents that are unlikely to cause incidents. To enable this, we propose the concept of an agent's \textit{Exchange Value}, which quantifies an individual agent's contribution to the collective desirability score. The Exchange Value is the expected change in desirability score when substituting the agent for a randomly selected agent. We propose additional methods for estimating Exchange Values from real-world datasets, enabling us to learn desired imitation policies that outperform relevant baselines. The project website can be found at https://tinyurl.com/select-to-perfect.

Autonomous agents deployed in the real world need to be robust against adversarial attacks on sensory inputs. Robustifying agent policies requires anticipating the strongest attacks possible. We demonstrate that existing observation-space attacks on reinforcement learning agents have a common weakness: while effective, their lack of information-theoretic detectability constraints makes them \textit{detectable} using automated means or human inspection. Detectability is undesirable to adversaries as it may trigger security escalations. We introduce \textit{\eattacks{}}, a novel form of adversarial attack on sequential decision-makers that is both effective and of $\epsilon-$bounded statistical detectability. We propose a novel dual ascent algorithm to learn such attacks end-to-end. Compared to existing attacks, we empirically find \eattacks{} to be significantly harder to detect with automated methods, and a small study with human participants\footnote{IRB approval under reference R84123/RE001} suggests they are similarly harder to detect for humans. Our findings suggest the need for better anomaly detectors, as well as effective hardware- and system-level defenses. The project website can be found at https://tinyurl.com/illusory-attacks.

Large vision-language models (VLMs) such as GPT-4 have achieved exceptional performance across various multi-modal tasks. However, the deployment of VLMs necessitates substantial energy consumption and computational resources. Once attackers maliciously induce high energy consumption and latency time (energy-latency cost) during inference of VLMs, it will exhaust computational resources. In this paper, we explore this attack surface about availability of VLMs and aim to induce high energy-latency cost during inference of VLMs. We find that high energy-latency cost during inference of VLMs can be manipulated by maximizing the length of generated sequences. To this end, we propose verbose images, with the goal of crafting an imperceptible perturbation to induce VLMs to generate long sentences during inference. Concretely, we design three loss objectives. First, a loss is proposed to delay the occurrence of end-of-sequence (EOS) token, where EOS token is a signal for VLMs to stop generating further tokens. Moreover, an uncertainty loss and a token diversity loss are proposed to increase the uncertainty over each generated token and the diversity among all tokens of the whole generated sequence, respectively, which can break output dependency at token-level and sequence-level. Furthermore, a temporal weight adjustment algorithm is proposed, which can effectively balance these losses. Extensive experiments demonstrate that our verbose images can increase the length of generated sequences by 7.87× and 8.56× compared to original images on MS-COCO and ImageNet datasets, which presents potential challenges for various applications.

When a small number of poisoned samples are injected into the training dataset of a deep neural network, the network can be induced to exhibit malicious behavior during inferences, which poses potential threats to real-world applications. While they have been intensively studied in classification, backdoor attacks on semantic segmentation have been largely overlooked. Unlike classification, semantic segmentation aims to classify every pixel within a given image. In this work, we explore backdoor attacks on segmentation models to misclassify all pixels of a victim class by injecting a specific trigger on non-victim pixels during inferences, which is dubbed Influencer Backdoor Attack (IBA). IBA is expected to maintain the classification accuracy of non-victim pixels and mislead classifications of all victim pixels in every single inference and could be easily applied to real-world scenes. Based on the context aggregation ability of segmentation models, we proposed a simple, yet effective, Nearest-Neighbor trigger injection strategy. We also introduce an innovative Pixel Random Labeling strategy which maintains optimal performance even when the trigger is placed far from the victim pixels. Our extensive experiments reveal that current segmentation models do suffer from backdoor attacks, demonstrate IBA real-world applicability, and show that our proposed techniques can further increase attack performance.

Different from traditional task-specific vision models, recent large VLMs can readily adapt to different vision tasks by simply using different textual instructions, i.e., prompts. However, a well-known concern about traditional task-specific vision models is that they can be misled by imperceptible adversarial perturbations. Furthermore, the concern is exacerbated by the phenomenon that the same adversarial perturbations can fool different task-specific models. Given that VLMs rely on prompts to adapt to different tasks, an intriguing question emerges: Can a single adversarial image mislead all predictions of VLMs when a thousand different prompts are given? This question essentially introduces a novel perspective on adversarial transferability: cross-prompt adversarial transferability. In this work, we propose the Cross-Prompt Attack (CroPA). This proposed method updates the visual adversarial perturbation with learnable textual prompts, which are designed to counteract the misleading effects of the adversarial image. By doing this, CroPA significantly improves the transferability of adversarial examples across prompts. Extensive experiments are conducted to verify the strong cross-prompt adversarial transferability of CroPA with prevalent VLMs including Flamingo, BLIP-2, and InstructBLIP in various different tasks.

Context-based fine-tuning methods, including prompting, in-context learning, soft prompting (also known as prompt tuning), and prefix-tuning, have gained popularity due to their ability to often match the performance of full fine-tuning with a fraction of the parameters. Despite their empirical successes, there is little theoretical understanding of how these techniques influence the internal computation of the model and their expressiveness limitations. We show that despite the continuous embedding space being more expressive than the discrete token space, soft-prompting and prefix-tuning are potentially less expressive than full fine-tuning, even with the same number of learnable parameters. Concretely, context-based fine-tuning cannot change the relative attention pattern over the content and can only bias the outputs of an attention layer in a fixed direction. This suggests that while techniques like prompting, in-context learning, soft prompting, and prefix-tuning can effectively elicit skills present in the pretrained model, they may not be able to learn novel tasks that require new attention patterns.

Synthetic training data has gained prominence in numerous learning tasks and scenarios, offering advantages such as dataset augmentation, generalization evaluation, and privacy preservation. Despite these benefits, the efficiency of synthetic data generated by current methodologies remains inferior when training advanced deep models exclusively, limiting its practical utility. To address this challenge, we analyze the principles underlying training data synthesis for supervised learning and elucidate a principled theoretical framework from the distribution-matching perspective that explicates the mechanisms governing synthesis efficacy. Through extensive experiments, we demonstrate the effectiveness of our synthetic data across diverse image classification tasks, both as a replacement for and augmentation to real datasets, while also benefits such as out-of-distribution generalization, privacy preservation, and scalability. Specifically, we achieve 70.9% top1 classification accuracy on ImageNet1K when training solely with synthetic data equivalent to 1 × the original real data size, which increases to 76.0% when scaling up to 10 × synthetic data.

We propose and study a realistic Continual Learning (CL) setting where learning algorithms are granted a restricted computational budget per time step while training. We apply this setting to large-scale semi-supervised Continual Learning scenarios with sparse label rate. Previous proficient CL methods perform very poorly in this challenging setting. Overfitting to the sparse labeled data and insufficient computational budget are the two main culprits for such a poor performance. Our new setting encourages learning methods to effectively and efficiently utilize the unlabeled data during training. To that end, we propose a simple but highly effective baseline, DietCL, which utilizes both unlabeled and labeled data jointly. DietCL meticulously allocates computational budget for both types of data. We validate our baseline, at scale, on several datasets, e.g., CLOC, ImageNet10K, and CGLM, under constraint budget setup. DietCL outperforms, by a large margin, all existing supervised CL algorithms as well as more recent continual semi-supervised methods. Our extensive analysis and ablations demonstrate that DietCL is stable under a full spectrum of label sparsity, computational budget and various other ablations.

Recent work provides promising evidence that Physics-Informed Neural Networks (PINN) can efficiently solve partial differential equations (PDE). However, previous works have failed to provide guarantees on the worst-case residual error of a PINN across the spatio-temporal domain - a measure akin to the tolerance of numerical solvers - focusing instead on point-wise comparisons between their solution and the ones obtained by a solver on a set of inputs. In real-world applications, one cannot consider tests on a finite set of points to be sufficient grounds for deployment, as the performance could be substantially worse on a different set. To alleviate this issue, we establish guaranteed error-based conditions for PINNs over their continuous applicability domain. To verify the extent to which they hold, we introduce $\partial$-CROWN: a general, efficient and scalable post-training framework to bound PINN residual errors. We demonstrate its effectiveness in obtaining tight certificates by applying it to two classically studied PINNs – Burgers’ and Schrödinger’s equations –, and two more challenging ones with real-world applications – the Allan-Cahn and Diffusion-Sorption equations.

In the next few years, applications of Generative AI are expected to revolutionize a number of different areas, ranging from science & medicine to education. The potential for these seismic changes has triggered a lively debate about potential risks and resulted in calls for tighter regulation, in particular from some of the major tech companies who are leading in AI development. While regulation is important, it is key that it does not put at risk the budding field of open-source Generative AI. We argue for the responsible open sourcing of generative AI models in the near and medium term. To set the stage, we first introduce an AI openness taxonomy system and apply it to 40 current large language models. We then outline differential benefits and risks of open versus closed source AI and present potential risk mitigation, ranging from best practices to calls for technical and scientific contributions. We hope that this report will add a much needed missing voice to the current public discourse on near to mid-term AI safety and other societal impact.

Despite the widespread adoption of prompting, prompt tuning and prefix-tuning of transformer models, our theoretical understanding of these fine-tuning methods remains limited. A key question is whether one can arbitrarily modify the behavior of a pretrained model by prompting or prefix-tuning it. Formally, whether prompting and prefix-tuning a pretrained model can universally approximate sequence-to-sequence functions. This paper answers in the affirmative and demonstrates that much smaller pretrained models than previously thought can be universal approximators when prefixed. In fact, prefix-tuning a single attention head is sufficient to approximate any continuous function making the attention mechanism uniquely suited for universal approximation. Moreover, any sequence-to-sequence function can be approximated by prefixing a transformer with depth linear in the sequence length. Beyond these density-type results, we also offer Jackson-type bounds on the length of the prefix needed to approximate a function to a desired precision.

Vision-Language Models (VLMs) have made remarkable progress in document-based Visual Question Answering (i.e., responding to queries about the contents of an input document provided as an image). In this work, we show these models can memorize responses for training samples and regurgitate them even when the relevant visual information has been removed. This includes Personal Identifiable Information (PII) repeated once in the training set, indicating these models could divulge memorised sensitive information and therefore pose a privacy risk. We quantitatively measure the extractability of information in controlled experiments and differentiate between cases where it arises from generalization capabilities or from memorization. We further investigate the factors that influence memorization across multiple state-of-the-art models and propose an effective heuristic countermeasure that empirically prevents the extractability of PII.

Relieving the reliance of neural network training on a global back-propagation (BP) has emerged as a notable research topic due to the biological implausibility and huge memory consumption caused by BP. Among the existing solutions, local learning optimizes gradient-isolated modules of a neural network with local errors and has been proved to be effective even on large-scale datasets. However, the reconciliation among local errors has never been investigated. In this paper, we first theoretically study non-greedy layer-wise training and show that the convergence cannot be assured when the local gradient in a module w.r.t. its input is not reconciled with the local gradient in the previous module w.r.t. its output. Inspired by the theoretical result, we further propose a local training strategy that successively regularizes the gradient reconciliation between neighboring modules without breaking gradient isolation or introducing any learnable parameters. Our method can be integrated into both local-BP and BP-free settings. In experiments, we achieve significant performance improvements compared to previous methods. Particularly, our method for CNN and Transformer architectures on ImageNet is able to attain a competitive performance with global BP, saving more than 40% memory consumption.

Neural image classifiers are known to undergo severe performance degradation when exposed to inputs that are sampled from environmental conditions that differ from their training data. Given the recent progress in Text-to-Image (T2I) generation, a natural question is how modern T2I generators can be used to simulate arbitrary interventions over such environmental factors in order to augment training data and improve the robustness of downstream classifiers. We experiment across a diverse collection of benchmarks in single domain generalization (SDG) and reducing reliance on spurious features (RRSF), ablating across key dimensions of T2I generation, including interventional prompting strategies, conditioning mechanisms, and post-hoc filtering, showing that modern T2I generators like Stable Diffusion can indeed be used to implement a powerful interventional data augmentation (IDA) mechanism, outperforming previously state-of-the-art data augmentation techniques regardless of how each dimension is configured.

We propose a neural network weight encoding method for network property prediction that utilizes set-to-set and set-to-vector functions to efficiently encode neural network parameters. Our approach is capable of encoding neural networks in a model zoo of mixed architecture and different parameter sizes as opposed to previous approaches that require custom encoding models for different architectures. Furthermore, our \textbf{S}et-based \textbf{N}eural network \textbf{E}ncoder (SNE) takes into consideration the hierarchical computational structure of neural networks. To respect symmetries inherent in network weight space, we utilize Logit Invariance to learn the required minimal invariance properties. Additionally, we introduce a \textit{pad-chunk-encode} pipeline to efficiently encode neural network layers that is adjustable to computational and memory constraints. We also introduce two new tasks for neural network property prediction: cross-dataset and cross-architecture. In cross-dataset property prediction, we evaluate how well property predictors generalize across model zoos trained on different datasets but of the same architecture. In cross-architecture property prediction, we evaluate how well property predictors transfer to model zoos of different architecture not seen during training. We show that SNE outperforms the relevant baselines on standard benchmarks.

Online continual learning, the process of training models on streaming data, has gained increasing attention in recent years. However, a critical aspect often overlooked is the label delay, where new data may not be labeled due to slow and costly annotation processes. We introduce a new continual learning framework with explicit modeling of the label delay between data and label streams over time steps. In each step, the framework reveals both unlabeled data from the current time step t and labels delayed with d steps, from the time step t−d. In our extensive experiments amounting to 1060 GPU days, we show that merely augmenting the computational resources is insufficient to tackle this challenge. Our findings underline a notable performance decline when solely relying on labeled data when the label delay becomes significant. More surprisingly, when using state-of-the-art SSL and TTA techniques to utilize the newer, unlabeled data, they fail to surpass the performance of a naïve method that simply trains on the delayed supervised stream. To this end, we introduce a simple, efficient baseline that rehearses from the labeled memory samples that are most similar to the new unlabeled samples. This method bridges the accuracy gap caused by label delay without significantly increasing computational complexity. We show experimentally that our method is the least affected by the label delay factor and in some cases successfully recovers the accuracy of the non-delayed counterpart. We conduct various ablations and sensitivity experiments, demonstrating the effectiveness of our approach.

Despite the importance of shape perception in human vision, early neural image classifiers relied less on shape information for object recognition than other (often spurious) features. While recent research suggests that current large Vision-Language Models (VLMs) exhibit more reliance on shape, we find them to still be seriously limited in this regard. To quantify such limitations, we introduce IllusionBench, a dataset that challenges current cutting-edge VLMs to decipher shape information when the shape is represented by an arrangement of visual elements in a scene. Our extensive evaluations reveal that, while these shapes are easily detectable by human annotators, current VLMs struggle to recognize them, indicating important avenues for future work in developing more robust visual perception systems. The full dataset and codebase are available at: https://arshiahemmat.github.io/illusionbench/

Safety fine-tuning helps align Large Language Models (LLMs) with human preferences for their safe deployment. To better understand the underlying factors that make models safe via safety fine-tuning, we design a synthetic data generation framework that captures salient aspects of an unsafe input by modeling the interaction between the task the model is asked to perform (e.g., “design”) versus the specific concepts the task is asked to be performed upon (e.g., a “cycle” vs. a “bomb”). Using this, we investigate three well-known safety fine-tuning methods—supervised safety fine-tuning, direct preference optimization, and unlearning—and provide significant evidence demonstrating that these methods minimally transform MLP weights to specifically align unsafe inputs into its weights’ null space. This yields a clustering of inputs based on whether the model deems them safe or not. Correspondingly, when an adversarial input (e.g., a jailbreak) is provided, its activations are closer to safer samples, leading to the model processing such an input as if it were safe. Code is available at https://github.com/fiveai/understanding_safety_finetuning.

Reinforcement learning from human feedback (RLHF) is widely used to train large language models (LLMs). However, it is unclear whether LLMs accurately learn the underlying preferences in human feedback data. We coin the term **Learned Feedback Pattern** (LFP) for patterns in an LLM's activations learned during RLHF that improve its performance on the fine-tuning task. We hypothesize that LLMs with LFPs accurately aligned to the fine-tuning feedback exhibit consistent activation patterns for outputs that would have received similar feedback during RLHF. To test this, we train probes to estimate the feedback signal implicit in the activations of a fine-tuned LLM. We then compare these estimates to the true feedback, measuring how accurate the LFPs are to the fine-tuning feedback. Our probes are trained on a condensed, sparse and interpretable representation of LLM activations, making it easier to correlate features of the input with our probe's predictions. We validate our probes by comparing the neural features they correlate with positive feedback inputs against the features GPT-4 describes and classifies as related to LFPs. Understanding LFPs can help minimize discrepancies between LLM behavior and training objectives, which is essential for the **safety** and **alignment** of LLMs.

Recent advancements in generative AI suggest the potential for large-scale interaction between autonomous agents and humans across platforms such as the internet. While such interactions could foster productive cooperation, the ability of AI agents to circumvent security oversight raises critical multi-agent security problems, particularly in the form of unintended information sharing or undesirable coordination. In our work, we establish the subfield of secret collusion, a form of multi-agent deception, in which two or more agents employ steganographic methods to conceal the true nature of their interactions, be it communicative or otherwise, from oversight. We propose a formal threat model for AI agents communicating steganographically and derive rigorous theoretical insights about the capacity and incentives of large language models (LLMs) to perform secret collusion, in addition to the limitations of threat mitigation measures. We complement our findings with empirical evaluations demonstrating rising steganographic capabilities in frontier single and multi-agent LLM setups and examining potential scenarios where collusion may emerge, revealing limitations in countermeasures such as monitoring, paraphrasing, and parameter optimization. Our work is the first to formalize and investigate secret collusion among frontier foundation models, identifying it as a critical area in AI Safety and outlining a comprehensive research agenda to mitigate future risks of collusion between generative AI systems.

Zero-shot and in-context learning enable solving tasks without model fine-tuning, making them essential for developing generative model solutions. Therefore, it is crucial to understand whether a pretrained model can be prompted to approximate any function, i.e., whether it is a universal in-context approximator. While it was recently shown that transformer models do possess this property, these results rely on their attention mechanism. Hence, these findings do not apply to fully recurrent architectures like RNNs, LSTMs, and the increasingly popular SSMs. We demonstrate that RNNs, LSTMs, GRUs, Linear RNNs, and linear gated architectures such as Mamba and Hawk/Griffin can also serve be universal in-context approximators. To streamline our argument, we introduce a programming language called LSRL that compiles to these fully recurrent architectures. LSRL may be of independent interest for further studies of fully recurrent models, such as constructing interpretability benchmarks. We also study the role of multiplicative gating and observe that architectures incorporating such gating (e.g., LSTMs, GRUs, Hawk/Griffin) can implement certain operations more stably, making them more viable candidates for practical in-context universal approximation.

Continual learning has primarily focused on the issue of catastrophic forgetting and the associated stability-plasticity tradeoffs. However, little attention has been paid to the efficacy of continually learned representations, as representations are learned alongside classifiers throughout the learning process. Our primary contribution is empirically demonstrating that existing online continually trained deep networks produce inferior representations compared to a simple pre-defined random transforms. Our approach embeds raw pixels using a fixed random transform, approximating an RBF-Kernel initialized before any data is seen. We then train a simple linear classifier on top without storing any exemplars, processing one sample at a time in an online continual learning setting. This method, called RanDumb, significantly outperforms state-of-the-art continually learned representations across all standard online continual learning benchmarks. Our study reveals the significant limitations of representation learning, particularly in low-exemplar and online continual learning scenarios. Extending our investigation to popular exemplar-free scenarios with pretrained models, we find that training only a linear classifier on top of pretrained representations surpasses most continual fine-tuning and prompt-tuning strategies. Overall, our investigation challenges the prevailing assumptions about effective representation learning in the online continual learning.

Standardized benchmarks drive progress in machine learning. However, with repeated testing, the risk of overfitting grows as algorithms over-exploit benchmark idiosyncrasies. In our work, we seek to mitigate this challenge by compiling \textit{ever-expanding} large-scale benchmarks called \textit{Lifelong Benchmarks}. As exemplars of our approach, we create \textit{Lifelong-CIFAR10} and \textit{Lifelong-ImageNet}, containing (for now) 1.69M and 1.98M test samples, respectively. While reducing overfitting, lifelong benchmarks introduce a key challenge: the high cost of evaluating a growing number of models across an ever-expanding sample set. To address this challenge, we also introduce an efficient evaluation framework: \textit{Sort \& Search (S\&S)}, which reuses previously evaluated models by leveraging dynamic programming algorithms to selectively rank and sub-select test samples, enabling cost-effective lifelong benchmarking. Extensive empirical evaluations across $\sim$31,000 models demonstrate that \textit{S\&S} achieves highly-efficient approximate accuracy measurement, reducing compute cost from 180 GPU days to 5 GPU hours ($\sim$1000x reduction) on a single A100 GPU, with low approximation error. As such, lifelong benchmarks offer a robust, practical solution to the ``benchmark exhaustion'' problem.

Web-crawled pretraining datasets underlie the impressive "zero-shot" evaluation performance of multimodal models, such as CLIP for classification and Stable-Diffusion for image generation. However, it is unclear how meaningful the notion of "zero-shot" generalization is for such multimodal models, as it is not known to what extent their pretraining datasets encompass the downstream concepts targeted for during "zero-shot" evaluation. In this work, we ask: How is the performance of multimodal models on downstream concepts influenced by the frequency of these concepts in their pretraining datasets? We comprehensively investigate this question across 34 models and 5 standard pretraining datasets (CC-3M, CC-12M, YFCC-15M, LAION-400M, LAION-Aesthetics), generating over 300GB of data artifacts. We consistently find that, far from exhibiting "zero-shot" generalization, multimodal models require exponentially more data to achieve linear improvements in downstream "zero-shot" performance, following a sample inefficient log-linear scaling trend. This trend persists even when controlling for sample-level similarity between pretraining and downstream datasets, and testing on purely synthetic data distributions. Furthermore, upon benchmarking models on long-tailed data sampled based on our analysis, we demonstrate that multimodal models across the board perform poorly. We contribute this long-tail test set as the Let it Wag! benchmark to further research in this direction. Taken together, our study reveals an exponential need for training data which implies that the key to "zero-shot" generalization capabilities under large-scale training data and compute paradigms remains to be found.

Generating high-quality 3D assets from text and images has long been challenging, primarily due to the absence of scalable 3D representations capable of capturing intricate geometry distributions. In this work, we introduce Direct3D, a native 3D generative model scalable to in-the-wild input images, without requiring a multi-view diffusion model or SDS optimization. Our approach comprises two primary components: a Direct 3D Variational Auto-Encoder (D3D-VAE) and a Direct 3D Diffusion Transformer (D3D-DiT). D3D-VAE efficiently encodes high-resolution 3D shapes into a compact and continuous latent triplane space. Notably, our method directly supervises the decoded geometry using a semi-continuous surface sampling strategy, diverging from previous methods relying on rendered images as supervision signals. D3D-DiT models the distribution of encoded 3D latents and is specifically designed to fuse positional information from the three feature maps of the triplane latent, enabling a native 3D generative model scalable to large-scale 3D datasets. Additionally, we introduce an innovative image-to-3D generation pipeline incorporating semantic and pixel-level image conditions, allowing the model to produce 3D shapes consistent with the provided conditional image input. Extensive experiments demonstrate the superiority of our large-scale pre-trained Direct3D over previous image-to-3D approaches, achieving significantly better generation quality and generalization ability, thus establishing a new state-of-the-art for 3D content creation. Project page: https://www.neural4d.com/research/direct3d.

Large Language Model (LLM) agents have been increasingly adopted as simulation tools to model humans in social science and role-playing applications. However, one fundamental question remains: can LLM agents really simulate human behavior? In this paper, we focus on one critical and elemental behavior in human interactions, trust, and investigate whether LLM agents can simulate human trust behavior. We first find that LLM agents generally exhibit trust behavior, referred to as agent trust, under the framework of Trust Games, which are widely recognized in behavioral economics. Then, we discover that GPT-4 agents manifest high behavioral alignment with humans in terms of trust behavior, indicating the feasibility of simulating human trust behavior with LLM agents. In addition, we probe the biases of agent trust and differences in agent trust towards other LLM agents and humans. We also explore the intrinsic properties of agent trust under conditions including external manipulations and advanced reasoning strategies. Our study provides new insights into the behaviors of LLM agents and the fundamental analogy between LLMs and humans beyond value alignment. We further illustrate broader implications of our discoveries for applications where trust is paramount.

Recent work has highlighted the label alignment property (LAP) in supervised learning, where the vector of all labels in the dataset is mostly in the span of the top few singular vectors of the data matrix. Drawing inspiration from this observation, we propose a regularization method for unsupervised domain adaptation that encourages alignment between the predictions in the target domain and its top singular vectors. Unlike conventional domain adaptation approaches that focus on regularizing representations, we instead regularize the classifier to align with the unsupervised target data, guided by the LAP in both the source and target domains. Theoretical analysis demonstrates that, under certain assumptions, our solution resides within the span of the top right singular vectors of the target domain data and aligns with the optimal solution. By removing the reliance on the commonly used optimal joint risk assumption found in classic domain adaptation theory, we showcase the effectiveness of our method on addressing problems where traditional domain adaptation methods often fall short due to high joint error. Additionally, we report improved performance over domain adaptation baselines in well-known tasks such as MNIST-USPS domain adaptation and cross-lingual sentiment analysis. An implementation is available at https://github.com/EhsanEI/lar/.

Tight and efficient neural network bounding is crucial to the scaling of neural network verification systems. Many efficient bounding algorithms have been presented recently, but they are often too loose to verify more challenging properties. This is due to the weakness of the employed relaxation, which is usually a linear program of size linear in the number of neurons. While a tighter linear relaxation for piecewise-linear activations exists, it comes at the cost of exponentially many constraints and currently lacks an efficient customized solver. We alleviate this deficiency by presenting two novel dual algorithms: one operates a subgradient method on a small active set of dual variables, the other exploits the sparsity of Frank-Wolfe type optimizers to incur only a linear memory cost. Both methods recover the strengths of the new relaxation: tightness and a linear separation oracle. At the same time, they share the benefits of previous dual approaches for weaker relaxations: massive parallelism, GPU implementation, low cost per iteration and valid bounds at any time. As a consequence, we can obtain better bounds than off-the-shelf solvers in only a fraction of their running time, attaining significant formal verification speed-ups.

In this paper, we address panoramic semantic segmentation which is under-explored due to two critical challenges: (1) image distortions and object deformations on panoramas; (2) lack of semantic annotations in the$360^\circ$imagery. To tackle these problems, first, we propose the upgraded Transformer for Panoramic Semantic Segmentation, ie, Trans4PASS+, equipped withDeformable Patch Embedding (DPE)andDeformable MLP (DMLPv2)modules for handling object deformations and image distortions whenever (before or after adaptation) and wherever (shallow or deep levels). Second, we enhance theMutual Prototypical Adaptation (MPA)strategy via pseudo-label rectification for unsupervised domain adaptive panoramic segmentation. Third, aside from Pinhole-to-Panoramic (Pin2Pan) adaptation, we create a new dataset (SynPASS) with 9,080 panoramic images, facilitating Synthetic-to-Real (Syn2Real) adaptation scheme in$360^\circ$imagery. Extensive experiments are conducted, which cover indoor and outdoor scenarios, and each of them is investigated withPin2PanandSyn2Realregimens. Trans4PASS+ achieves state-of-the-art performances on four domain adaptive panoramic semantic segmentation benchmarks. Code is available athttps://github.com/jamycheung/Trans4PASS.

The 2D image-based virtual try-on has aroused increased interest from the multimedia and computer vision fields due to its enormous commercial value. Nevertheless, most existing image-based virtual try-on approaches directly combine the person-identity representation and the in-shop clothing items without taking their mutual correlations into consideration. Moreover, these methods are commonly established on pure convolutional neural networks (CNNs) architectures which are not simple to capture the long-range correlations among the input pixels. As a result, it generally results in inconsistent results. To alleviate these issues, in this article, we propose a novel two-stage cloth interactive transformer (CIT) method for the virtual try-on task. During the first stage, we design a CIT matching block, aiming at precisely capturing the long-range correlations between the cloth-agnostic person information and the in-shop cloth information. Consequently, it makes the warped in-shop clothing items look more natural in appearance. In the second stage, we put forth a CIT reasoning block for establishing global mutual interactive dependencies among person representation, the warped clothing item, and the corresponding warped cloth mask. The empirical results, based on mutual dependencies, demonstrate that the final try-on results are more realistic. Substantial empirical results on a public fashion dataset illustrate that the suggested CIT attains competitive virtual try-on performance.

Referring image segmentation segments an image from a language expression. With the aim of producing high-quality masks, existing methods often adopt iterative learning approaches that rely on RNNs or stacked attention layers to refine vision-language features. Despite their complexity, RNN-based methods are subject to specific encoder choices, while attention-based methods offer limited gains. In this work, we introduce a simple yet effective alternative for progressively learning discriminative multi-modal features. The core idea of our approach is to leverage a continuously updated query as the representation of the target object and at each iteration, strengthen multi-modal features strongly correlated to the query while weakening less related ones. As the query is initialized by language features and successively updated by object features, our algorithm gradually shifts from being localization-centric to segmentation-centric. This strategy enables the incremental recovery of missing object parts and/or removal of extraneous parts through iteration. Compared to its counterparts, our method is more versatile—it can be plugged into prior arts straightforwardly and consistently bring improvements. Experimental results on the challenging datasets of RefCOCO, RefCOCO+, and G-Ref demonstrate its advantage with respect to the state-of-the-art methods.

Recent research on Generalized Zero-Shot Learning (GZSL) has focused primarily on generation-based methods. However, current literature has overlooked the fundamental principles of these methods and has made limited progress in a complex manner. In this paper, we aim to deconstruct the generator-classifier framework and provide guidance for its improvement and extension. We begin by breaking down the generator-learned unseen class distribution into class-level and instance-level distributions. Through our analysis of the role of these two types of distributions in solving the GZSL problem, we generalize the focus of the generation-based approach, emphasizing the importance of (i) attribute generalization in generator learning and (ii) independent classifier learning with partially biased data. We present a simple method based on this analysis that outperforms SotAs on four public GZSL datasets, demonstrating the validity of our deconstruction. Furthermore, our proposed method remains effective even without a generative model, representing a step towards simplifying the generator-classifier structure. Our code is available at https://github.com/cdb342/DGZ.

It is now well known that neural networks can be wrong with high confidence in their predictions, leading to poor calibration. The most common post-hoc approach to compensate for this is to perform temperature scaling, which adjusts the confidences of the predictions on any input by scaling the logits by a fixed value. Whilst this approach typically improves the average calibration across the whole test dataset, this improvement typically reduces the individual confidences of the predictions irrespective of whether the classification of a given input is correct or incorrect. With this insight, we base our method on the observation that different samples contribute to the calibration error by varying amounts, with some needing to increase their confidence and others needing to decrease it. Therefore, for each input, we propose to predict a different temperature value, allowing us to adjust the mismatch between confidence and accuracy at a finer granularity. Our method is applied post-hoc, enabling it to be very fast with a negligible memory footprint and is applied to off-the-shelf pre-trained classifiers. We test our method on the ResNet50 and WideResNet28-10 architectures using the CIFAR10/100 and Tiny-ImageNet datasets, showing that producing per-data-point temperatures improves the expected calibration error across the whole test set.

Current evaluations of Continual Learning (CL) methods typically assume that there is no constraint on training time and computation. This is an unrealistic assumption for any real-world setting, which motivates us to propose: a practical real-time evaluation of continual learning, in which the stream does not wait for the model to complete training before revealing the next data for predictions. To do this, we evaluate current CL methods with respect to their computational costs. We conduct extensive experiments on CLOC, a large-scale dataset containing 39 million time-stamped images with geolocation labels. We show that a simple baseline outperforms state-of-the-art CL methods under this evaluation, questioning the applicability of existing methods in realistic settings. In addition, we explore various CL components commonly used in the literature, including memory sampling strategies and regularization approaches. We find that all considered methods fail to be competitive against our simple baseline. This surprisingly suggests that the majority of existing CL literature is tailored to a specific class of streams that is not practical. We hope that the evaluation we provide will be the first step towards a paradigm shift to consider the computational cost in the development of online continual learning methods.

Motivated by the increasing popularity of transformers in computer vision, in recent times there has been a rapid development of novel architectures. While in-domain performance follows a constant, upward trend, properties like robustness or uncertainty estimation are less explored-leaving doubts about advances in model reliability. Studies along these axes exist, but they are mainly limited to classification models. In contrast, we carry out a study on semantic segmentation, a relevant task for many real-world applications where model reliability is paramount. We analyze a broad variety of models, spanning from older ResNet-based architectures to novel transformers and assess their reliability based on four metrics: robustness, calibration, misclassification detection and out-of-distribution (OOD) detection. We find that while recent models are significantly more robust, they are not overall more reliable in terms of uncertainty estimation. We further explore methods that can come to the rescue and show that improving calibration can also help with other uncertainty metrics such as misclassification or OOD detection. This is the first study on modern segmentation models focused on both robustness and uncertainty estimation and we hope it will help practitioners and researchers interested in this fundamental vision task11Code available at https://github.com/naver/relis.

High-quality 3D ground-truth shapes are critical for 3D object reconstruction evaluation. However, it is difficult to create a replica of an object in reality, and even 3D reconstructions generated by 3D scanners have artefacts that cause biases in evaluation. To address this issue, we introduce a novel multi-view RGBD dataset captured using a mobile device, which includes highly precise 3D ground-truth annotations for 153 object models featuring a diverse set of 3D structures. We obtain precise 3D ground-truth shape without relying on high-end 3D scanners by utilising LEGO models with known geometry as the 3D structures for image capture. The distinct data modality offered by high-resolution RGB images and low-resolution depth maps captured on a mobile device, when combined with precise 3D geometry annotations, presents a unique opportunity for future research on high-fidelity 3D reconstruction. Furthermore, we evaluate a range of 3D reconstruction algorithms on the proposed dataset.

Reliable uncertainty from deterministic single-forward pass models is sought after because conventional methods of uncertainty quantification are computationally expensive. We take two complex single-forward-pass uncertainty approaches, DUQ and SNGP, and examine whether they mainly rely on a well-regularized feature space. Crucially, without using their more complex methods for estimating uncertainty, we find that a single softmax neural net with such a regularized feature-space, achieved via residual connections and spectral normalization, outperforms DUQ and SNGP's epistemic uncertainty predictions using simple Gaussian Discriminant Analysis post-training as a separate feature-space density estimator-without fine-tuning on OoD data, feature ensembling, or input pre-procressing. Our conceptually simple Deep Deterministic Uncertainty (DDU) baseline can also be used to disentangle aleatoric and epistemic uncertainty and performs as well as Deep Ensembles, the state-of-the art for uncertainty prediction, on several OoD bench-marks (CIFAR-10/100 vs SVHN/Tiny-ImageNet, ImageNet vs ImageNet-O), active learning settings across different model architectures, as well as in large scale vision tasks like semantic segmentation, while being computationally cheaper.

We introduce Patch Aligned Contrastive Learning (PACL), a modified compatibility function for CLIP's contrastive loss, intending to train an alignment between the patch tokens of the vision encoder and the CLS token of the text encoder. With such an alignment, a model can identify regions of an image corresponding to a given text input, and therefore transfer seamlessly to the task of open vocabulary semantic segmentation without requiring any segmentation annotations during training. Using pre-trained CLIP encoders with PACL, we are able to set the state-of-the-art on the task of open vocabulary zero-shot segmentation on 4 different segmentation benchmarks: Pascal VOC, Pascal Context, COCO Stuff and ADE20K. Furthermore, we show that PACL is also applicable to image-level predictions and when used with a CLIP backbone, provides a general improvement in zero-shot classification accuracy compared to CLIP, across a suite of 12 image classification datasets.

When deploying a machine learning model to a new environment, we often encounter the distribution shift problem - meaning the target data distribution is different from the model's training distribution. In this paper, we assume that labels are not provided for this new domain, and that we do not store the source data (e.g., for privacy reasons). It has been shown that even small shifts in the data distribution can affect the model's performance severely. Test Time Adaptation offers a means to combat this problem, as it allows the model to adapt during test time to the new data distribution, using only unlabeled test data batches. To achieve this, the predominant approach is to optimize a surrogate loss on the test-time unlabeled target data. In particular, minimizing the prediction's entropy on target samples [34] has received much interest as it is task-agnostic and does not require altering the model's training phase (e.g., does not require adding a self-supervised task during training on the source domain). However, as the target data's batch size is often small in real-world scenarios (e.g., autonomous driving models process each few frames in real-time), we argue that this surrogate loss is not optimal since it often collapses with small batch sizes. To tackle this problem, in this paper, we propose to use an invariance regularizer as the surrogate loss during test-time adaptation, motivated by our theoretical results regarding the model's performance under input transformations. The resulting method (TIPI-Test time adaPtation with transformation Invariance) is validated with extensive experiments in various benchmarks (Cifar10-C, Cifar100-C, ImageNet-C, DIGITS, and VisDA17). Remarkably, TIPI is robust against small batch sizes (as small as 2 in our experiments), and consistently outperforms TENT [34] in all settings. Our code is released at https://github.com/atuannguyen/TIPI.

Continual Learning (CL) aims to sequentially train models on streams of incoming data that vary in distribution by preserving previous knowledge while adapting to new data. Current CL literature focuses on restricted access to previously seen data, while imposing no constraints on the computational budget for training. This is unreasonable for applications in-the-wild, where systems are primarily constrained by computational and time budgets, not storage. We revisit this problem with a large-scale benchmark and analyze the performance of traditional CL approaches in a compute-constrained setting, where effective memory samples used in training can be implicitly restricted as a consequence of limited computation. We conduct experiments evaluating various CL sampling strategies, distillation losses, and partial fine-tuning on two large-scale datasets, namely ImageNet2K and Continual Google Landmarks V2 in data incremental, class incremental, and time incremental settings. Through extensive experiments amounting to a total of over 1500 GPU-hours, we find that, under compute-constrained setting, traditional CL approaches, with no exception, fail to outperform a simple minimal baseline that samples uniformly from memory. Our conclusions are consistent in a different number of stream time steps, e.g., 20 to 200, and under several computational budgets. This suggests that most existing CL methods are particularly too computationally expensive for realistic budgeted deployment. Code for this project is available at: https://github.com/drimpossible/BudgetCL.

Video-language pre-trained models have shown remarkable success in guiding video question-answering (VideoQA) tasks. However, due to the length of video sequences, training large-scale video-based models incurs considerably higher costs than training image-based ones. This motivates us to leverage the knowledge from image-based pretraining, despite the obvious gaps between image and video domains. To bridge these gaps, in this paper, we propose Tem-adapter, which enables the learning of temporal dynamics and complex semantics by a visual Temporal Aligner and a textual Semantic Aligner. Unlike conventional pretrained knowledge adaptation methods that only concentrate on the downstream task objective, the Temporal Aligner introduces an extra language-guided autoregressive task aimed at facilitating the learning of temporal dependencies, with the objective of predicting future states based on historical clues and language guidance that describes event progression. Besides, to reduce the semantic gap and adapt the textual representation for better event description, we introduce a Semantic Aligner that first designs a template to fuse question and answer pairs as event descriptions and then learns a Transformer decoder with the whole video sequence as guidance for refinement. We evaluate Tem-adapter and different pre-train transferring methods on two VideoQA benchmarks, and the significant performance improvement demonstrates the effectiveness of our method.1

Video object segmentation (VOS) aims at segmenting a particular object throughout the entire video clip sequence. The state-of-the-art VOS methods have achieved excellent performance (e.g., 90+% $\mathcal{J}$ & $\mathcal{F}$) on existing datasets. However, since the target objects in these existing datasets are usually relatively salient, dominant, and isolated, VOS under complex scenes has rarely been studied. To revisit VOS and make it more applicable in the real world, we collect a new VOS dataset called coMplex video Object SEgmentation (MOSE) to study the tracking and segmenting objects in complex scenarios. MOSE contains 2,149 video clips and 5,200 objects from 36 categories, with 431,725 high-quality object segmentation masks. The most notable feature of MOSE dataset is complex scenes with crowded and occluded objects. The target objects in the videos are commonly occluded by others and disappear in some frames. To analyze the proposed MOSE dataset, we benchmark 18 existing VOS methods under 4 different settings on the proposed MOSE dataset and conduct comprehensive comparisons. The experiments show that current VOS algorithms cannot well perceive objects in complex scenes. For example, under the semi-supervised VOS setting, the highest $\mathcal{J}$ & $\mathcal{F}$ by existing state-of-the-art VOS methods is only 59.4% on MOSE, much lower than their ∼90% $\mathcal{J}$ & $\mathcal{F}$ performance on DAVIS. The results reveal that although excellent performance has been achieved on existing benchmarks, there are unresolved challenges under complex scenes and more efforts are desired to explore these challenges in the future.

We revisit the common practice of evaluating adaptation of Online Continual Learning (OCL) algorithms through the metric of online accuracy, which measures the accuracy of the model on the immediate next few samples. However, we show that this metric is unreliable, as even vacuous blind classifiers, which do not use input images for prediction, can achieve unrealistically high online accuracy by exploiting spurious label correlations in the data stream. Our study reveals that existing OCL algorithms can also achieve high online accuracy, but perform poorly in retaining useful information, suggesting that they unintentionally learn spurious label correlations. To address this issue, we propose a novel metric for measuring adaptation based on the accuracy on the near-future samples, where spurious correlations are removed. We benchmark existing OCL approaches using our proposed metric on large-scale datasets under various computational budgets and find that better generalization can be achieved by retaining and reusing past seen information. We believe that our proposed metric can aid in the development of truly adaptive OCL methods. We provide code to reproduce our results at https://github.com/drimpossible/EvalOCL.

This paper presents OxfordTVG-HIC (Humorous Image Captions), a large-scale dataset for humour generation and understanding. Humour is an abstract, subjective, and context-dependent cognitive construct involving several cognitive factors, making it a challenging task to generate and interpret. Hence, humour generation and understanding can serve as a new task for evaluating the ability of deep-learning methods to process abstract and subjective information. Due to the scarcity of data, humourrelated generation tasks such as captioning remain underexplored. To address this gap, OxfordTVG-HIC offers approximately 2.9M image-text pairs with humour scores to train a generalizable humour captioning model. Contrary to existing captioning datasets, OxfordTVG-HIC features a wide range of emotional and semantic diversity resulting in out-of-context examples that are particularly conducive to generating humour. Moreover, OxfordTVG-HIC is curated devoid of offensive content. We also show how OxfordTVG-HIC can be leveraged for evaluating the humour of a generated text. Through explainability analysis of the trained models, we identify the visual and linguistic cues influential for evoking humour prediction (and generation). We observe qualitatively that these cues are aligned with the benign violation theory of humour in cognitive psychology.

Transformers for graph data are increasingly widely studied and successful in numerous learning tasks. Graph inductive biases are crucial for Graph Transformers, and previous works incorporate them using message-passing modules and/or positional encodings. However, Graph Transformers that use message-passing inherit known issues of message-passing, and differ significantly from Transformers used in other domains, thus making transfer of research advances more difficult. On the other hand, Graph Transformers without message-passing often perform poorly on smaller datasets, where inductive biases are more crucial. To bridge this gap, we propose the Graph Inductive bias Transformer (GRIT) --- a new Graph Transformer that incorporates graph inductive biases without using message passing. GRIT is based on several architectural changes that are each theoretically and empirically justified, including: learned relative positional encodings initialized with random walk probabilities, a flexible attention mechanism that updates node and node-pair representations, and injection of degree information in each layer. We prove that GRIT is expressive --- it can express shortest path distances and various graph propagation matrices. GRIT achieves state-of-the-art empirical performance across a variety of graph datasets, thus showing the power that Graph Transformers without message-passing can deliver.

Improving and guaranteeing the robustness of deep learning models has been a topic of intense research. Ensembling, which combines several classifiers to provide a better model, has been shown to be beneficial for generalisation, uncertainty estimation, calibration, and mitigating the effects of concept drift. However, the impact of ensembling on certified robustness is less well understood. In this work, we generalise Lipschitz continuity by introducing S-Lipschitz classifiers, which we use to analyse the theoretical robustness of ensembles. Our results are precise conditions when ensembles of robust classifiers are more robust than any constituent classifier, as well as conditions when they are less robust.

Referring expression comprehension aims to localize a natural language description in an image. Using location priors to help reduce inaccuracies in cross-modal alignments is the state of the art for CNN-based methods tackling this problem. Recent Transformer-based models cast aside this idea, making the case for steering away from hand-designed components. In this work, we propose LUNA, which uses language as continuing anchors to guide box prediction in a Transformer decoder, and thus show that language-guided location priors can be effectively exploited in a Transformer-based architecture. Our method first initializes an anchor box from the input expression via a small "proto-decoder,'' and then uses this anchor and its refined successors as location guidance in a modified Transformer decoder. At each decoder layer, the anchor box is first used as a query for gathering multi-modal context, and then updated based on the gathered context (producing the next, refined anchor). In the end, a lightweight assessment pathway evaluates the quality of all produced anchors, yielding the final prediction in a dynamic way. This approach allows box decoding to be conditioned on learned anchors, which facilitates accurate grounding, as we shown in the experiments. Our method outperforms existing state-of-the-art methods on the datasets of ReferIt Game, RefCOCO/+/g, and Flickr30K Entities.

Various adaptation methods, such as LoRA, prompts, and adapters, have been proposed to enhance the performance of pre-trained vision-language models in specific domains. As test samples in real-world applications usually differ from adaptation data, the robustness of these adaptation methods against distribution shifts are essential. In this study, we assess the robustness of 11 widely-used adaptation methods across 4 vision-language datasets under multimodal corruptions. Concretely, we introduce 7 benchmark datasets, including 96 visual and 87 textual corruptions, to investigate the robustness of different adaptation methods, the impact of available adaptation examples, and the influence of trainable parameter size during adaptation. Our analysis reveals that: 1) Adaptation methods are more sensitive to text corruptions than visual corruptions. 2) Full fine-tuning does not consistently provide the highest robustness; instead, adapters can achieve better robustness with comparable clean performance. 3) Contrary to expectations, our findings indicate that increasing the number of adaptation data and parameters does not guarantee enhanced robustness; instead, it results in even lower robustness. We hope this study could benefit future research in the development of robust multimodal adaptation methods. The benchmark, code, and dataset used in this study can be accessed at https://adarobustness.github.io.

Recent language models have shown impressive multilingual performance, even when not explicitly trained for it. Despite this, there are concerns about the quality of their outputs across different languages. In this paper, we show how disparity in the treatment of different languages arises at the tokenization stage, well before a model is even invoked. The same text translated into different languages can have drastically different tokenization lengths, with differences up to 15 times in some cases. These disparities persist even for tokenizers that are intentionally trained for multilingual support. Character-level and byte-level models also exhibit over 4 times the difference in the encoding length for some language pairs. This induces unfair treatment for some language communities in regard to the cost of accessing commercial language services, the processing time and latency, as well as the amount of content that can be provided as context to the models. Therefore, we make the case that we should train future language models using multilingually fair subword tokenizers.

This paper presents a new mechanism to facilitate the training of mask transformers for efficient panoptic segmentation, democratizing its deployment. We observe that due to the high complexity in the training objective of panoptic segmentation, it will inevitably lead to much higher penalization on false positive. Such unbalanced loss makes the training process of the end-to-end mask-transformer based architectures difficult, especially for efficient models. In this paper, we present ReMaX that adds relaxation to mask predictions and class predictions during the training phase for panoptic segmentation. We demonstrate that via these simple relaxation techniques during training, our model can be consistently improved by a clear margin without any extra computational cost on inference. By combining our method with efficient backbones like MobileNetV3-Small, our method achieves new state-of-the-art results for efficient panoptic segmentation on COCO, ADE20K and Cityscapes. Code and pre-trained checkpoints will be available at https://github.com/google-research/deeplab2.

Semantic segmentation datasets often exhibit two types of imbalance: \textit{class imbalance}, where some classes appear more frequently than others and \textit{size imbalance}, where some objects occupy more pixels than others. This causes traditional evaluation metrics to be biased towards \textit{majority classes} (e.g. overall pixel-wise accuracy) and \textit{large objects} (e.g. mean pixel-wise accuracy and per-dataset mean intersection over union). To address these shortcomings, we propose the use of fine-grained mIoUs along with corresponding worst-case metrics, thereby offering a more holistic evaluation of segmentation techniques. These fine-grained metrics offer less bias towards large objects, richer statistical information, and valuable insights into model and dataset auditing. Furthermore, we undertake an extensive benchmark study, where we train and evaluate 15 modern neural networks with the proposed metrics on 12 diverse natural and aerial segmentation datasets. Our benchmark study highlights the necessity of not basing evaluations on a single metric and confirms that fine-grained mIoUs reduce the bias towards large objects. Moreover, we identify the crucial role played by architecture designs and loss functions, which lead to best practices in optimizing fine-grained metrics. The code is available at \href{https://github.com/zifuwanggg/JDTLosses}{https://github.com/zifuwanggg/JDTLosses}.

As modern networks have been proven to be unprotected from adversarial attacks and are applied in safety-critical applications, defense against them is very crucial. Many works were dedicated to this topic, but randomized smoothing has been recently proven to be an effective approach for the certified defense of deep neural networks and getting robust classifiers. Some prior results were obtained utilizing the techniques of adding extra parameters to extend the limits of the certification regions. In this way, sample-wise optimization was proposed to maximize the certification radius per input. The idea was further extended with the generalized anisotropic counterparts of ℓ1and ℓ2certificates which allow achieving larger certified region volume avoiding worst-case certification near potentially larger safe regions. However, anisotropic certification is limited by the aligned axis lacking the freedom to extend in any direction. To mitigate this constraint, in this work, we (i) revisit the anisotropic certification, provide an analysis of its non-axis aligned counterpart and propose its rotation-free extension, (ii) conduct experiments on the CIFAR-10 dataset to report the improved performance.

Abstract Remote photoplethysmography (rPPG), which aims at measuring heart activities and physiological signals from facial video without any contact, has great potential in many applications (e.g., remote healthcare and affective computing). Recent deep learning approaches focus on mining subtle rPPG clues using convolutional neural networks with limited spatio-temporal receptive fields, which neglect the long-range spatio-temporal perception and interaction for rPPG modeling. In this paper, we propose two end-to-end video transformer based architectures, namely PhysFormer and PhysFormer++, to adaptively aggregate both local and global spatio-temporal features for rPPG representation enhancement. As key modules in PhysFormer, the temporal difference transformers first enhance the quasi-periodic rPPG features with temporal difference guided global attention, and then refine the local spatio-temporal representation against interference. To better exploit the temporal contextual and periodic rPPG clues, we also extend the PhysFormer to the two-pathway SlowFast based PhysFormer++ with temporal difference periodic and cross-attention transformers. Furthermore, we propose the label distribution learning and a curriculum learning inspired dynamic constraint in frequency domain, which provide elaborate supervisions for PhysFormer and PhysFormer++ and alleviate overfitting. Comprehensive experiments are performed on four benchmark datasets to show our superior performance on both intra- and cross-dataset testings. Unlike most transformer networks needed pretraining from large-scale datasets, the proposed PhysFormer family can be easily trained from scratch on rPPG datasets, which makes it promising as a novel transformer baseline for the rPPG community.

In this paper, we address the task of semantic-guided image generation. One challenge common to most existing image-level generation methods is the difficulty in generating small objects and detailed local textures. To address this, in this work we consider generating images using local context. As such, we design a local class-specific generative network using semantic maps as guidance, which separately constructs and learns subgenerators for different classes, enabling it to capture finer details. To learn more discriminative class-specific feature representations for the local generation, we also propose a novel classification module. To combine the advantages of both global image-level and local class-specific generation, a joint generation network is designed with an attention fusion module and a dual-discriminator structure embedded. Lastly, we propose a novel semantic-aware upsampling method, which has a larger receptive field and can take far-away pixels that are semantically related for feature upsampling, enabling it to better preserve semantic consistency for instances with the same semantic labels. Extensive experiments on two image generation tasks show the superior performance of the proposed method. State-of-the-art results are established by large margins on both tasks and on nine challenging public benchmarks. The source code and trained models are available at https://github.com/Ha0Tang/LGGAN.

Self-attention mechanisms and non-local blocks have become crucial building blocks for state-of-the-art neural architectures thanks to their unparalleled ability in capturing long-range dependencies in the input. However their cost is quadratic with the number of spatial positions hence making their use impractical in many real case applications. In this work, we analyze these methods through a polynomial lens, and we show that self-attention can be seen as a special case of a 3 rd order polynomial. Within this polynomial framework, we are able to design polynomial operators capable of accessing the same data pattern of non-local and self-attention blocks while reducing the complexity from quadratic to linear. As a result, we propose two modules (Poly-NL and Poly-SA) that can be used as "drop-in" replacements for more-complex non-local and self-attention layers in state-of-the-art CNNs and ViT architectures. Our modules can achieve comparable, if not better, performance across a wide range of computer vision tasks while keeping a complexity equivalent to a standard linear layer.

The goal of this special section is to provide a platform to summarize what we have achieved and where we are moving toward this methodological research direction. It gathers the latest advances in learning with graph-structured data in computer vision, as well as interdisciplinary efforts on graph-based pattern analysis in sociology, physics, chemistry, finance, biology, etc.

Empowered by large datasets, e.g., ImageNet and MS COCO, unsupervised learning on large-scale data has enabled significant advances for classification tasks. However, whether the large-scale unsupervised semantic segmentation can be achieved remains unknown. There are two major challenges: i) we need a large-scale benchmark for assessing algorithms; ii) we need to develop methods to simultaneously learn category and shape representation in an unsupervised manner. In this work, we propose a new problem of large-scale unsupervised semantic segmentation (LUSS) with a newly created benchmark dataset to help the research progress. Building on the ImageNet dataset, we propose the ImageNet-S dataset with 1.2 million training images and 50k high-quality semantic segmentation annotations for evaluation. Our benchmark has a high data diversity and a clear task objective. We also present a simple yet effective method that works surprisingly well for LUSS. In addition, we benchmark related un/weakly/fully supervised methods accordingly, identifying the challenges and possible directions of LUSS. The benchmark and source code is publicly available at https://github.com/LUSSeg.

In this article, we introduce SiamMask, a framework to perform both visual object tracking and video object segmentation, in real-time, with the same simple method. We improve the offline training procedure of popular fully-convolutional Siamese approaches by augmenting their losses with a binary segmentation task. Once the offline training is completed, SiamMask only requires a single bounding box for initialization and can simultaneously carry out visual object tracking and segmentation at high frame-rates. Moreover, we show that it is possible to extend the framework to handle multiple object tracking and segmentation by simply re-using the multi-task model in a cascaded fashion. Experimental results show that our approach has high processing efficiency, at around 55 frames per second. It yields real-time state-of-the art results on visual-object tracking benchmarks, while at the same time demonstrating competitive performance at a high speed for video object segmentation benchmarks.

We introduce a new image segmentation task, called Entity Segmentation (ES), which aims to segment all visual entities (objects and stuffs) in an image without predicting their semantic labels. By removing the need of class label prediction, the models trained for such task can focus more on improving segmentation quality. It has many practical applications such as image manipulation and editing where the quality of segmentation masks is crucial but class labels are less important. We conduct the first-ever study to investigate the feasibility of convolutional center-based representation to segment things and stuffs in a unified manner, and show that such representation fits exceptionally well in the context of ES. More specifically, we propose a CondInst-like fully-convolutional architecture with two novel modules specifically designed to exploit the class-agnostic and non-overlapping requirements of ES. Experiments show that the models designed and trained for ES significantly outperforms popular class-specific panoptic segmentation models in terms of segmentation quality. Moreover, an ES model can be easily trained on a combination of multiple datasets without the need to resolve label conflicts in dataset merging, and the model trained for ES on one or more datasets can generalize very well to other test datasets of unseen domains. The code has been released at https://github.com/dvlab-research/Entity.

The computational complexity of transformers limits it to be widely deployed onto frameworks for visual recognition. Recent work [9] significantly accelerates the network processing speed by reducing the resolution at the beginning of the network, however, it is still hard to be directly generalized onto other downstream tasks e.g. object detection and segmentation like CNN. In this paper, we present a transformer-based architecture retaining both the local and global interactions within the network, and can be transferable to other downstream tasks. The proposed architecture reforms the original full spatial self-attention into pixel-wise local attention and patch-wise global attention. Such factorization saves the computational cost while retaining the information of different granularities, which helps generate multi-scale features required by different tasks. By exploiting the factorized attention, we construct a Separable Transformer (SeT) for visual modeling. Experimental results show that SeT outperforms the previous state-of-the-art transformer-based approaches and its CNN counterparts on three major tasks including image classification, object detection and instance segmentation.

Recurrent models are a popular choice for video enhancement tasks such as video denoising or super-resolution. In this work, we focus on their stability as dynamical systems and show that they tend to fail catastrophically at inference time on long video sequences. To address this issue, we (1) introduce a diagnostic tool which produces input sequences optimized to trigger instabilities and that can be interpreted as visualizations of temporal receptive fields, and (2) propose two approaches to enforce the stability of a model during training: constraining the spectral norm or constraining the stable rank of its convolutional layers. We then introduce Stable Rank Normalization for Convolutional layers (SRN-C), a new algorithm that enforces these constraints. Our experimental results suggest that SRN-C successfully enforces stablility in recurrent video processing models without a significant performance loss.

We propose a novel model named Multi-Channel Attention Selection Generative Adversarial Network (SelectionGAN) for guided image-to-image translation, where we translate an input image into another while respecting an external semantic guidance. The proposed SelectionGAN explicitly utilizes the semantic guidance information and consists of two stages. In the first stage, the input image and the conditional semantic guidance are fed into a cycled semantic-guided generation network to produce initial coarse results. In the second stage, we refine the initial results by using the proposed multi-scale spatial pooling & channel selection module and the multi-channel attention selection module. Moreover, uncertainty maps automatically learned from attention maps are used to guide the pixel loss for better network optimization. Exhaustive experiments on four challenging guided image-to-image translation tasks (face, hand, body, and street view) demonstrate that our SelectionGAN is able to generate significantly better results than the state-of-the-art methods. Meanwhile, the proposed framework and modules are unified solutions and can be applied to solve other generation tasks such as semantic image synthesis. The code is available at https://github.com/Ha0Tang/SelectionGAN.

This article presents Holistically-Attracted Wireframe Parsing (HAWP), a method for geometric analysis of 2D images containing wireframes formed by line segments and junctions. HAWP utilizes a parsimonious Holistic Attraction (HAT) field representation that encodes line segments using a closed-form 4D geometric vector field. The proposed HAWP consists of three sequential components empowered by end-to-end and HAT-driven designs: 1) generating a dense set of line segments from HAT fields and endpoint proposals from heatmaps, 2) binding the dense line segments to sparse endpoint proposals to produce initial wireframes, and 3) filtering false positive proposals through a novel endpoint-decoupled line-of-interest aligning (EPD LOIAlign) module that captures the co-occurrence between endpoint proposals and HAT fields for better verification. Thanks to our novel designs, HAWPv2 shows strong performance in fully supervised learning, while HAWPv3 excels in self-supervised learning, achieving superior repeatability scores and efficient training (24 GPU hours on a single GPU). Furthermore, HAWPv3 exhibits a promising potential for wireframe parsing in out-of-distribution images without providing ground truth labels of wireframes.

Modelling long-range dependencies is critical for scene understanding tasks in computer vision. Although convolution neural networks (CNNs) have excelled in many vision tasks, they are still limited in capturing long-range structured relationships as they typically consist of layers of local kernels. A fully-connected graph, such as the self-attention operation in Transformers, is beneficial for such modelling, however, its computational overhead is prohibitive. In this paper, we propose a dynamic graph message passing network, that significantly reduces the computational complexity compared to related works modelling a fully-connected graph. This is achieved by adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically predict node-dependent filter weights and the affinity matrix for propagating information between them. This formulation allows us to design a self-attention module, and more importantly a new Transformer-based backbone network, that we use for both image classification pretraining, and for addressing various downstream tasks (e.g. object detection, instance and semantic segmentation). Using this model, we show significant improvements with respect to strong, state-of-the-art baselines on four different tasks. Our approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters. Code and models will be made publicly available at https://github.com/fudan-zvg/DGMN2.

State-of-the-art methods in the image-to-image translation are capable of learning a mapping from a source domain to a target domain with unpaired image data. Though the existing methods have achieved promising results, they still produce visual artifacts, being able to translate low-level information but not high-level semantics of input images. One possible reason is that generators do not have the ability to perceive the most discriminative parts between the source and target domains, thus making the generated images low quality. In this article, we propose a new Attention-Guided Generative Adversarial Networks (AttentionGAN) for the unpaired image-to-image translation task. AttentionGAN can identify the most discriminative foreground objects and minimize the change of the background. The attention-guided generators in AttentionGAN are able to produce attention masks, and then fuse the generation output with the attention masks to obtain high-quality target images. Accordingly, we also design a novel attention-guided discriminator which only considers attended regions. Extensive experiments are conducted on several generative tasks with eight public datasets, demonstrating that the proposed method is effective to generate sharper and more realistic images compared with existing competitive models. The code is available at https://github.com/Ha0Tang/AttentionGAN.

Deep neural networks are vulnerable to input deformations in the form of vector fields of pixel displacements and to other parameterized geometric deformations e.g. translations, rotations, etc. Current input deformation certification methods either (i) do not scale to deep networks on large input datasets, or (ii) can only certify a specific class of deformations, e.g. only rotations. We reformulate certification in randomized smoothing setting for both general vector field and parameterized deformations and propose DeformRS-VF and DeformRS-Par, respectively. Our new formulation scales to large networks on large input datasets. For instance, DeformRS-Par certifies rich deformations, covering translations, rotations, scaling, affine deformations, and other visually aligned deformations such as ones parameterized by Discrete-Cosine-Transform basis. Extensive experiments on MNIST, CIFAR10, and ImageNet show competitive performance of DeformRS-Par achieving a certified accuracy of 39\% against perturbed rotations in the set [-10 degree, 10 degree] on ImageNet.

Deep neural networks are vulnerable to small input perturbations known as adversarial attacks. Inspired by the fact that these adversaries are constructed by iteratively minimizing the confidence of a network for the true class label, we propose the anti-adversary layer, aimed at countering this effect. In particular, our layer generates an input perturbation in the opposite direction of the adversarial one and feeds the classifier a perturbed version of the input. Our approach is training-free and theoretically supported. We verify the effectiveness of our approach by combining our layer with both nominally and robustly trained models and conduct large-scale experiments from black-box to adaptive attacks on CIFAR10, CIFAR100, and ImageNet. Our layer significantly enhances model robustness while coming at no cost on clean accuracy.

Naive approaches to amortized inference in probabilistic programs with unbounded loops can produce estimators with infinite variance. This is particularly true of importance sampling inference in programs that explicitly include rejection sampling as part of the user-programmed generative procedure. In this paper we develop a new and efficient amortized importance sampling estimator. We prove finite variance of our estimator and empirically demonstrate our method’s correctness and efficiency compared to existing alternatives on generative programs containing rejection sampling loops and discuss how to implement our method in a generic probabilistic programming framework.

Many video understanding tasks require analyzing multishot videos, but existing datasets for video object segmentation (VOS) only consider single-shot videos. To address this challenge, we collected a new dataset-YouMVaS-of 200 popular YouTube videos spanning ten genres, where each video is on average five minutes long and with 75 shots. We selected recurring actors and annotated 431K segmentation masks at a frame rate of six, exceeding previous datasets in average video duration, object variation, and narrative structure complexity. We incorporated good practices of model architecture design, memory management, and multi-shot tracking into an existing video segmentation method to build competitive baseline methods. Through error analysis, we found that these baselines still fail to cope with cross-shot appearance variation on our YouMVOS dataset. Thus, our dataset poses new challenges in multi-shot segmentation towards better video analysis. Data, code, and pre-trained models are available at https://donglaiw.github.io/proj/youMVOS

Referring image segmentation is a fundamental vision-language task that aims to segment out an object referred to by a natural language expression from an image. One of the key challenges behind this task is leveraging the referring expression for highlighting relevant positions in the image. A paradigm for tackling this problem is to leverage a powerful vision-language (“cross-madal”) decoder to fuse features independently extracted from a vision encoder and a language encoder. Recent methods have made remarkable advancements in this paradigm by exploiting Transformers as cross-modal decoders, concurrent to the Transformer's overwhelming success in many other vision-language tasks. Adopting a different approach in this work, we show that significantly better cross-modal alignments can be achieved through the early fusion of linguistic and visual features in intermediate layers of a vision Transformer encoder network. By conducting cross-modal feature fusion in the visual feature encoding stage, we can leverage the well-proven correlation modeling power of a Transformer encoder for excavating helpful multi-modal context. This way, accurate segmentation results are readily harvested with a light-weight mask predictor. Without bells and whistles, our method surpasses the previous state-of-the-art methods on Ref CoCo, RefCOCO+, and G-Ref by large margins.

Mixup-based augmentation has been found to be effective for generalizing models during training, especially for Vision Transformers (ViTs) since they can easily overfit. However, previous mixup-based methods have an underlying prior knowledge that the linearly interpolated ratio of targets should be kept the same as the ratio proposed in input interpolation. This may lead to a strange phenomenon that sometimes there is no valid object in the mixed image due to the random process in augmentation but there is still response in the label space. To bridge such gap between the input and label spaces, we propose TransMix, which mixes labels based on the attention maps of Vision Transformers. The confidence of the label will be larger if the corresponding input image is weighted higher by the attention map. TransMix is embarrassingly simple and can be implemented in just a few lines of code without introducing any extra parameters and FLOPs to ViT-based models. Experimental results show that our method can consistently improve various ViT-based models at scales on ImageNet classification. After pre-trained with TransMix on ImageNet, the ViT-based models also demonstrate better transferability to semantic segmentation, object detection and instance segmentation. TransMix also exhibits to be more robust when evaluating on 4 different benchmarks. Code is publicly available at https://github.com/Beckschen/TransMix.

Dense 3D reconstruction from a stream of depth images is the key to many mixed reality and robotic applications. Although methods based on Truncated Signed Distance Function (TSDF) Fusion have advanced the field over the years, the TSDF volume representation is confronted with striking a balance between the robustness to noisy measurements and maintaining the level of detail. We present Bi-level Neural Volume Fusion (BNV-Fusion), which leverages recent advances in neural implicit representations and neural rendering for dense 3D reconstruction. In order to incrementally integrate new depth maps into a global neural implicit representation, we propose a novel bi-level fusion strategy that considers both efficiency and reconstruction quality by design. We evaluate the proposed method on multiple datasets quantitatively and qualitatively, demonstrating a significant improvement over existing methods.

Class Incremental Learning (CIL) aims at learning a classifier in a phase-by-phase manner, in which only data of a subset of the classes are provided at each phase. Previous works mainly focus on mitigating forgetting in phases after the initial one. However, we find that improving CIL at its initial phase is also a promising direction. Specifically, we experimentally show that directly encouraging CIL Learner at the initial phase to output similar representations as the model jointly trained on all classes can greatly boost the CIL performance. Motivated by this, we study the differ-ence between a naively-trained initial-phase model and the oracle model. Specifically, since one major difference be-tween these two models is the number of training classes, we investigate how such difference affects the model rep-resentations. We find that, with fewer training classes, the data representations of each class lie in a long and narrow region; with more training classes, the representations of each class scatter more uniformly. Inspired by this obser-vation, we propose Class-wise Decorrelation (CwD) that ef-fectively regularizes representations of each class to scatter more uniformly, thus mimicking the model jointly trained with all classes (i.e., the oracle model). Our CwD is simple to implement and easy to plug into existing methods. Ex-tensive experiments on various benchmark datasets show that CwD consistently and significantly improves the per-formance of existing state-of-the-art methods by around 1% to 3%. Code: https://github.com/Yujun-Shi/CwD.

The resurgence of unsupervised learning can be attributed to the remarkable progress of self-supervised learning, which includes generative$(\mathcal{G})$and discriminative$(\mathcal{D})$models. In computer vision, the mainstream self-supervised learning algorithms are$\mathcal{D}$models. However, designing a$\mathcal{D}$model could be over-complicated; also, some studies hinted that a$\mathcal{D}$model might not be as general and interpretable as a$\mathcal{G}$model. In this paper, we switch from$\mathcal{D}$models to$\mathcal{G}$models using the classical auto-encoder$(AE)$. Note that a vanilla$\mathcal{G}$model was far less efficient than a$\mathcal{D}$model in self-supervised computer vision tasks, as it wastes model capability on overfitting semantic-agnostic high-frequency details. Inspired by perceptual learning that could use cross-view learning to perceive concepts and semantics11Following [26], we refer to semantics as visual concepts, e.g., a semantic-ware model indicates the model can perceive visual concepts, and the learned features are efficient in object recognition, detection, etc., we propose a novel$AE$that could learn semantic-aware representation via cross-view image reconstruction. We use one view of an image as the input and another view of the same image as the reconstruction target. This kind of$AE$has rarely been studied before, and the optimization is very difficult. To enhance learning ability and find a feasible solution, we propose a semantic aligner that uses geometric transformation knowledge to align the hidden code of$AE$to help optimization. These techniques significantly improve the representation learning ability of$AE$and make selfsupervised learning with$\mathcal{G}$models possible. Extensive experiments on many large-scale benchmarks (e.g., ImageNet, COCO 2017, and SYSU-30k) demonstrate the effectiveness of our methods. Code is available at https://github.com/wanggrun/Semantic-Aware-AE.

Remote photoplethysmography (rPPG), which aims at measuring heart activities and physiological signals from facial video without any contact, has great potential in many applications. Recent deep learning approaches focus on mining subtle rPPG clues using convolutional neural networks with limited spatio-temporal receptive fields, which neglect the long-range spatio-temporal perception and interaction for rPPG modeling. In this paper, we propose the PhysFormer, an end-to-end video transformer based architecture, to adaptively aggregate both local and global spatio-temporal features for rPPG representation enhancement. As key modules in PhysFormer, the temporal difference transformers first enhance the quasi-periodic rPPG features with temporal difference guided global attention, and then refine the local spatio-temporal representation against interference. Furthermore, we also propose the label distribution learning and a curriculum learning inspired dynamic constraint in frequency domain, which provide elaborate supervisions for PhysFormer and alleviate overfitting. Comprehensive experiments are performed on four benchmark datasets to show our superior performance on both intra- and cross-dataset testings. One highlight is that, unlike most transformer networks needed pretraining from large-scale datasets, the proposed PhysFormer can be easily trained from scratch on rPPG datasets, which makes it promising as a novel transformer baseline for the rPPG community. The codes are available at https://github.com/ZitongYu/PhysFormer.

We propose ADIOS, a masked image model (MIM) framework for self-supervised learning, which simultaneously learns a masking function and an image encoder using an adversarial objective. The image encoder is trained to minimise the distance between representations of the original and that of a masked image. The masking function, conversely, aims at maximising this distance. ADIOS consistently improves on state-of-the-art self-supervised learning (SSL) methods on a variety of tasks and datasets—including classification on ImageNet100 and STL10, transfer learning on CIFAR10/100, Flowers102 and iNaturalist, as well as robustness evaluated on the backgrounds challenge (Xiao et al., 2021)—while generating semantically meaningful masks. Unlike modern MIM models such as MAE, BEiT and iBOT, ADIOS does not rely on the image-patch tokenisation construction of Vision Transformers, and can be implemented with convolutional backbones. We further demonstrate that the masks learned by ADIOS are more effective in improving representation learning of SSL methods than masking schemes used in popular MIM models.

We consider the problem of communicating exogenous information by means of Markov decision process trajectories. This setting, which we call a Markov coding game (MCG), generalizes both source coding and a large class of referential games. MCGs also isolate a problem that is important in decentralized control settings in which cheap-talk is not available—namely, they require balancing communication with the associated cost of communicating. We contribute a theoretically grounded approach to MCGs based on maximum entropy reinforcement learning and minimum entropy coupling that we call MEME. Due to recent breakthroughs in approximation algorithms for minimum entropy coupling, MEME is not merely a theoretical algorithm, but can be applied to practical settings. Empirically, we show both that MEME is able to outperform a strong baseline on small MCGs and that MEME is able to achieve strong performance on extremely large MCGs. To the latter point, we demonstrate that MEME is able to losslessly communicate binary images via trajectories of Cartpole and Pong, while simultaneously achieving the maximal or near maximal expected returns, and that it is even capable of performing well in the presence of actuator noise.

Semantic-descriptor-based Generalized Zero-Shot Learning (GZSL) poses challenges in recognizing novel classes in the test phase. The development of generative models enables current GZSL techniques to probe further into the semantic-visual link, culminating in a two-stage form that includes a generator and a classifier. However, existing generation-based methods focus on enhancing the generator's effect while neglecting the improvement of the classifier. In this paper, we first analyze of two properties of the generated pseudo unseen samples: bias and homogeneity. Then, we perform variational Bayesian inference to back-derive the evaluation metrics, which reflects the balance of the seen and unseen classes. As a consequence of our derivation, the aforementioned two properties are incorporated into the classifier training as seen-unseen priors via logit adjustment. The Zero-Shot Logit Adjustment further puts semantic-based classifiers into effect in generation-based GZSL. Our experiments demonstrate that the proposed technique achieves state-of-the-art when combined with the basic generator, and it can improve various generative Zero-Shot Learning frameworks. Our codes are available on https://github.com/cdb342/IJCAI-2022-ZLA.

Learning to hash pictures a list-wise sorting problem. Its testing metrics, e.g., mean-average precision, count on a sorted candidate list ordered by pair-wise code similarity. However, scarcely does one train a deep hashing model with the sorted results end-to-end because of the non-differentiable nature of the sorting operation. This inconsistency in the objectives of training and test may lead to sub-optimal performance since the training loss often fails to reflect the actual retrieval metric. In this paper, we tackle this problem by introducing Naturally-Sorted Hashing (NSH). We sort the Hamming distances of samples' hash codes and accordingly gather their latent representations for self-supervised training. Thanks to the recent advances in differentiable sorting approximations, the hash head receives gradients from the sorter so that the hash encoder can be optimized along with the training procedure. Additionally, we describe a novel Sorted Noise-Contrastive Estimation (SortedNCE) loss that selectively picks positive and negative samples for contrastive learning, which allows NSH to mine data semantic relations during training in an unsupervised manner. Our extensive experiments show the proposed NSH model significantly outperforms the existing unsupervised hashing methods on three benchmarked datasets.

Story visualization aims to generate a series of images, semantically matching a given sequence of sentences, one for each, and different output images within a story should be consistent with each other. Current methods generate story images by using a heavy architecture with two generative adversarial networks (GANs), one for image quality, and one for story consistency, and also rely on additional segmentation masks or auxiliary captioning networks. In this paper, we aim to build a concise and single-GAN-based network, neither depending on additional semantic information nor captioning networks. To achieve this, we propose a contrastive-learning- and clustering-learning-based approach for story visualization. Our network utilizes contrastive losses between language and visual information to maximize the mutual information between them, and further extends it with clustering learning in the training process to capture semantic similarity across modalities. So, the discriminator in our approach provides comprehensive feedback to the generator, regarding both image quality and story consistency at the same time, allowing to have a single-GAN-based network to produce high-quality synthetic results. Extensive experiments on two datasets demonstrate that our single-GAN-based network has a smaller number of total parameters in the network, but achieves a major step up from previous methods, which improves FID from 78.64 to 39.17, and FSD from 94.53 to 41.18 on Pororo-SV, and establishes a strong benchmark FID of 76.51 and FSD of 19.74 on Abstract Scenes.

Recently, Wong et al. (2020) showed that adversarial training with single-step FGSM leads to a characteristic failure mode named catastrophic overfitting (CO), in which a model becomes suddenly vulnerable to multi-step attacks. Experimentally they showed that simply adding a random perturbation prior to FGSM (RS-FGSM) could prevent CO. However, Andriushchenko & Flammarion (2020) observed that RS-FGSM still leads to CO for larger perturbations, and proposed a computationally expensive regularizer (GradAlign) to avoid it. In this work, we methodically revisit the role of noise and clipping in single-step adversarial training. Contrary to previous intuitions, we find that using a stronger noise around the clean sample combined with \textit{not clipping} is highly effective in avoiding CO for large perturbation radii. We then propose Noise-FGSM (N-FGSM) that, while providing the benefits of single-step adversarial training, does not suffer from CO. Empirical analyses on a large suite of experiments show that N-FGSM is able to match or surpass the performance of previous state of-the-art GradAlign while achieving 3$\times$ speed-up.

3D Garment modeling is a critical and challenging topic in the area of computer vision and graphics, with increasing attention focused on garment representation learning, garment reconstruction, and controllable garment manipulation, whereas existing methods were constrained to model garments under specific categories or with relatively simple topologies. In this paper, we propose a novel Neural Sewing Machine (NSM), a learning-based framework for structure-preserving 3D garment modeling, which is capable of learning representations for garments with diverse shapes and topologies and is successfully applied to 3D garment reconstruction and controllable manipulation. To model generic garments, we first obtain sewing pattern embedding via a unified sewing pattern encoding module, as the sewing pattern can accurately describe the intrinsic structure and the topology of the 3D garment. Then we use a 3D garment decoder to decode the sewing pattern embedding into a 3D garment using the UV-position maps with masks. To preserve the intrinsic structure of the predicted 3D garment, we introduce an inner-panel structure-preserving loss, an inter-panel structure-preserving loss, and a surface-normal loss in the learning process of our framework. We evaluate NSM on the public 3D garment dataset with sewing patterns with diverse garment shapes and categories. Extensive experiments demonstrate that the proposed NSM is capable of representing 3D garments under diverse garment shapes and topologies, realistically reconstructing 3D garments from 2D images with the preserved structure, and accurately manipulating the 3D garment categories, shapes, and topologies, outperforming the state-of-the-art methods by a clear margin.

We study how an autonomous agent learns to perform a task from demonstrations in a different domain, such as a different environment or different agent. Such cross-domain imitation learning is required to, for example, train an artificial agent from demonstrations of a human expert. We propose a scalable framework that enables cross-domain imitation learning without access to additional demonstrations or further domain knowledge. We jointly train the learner agent's policy and learn a mapping between the learner and expert domains with adversarial training. We effect this by using a mutual information criterion to find an embedding of the expert's state space that contains task-relevant information and is invariant to domain specifics. This step significantly simplifies estimating the mapping between the learner and expert domains and hence facilitates end-to-end learning. We demonstrate successful transfer of policies between considerably different domains, without extra supervision such as additional demonstrations, and in situations where other methods fail.

Federated Learning (FL) refers to the decentralized and privacy-preserving machine learning framework in which multiple clients collaborate (with the help of a central server) to train a global model without sharing their data. However, most existing FL methods only focus on maximizing the model's performance on the source clients' data (e.g., mobile users) without considering its generalization ability to unknown target data (e.g., a new user). In this paper, we incorporate the problem of Domain Generalization (DG) into Federated Learning to tackle the aforementioned issue. However, virtually all existing DG methods require a centralized setting where data is shared across the domains, which violates the principles of decentralized FL and hence not applicable. To this end, we propose a simple yet novel representation learning framework, namely FedSR, which enables domain generalization while still respecting the decentralized and privacy-preserving natures of this FL setting. Motivated by classical machine learning algorithms, we aim to learn a simple representation of the data for better generalization. In particular, we enforce an L2-norm regularizer on the representation and a conditional mutual information (between the representation and the data given the label) regularizer to encourage the model to only learn essential information (while ignoring spurious correlations such as the background). Furthermore, we provide theoretical connections between the above two objectives and representation alignment in domain generalization. Extensive experimental results suggest that our method significantly outperforms relevant baselines in this particular problem.

We show that the effectiveness of the well celebrated Mixup can be further improved if instead of using it as the sole learning objective, it is utilized as an additional regularizer to the standard cross-entropy loss. This simple change not only improves accuracy but also significantly improves the quality of the predictive uncertainty estimation of Mixup in most cases under various forms of covariate shifts and out-of-distribution detection experiments. In fact, we observe that Mixup otherwise yields much degraded performance on detecting out-of-distribution samples possibly, as we show empirically, due to its tendency to learn models exhibiting high-entropy throughout; making it difficult to differentiate in-distribution samples from out-of-distribution ones. To show the efficacy of our approach (RegMixup), we provide thorough analyses and experiments on vision datasets (ImageNet & CIFAR-10/100) and compare it with a suite of recent approaches for reliable uncertainty estimation.

Binary pointwise labels (aka implicit feedback) are heavily leveraged by deep learning based recommendation algorithms nowadays. In this paper we discuss the limited expressiveness of these labels may fail to accommodate varying degrees of user preference, and thus lead to conflicts during model training, which we call annotation bias. To solve this issue, we find the soft-labeling property of pairwise labels could be utilized to alleviate the bias of pointwise labels. To this end, we propose a momentum contrast framework (\method ) that combines pointwise and pairwise learning for recommendation. \method has a three-tower network structure: one user network and two item networks. The two item networks are used for computing pointwise and pairwise loss respectively. To alleviate the influence of the annotation bias, we perform a momentum update to ensure a consistent item representation. Extensive experiments on real-world datasets demonstrate the superiority of our method against state-of-the-art recommendation algorithms.

Randomized smoothing is a recent technique that achieves state-of-art performance in training certifiably robust deep neural networks. While the smoothing family of distributions is often connected to the choice of the norm used for certification, the parameters of these distributions are always set as global hyper parameters independent from the input data on which a network is certified. In this work, we revisit Gaussian randomized smoothing and show that the variance of the Gaussian distribution can be optimized at each input so as to maximize the certification radius for the construction of the smooth classifier. Since the data dependent classifier does not directly enjoy sound certification with existing approaches, we propose a memory-enhanced data dependent smooth classifier that is certifiable by construction. This new approach is generic, parameter-free, and easy to implement. In fact, we show that our data dependent framework can be seamlessly incorporated into 3 randomized smoothing approaches, leading to consistent improved certified accuracy. When this framework is used in the training routine of these approaches followed by a data dependent certification, we achieve 9% and 6% improvement over the certified accuracy of the strongest baseline for a radius of 0.5 on CIFAR10 and ImageNet.

Abstract Can our video understanding systems perceive objects when a heavy occlusion exists in a scene? To answer this question, we collect a large-scale dataset called OVIS for occluded video instance segmentation, that is, to simultaneously detect, segment, and track instances in occluded scenes. OVIS consists of 296k high-quality instance masks from 25 semantic categories, where object occlusions usually occur. While our human vision systems can understand those occluded instances by contextual reasoning and association, our experiments suggest that current video understanding systems cannot. On the OVIS dataset, the highest AP achieved by state-of-the-art algorithms is only 16.3, which reveals that we are still at a nascent stage for understanding objects, instances, and videos in a real-world scenario. We also present a simple plug-and-play module that performs temporal feature calibration to complement missing object cues caused by occlusion. Built upon MaskTrack R-CNN and SipMask, we obtain a remarkable AP improvement on the OVIS dataset. The OVIS dataset and project code are available at http://songbai.site/ovis .

In this paper, we propose a geometric neural network with edge-aware refinement (GeoNet++) to jointly predict both depth and surface normal maps from a single image. Building on top of two-stream CNNs, GeoNet++ captures the geometric relationships between depth and surface normals with the proposed depth-to-normal and normal-to-depth modules. In particular, the "depth-to-normal" module exploits the least square solution of estimating surface normals from depth to improve their quality, while the "normal-to-depth" module refines the depth map based on the constraints on surface normals through kernel regression. Boundary information is exploited via an edge-aware refinement module. GeoNet++ effectively predicts depth and surface normals with high 3D consistency and sharp boundaries resulting in better reconstructed 3D scenes. Note that GeoNet++ is generic and can be used in other depth/normal prediction frameworks to improve 3D reconstruction quality and pixel-wise accuracy of depth and surface normals. Furthermore, we propose a new 3D geometric metric (3DGM) for evaluating depth prediction in 3D. In contrast to current metrics that focus on evaluating pixel-wise error/accuracy, 3DGM measures whether the predicted depth can reconstruct high quality 3D surface normals. This is a more natural metric for many 3D application domains. Our experiments on NYUD-V2 [1] and KITTI [2] datasets verify that GeoNet++ produces fine boundary details and the predicted depth can be used to reconstruct high quality 3D surfaces.

In continual learning, the learner faces a stream of data whose distribution changes over time. Modern neural networks are known to suffer under this setting, as they quickly forget previously acquired knowledge. To address such catastrophic forgetting, many continual learning methods implement different types of experience replay, re-learning on past data stored in a small buffer known as episodic memory. In this work, we complement experience replay with a new objective that we call ``anchoring'', where the learner uses bilevel optimization to update its knowledge on the current task, while keeping intact the predictions on some anchor points of past tasks. These anchor points are learned using gradient-based optimization to maximize forgetting, which is approximated by fine-tuning the currently trained model on the episodic memory of past tasks. Experiments on several supervised learning benchmarks for continual learning demonstrate that our approach improves the standard experience replay in terms of both accuracy and forgetting metrics and for various sizes of episodic memory.

Quantizing large Neural Networks (NN) while maintaining the performance is highly desirable for resource-limited devices due to reduced memory and time complexity. It is usually formulated as a constrained optimization problem and optimized via a modified version of gradient descent. In this work, by interpreting the continuous parameters (unconstrained) as the dual of the quantized ones, we introduce a Mirror Descent (MD) framework for NN quantization. Specifically, we provide conditions on the projections (i.e., mapping from continuous to quantized ones) which would enable us to derive valid mirror maps and in turn the respective MD updates. Furthermore, we present a numerically stable implementation of MD that requires storing an additional set of auxiliary variables (unconstrained), and show that it is strikingly analogous to the Straight Through Estimator (STE) based method which is typically viewed as a “trick” to avoid vanishing gradients issue. Our experiments on CIFAR-10/100, TinyImageNet, and ImageNet classification datasets with VGG-16, ResNet-18, and MobileNetV2 architectures show that our MD variants yield state-of-the-art performance.

Current developments in temporal event or action localization usually target actions captured by a single camera. However, extensive events or actions in the wild may be captured as a sequence of shots by multiple cameras at different positions. In this paper, we propose a new and challenging task called multi-shot temporal event localization, and accordingly, collect a large-scale dataset called MUlti-Shot EventS (MUSES). MUSES has 31,477 event instances for a total of 716 video hours. The core nature of MUSES is the frequent shot cuts, for an average of 19 shots per instance and 176 shots per video, which induces large intra-instance variations. Our comprehensive evaluations show that the state-of-the-art method in temporal action localization only achieves an mAP of 13.1% at IoU=0.5. As a minor contribution, we present a simple baseline approach for handling the intra-instance variations, which reports an mAP of 18.9% on MUSES and 56.9% on THUMOS14 at IoU=0.5. To facilitate research in this direction, we release the dataset and the project code at https://songbai.site/muses/.

The majority of existing few-shot learning methods describe image relations with binary labels. However, such binary relations are insufficient to teach the network complicated real-world relations, due to the lack of decision smoothness. Furthermore, current few-shot learning models capture only the similarity via relation labels, but they are not exposed to class concepts associated with objects, which is likely detrimental to the classification performance due to underutilization of the available class labels. For instance, children learn the concept of tiger from a few of actual examples as well as from comparisons of tiger to other animals. Thus, we hypothesize that both similarity and class concept learning must be occurring simultaneously. With these observations at hand, we study the fundamental problem of simplistic class modeling in current few-shot learning methods. We rethink the relations between class concepts, and propose a novel Absolute-relative Learning paradigm to fully take advantage of label information to refine the image an relation representations in both supervised and unsupervised scenarios. Our proposed paradigm improves the performance of several state-of-the-art models on publicly available datasets.

Most recent semantic segmentation methods adopt a fully-convolutional network (FCN) with an encoder-decoder architecture. The encoder progressively reduces the spatial resolution and learns more abstract/semantic visual concepts with larger receptive fields. Since context modeling is critical for segmentation, the latest efforts have been focused on increasing the receptive field, through either dilated/atrous convolutions or inserting attention modules. However, the encoder-decoder based FCN architecture remains unchanged. In this paper, we aim to provide an alternative perspective by treating semantic segmentation as a sequence-to-sequence prediction task. Specifically, we deploy a pure transformer (i.e., without convolution and resolution reduction) to encode an image as a sequence of patches. With the global context modeled in every layer of the transformer, this encoder can be combined with a simple decoder to provide a powerful segmentation model, termed SEgmentation TRansformer (SETR). Extensive experiments show that SETR achieves new state of the art on ADE20K (50.28% mIoU), Pascal Context (55.83% mIoU) and competitive results on Cityscapes. Particularly, we achieve the first position in the highly competitive ADE20K test server leaderboard on the day of submission.

The 2-D median filter, one of the oldest and most well-established image-filtering techniques, still sees widespread use throughout computer vision. Despite its relative algorithmic simplicity, accelerating the 2-D median filter via a hardware implementation becomes increasingly challenging as the window size increases, since the resources required grow quartically with the window size. Previous works, in a non-FPGA context, have shown that separately applying several directional median filters to an image, and then taking the median of their results, yields performance that is competitive with, and in some cases even better than, that of a classic 2-D window median. Inspired by these approaches, we propose a novel way of substituting a 2-D median filter on an FPGA with a sequence of directional median filters, in our case arranged as a pipeline, in the pursuit of an FPGA implementation that achieves better scalability and hardware efficiency without sacrificing accuracy. We empirically show that the combination of three particular directional filters, in any order, achieves this, whilst requiring quadratically fewer resources on the FPGA and allowing for much higher throughput.

Aggregating features from different depths of a network is widely adopted to improve the network capability. Lots of modern architectures are equipped with skip connections, which actually makes the feature aggregation happen in all these networks. Since different features tell different semantic meanings, there are inconsistencies and incompatibilities to be solved. However, existing works naively blend deep features via element-wise summation or concatenation with a convolution behind. Better feature aggregation method beyond summation or concatenation is rarely explored. In this paper, given two layers of features to be aggregated together, we first detect and identify where and what needs to be updated in one layer, then replace the feature at the identified location with the information of the other layer This process, which we call DEtect-rePLAce (DEPLA), enables us to avoid inconsistent patterns while keeping useful information in the merged outputs. Experimental results demonstrate our method largely boosts multiple baselines e.g. ResNet, FishNet and FPN on three major vision tasks including ImageNet classification, MS COCO object detection and instance segmentation.

Self-supervised learning (especially contrastive learning) has attracted great interest due to its huge potential in learning discriminative representations in an unsupervised manner. Despite the acknowledged successes, existing contrastive learning methods suffer from very low learning efficiency, e.g., taking about ten times more training epochs than supervised learning for comparable recognition accuracy. In this paper, we reveal two contradictory phenomena in contrastive learning that we call under-clustering and over-clustering problems, which are major obstacles to learning efficiency. Under-clustering means that the model cannot efficiently learn to discover the dissimilarity between inter-class samples when the negative sample pairs for contrastive learning are insufficient to differentiate all the actual object classes. Over-clustering implies that the model cannot efficiently learn features from excessive negative sample pairs, forcing the model to over-cluster samples of the same actual classes into different clusters. To simultaneously overcome these two problems, we propose a novel self-supervised learning framework using a truncated triplet loss. Precisely, we employ a triplet loss tending to maximize the relative distance between the positive pair and negative pairs to address the under-clustering problem; and we construct the negative pair by selecting a negative sample deputy from all negative samples to avoid the over-clustering problem, guaranteed by the Bernoulli Distribution model. We extensively evaluate our framework in several large-scale benchmarks (e.g., ImageNet, SYSU-30k, and COCO). The results demonstrate our model’s superiority (e.g., the learning efficiency) over the latest state-of-the-art methods by a clear margin. See Codes1.

Transformers with powerful global relation modeling abilities have been introduced to fundamental computer vision tasks recently. As a typical example, the Vision Trans-former (ViT) directly applies a pure transformer architecture on image classification, by simply splitting images into tokens with a fixed length, and employing transformers to learn relations between these tokens. However, such naive tokenization could destruct object structures, assign grids to uninterested regions such as background, and introduce interference signals. To mitigate the above issues, in this paper, we propose an iterative and progressive sampling strategy to locate discriminative regions. At each iteration, embeddings of the current sampling step are fed into a transformer encoder layer, and a group of sampling off-sets is predicted to update the sampling locations for the next step. The progressive sampling is differentiable. When combined with the Vision Transformer, the obtained PS-ViT network can adaptively learn where to look. The proposed PS-ViT is both effective and efficient. When trained from scratch on ImageNet, PS-ViT performs 3.8% higher than the vanilla ViT in terms of top-1 accuracy with about 4× fewer parameters and 10× fewer FLOPs. Code is available at https://github.com/yuexy/PS-ViT.

Full-motion cost volumes play a central role in current state-of-the-art optical flow methods. However, constructed using simple feature correlations, they lack the ability to encapsulate prior, or even non-local knowledge. This creates artifacts in poorly constrained ambiguous regions, such as occluded and textureless areas. We propose a separable cost volume module, a drop-in replacement to correlation cost volumes, that uses non-local aggregation layers to exploit global context cues and prior knowledge, in order to disambiguate motions in these regions. Our method leads both the now standard Sintel and KITTI optical flow benchmarks in terms of accuracy, and is also shown to generalize better from synthetic to real data.

Self-attention networks have revolutionized natural language processing and are making impressive strides in image analysis tasks such as image classification and object detection. Inspired by this success, we investigate the application of self-attention networks to 3D point cloud processing. We design self-attention layers for point clouds and use these to construct self-attention networks for tasks such as semantic scene segmentation, object part segmentation, and object classification. Our Point Transformer design improves upon prior work across domains and tasks. For example, on the challenging S3DIS dataset for large-scale semantic scene segmentation, the Point Transformer attains an mIoU of 70.4% on Area 5, outperforming the strongest prior model by 3.3 absolute percentage points and crossing the 70% mIoU threshold for the first time.

Recent progress in neural network verification has challenged the notion of a convex barrier, that is, an inherent weakness in the convex relaxation of the output of a neural network. Specifically, there now exists a tight relaxation for verifying the robustness of a neural network to $\ell_\infty$ input perturbations, as well as efficient primal and dual solvers for the relaxation. Buoyed by this success, we consider the problem of developing similar techniques for verifying robustness to input perturbations within the probability simplex. We prove a somewhat surprising result that, in this case, not only can one design a tight relaxation that overcomes the convex barrier, but the size of the relaxation remains linear in the number of neurons, thereby leading to simpler and more efficient algorithms. We establish the scalability of our overall approach via the specification of $\ell_1$ robustness for CIFAR-10 and MNIST classification, where our approach improves the state of the art verified accuracy by up to $14.4\%$. Furthermore, we establish its accuracy on a novel and highly challenging task of verifying the robustness of a multi-modal (text and image) classifier to arbitrary changes in its textual input.

We propose FACtored Multi-Agent Centralised policy gradients (FACMAC), a new method for cooperative multi-agent reinforcement learning in both discrete and continuous action spaces. Like MADDPG, a popular multi-agent actor-critic method, our approach uses deep deterministic policy gradients to learn policies. However, FACMAC learns a centralised but factored critic, which combines per-agent utilities into the joint action-value function via a non-linear monotonic function, as in QMIX, a popular multi-agent $Q$-learning algorithm. However, unlike QMIX, there are no inherent constraints on factoring the critic. We thus also employ a nonmonotonic factorisation and empirically demonstrate that its increased representational capacity allows it to solve some tasks that cannot be solved with monolithic, or monotonically factored critics. In addition, FACMAC uses a centralised policy gradient estimator that optimises over the entire joint action space, rather than optimising over each agent's action space separately as in MADDPG. This allows for more coordinated policy changes and fully reaps the benefits of a centralised critic. We evaluate FACMAC on variants of the multi-agent particle environments, a novel multi-agent MuJoCo benchmark, and a challenging set of StarCraft II micromanagement tasks. Empirical results demonstrate FACMAC's superior performance over MADDPG and other baselines on all three domains.

Recent advances in self-supervised learning with instance-level contrastive objectives facilitate unsupervised clustering. However, a standalone datum is not perceiving the context of the holistic cluster, and may undergo sub-optimal assignment. In this paper, we extend the mainstream contrastive learning paradigm to a cluster-level scheme, where all the data subjected to the same cluster contribute to a unified representation that encodes the context of each data group. Contrastive learning with this representation then rewards the assignment of each datum. To implement this vision, we propose twin-contrast clustering (TCC). We define a set of categorical variables as clustering assignment confidence, which links the instance-level learning track with the cluster-level one. On one hand, with the corresponding assignment variables being the weight, a weighted aggregation along the data points implements the set representation of a cluster. We further propose heuristic cluster augmentation equivalents to enable cluster-level contrastive learning. On the other hand, we derive the evidence lower-bound of the instance-level contrastive objective with the assignments. By reparametrizing the assignment variables, TCC is trained end-to-end, requiring no alternating steps. Extensive experiments show that TCC outperforms the state-of-the-art on benchmarked datasets.

Automated data augmentation (ADA) techniques have played an important role in boosting the performance of deep models. Such techniques mostly aim to optimize a parameterized distribution over a discrete augmentation space. Thus, are restricted by the discretization of the search space which normally is handcrafted. To overcome the limitations, we take the first step to constructing a continuous mapping from $\mathbb{R}^d$ to image transformations (an augmentation space). Using this mapping, we take a novel approach where 1) we pose the ADA as a continuous optimization problem over the parameters of the augmentation distribution; and 2) use Stochastic Gradient Langevin Dynamics to learn and sample augmentations. This allows us to potentially explore the space of infinitely many possible augmentations, which otherwise was not possible due to the discretization of the space. This view of ADA is radically different from the standard discretization based view of ADA, and it opens avenues for utilizing the vast efficient gradient-based algorithms available for continuous optimization problems. Results over multiple benchmarks demonstrate the efficiency improvement of this work compared with previous methods.

Tracking objects of interest in a video is one of the most popular and widely applicable problems in computer vision. However, with the years, a Cambrian explosion of use cases and benchmarks has fragmented the problem in a multitude of different experimental setups. As a consequence, the literature has fragmented too, and now novel approaches proposed by the community are usually specialised to fit only one specific setup. To understand to what extent this specialisation is necessary, in this work we present UniTrack, a solution to address five different tasks within the same framework. UniTrack consists of a single and task-agnostic appearance model, which can be learned in a supervised or self-supervised fashion, and multiple ``heads'' that address individual tasks and do not require training. We show how most tracking tasks can be solved within this framework, and that the same appearance model can be successfully used to obtain results that are competitive against specialised methods for most of the tasks considered. The framework also allows us to analyse appearance models obtained with the most recent self-supervised methods, thus extending their evaluation and comparison to a larger variety of important problems.

We present an approach to contrastive representation learning for semantic segmentation. Our approach leverages the representational power of existing feature extractors to find corresponding regions across images. These cross-image correspondences are used as auxiliary labels to guide the pixel-level selection of positive and negative samples for more effective contrastive learning in semantic segmentation. We show that auxiliary labels can be generated from a variety of feature extractors, ranging from image classification networks that have been trained using unsupervised contrastive learning to segmentation models that have been trained on a small amount of labeled data. We additionally introduce a novel metric for rapidly judging the quality of a given auxiliary-labeling strategy, and empirically analyze various factors that influence the performance of contrastive learning for semantic segmentation. We demonstrate the effectiveness of our method both in the low-data as well as the high-data regime on various datasets. Our experiments show that contrastive learning with our auxiliary-labeling approach consistently boosts semantic segmentation accuracy when compared to standard ImageNet pretraining and outperforms existing approaches of contrastive and semi-supervised semantic segmentation.

Multi-object tracking (MOT) has been notoriously difficult to evaluate. Previous metrics overemphasize the importance of either detection or association. To address this, we present a novel MOT evaluation metric, higher order tracking accuracy (HOTA), which explicitly balances the effect of performing accurate detection, association and localization into a single unified metric for comparing trackers. HOTA decomposes into a family of sub-metrics which are able to evaluate each of five basic error types separately, which enables clear analysis of tracking performance. We evaluate the effectiveness of HOTA on the MOTChallenge benchmark, and show that it is able to capture important aspects of MOT performance not previously taken into account by established metrics. Furthermore, we show HOTA scores better align with human visual evaluation of tracking performance.

Person re-identification (re-ID) has attracted much attention recently due to its great importance in video surveillance. In general, distance metrics used to identify two person images are expected to be robust under various appearance changes. However, our work observes the extreme vulnerability of existing distance metrics to adversarial examples, generated by simply adding human-imperceptible perturbations to person images. Hence, the security danger is dramatically increased when deploying commercial re-ID systems in video surveillance. Although adversarial examples have been extensively applied for classification analysis, it is rarely studied in metric analysis like person re-identification. The most likely reason is the natural gap between the training and testing of re-ID networks, that is, the predictions of a re-ID network cannot be directly used during testing without an effective metric. In this work, we bridge the gap by proposing Adversarial Metric Attack, a parallel methodology to adversarial classification attacks. Comprehensive experiments clearly reveal the adversarial effects in re-ID systems. Meanwhile, we also present an early attempt of training a metric-preserving network, thereby defending the metric against adversarial attacks. At last, by benchmarking various adversarial settings, we expect that our work can facilitate the development of adversarial attack and defense in metric-based applications.

Representing features at multiple scales is of great importance for numerous vision tasks. Recent advances in backbone convolutional neural networks (CNNs) continually demonstrate stronger multi-scale representation ability, leading to consistent performance gains on a wide range of applications. However, most existing methods represent the multi-scale features in a layer-wise manner. In this paper, we propose a novel building block for CNNs, namely Res2Net, by constructing hierarchical residual-like connections within one single residual block. The Res2Net represents multi-scale features at a granular level and increases the range of receptive fields for each network layer. The proposed Res2Net block can be plugged into the state-of-the-art backbone CNN models, e.g., ResNet, ResNeXt, and DLA. We evaluate the Res2Net block on all these models and demonstrate consistent performance gains over baseline models on widely-used datasets, e.g., CIFAR-100 and ImageNet. Further ablation studies and experimental results on representative computer vision tasks, i.e., object detection, class activation mapping, and salient object detection, further verify the superiority of the Res2Net over the state-of-the-art baseline methods. The source code and trained models are available on https://mmcheng.net/res2net/.

This paper presents regional attraction of line segment maps, and hereby poses the problem of line segment detection (LSD) as a problem of region coloring. Given a line segment map, the proposed regional attraction first establishes the relationship between line segments and regions in the image lattice. Based on this, the line segment map is equivalently transformed to an attraction field map (AFM), which can be remapped to a set of line segments without loss of information. Accordingly, we develop an end-to-end framework to learn attraction field maps for raw input images, followed by a squeeze module to detect line segments. Apart from existing works, the proposed detector properly handles the local ambiguity and does not rely on the accurate identification of edge pixels. Comprehensive experiments on the Wireframe dataset and the YorkUrban dataset demonstrate the superiority of our method. In particular, we achieve an F-measure of 0.831 on the Wireframe dataset, advancing the state-of-the-art performance by 10.3 percent.

In this paper, we address the task of semantic-guided scene generation. One open challenge widely observed in global image-level generation methods is the difficulty of generating small objects and detailed local texture. To tackle this issue, in this work we consider learning the scene generation in a local context, and correspondingly design a local class-specific generative network with semantic maps as a guidance, which separately constructs and learns sub-generators concentrating on the generation of different classes, and is able to provide more scene details. To learn more discriminative class-specific feature representations for the local generation, a novel classification module is also proposed. To combine the advantage of both global image-level and the local class-specific generation, a joint generation network is designed with an attention fusion module and a dual-discriminator structure embedded. Extensive experiments on two scene image generation tasks show superior generation performance of the proposed model. State-of-the-art results are established by large margins on both tasks and on challenging public benchmarks. The source code and trained models are available at https://github.com/Ha0Tang/LGGAN.

We present an approach for estimating surface normals from in-the-wild color images of faces. While data-driven strategies have been proposed for single face images, limited available ground truth data makes this problem difficult. To alleviate this issue, we propose a method that can leverage all available image and normal data, whether paired or not, thanks to a novel cross-modal learning architecture. In particular, we enable additional training with single modality data, either color or normal, by using two encoder-decoder networks with a shared latent space. The proposed architecture also enables face details to be transferred between the image and normal domains, given paired data, through skip connections between the image encoder and normal decoder. Core to our approach is a novel module that we call deactivable skip connections, which allows integrating both the auto-encoded and image-to-normal branches within the same architecture that can be trained end-to-end. This allows learning of a rich latent space that can accurately capture the normal information. We compare against state-of-the-art methods and show that our approach can achieve significant improvements, both quantitative and qualitative, with natural face images.

The goal of our paper is to semantically edit parts of an image matching a given text that describes desired attributes (e.g., texture, colour, and background), while preserving other contents that are irrelevant to the text. To achieve this, we propose a novel generative adversarial network (ManiGAN), which contains two key components: text-image affine combination module (ACM) and detail correction module (DCM). The ACM selects image regions relevant to the given text and then correlates the regions with corresponding semantic words for effective manipulation. Meanwhile, it encodes original image features to help reconstruct text-irrelevant contents. The DCM rectifies mismatched attributes and completes missing contents of the synthetic image. Finally, we suggest a new metric for evaluating image manipulation results, in terms of both the generation of new attributes and the reconstruction of text-irrelevant contents. Extensive experiments on the CUB and COCO datasets demonstrate the superior performance of the proposed method.

We present an end-to-end network to bridge the gap between training and inference pipeline for panoptic segmentation, a task that seeks to partition an image into semantic regions for "stuff" and object instances for "things". In contrast to recent works, our network exploits a parametrised, yet lightweight panoptic segmentation submodule, powered by an end-to-end learnt dense instance affinity, to capture the probability that any pair of pixels belong to the same instance. This panoptic submodule gives rise to a novel propagation mechanism for panoptic logits and enables the network to output a coherent panoptic segmentation map for both "stuff" and "thing" classes, without any post-processing. Reaping the benefits of end-to-end training, our full system sets new records on the popular street scene dataset, Cityscapes, achieving 61.4 PQ with a ResNet-50 backbone using only the fine annotations. On the challenging COCO dataset, our ResNet-50-based network also delivers state-of-the-art accuracy of 43.4 PQ. Moreover, our network flexibly works with and without object mask cues, performing competitively under both settings, which is of interest for applications with computation budgets.

Generative Adversarial Networks (GANs) can generate realistic fake face images that can easily fool human beings. On the contrary, a common Convolutional Neural Network(CNN) discriminator can achieve more than99.9%accuracyin discerning fake/real images. In this paper, we conduct an empirical study on fake/real faces, and have two important observations: firstly, the texture of fake faces is substantially different from real ones; secondly, global texture statistics are more robust to image editing and transferable to fake faces from different GANs and datasets. Motivated by the above observations, we propose a new architecture coined as Gram-Net, which leverages global image texture representations for robust fake image detection. Experimental results on several datasets demonstrate that our Gram-Netoutperforms existing approaches. Especially, our Gram-Netis more robust to image editings, e.g. down-sampling, JPEGcompression, blur, and noise. More importantly, our Gram-Net generalizes significantly better in detecting fake faces from GAN models not seen in the training phase and can perform decently in detecting fake natural images.

We present Siam R-CNN, a Siamese re-detection architecture which unleashes the full power of two-stage object detection approaches for visual object tracking. We combine this with a novel tracklet-based dynamic programming algorithm, which takes advantage of re-detections of both the first-frame template and previous-frame predictions, to model the full history of both the object to be tracked and potential distractor objects. This enables our approach to make better tracking decisions, as well as to re-detect tracked objects after long occlusion. Finally, we propose a novel hard example mining strategy to improve Siam R-CNN's robustness to similar looking objects. Siam R-CNN achieves the current best performance on ten tracking benchmarks, with especially strong results for long-term tracking. We make our code and models available at www.vision.rwth-aachen.de/page/siamrcnn.

This paper presents a fast and parsimonious parsing method to accurately and robustly detect a vectorized wireframe in an input image with a single forward pass. The proposed method is end-to-end trainable, consisting of three components: (i) line segment and junction proposal generation, (ii) line segment and junction matching, and (iii) line segment and junction verification. For computing line segment proposals, a novel exact dual representation is proposed which exploits a parsimonious geometric reparameterization for line segments and forms a holistic 4-dimensional attraction field map for an input image. Junctions can be treated as the “basins” in the attraction field. The proposed method is thus called Holistically-Attracted Wireframe Parser (HAWP). In experiments, the proposed method is tested on two benchmarks, the Wireframe dataset [14] and the YorkUrban dataset [8]. On both benchmarks, it obtains state-of-the-art performance in terms of accuracy and efficiency. For example, on the Wireframe dataset, compared to the previous state-of-the-art method L-CNN [36], it improves the challenging mean structural average precision (msAP) by a large margin (2.8% absolute improvements), and achieves 29.5 FPS on a single GPU (89% relative improvement). A systematic ablation study is performed to further justify the proposed method.

Modelling long-range dependencies is critical for scene understanding tasks in computer vision. Although CNNs have excelled in many vision tasks, they are still limited in capturing long-range structured relationships as they typically consist of layers of local kernels. A fully-connected graph is beneficial for such modelling, however, its computational overhead is prohibitive. We propose a dynamic graph message passing network, that significantly reduces the computational complexity compared to related works modelling a fully-connected graph. This is achieved by adaptively sampling nodes in the graph, conditioned on the input, for message passing. Based on the sampled nodes, we dynamically predict node-dependent filter weights and the affinity matrix for propagating information between them. Using this model, we show significant improvements with respect to strong, state-of-the-art baselines on three different tasks and backbone architectures. Our approach also outperforms fully-connected graphs while using substantially fewer floating-point operations and parameters.

The 2-D median filter, one of the oldest and most well-established image-filtering techniques, still sees widespread use throughout computer vision. Despite its relative algorithmic simplicity, accelerating the 2-D median filter via a hardware implementation becomes increasingly challenging as the window size increases, since the resources required grow quadratically with the window size. Previous works, in a non-FPGA context, have shown that applying a sequence of multiple directional median filters to an image yields results that are competitive with, and in some cases even better than, those of a classic 2-D window median. Inspired by these approaches, we propose a novel way of substituting a 2-D median filter on an FPGA with a sequence of directional median filters, in our case in the pursuit of an FPGA implementation that achieves better scalability and hardware efficiency without sacrificing accuracy. We empirically show that the combination of three particular directional filters, in any order, achieves this, whilst requiring quadratically fewer resources on the FPGA. Our approach allows for much higher throughput and is easier to implement as a pipeline.

Image inpainting aims at synthesizing the missing content of damaged or corrupted images to produce visually realistic restorations; typical applications are in image restoration, automatic scene editing, super-resolution, and dynamic object removal. In this paper, we propose Semantic and Edge Conditioned Inpainting Generative Adversarial Network (SECI-GAN), an architecture that jointly exploits the high-level cues extracted by semantic segmentation and the fine-grained details captured by edge extraction to condition the image inpainting process. SECI-GAN is designed with a particular focus on recovering big regions belonging to the same object (e.g. cars or pedestrians) in the context of dynamic object removal from complex street views. To demonstrate the effectiveness of SECI-GAN, we evaluate our results on the Cityscapes dataset, showing that SECI-GAN is better than competing state-of-the-art models at recovering the structure and the content of the missing parts while producing consistent predictions.

We propose a robust baseline method for instance segmentation which are specially designed for large-scale outdoor LiDAR point clouds. Our method includes a novel dense feature encoding technique, allowing the localization and segmentation of small, far-away objects, a simple but effective solution for single-shot instance prediction and effective strategies for handling severe class imbalances. Since there is no public dataset for the study of LiDAR instance segmentation, we also build a new publicly available LiDAR point cloud dataset to include both precise 3D bounding box and point-wise labels for instance segmentation, while still being about 3~20 times as large as other existing LiDAR datasets. The dataset will be published at https://github.com/feihuzhang/LiDARSeg.

Personal robots and driverless cars need to be able to operate in novel environments and thus quickly and efficiently learn to recognise new object classes. We address this problem by considering the task of video object segmentation. Previous accurate methods for this task finetune a model using the first annotated frame, and/or use additional inputs such as optical flow and complex post-processing. In contrast, we develop a fast, causal algorithm that requires no finetuning, auxiliary inputs or post-processing, and segments a variable number of objects in a single forward-pass. We represent an object with clusters, or "visual words", in the embedding space, which correspond to object parts in the image space. This allows us to robustly match to the reference objects throughout the video, because although the global appearance of an object changes as it undergoes occlusions and deformations, the appearance of more local parts may stay consistent. We learn these visual words in an unsupervised manner, using meta-learning to ensure that our training objective matches our inference procedure. We achieve comparable accuracy to finetuning based methods (whilst being 1 to 2 orders of magnitude faster), and state-of-the-art in terms of speed/accuracy trade-offs on four video segmentation datasets. Code is available at https://github.com/harkiratbehl/MetaVOS.

In continual learning (CL), a learner is faced with a sequence of tasks, arriving one after the other, and the goal is to remember all the tasks once the continual learning experience is finished. The prior art in CL uses episodic memory, parameter regularization or extensible network structures to reduce interference among tasks, but in the end, all the approaches learn different tasks in a joint vector space. We believe this invariably leads to interference among different tasks. We propose to learn tasks in different (low-rank) vector subspaces that are kept orthogonal to each other in order to minimize interference. Further, to keep the gradients of different tasks coming from these subspaces orthogonal to each other, we learn isometric mappings by posing network training as an optimization problem over the Stiefel manifold. To the best of our understanding, we report, for the first time, strong results over experience-replay baseline with and without memory on standard classification benchmarks in continual learning.

Training Neural Ordinary Differential Equations (ODEs) is often computationally expensive. Indeed, computing the forward pass of such models involves solving an ODE which can become arbitrarily complex during training. Recent works have shown that regularizing the dynamics of the ODE can partially alleviate this. In this paper we propose a new regularization technique: randomly sampling the end time of the ODE during training. The proposed regularization is simple to implement, has negligible overhead and is effective across a wide variety of tasks. Further, the technique is orthogonal to several other methods proposed to regularize the dynamics of ODEs and as such can be used in conjunction with them. We show through experiments on normalizing flows, time series models and image recognition that the proposed regularization can significantly decrease training time and even improve performance over baseline models.

We propose a novel lightweight generative adversarial network for efficient image manipulation using natural language descriptions. To achieve this, a new word-level discriminator is proposed, which provides the generator with fine-grained training feedback at word-level, to facilitate training a lightweight generator that has a small number of parameters, but can still correctly focus on specific visual attributes of an image, and then edit them without affecting other contents that are not described in the text. Furthermore, thanks to the explicit training signal related to each word, the discriminator can also be simplified to have a lightweight structure. Compared with the state of the art, our method has a much smaller number of parameters, but still achieves a competitive manipulation performance. Extensive experimental results demonstrate that our method can better disentangle different visual attributes, then correctly map them to corresponding semantic words, and thus achieve a more accurate image modification using natural language descriptions.

Miscalibration -- a mismatch between a model's confidence and its correctness -- of Deep Neural Networks (DNNs) makes their predictions hard to rely on. Ideally, we want networks to be accurate, calibrated and confident. We show that, as opposed to the standard cross-entropy loss, focal loss (Lin et al., 2017) allows us to learn models that are already very well calibrated. When combined with temperature scaling, whilst preserving accuracy, it yields state-of-the-art calibrated models. We provide a thorough analysis of the factors causing miscalibration, and use the insights we glean from this to justify the empirically excellent performance of focal loss. To facilitate the use of focal loss in practice, we also provide a principled approach to automatically select the hyperparameter involved in the loss function. We perform extensive experiments on a variety of computer vision and NLP datasets, and with a wide variety of network architectures, and show that our approach achieves state-of-the-art calibration without compromising on accuracy in almost all cases. Code is available at https://github.com/torrvision/focal_calibration.

A fundamental component of neural network verification is the computation of bounds on the values their outputs can take. Previous methods have either used off-the-shelf solvers, discarding the problem structure, or relaxed the problem even further, making the bounds unnecessarily loose. We propose a novel approach based on Lagrangian Decomposition. Our formulation admits an efficient supergradient ascent algorithm, as well as an improved proximal algorithm. Both the algorithms offer three advantages: (i) they yield bounds that are provably at least as tight as previous dual algorithms relying on Lagrangian relaxations; (ii) they are based on operations analogous to forward/backward pass of neural networks layers and are therefore easily parallelizable, amenable to GPU implementation and able to take advantage of the convolutional structure of problems; and (iii) they allow for anytime stopping while still providing valid bounds. Empirically, we show that we obtain bounds comparable with off-the-shelf solvers in a fraction of their running time, and obtain tighter bounds in the same time as previous dual algorithms. This results in an overall speed-up when employing the bounds for formal verification. Code for our algorithms is available at https://github.com/oval-group/decomposition-plnn-bounds.

Abstract Deep generative modelling for human body analysis is an emerging problem with many interesting applications. However, the latent space learned by such approaches is typically not interpretable, resulting in less flexibility. In this work, we present deep generative models for human body analysis in which the body pose and the visual appearance are disentangled. Such a disentanglement allows independent manipulation of pose and appearance, and hence enables applications such as pose-transfer without specific training for such a task. Our proposed models, the Conditional-DGPose and the Semi-DGPose, have different characteristics. In the first, body pose labels are taken as conditioners, from a fully-supervised training set. In the second, our structured semi-supervised approach allows for pose estimation to be performed by the model itself and relaxes the need for labelled data. Therefore, the Semi-DGPose aims for the joint understanding and generation of people in images. It is not only capable of mapping images to interpretable latent representations but also able to map these representations back to the image space. We compare our models with relevant baselines, the ClothNet-Body and the Pose Guided Person Generation networks, demonstrating their merits on the Human3.6M, ChictopiaPlus and DeepFashion benchmarks.

The success of Deep Learning and its potential use in many safety-critical applicationshas motivated research on formal verification of Neural Network (NN) models. In thiscontext, verification involves proving or disproving that an NN model satisfies certaininput-output properties. Despite the reputation of learned NN models as black boxes,and the theoretical hardness of proving useful properties about them, researchers havebeen successful in verifying some classes of models by exploiting their piecewise linearstructure and taking insights from formal methods such as Satisifiability Modulo Theory.However, these methods are still far from scaling to realistic neural networks. To facilitateprogress on this crucial area, we exploit the Mixed Integer Linear Programming (MIP) formulation of verification to propose a family of algorithms based on Branch-and-Bound (BaB). We show that our family contains previous verification methods as special cases.With the help of the BaB framework, we make three key contributions. Firstly, we identifynew methods that combine the strengths of multiple existing approaches, accomplishingsignificant performance improvements over previous state of the art. Secondly, we introducean effective branching strategy on ReLU non-linearities. This branching strategy allows usto efficiently and successfully deal with high input dimensional problems with convolutionalnetwork architecture, on which previous methods fail frequently. Finally, we proposecomprehensive test data sets and benchmarks which includes a collection of previouslyreleased testcases. We use the data sets to conduct a thorough experimental comparison ofexisting and new algorithms and to provide an inclusive analysis of the factors impactingthe hardness of verification problems.

Camera pose estimation is an important problem in computer vision, with applications as diverse as simultaneous localisation and mapping, virtual/augmented reality and navigation. Common techniques match the current image against keyframes with known poses coming from a tracker, directly regress the pose, or establish correspondences between keypoints in the current image and points in the scene in order to estimate the pose. In recent years, regression forests have become a popular alternative to establish such correspondences. They achieve accurate results, but have traditionally needed to be trained offline on the target scene, preventing relocalisation in new environments. Recently, we showed how to circumvent this limitation by adapting a pre-trained forest to a new scene on the fly. The adapted forests achieved relocalisation performance that was on par with that of offline forests, and our approach was able to estimate the camera pose in close to real time, which made it desirable for systems that require online relocalisation. In this paper, we present an extension of this work that achieves significantly better relocalisation performance whilst running fully in real time. To achieve this, we make several changes to the original approach: (i) instead of simply accepting the camera pose hypothesis produced by RANSAC without question, we make it possible to score the final few hypotheses it considers using a geometric approach and select the most promising one; (ii) we chain several instantiations of our relocaliser (with different parameter settings) together in a cascade, allowing us to try faster but less accurate relocalisation first, only falling back to slower, more accurate relocalisation as necessary; and (iii) we tune the parameters of our cascade, and the individual relocalisers it contains, to achieve effective overall performance. Taken together, these changes allow us to significantly improve upon the performance our original state-of-the-art method was able to achieve on the well-known 7-Scenes and Stanford 4 Scenes benchmarks. As additional contributions, we present a novel way of visualising the internal behaviour of our forests, and use the insights gleaned from this to show how to entirely circumvent the need to pre-train a forest on a generic scene.

Rotoscoping, the detailed delineation of scene elements through a video shot, is a painstaking task of tremendous importance in professional post-production pipelines. While pixel-wise segmentation techniques can help for this task, professional rotoscoping tools rely on parametric curves that offer the artists a much better interactive control on the definition, editing and manipulation of the segments of interest. Sticking to this prevalent rotoscoping paradigm, we propose a novel framework to capture and track the visual aspect of an arbitrary object in a scene, given an initial closed outline of this object. This model combines a collection of local foreground/background appearance models spread along the outline, a global appearance model of the enclosed object and a set of distinctive foreground landmarks. The structure of this rich appearance model allows simple initialization, efficient iterative optimization with exact minimization at each step, and on-line adaptation in videos. We further extend this model by so-called trimaps which serve as an input to alpha-matting algorithms to allow truly seamless compositing. To this end, we leverage local classifiers attached to the roto-curves to define a confidence measure that is well-suited to define trimaps with adaptive band-widths. The resulting trimaps are parametric, temporally consistent and remain fully editable by the artist. We demonstrate qualitatively and quantitatively the merit of this framework through comparisons with tools based on either dynamic segmentation with a closed curve or pixel-wise binary labelling.

Many applications require a camera to be relocalised online, without expensive offline training on the target scene. Whilst both keyframe and sparse keypoint matching methods can be used online, the former often fail away from the training trajectory, and the latter can struggle in textureless regions. By contrast, scene coordinate regression (SCoRe) methods generalise to novel poses and can leverage dense correspondences to improve robustness, and recent work has shown how to adapt SCoRe forests between scenes, allowing their state-of-the-art performance to be leveraged online. However, because they use features hand-crafted for indoor use, they do not generalise well to harder outdoor scenes. Whilst replacing the forest with a neural network and learning suitable features for outdoor use is possible, the techniques used to adapt forests between scenes are unfortunately harder to transfer to a network context. In this paper, we address this by proposing a novel way of leveraging a network trained on one scene to predict points in another scene. Our approach replaces the appearance clustering performed by the branching structure of a regression forest with a two-step process that first uses the network to predict points in the original scene, and then uses these predicted points to look up clusters of points from the new scene. We show experimentally that our online approach achieves state-of-the-art performance on both the 7-Scenes and Cambridge Landmarks datasets, whilst running in under 300ms, making it highly effective in live scenarios.

This work studies the unsupervised re-ranking procedure for object retrieval and person re-identification with a specific concentration on an ensemble of multiple metrics (or similarities). While the re-ranking step is involved by running a diffusion process on the underlying data manifolds, the fusion step can leverage the complementarity of multiple metrics. We give a comprehensive summary of existing fusion with diffusion strategies, and systematically analyze their pros and cons. Based on the analysis, we propose a unified yet robust algorithm which inherits their advantages and discards their disadvantages. Hence, we call it Unified Ensemble Diffusion (UED). More interestingly, we derive that the inherited properties indeed stem from a theoretical framework, where the relevant works can be elegantly summarized as special cases of UED by imposing additional constraints on the objective function and varying the solver of similarity propagation. Extensive experiments with 3D shape retrieval, image retrieval and person re-identification demonstrate that the proposed framework outperforms the state of the arts, and at the same time suggest that re-ranking via metric fusion is a promising tool to further improve the retrieval performance of existing algorithms.

We present in this work the first end-to-end deep learning based method that predicts both 3D hand shape and pose from RGB images in the wild. Our network consists of the concatenation of a deep convolutional encoder, and a fixed model-based decoder. Given an input image, and optionally 2D joint detections obtained from an independent CNN, the encoder predicts a set of hand and view parameters. The decoder has two components: A pre-computed articulated mesh deformation hand model that generates a 3D mesh from the hand parameters, and a re-projection module controlled by the view parameters that projects the generated hand into the image domain. We show that using the shape and pose prior knowledge encoded in the hand model within a deep learning framework yields state-of-the-art performance in 3D pose prediction from images on standard benchmarks, and produces geometrically valid and plausible 3D reconstructions. Additionally, we show that training with weak supervision in the form of 2D joint annotations on datasets of images in the wild, in conjunction with full supervision in the form of 3D joint annotations on limited available datasets allows for good generalization to 3D shape and pose predictions on images in the wild.

Deep networks consume a large amount of memory by their nature. A natural question arises can we reduce that memory requirement whilst maintaining performance. In particular, in this work we address the problem of memory efficient learning for multiple tasks. To this end, we propose a novel network architecture producing multiple networks of different configurations, termed deep virtual networks (DVNs), for different tasks. Each DVN is specialized for a single task and structured hierarchically. The hierarchical structure, which contains multiple levels of hierarchy corresponding to different numbers of parameters, enables multiple inference for different memory budgets. The building block of a deep virtual network is based on a disjoint collection of parameters of a network, which we call a unit. The lowest level of hierarchy in a deep virtual network is a unit, and higher levels of hierarchy contain lower levels' units and other additional units. Given a budget on the number of parameters, a different level of a deep virtual network can be chosen to perform the task. A unit can be shared by different DVNs, allowing multiple DVNs in a single network. In addition, shared units provide assistance to the target task with additional knowledge learned from another tasks. This cooperative configuration of DVNs makes it possible to handle different tasks in a memory-aware manner. Our experiments show that the proposed method outperforms existing approaches for multiple tasks. Notably, ours is more efficient than others as it allows memory-aware inference for all tasks.

Real world applications of stereo depth estimation require models that are robust to dynamic variations in the environment. Even though deep learning based stereo methods are successful, they often fail to generalize to unseen variations in the environment, making them less suitable for practical applications such as autonomous driving. In this work, we introduce a ``learning-to-adapt'' framework that enables deep stereo methods to continuously adapt to new target domains in an unsupervised manner. Specifically, our approach incorporates the adaptation procedure into the learning objective to obtain a base set of parameters that are better suited for unsupervised online adaptation. To further improve the quality of the adaptation, we learn a confidence measure that effectively masks the errors introduced during the unsupervised adaptation. We evaluate our method on synthetic and real-world stereo datasets and our experiments evidence that learning-to-adapt is, indeed beneficial for online adaptation on vastly different domains.

In this paper we illustrate how to perform both visual object tracking and semi-supervised video object segmentation, in real-time, with a single simple approach. Our method, dubbed SiamMask, improves the offline training procedure of popular fully-convolutional Siamese approaches for object tracking by augmenting their loss with a binary segmentation task. Once trained, SiamMask solely relies on a single bounding box initialisation and operates online, producing class-agnostic object segmentation masks and rotated bounding boxes at 55 frames per second. Despite its simplicity, versatility and fast speed, our strategy allows us to establish a new state-of-the-art among real-time trackers on VOT-2018, while at the same time demonstrating competitive performance and the best speed for the semi-supervised video object segmentation task on DAVIS-2016 and DAVIS-2017.

In the stereo matching task, matching cost aggregation is crucial in both traditional methods and deep neural network models in order to accurately estimate disparities. We propose two novel neural net layers, aimed at capturing local and the whole-image cost dependencies respectively. The first is a semi-global aggregation layer which is a differentiable approximation of the semi-global matching, the second is the local guided aggregation layer which follows a traditional cost filtering strategy to refine thin structures. These two layers can be used to replace the widely used 3D convolutional layer which is computationally costly and memory-consuming as it has cubic computational/memory complexity. In the experiments, we show that nets with a two-layer guided aggregation block easily outperform the state-of-the-art GC-Net which has nineteen 3D convolutional layers. We also train a deep guided aggregation network (GA-Net) which gets better accuracies than state-of-the-art methods on both Scene Flow dataset and KITTI benchmarks.

Compressing large Neural Networks (NN) by quantizing the parameters, while maintaining the performance is highly desirable due to reduced memory and time complexity. In this work, we cast NN quantization as a discrete labelling problem, and by examining relaxations, we design an efficient iterative optimization procedure that involves stochastic gradient descent followed by a projection. We prove that our simple projected gradient descent approach is, in fact, equivalent to a proximal version of the well-known mean-field method. These findings would allow the decades-old and theoretically grounded research on MRF optimization to be used to design better network quantization schemes. Our experiments on standard classification datasets (MNIST, CIFAR10/100, TinyImageNet) with convolutional and residual architectures show that our algorithm obtains fully-quantized networks with accuracies very close to the floating-point reference networks.

We propose an interactive GAN-based sketch-to-image translation method that helps novice users easily create images of simple objects. The user starts with a sparse sketch and a desired object category, and the network then recommends its plausible completion(s) and shows a corresponding synthesized image. This enables a feedback loop, where the user can edit the sketch based on the network's recommendations, while the network is able to better synthesize the image that the user might have in mind. In order to use a single model for a wide array of object classes, we introduce a gating-based approach for class conditioning, which allows us to generate distinct classes without feature mixing, from a single generator network.

Unsupervised video object segmentation has often been tackled by methods based on recurrent neural networks and optical flow. Despite their complexity, these kinds of approach tend to favour short-term temporal dependencies and are thus prone to accumulating inaccuracies, which cause drift over time. Moreover, simple (static) image segmentation models, alone, can perform competitively against these methods, which further suggests that the way temporal dependencies are modelled should be reconsidered. Motivated by these observations, in this paper we explore simple yet effective strategies to model long-term temporal dependencies. Inspired by the non-local operators, we introduce a technique to establish dense correspondences between pixel embeddings of a reference "anchor" frame and the current one. This allows the learning of pairwise dependencies at arbitrarily long distances without conditioning on intermediate frames. Without online supervision, our approach can suppress the background and precisely segment the foreground object even in challenging scenarios, while maintaining consistent performance over time. With a mean IoU of 81.7%, our method ranks first on the DAVIS-2016 leaderboard of unsupervised methods, while still being competitive against state-of-the-art online semi-supervised approaches. We further evaluate our method on the FBMS dataset and the video saliency dataset ViSal, showing results competitive with the state of the art.

We present a novel probabilistic programming framework that couples directly to existing large-scale simulators through a cross-platform probabilistic execution protocol, which allows general-purpose inference engines to record and control random number draws within simulators in a language-agnostic way. The execution of existing simulators as probabilistic programs enables highly interpretable posterior inference in the structured model defined by the simulator code base. We demonstrate the technique in particle physics, on a scientifically accurate simulation of the tau lepton decay, which is a key ingredient in establishing the properties of the Higgs boson. Inference efficiency is achieved via inference compilation where a deep recurrent neural network is trained to parameterize proposal distributions and control the stochastic simulator in a sequential importance sampling scheme, at a fraction of the computational cost of a Markov chain Monte Carlo baseline.

In this paper, we propose a novel controllable text-to-image generative adversarial network (ControlGAN), which can effectively synthesise high-quality images and also control parts of the image generation according to natural language descriptions. To achieve this, we introduce a word-level spatial and channel-wise attention-driven generator that can disentangle different visual attributes, and allow the model to focus on generating and manipulating subregions corresponding to the most relevant words. Also, a word-level discriminator is proposed to provide fine-grained supervisory feedback by correlating words with image regions, facilitating training an effective generator which is able to manipulate specific visual attributes without affecting the generation of other content. Furthermore, perceptual loss is adopted to reduce the randomness involved in the image generation, and to encourage the generator to manipulate specific attributes required in the modified text. Extensive experiments on benchmark datasets demonstrate that our method outperforms existing state of the art, and is able to effectively manipulate synthetic images using natural language descriptions.

Learning generative models that span multiple data modalities, such as vision and language, is often motivated by the desire to learn more useful, generalisable representations that faithfully capture common underlying factors between the modalities. In this work, we characterise successful learning of such models as the fulfilment of four criteria: i) implicit latent decomposition into shared and private subspaces, ii) coherent joint generation over all modalities, iii) coherent cross-generation across individual modalities, and iv) improved model learning for individual modalities through multi-modal integration. Here, we propose a mixture-of-experts multi-modal variational autoencoder (MMVAE) for learning of generative models on different sets of modalities, including a challenging image <-> language dataset, and demonstrate its ability to satisfy all four criteria, both qualitatively and quantitatively.

Cooperative multi-agent reinforcement learning often requires decentralised policies, which severely limit the agents' ability to coordinate their behaviour. In this paper, we show that common knowledge between agents allows for complex decentralised coordination. Common knowledge arises naturally in a large number of decentralised cooperative multi-agent tasks, for example, when agents can reconstruct parts of each others' observations. Since agents can independently agree on their common knowledge, they can execute complex coordinated policies that condition on this knowledge in a fully decentralised fashion. We propose multi-agent common knowledge reinforcement learning (MACKRL), a novel stochastic actor-critic algorithm that learns a hierarchical policy tree. Higher levels in the hierarchy coordinate groups of agents by conditioning on their common knowledge, or delegate to lower levels with smaller subgroups but potentially richer common knowledge. The entire policy tree can be executed in a fully decentralised fashion. As the lowest policy tree level consists of independent policies for each agent, MACKRL reduces to independently learnt decentralised policies as a special case. We demonstrate that our method can exploit common knowledge for superior performance on complex decentralised coordination tasks, including a stochastic matrix game and challenging problems in StarCraft II unit micromanagement.

Probabilistic programming languages (PPLs) are receiving widespread attention for performing Bayesian inference in complex generative models. However, applications to science remain limited because of the impracticability of rewriting complex scientific simulators in a PPL, the computational cost of inference, and the lack of scalable implementations. To address these, we present a novel PPL framework that couples directly to existing scientific simulators through a cross-platform probabilistic execution protocol and provides Markov chain Monte Carlo (MCMC) and deep-learning-based inference compilation (IC) engines for tractable inference. To guide IC inference, we perform distributed training of a dynamic 3DCNN-LSTM architecture with a PyTorch-MPI-based framework on 1,024 32-core CPU nodes of the Cori supercomputer with a global mini-batch size of 128k: achieving a performance of 450 Tflop/s through enhancements to PyTorch. We demonstrate a Large Hadron Collider (LHC) use-case with the C++ Sherpa simulator and achieve the largest-scale posterior inference in a Turing-complete PPL.

We propose a deep generative model of humans in natural images which keeps 2D pose separated from other latent factors of variation, such as background scene and clothing. In contrast to methods that learn generative models of low-dimensional representations, e.g., segmentation masks and 2D skeletons, our single-stage end-to-end conditional-VAEGAN learns directly on the image space. The flexibility of this approach allows the sampling of people with independent variations of pose and appearance. Moreover, it enables the reconstruction of images conditioned to a given posture, allowing, for instance, pose-transfer from one person to another. We validate our method on the Human3.6M dataset and achieve state-of-the-art results on the ChictopiaPlus benchmark. Our model, named Conditional-DGPose, outperforms the closest related work in the literature. It generates more realistic and accurate images regarding both, body posture and image quality, learning the underlying factors of pose and appearance variation.

Training a generic objectness measure to produce object proposals has recently become of significant interest. We observe that generic objects with well-defined closed boundaries can be detected by looking at the norm of gradients, with a suitable resizing of their corresponding image windows to a small fixed size. Based on this observation and computational reasons, we propose to resize the window to 8 × 8 and use the norm of the gradients as a simple 64D feature to describe it, for explicitly training a generic objectness measure. We further show how the binarized version of this feature, namely binarized normed gradients (BING), can be used for efficient objectness estimation, which requires only a few atomic operations (e.g., add, bitwise shift, etc.). To improve localization quality of the proposals while maintaining efficiency, we propose a novel fast segmentation method and demonstrate its effectiveness for improving BING’s localization performance, when used in multi-thresholding straddling expansion (MTSE) post-processing. On the challenging PASCAL VOC2007 dataset, using 1000 proposals per image and intersection-over-union threshold of 0.5, our proposal method achieves a 95.6% object detection rate and 78.6% mean average best overlap in less than 0.005 second per image.

Recent progress on salient object detection is substantial, benefiting mostly from the explosive development of Convolutional Neural Networks (CNNs). Semantic segmentation and salient object detection algorithms developed lately have been mostly based on Fully Convolutional Neural Networks (FCNs). There is still a large room for improvement over the generic FCN models that do not explicitly deal with the scale-space problem. The Holistically-Nested Edge Detector (HED) provides a skip-layer structure with deep supervision for edge and boundary detection, but the performance gain of HED on saliency detection is not obvious. In this paper, we propose a new salient object detection method by introducing short connections to the skip-layer structures within the HED architecture. Our framework takes full advantage of multi-level and multi-scale features extracted from FCNs, providing more advanced representations at each layer, a property that is critically needed to perform segment detection. Our method produces state-of-the-art results on 5 widely tested salient object detection benchmarks, with advantages in terms of efficiency (0.08 seconds per image), effectiveness, and simplicity over the existing algorithms. Beyond that, we conduct an exhaustive analysis of the role of training data on performance. We provide a training set for future research and fair comparisons.

Dense conditional random fields (CRFs) have become a popular framework for modeling several problems in computer vision such as stereo correspondence and multiclass semantic segmentation. By modeling long-range interactions, dense CRFs provide a labeling that captures finer detail than their sparse counterparts. Currently, the state-of-the-art algorithm performs mean-field inference using a filter-based method but fails to provide a strong theoretical guarantee on the quality of the solution. A question naturally arises as to whether it is possible to obtain a maximum a posteriori (MAP) estimate of a dense CRF using a principled method. Within this paper, we show that this is indeed possible. Specifically, we will show that, by using a filter-based method, continuous relaxations of the MAP problem can be optimized efficiently using state-of-the-art algorithms. Specifically, we will solve a quadratic programming relaxation using the Frank--Wolfe algorithm and a linear programming relaxation by developing a proximal minimization framework. By exploiting labeling consistency in the higher-order potentials and utilizing the filter-based method, we are able to formulate the above algorithms such that each iteration has a complexity linear in the number of classes and random variables. The presented algorithms can be applied to any labeling problem using a dense CRF with sparse higher-order potentials. In this paper, we use semantic segmentation as an example application as it demonstrates the ability of the algorithm to scale to dense CRFs with large dimensions. We perform experiments on the Pascal dataset to indicate that the presented algorithms are able to attain lower energies than the mean-field inference method.

We propose a 2D multi-level appearance representation of the human body in RGB images, spatially modelled using a fully-connected graphical model. The appearance model is based on a CNN body part detector, which uses shared features in a cascade architecture to simultaneously detect body parts with different levels of granularity. We use a fully-connected Conditional Random Field (CRF) as our spatial model, over which approximate inference is efficiently performed using the Mean-Field algorithm, implemented as a Recurrent Neural Network (RNN). The stronger visual support from body parts with different levels of granularity, along with the fully-connected pairwise spatial relations, which have their weights learnt by the model, improve the performance of the bottom-up part detector. We adopt an end-to-end training strategy to leverage the potential of both our appearance and spatial models, and achieve competitive results on the MPII and LSP datasets.

Deep Neural Networks (DNNs) have demonstrated exceptional performance on most recognition tasks such as image classification and segmentation. However, they have also been shown to be vulnerable to adversarial examples. This phenomenon has recently attracted a lot of attention but it has not been extensively studied on multiple, large-scale datasets and structured prediction tasks such as semantic segmentation which often require more specialised networks with additional components such as CRFs, dilated convolutions, skip-connections and multiscale processing. In this paper, we present what to our knowledge is the first rigorous evaluation of adversarial attacks on modern semantic segmentation models, using two large-scale datasets. We analyse the effect of different network architectures, model capacity and multiscale processing, and show that many observations made on the task of classification do not always transfer to this more complex task. Furthermore, we show how mean-field inference in deep structured models, multiscale processing (and more generally, input transformations) naturally implement recently proposed adversarial defenses. Our observations will aid future efforts in understanding and defending against adversarial examples. Moreover, in the shorter term, we show how to effectively benchmark robustness and show which segmentation models should currently be preferred in safety-critical applications due to their inherent robustness.

We propose MAD-GAN, an intuitive generalization to the Generative Adversarial Networks (GANs) and its conditional variants to address the well known problem of mode collapse. First, MAD-GAN is a multi-agent GAN architecture incorporating multiple generators and one discriminator. Second, to enforce that different generators capture diverse high probability modes, the discriminator of MAD-GAN is designed such that along with finding the real and fake samples, it is also required to identify the generator that generated the given fake sample. Intuitively, to succeed in this task, the discriminator must learn to push different generators towards different identifiable modes. We perform extensive experiments on synthetic and real datasets and compare MAD-GAN with different variants of GAN. We show high quality diverse sample generations for challenging tasks such as image-to-image translation and face generation. In addition, we also show that MAD-GAN is able to disentangle different modalities when trained using highly challenging diverse-class dataset (e.g. dataset with images of forests, icebergs, and bedrooms). In the end, we show its efficacy on the unsupervised feature representation task.

We present FLIPDIAL, a generative model for Visual Dialogue that simultaneously plays the role of both participants in a visually-grounded dialogue. Given context in the form of an image and an associated caption summarising the contents of the image, FLIPDIAL learns both to answer questions and put forward questions, capable of generating entire sequences of dialogue (question-answer pairs) which are diverse and relevant to the image. To do this, FLIPDIAL relies on a simple but surprisingly powerful idea: it uses convolutional neural networks (CNNs) to encode entire dialogues directly, implicitly capturing dialogue context, and conditional VAEs to learn the generative model, FLIPDIAL outperforms the state-of-the-art model in the sequential answering task (1VD) on the VisDial dataset by 5 points in Mean Rank using the generated answers. We are the first to extend this paradigm to full two-way visual dialogue (2VD), where our model is capable of generating both questions and answers in sequence based on a visual input, for which we propose a set of novel evaluation measures and metrics.

We present a conceptually simple, flexible, and general framework for few-shot learning, where a classifier must learn to recognise new classes given only few examples from each. Our method, called the Relation Network (RN), is trained end-to-end from scratch. During meta-learning, it learns to learn a deep distance metric to compare a small number of images within episodes, each of which is designed to simulate the few-shot setting. Once trained, a RN is able to classify images of new classes by computing relation scores between query images and the few examples of each new class without further updating the network. Besides providing improved performance on few-shot learning, our framework is easily extended to zero-shot learning. Extensive experiments on five benchmarks demonstrate that our simple approach provides a unified and effective approach for both of these two tasks.

Stereo depth estimation is used for many computer vision applications. Though many popular methods strive solely for depth quality, for real-time mobile applications (e.g. prosthetic glasses or micro-UAVs), speed and power efficiency are equally, if not more, important. Many real-world systems rely on Semi-Global Matching (SGM) to achieve a good accuracy vs. speed balance, but power efficiency is hard to achieve with conventional hardware, making the use of embedded devices such as FPGAs attractive for low-power applications. However, the full SGM algorithm is ill-suited to deployment on FPGAs, and so most FPGA variants of it are partial, at the expense of accuracy. In a non-FPGA context, the accuracy of SGM has been improved by More Global Matching (MGM), which also helps tackle the streaking artifacts that afflict SGM. In this paper, we propose a novel, resource-efficient method that is inspired by MGM's techniques for improving depth quality, but which can be implemented to run in real time on a low-power FPGA. Through evaluation on multiple datasets (KITTI and Middlebury), we show that in comparison to other real-time capable stereo approaches, we can achieve a state-of-the-art balance between accuracy, power efficiency and speed, making our approach highly desirable for use in real-time systems with limited power.

The success of Deep Learning and its potential use in many safety-critical applications has motivated research on formal verification of Neural Network (NN) models. Despite the reputation of learned NN models to behave as black boxes and the theoretical hardness of proving their properties, researchers have been successful in verifying some classes of models by exploiting their piecewise linear structure and taking insights from formal methods such as Satisifiability Modulo Theory. These methods are however still far from scaling to realistic neural networks. To facilitate progress on this crucial area, we make two key contributions. First, we present a unified framework that encompasses previous methods. This analysis results in the identification of new methods that combine the strengths of multiple existing approaches, accomplishing a speedup of two orders of magnitude compared to the previous state of the art. Second, we propose a new data set of benchmarks which includes a collection of previously released testcases. We use the benchmark to provide the first experimental comparison of existing algorithms and identify the factors impacting the hardness of verification problems.

The vulnerability of deep image classification networks to adversarial attack is now well known, but less well understood. Via a novel experimental analysis, we illustrate some facts about deep convolutional networks for image classification that shed new light on their behaviour and how it connects to the problem of adversaries. In short, the celebrated performance of these networks and their vulnerability to adversarial attack are simply two sides of the same coin: the input image-space directions along which the networks are most vulnerable to attack are the same directions which they use to achieve their classification performance in the first place. We develop this result in two main steps. The first uncovers the fact that classes tend to be associated with specific image-space directions. This is shown by an examination of the class-score outputs of nets as functions of 1D movements along these directions. This provides a novel perspective on the existence of universal adversarial perturbations. The second is a clear demonstration of the tight coupling between classification performance and vulnerability to adversarial attack within the spaces spanned by these directions. Thus, our analysis resolves the apparent contradiction between accuracy and vulnerability. It provides a new perspective on much of the prior art and reveals profound implications for efforts to construct neural nets that are both accurate and robust to adversarial attack.

A key challenge in feature correspondence is the difficulty in differentiating true and false matches at a local descriptor level. This forces adoption of strict similarity thresholds that discard many true matches. However, if analyzed at a global level, false matches are usually randomly scattered while true matches tend to be coherent (clustered around a few dominant motions), thus creating a coherence based separability constraint. This paper proposes a non-linear regression technique that can discover such a coherence based separability constraint from highly noisy matches and embed it into a correspondence likelihood model. Once computed, the model can filter the entire set of nearest neighbor matches (which typically contains over 90 percent false matches) for true matches. We integrate our technique into a full feature correspondence system which reliably generates large numbers of good quality correspondences over wide baselines where previous techniques provide few or no matches.

We are motivated by the need for a generic object proposal generation algorithm which achieves good balance between object detection recall, proposal localization quality and computational efficiency. We propose a novel object proposal algorithm, BING++, which inherits the virtue of good computational efficiency of BING [1] but significantly improves its proposal localization quality. At high level we formulate the problem of object proposal generation from a novel probabilistic perspective, based on which our BING++ manages to improve the localization quality by employing edges and segments to estimate object boundaries and update the proposals sequentially. We propose learning the parameters efficiently by searching for approximate solutions in a quantized parameter space for complexity reduction. We demonstrate the generalization of BING++ with the same fixed parameters across different object classes and datasets. Empirically our BING++ can run at half speed of BING on CPU, but significantly improve the localization quality by 18.5 and 16.7 percent on both VOC2007 and Microhsoft COCO datasets, respectively. Compared with other state-of-the-art approaches, BING++ can achieve comparable performance, but run significantly faster.

Semantic segmentation and other pixel-level labeling tasks have made significant progress recently due to the deep learning paradigm. Many state-of-the-art structured prediction methods also include a random field model with a hand-crafted Gaussian potential to model spatial priors and label consistencies and feature-based image conditioning. These random field models with image conditioning typically require computationally demanding filtering techniques during inference. In this paper, we present a new inference and learning framework which can learn arbitrary pairwise conditional random field (CRF) potentials. Both standard spatial and high-dimensional bilateral kernels are considered. In addition, we introduce a new type of potential function which is image-dependent like the bilateral kernel, but an order of magnitude faster to compute since only spatial convolutions are employed. It is empirically demonstrated that such learned potentials can improve segmentation accuracy and that certain label-class interactions are indeed better modeled by a non-Gaussian potential. Our framework is evaluated on several public benchmarks for semantic segmentation with improved performance compared to previous state-of-the-art CNN+CRF models.

Reconstructing dense, volumetric models of real-world 3D scenes is important for many tasks, but capturing large scenes can take significant time, and the risk of transient changes to the scene goes up as the capture time increases. These are good reasons to want instead to capture several smaller sub-scenes that can be joined to make the whole scene. Achieving this has traditionally been difficult: joining sub-scenes that may never have been viewed from the same angle requires a high-quality camera relocaliser that can cope with novel poses, and tracking drift in each sub-scene can prevent them from being joined to make a consistent overall scene. Recent advances, however, have significantly improved our ability to capture medium-sized sub-scenes with little to no tracking drift: real-time globally consistent reconstruction systems can close loops and re-integrate the scene surface on the fly, whilst new visual-inertial odometry approaches can significantly reduce tracking drift during live reconstruction. Moreover, high-quality regression forest-based relocalisers have recently been made more practical by the introduction of a method to allow them to be trained and used online. In this paper, we leverage these advances to present what to our knowledge is the first system to allow multiple users to collaborate interactively to reconstruct dense, voxel-based models of whole buildings using only consumer-grade hardware, a task that has traditionally been both time-consuming and dependent on the availability of specialised hardware. Using our system, an entire house or lab can be reconstructed in under half an hour and at a far lower cost than was previously possible.

In recent years, major advances have been made in 3D scene reconstruction. However, much less progress has been made for objects, which can exhibit far fewer unambiguous geometric/texture cues than a full scene, and thus are much harder to track against. In this work we present a novel probabilistic object reconstruction framework that simultaneously allows for online, implicit deformation of the objects surface to reduce tracking drift and handle loop closure events. Coupled with our probabilistic formulation is the use of a multi subsegment representation of the object, used to enforce global consistency, with segmentation of the object built in to the formulation. Finally, we employ a CRF framework to refine the overall segmentation, defined by a probability field over the object. We present compelling results over the current state-of-the-art object reconstruction work and demonstrate robustness and consistency w.r.t. established dense SLAM frameworks.

The fully connected conditional random field (CRF) with Gaussian pairwise potentials has proven popular and effective for multi-class semantic segmentation. While the energy of a dense CRF can be minimized accurately using a linear programming (LP) relaxation, the state-of-the-art algorithm is too slow to be useful in practice. To alleviate this deficiency, we introduce an efficient LP minimization algorithm for dense CRFs. To this end, we develop a proximal minimization framework, where the dual of each proximal problem is optimized via block coordinate descent. We show that each block of variables can be efficiently optimized. Specifically, for one block, the problem decomposes into significantly smaller subproblems, each of which is defined over a single pixel. For the other block, the problem is optimized via conditional gradient descent. This has two advantages: 1) the conditional gradient can be computed in a time linear in the number of pixels and labels, and 2) the optimal step size can be computed analytically. Our experiments on standard datasets provide compelling evidence that our approach outperforms all existing baselines including the previous LP based approach for dense CRFs.

Semantic segmentation and object detection research have recently achieved rapid progress. However, the former task has no notion of different instances of the same object, and the latter operates at a coarse, bounding-box level. We propose an Instance Segmentation system that produces a segmentation map where each pixel is assigned an object class and instance identity label. Most approaches adapt object detectors to produce segments instead of boxes. In contrast, our method is based on an initial semantic segmentation module, which feeds into an instance subnetwork. This subnetwork uses the initial category-level segmentation, along with cues from the output of an object detector, within an end-to-end CRF to predict instances. This part of our model is dynamically instantiated to produce a variable number of instances per image. Our end-to-end approach requires no post-processing and considers the image holistically, instead of processing independent proposals. Therefore, unlike some related work, a pixel cannot belong to multiple instances. Furthermore, far more precise segmentations are achieved, as shown by our substantial improvements at high APr thresholds.

Camera relocalisation is an important problem in computer vision, with applications in simultaneous localisation and mapping, virtual/augmented reality and navigation. Common techniques either match the current image against keyframes with known poses coming from a tracker, or establish 2D-to-3D correspondences between keypoints in the current image and points in the scene in order to estimate the camera pose. Recently, regression forests have become a popular alternative to establish such correspondences. They achieve accurate results, but must be trained offline on the target scene, preventing relocalisation in new environments. In this paper, we show how to circumvent this limitation by adapting a pre-trained forest to a new scene on the fly. Our adapted forests achieve relocalisation performance that is on par with that of offline forests, and our approach runs in under 150ms, making it desirable for real-time systems that require online relocalisation.

Recent progress on saliency detection is substantial, benefiting mostly from the explosive development of Convolutional Neural Networks (CNNs). Semantic segmentation and saliency detection algorithms developed lately have been mostly based on Fully Convolutional Neural Networks (FCNs). There is still a large room for improvement over the generic FCN models that do not explicitly deal with the scale-space problem. Holisitcally-Nested Edge Detector (HED) provides a skip-layer structure with deep supervision for edge and boundary detection, but the performance gain of HED on saliency detection is not obvious. In this paper, we propose a new saliency method by introducing short connections to the skip-layer structures within the HED architecture. Our framework provides rich multi-scale feature maps at each layer, a property that is critically needed to perform segment detection. Our method produces state-of-the-art results on 5 widely tested salient object detection benchmarks, with advantages in terms of efficiency (0.08 seconds per image), effectiveness, and simplicity over the existing algorithms.

We introduce a Deep Stochastic IOC RNN Encoder-decoder framework, DESIRE, for the task of future predictions of multiple interacting agents in dynamic scenes. DESIRE effectively predicts future locations of objects in multiple scenes by 1) accounting for the multi-modal nature of the future prediction (i.e., given the same context, future may vary), 2) foreseeing the potential future outcomes and make a strategic prediction based on that, and 3) reasoning not only from the past motion history, but also from the scene context as well as the interactions among the agents. DESIRE achieves these in a single end-to-end trainable neural network model, while being computationally efficient. The model first obtains a diverse set of hypothetical future prediction samples employing a conditional variational auto-encoder, which are ranked and refined by the following RNN scoring-regression module. Samples are scored by accounting for accumulated future rewards, which enables better long-term strategic decisions similar to IOC frameworks. An RNN scene context fusion module jointly captures past motion histories, the semantic scene context and interactions among multiple agents. A feedback mechanism iterates over the ranking and refinement to further boost the prediction accuracy. We evaluate our model on two publicly available datasets: KITTI and Stanford Drone Dataset. Our experiments show that the proposed model significantly improves the prediction accuracy compared to other baseline methods.

Rotoscoping, the detailed delineation of scene elements through a video shot, is a painstaking task of tremendous importance in professional post-production pipelines. While pixel-wise segmentation techniques can help for this task, professional rotoscoping tools rely on parametric curves that offer the artists a much better interactive control on the definition, editing and manipulation of the segments of interest. Sticking to this prevalent rotoscoping paradigm, we propose a novel framework to capture and track the visual aspect of an arbitrary object in a scene, given a first closed outline of this object. This model combines a collection of local foreground/background appearance models spread along the outline, a global appearance model of the enclosed object and a set of distinctive foreground landmarks. The structure of this rich appearance model allows simple initialization, efficient iterative optimization with exact minimization at each step, and on-line adaptation in videos. We demonstrate qualitatively and quantitatively the merit of this framework through comparisons with tools based on either dynamic segmentation with a closed curve or pixel-wise binary labelling.

The Correlation Filter is an algorithm that trains a linear template to discriminate between images and their translations. It is well suited to object tracking because its formulation in the Fourier domain provides a fast solution, enabling the detector to be re-trained once per frame. Previous works that use the Correlation Filter, however, have adopted features that were either manually designed or trained for a different task. This work is the first to overcome this limitation by interpreting the Correlation Filter learner, which has a closed-form solution, as a differentiable layer in a deep neural network. This enables learning deep features that are tightly coupled to the Correlation Filter. Experiments illustrate that our method has the important practical benefit of allowing lightweight architectures to achieve state-of-the-art performance at high framerates.

We present a deep-learning framework for real-time multiple spatio-temporal (S/T) action localisation and classification. Current state-of-the-art approaches work offline, and are too slow to be useful in real-world settings. To overcome their limitations we introduce two major developments. Firstly, we adopt real-time SSD (Single Shot Multi-Box Detector) CNNs to regress and classify detection boxes in each video frame potentially containing an action of interest. Secondly, we design an original and efficient online algorithm to incrementally construct and label ‘action tubes’ from the SSD frame level detections. As a result, our system is not only capable of performing S/T detection in real time, but can also perform early action prediction in an online fashion. We achieve new state-of-the-art results in both S/T action localisation and early action prediction on the challenging UCF101-24 and J-HMDB-21 benchmarks, even when compared to the top offline competitors. To the best of our knowledge, ours is the first real-time (up to 40fps) system able to perform online S/T action localisation on the untrimmed videos of UCF101-24.

Many real-world problems, such as network packet routing and urban traffic control, are naturally modeled as multi-agent reinforcement learning (RL) problems. However, existing multi-agent RL methods typically scale poorly in the problem size. Therefore, a key challenge is to translate the success of deep learning on single-agent RL to the multi-agent setting. A major stumbling block is that independent Q-learning, the most popular multi-agent RL method, introduces nonstationarity that makes it incompatible with the experience replay memory on which deep Q-learning relies. This paper proposes two methods that address this problem: 1) using a multi-agent variant of importance sampling to naturally decay obsolete data and 2) conditioning each agent’s value function on a fingerprint that disambiguates the age of the data sampled from the replay memory. Results on a challenging decentralised variant of StarCraft unit micromanagement confirm that these methods enable the successful combination of experience replay with multi-agent RL.

This work addresses the task of camera localization in a known 3D scene given a single input RGB image. State-of-the-art approaches accomplish this in two steps: firstly, regressing for every pixel in the image its 3D scene coordinate and subsequently, using these coordinates to estimate the final 6D camera pose via RANSAC. To solve the first step. Random Forests (RFs) are typically used. On the other hand. Neural Networks (NNs) reign in many dense regression tasks, but are not test-time efficient. We ask the question: which of the two is best for camera localization? To address this, we make two method contributions: (1) a test-time efficient NN architecture which we term a ForestNet that is derived and initialized from a RF, and (2) a new fully-differentiable robust averaging technique for regression ensembles which can be trained end-to-end with a NN. Our experimental findings show that for scene coordinate regression, traditional NN architectures are superior to test-time efficient RFs and ForestNets, however, this does not translate to final 6D camera pose accuracy where RFs and ForestNets perform slightly better. To summarize, our best method, a ForestNet with a robust average, which has an equivalent fast and lightweight RF, improves over the state-of-the-art for camera localization on the 7-Scenes dataset [1]. While this work focuses on scene coordinate regression for camera localization, our innovations may also be applied to other continuous regression tasks.

Variational autoencoders (VAEs) learn representations of data by jointly training a probabilistic encoder and decoder network. Typically these models encode all features of the data into a single variable. Here we are interested in learning disentangled representations that encode distinct aspects of the data into separate variables. We propose to learn such representations using model architectures that generalise from standard VAEs, employing a general graphical model structure in the encoder and decoder. This allows us to train partially-specified models that make relatively strong assumptions about a subset of interpretable variables and rely on the flexibility of neural networks to learn representations for the remaining variables. We further define a general objective for semi-supervised learning in this model class, which can be approximated using an importance sampling procedure. We evaluate our framework's ability to learn disentangled representations, both by qualitative exploration of its generative capacity, and quantitative evaluation of its discriminative ability on a variety of models and datasets.

In computer vision, many problems can be formulated as binary quadratic programs (BQPs), which are in general NP hard. Finding a solution when the problem is of large size to be of practical interest typically requires relaxation. Semidefinite relaxation usually yields tight bounds, but its computational complexity is high. In this work, we present a semidefinite programming (SDP) formulation for BQPs, with two desirable properties. First, it produces similar bounds to the standard SDP formulation. Second, compared with the conventional SDP formulation, the proposed SDP formulation leads to a considerably more efficient and scalable dual optimization approach. We then propose two solvers, namely, quasi-Newton and smoothing Newton methods, for the simplified dual problem. Both of them are significantly more efficient than standard interior-point methods. Empirically the smoothing Newton solver is faster than the quasi-Newton solver for dense or medium-sized problems, while the quasi-Newton solver is preferable for large sparse/structured problems.

In this paper, we present a novel, general, and efficient architecture for addressing computer vision problems that are approached from an 'Analysis by Synthesis' standpoint. Analysis by synthesis involves the minimization of reconstruction error, which is typically a non-convex function of the latent target variables. State-of-the-art methods adopt a hybrid scheme where discriminatively trained predictors like Random Forests or Convolutional Neural Networks are used to initialize local search algorithms. While these hybrid methods have been shown to produce promising results, they often get stuck in local optima. Our method goes beyond the conventional hybrid architecture by not only proposing multiple accurate initial solutions but by also defining a navigational structure over the solution space that can be used for extremely efficient gradient-free local search. We demonstrate the efficacy and generalizability of our approach on tasks as diverse as Hand Pose Estimation, RGB Camera Relocalization, and Image Retrieval.

Correlation Filter-based trackers have recently achieved excellent performance, showing great robustness to challenging situations exhibiting motion blur and illumination changes. However, since the model that they learn depends strongly on the spatial layout of the tracked object, they are notoriously sensitive to deformation. Models based on colour statistics have complementary traits: they cope well with variation in shape, but suffer when illumination is not consistent throughout a sequence. Moreover, colour distributions alone can be insufficiently discriminative. In this paper, we show that a simple tracker combining complementary cues in a ridge regression framework can operate faster than 80 FPS and outperform not only all entries in the popular VOT14 competition, but also recent and far more sophisticated trackers according to multiple benchmarks.

Learning attribute models for applications like Zero-Shot Learning (ZSL) and image search is challenging because they require attribute classifiers to generalize to test data that may be very different from the training data. A typical scenario is when the notion of an attribute may differ from one user to another, e.g. one user may find a shoe formal whereas another user may not. In this case, the distribution of labels at test time is different from that at training time. We argue that due to the uncertainty in what the test distribution might be, committing to one attribute model during training is not advisable. We propose a novel framework for attribute learning which involves training an ensemble of diverse models for attributes and identifying experts from them at test time given a small amount of personalized annotations from a user. Our approach for attribute personalization is not specific to any classification model and we show results using Random Forest and SVM ensembles. We experiment with 2 datasets: SUN Attributes and Shoes and show significant improvements over baselines.

One-shot learning is usually tackled by using generative models or discriminative embeddings. Discriminative methods based on deep learning, which are very effective in other learning scenarios, are ill-suited for one-shot learning as they need large amounts of training data. In this paper, we propose a method to learn the parameters of a deep model in one shot. We construct the learner as a second deep network, called a learnet, which predicts the parameters of a pupil network from a single exemplar. In this manner we obtain an efficient feed-forward one-shot learner, trained end-to-end by minimizing a one-shot classification objective in a learning to learn formulation. In order to make the construction feasible, we propose a number of factorizations of the parameters of the pupil network. We demonstrate encouraging results by learning characters from single exemplars in Omniglot, and by tracking visual objects from a single initial exemplar in the Visual Object Tracking benchmark.

This paper proposes an adaptive neural-compilation framework to address the problem of learning efficient program. Traditional code optimisation strategies used in compilers are based on applying pre-specified set of transformations that make the code faster to execute without changing its semantics. In contrast, our work involves adapting programs to make them more efficient while considering correctness only on a target input distribution. Our approach is inspired by the recent works on differentiable representations of programs. We show that it is possible to compile programs written in a low-level language to a differentiable representation. We also show how programs in this representation can be optimised to make them efficient on a target distribution of inputs. Experimental results demonstrate that our approach enables learning specifically-tuned algorithms for given data distributions with a high success rate.

Adaptive tracking-by-detection methods are widely used in computer vision for tracking arbitrary objects. Current approaches treat the tracking problem as a classification task and use online learning techniques to update the object model. However, for these updates to happen one needs to convert the estimated object position into a set of labelled training examples, and it is not clear how best to perform this intermediate step. Furthermore, the objective for the classifier (label prediction) is not explicitly coupled to the objective for the tracker (estimation of object position). In this paper, we present a framework for adaptive visual object tracking based on structured output prediction. By explicitly allowing the output space to express the needs of the tracker, we avoid the need for an intermediate classification step. Our method uses a kernelised structured output support vector machine (SVM), which is learned online to provide adaptive tracking. To allow our tracker to run at high frame rates, we (a) introduce a budgeting mechanism that prevents the unbounded growth in the number of support vectors that would otherwise occur during tracking, and (b) show how to implement tracking on the GPU. Experimentally, we show that our algorithm is able to outperform state-of-the-art trackers on various benchmark videos. Additionally, we show that we can easily incorporate additional features and kernels into our framework, which results in increased tracking performance.

Object proposal algorithms have shown great promise as a first step for object recognition and detection. Good object proposal generation algorithms require high object detection recall rate as well as low computational cost, because generating object proposals is usually utilized as a preprocessing step. The problem of how to accelerate the object proposal generation and evaluation process without decreasing recall is thus of great interest. In this paper, we propose a new object proposal generation method using two-stage cascade support vector machines (SVMs), where in the first stage linear filters are learned for predefined quantized scales/aspect-ratios independently, and in the second stage a global linear classifier is learned across all the quantized scales/aspect-ratios for calibration, so that all the windows from the first stage can be compared properly. The windows with highest scores from the second stage are kept as inputs to our new efficient proposal calibration algorithm to improve their localization quality significantly, resulting in our final object proposals. We explain our scale/aspect-ratio quantization scheme, and investigate the effects of combinations of ℓ1and ℓ2regularizers in cascade SVMs with/without ranking constraints in learning. Comprehensive experiments on VOC2007 dataset are conducted, and our method is comparable with the current state-of-the-art methods with much better computational efficiency.

It is not always possible to recognise objects and infer material properties for a scene from visual cues alone, since objects can look visually similar whilst being made of very different materials. In this paper, we therefore present an approach that augments the available dense visual cues with sparse auditory cues in order to estimate dense object and material labels. Since estimates of object class and material properties are mutually informative, we optimise our multi-output labelling jointly using a random-field framework. We evaluate our system on a new dataset with paired visual and auditory data that we make publicly available. We demonstrate that this joint estimation of object and material labels significantly outperforms the estimation of either category in isolation.

Recent works on zero-shot learning make use of side information such as visual attributes or natural language semantics to define the relations between output visual classes and then use these relationships to draw inference on new unseen classes at test time. In a novel extension to this idea, we propose the use of visual prototypical concepts as side information. For most real-world visual object categories, it may be difficult to establish a unique prototype. However, in cases such as traffic signs, brand logos, flags, and even natural language characters, these prototypical templates are available and can be leveraged for an improved recognition performance. The present work proposes a way to incorporate this prototypical information in a deep learning framework. Using prototypes as prior information, the deepnet pipeline learns the input image projections into the prototypical embedding space subject to minimization of the final classification loss. Based on our experiments with two different datasets of traffic signs and brand logos, prototypical embeddings incorporated in a conventional convolutional neural network improve the recognition performance. Recognition accuracy on the Belga logo dataset is especially noteworthy and establishes a new state-of-the-art. In zero-shot learning scenarios, the same system can be directly deployed to draw inference on unseen classes by simply adding the prototypical information for these new classes at test time. Thus, unlike earlier approaches, testing on seen and unseen classes is handled using the same pipeline, and the system can be tuned for a trade-off of seen and unseen class performance as per task requirement. Comparison with one of the latest works in the zero-shot learning domain yields top results on the two datasets mentioned above.

We present an augmented reality system for large scale 3D reconstruction and recognition in outdoor scenes. Unlike existing prior work, which tries to reconstruct scenes using active depth cameras, we use a purely passive stereo setup, allowing for outdoor use and extended sensing range. Our system not only produces a map of the 3D environment in real-time, it also allows the user to draw (or 'paint') with a laser pointer directly onto the reconstruction to segment the model into objects. Given these examples our system then learns to segment other parts of the 3D map during online acquisition. Unlike typical object recognition systems, ours therefore very much places the user 'in the loop' to segment particular objects of interest, rather than learning from predefined databases. The laser pointer additionally helps to 'clean up' the stereo reconstruction and final 3D map, interactively. Using our system, within minutes, a user can capture a full 3D map, segment it into objects of interest, and refine parts of the model during capture. We provide full technical details of our system to aid replication, as well as quantitative evaluation of system components. We demonstrate the possibility of using our system for helping the visually impaired navigate through spaces. Beyond this use, our system can be used for playing large-scale augmented reality games, shared online to augment streetview data, and used for more detailed car and person navigation.

In this paper we show that multiple object tracking (MOT) can be formulated in a framework, where the detection and data-association are performed simultaneously. Our method allows us to overcome the confinements of data association based MOT approaches; where the performance is dependent on the object detection results provided at input level. At the core of our method lies structured learning which learns a model for each target and infers the best location of all targets simultaneously in a video clip. The inference of our structured learning is done through a new Target Identity-aware Network Flow (TINF), where each node in the network encodes the probability of each target identity belonging to that node. The proposed Lagrangian relaxation optimization finds the high quality solution to the network. During optimization a soft spatial constraint is enforced between the nodes of the graph which helps reducing the ambiguity caused by nearby targets with similar appearance in crowded scenarios. We show that automatically detecting and tracking targets in a single framework can help resolve the ambiguities due to frequent occlusion and heavy articulation of targets. Our experiments involve challenging yet distinct datasets and show that our method can achieve results better than the state-of-art.

Recent advances in camera relocalization use predictions from a regression forest to guide the camera pose optimization procedure. In these methods, each tree associates one pixel with a point in the scene's 3D world coordinate frame. In previous work, these predictions were point estimates and the subsequent camera pose optimization implicitly assumed an isotropic distribution of these estimates. In this paper, we train a regression forest to predict mixtures of anisotropic 3D Gaussians and show how the predicted uncertainties can be taken into account for continuous pose optimization. Experiments show that our proposed method is able to relocalize up to 40% more frames than the state of the art.

Pixel-level labelling tasks, such as semantic segmentation, play a central role in image understanding. Recent approaches have attempted to harness the capabilities of deep learning techniques for image recognition to tackle pixel-level labelling tasks. One central issue in this methodology is the limited capacity of deep learning techniques to delineate visual objects. To solve this problem, we introduce a new form of convolutional neural network that combines the strengths of Convolutional Neural Networks (CNNs) and Conditional Random Fields (CRFs)-based probabilistic graphical modelling. To this end, we formulate Conditional Random Fields with Gaussian pairwise potentials and mean-field approximate inference as Recurrent Neural Networks. This network, called CRF-RNN, is then plugged in as a part of a CNN to obtain a deep network that has desirable properties of both CNNs and CRFs. Importantly, our system fully integrates CRF modelling with CNNs, making it possible to train the whole deep network end-to-end with the usual back-propagation algorithm, avoiding offline post-processing methods for object delineation. We apply the proposed method to the problem of semantic image segmentation, obtaining top results on the challenging Pascal VOC 2012 segmentation benchmark.

Zero-shot learning consists in learning how to recognize new concepts by just having a description of them. Many sophisticated approaches have been proposed to address the challenges this problem comprises. In this paper we describe a zero-shot learning approach that can be implemented in just one line of code, yet it is able to outperform state of the art approaches on standard datasets. The approach is based on a more general framework which models the relationships between features, attributes, and classes as a two linear layers network, where the weights of the top layer are not learned but are given by the environment. We further provide a learning bound on the generalization error of this kind of approaches, by casting them as domain adaptation methods. In experiments carried out on three standard real datasets, we found that our approach is able to perform significantly better than the state of art on all of them, obtaining a ratio of improvement up to 17%.

Our abilities in scene understanding, which allow us to perceive the 3D structure of our surroundings and intuitively recognise the objects we see, are things that we largely take for granted, but for robots, the task of understanding large scenes quickly remains extremely challenging. Recently, scene understanding approaches based on 3D reconstruction and semantic segmentation have become popular, but existing methods either do not scale, fail outdoors, provide only sparse reconstructions or are rather slow. In this paper, we build on a recent hash-based technique for large-scale fusion and an efficient mean-field inference algorithm for densely-connected CRFs to present what to our knowledge is the first system that can perform dense, large-scale, outdoor semantic reconstruction of a scene in (near) real time. We also present a `semantic fusion' approach that allows us to handle dynamic objects more effectively than previous approaches. We demonstrate the effectiveness of our approach on the KITTI dataset, and provide qualitative and quantitative results showing high-quality dense reconstruction and labelling of a number of scenes.

Aquiring reliable depth maps is an essential prerequisite for accurate and incremental 3D reconstruction used in a variety of robotics applications. Depth maps produced by affordable Kinect-like cameras have become a de-facto standard for indoor reconstruction and the driving force behind the success of many algorithms. However, Kinect-like cameras are less effective outdoors where one should rely on other sensors. Often, we use a combination of a stereo camera and lidar, however, process the acquired data in independent pipelines which generally leads to sub-optimal performance since both sensors suffer from different drawbacks. In this paper, we propose a probabilistic model that efficiently exploits complementarity between different depth-sensing modalities for incremental dense scene reconstruction. Our model uses a piecewise planarity prior assumption which is common in both the indoor and outdoor scenes. We demonstrate the effectiveness of our approach on the KITTI dataset, and provide qualitative and quantitative results showing high-quality dense reconstruction of a number of scenes.

Figure-ground segmentation from bounding box input, provided either automatically or manually, has been extremely popular in the last decade and influenced various applications. A lot of research has focused on high-quality segmentation, using complex formulations which often lead to slow techniques, and often hamper practical usage. In this paper we demonstrate a very fast segmentation technique which still achieves very high quality results. We propose to replace the time consuming iterative refinement of global colour models in traditional GrabCut formulation by a densely connected crf. To motivate this decision, we show that a dense crf implicitly models unnormalized global colour models for foreground and background. Such relationship provides insightful analysis to bridge between dense crf and GrabCut functional. We extensively evaluate our algorithm using two famous benchmarks. Our experimental results demonstrated that the proposed algorithm achieves an order of magnitude (10×) speed-up with respect to the closest competitor, and at the same time achieves a considerably higher accuracy.

Automatic estimation of salient object regions across images, without any prior assumption or knowledge of the contents of the corresponding scenes, enhances many computer vision and computer graphics applications. We introduce a regional contrast based salient object detection algorithm, which simultaneously evaluates global contrast differences and spatial weighted coherence scores. The proposed algorithm is simple, efficient, naturally multi-scale, and produces full-resolution, high-quality saliency maps. These saliency maps are further used to initialize a novel iterative version of GrabCut, namely SaliencyCut, for high quality unsupervised salient object segmentation. We extensively evaluated our algorithm using traditional salient object detection datasets, as well as a more challenging Internet image dataset. Our experimental results demonstrate that our algorithm consistently outperforms 15 existing salient object detection and segmentation methods, yielding higher precision and better recall rates. We also show that our algorithm can be used to efficiently extract salient object masks from Internet images, enabling effective sketch-based image retrieval (SBIR) via simple shape comparisons. Despite such noisy internet images, where the saliency regions are ambiguous, our saliency guided image retrieval achieves a superior retrieval rate compared with state-of-the-art SBIR methods, and additionally provides important target object region information.

We present a new interactive and online approach to 3D scene understanding. Our system, SemanticPaint , allows users to simultaneously scan their environment whilst interactively segmenting the scene simply by reaching out and touching any desired object or surface. Our system continuously learns from these segmentations, and labels new unseen parts of the environment. Unlike offline systems where capture, labeling, and batch learning often take hours or even days to perform, our approach is fully online. This provides users with continuous live feedback of the recognition during capture, allowing to immediately correct errors in the segmentation and/or learning—a feature that has so far been unavailable to batch and offline methods. This leads to models that are tailored or personalized specifically to the user's environments and object classes of interest, opening up the potential for new applications in augmented reality, interior design, and human/robot navigation. It also provides the ability to capture substantial labeled 3D datasets for training large-scale visual recognition systems.

Volumetric methods provide efficient, flexible and simple ways of integrating multiple depth images into a full 3D model. They provide dense and photorealistic 3D reconstructions, and parallelised implementations on GPUs achieve real-time performance on modern graphics hardware. To run such methods on mobile devices, providing users with freedom of movement and instantaneous reconstruction feedback, remains challenging however. In this paper we present a range of modifications to existing volumetric integration methods based on voxel block hashing, considerably improving their performance and making them applicable to tablet computer applications. We present (i) optimisations for the basic data structure, and its allocation and integration; (ii) a highly optimised raycasting pipeline; and (iii) extensions to the camera tracker to incorporate IMU data. In total, our system thus achieves frame rates up 47 Hz on a Nvidia Shield Tablet and 910 Hz on a Nvidia GTX Titan XGPU, or even beyond 1.1 kHz without visualisation.

Training a generic objectness measure to produce a small set of candidate object windows, has been shown to speed up the classical sliding window object detection paradigm. We observe that generic objects with well-defined closed boundary can be discriminated by looking at the norm of gradients, with a suitable resizing of their corresponding image windows in to a small fixed size. Based on this observation and computational reasons, we propose to resize the window to 8 × 8 and use the norm of the gradients as a simple 64D feature to describe it, for explicitly training a generic objectness measure. We further show how the binarized version of this feature, namely binarized normed gradients (BING), can be used for efficient objectness estimation, which requires only a few atomic operations (e.g. ADD, BITWISE SHIFT, etc.). Experiments on the challenging PASCAL VOC 2007 dataset show that our method efficiently (300fps on a single laptop CPU) generates a small set of category-independent, high quality object windows, yielding 96.2% object detection rate (DR) with 1, 000 proposals. Increasing the numbers of proposals and color spaces for computing BING features, our performance can be further improved to 99.5% DR.

The concepts of objects and attributes are both important for describing images precisely, since verbal descriptions often contain both adjectives and nouns (e.g. 'I see a shiny red chair'). In this paper, we formulate the problem of joint visual attribute and object class image segmentation as a dense multi-labelling problem, where each pixel in an image can be associated with both an object-class and a set of visual attributes labels. In order to learn the label correlations, we adopt a boosting-based piecewise training approach with respect to the visual appearance and co-occurrence cues. We use a filtering-based mean-field approximation approach for efficient joint inference. Further, we develop a hierarchical model to incorporate region-level object and attribute information. Experiments on the aPASCAL, CORE and attribute augmented NYU indoor scenes datasets show that the proposed approach is able to achieve state-of-the-art results.

This paper makes two contributions: the first is the proposal of a new model-The associative hierarchical random field (AHRF), and a novel algorithm for its optimization; the second is the application of this model to the problem of semantic segmentation. Most methods for semantic segmentation are formulated as a labeling problem for variables that might correspond to either pixels or segments such as super-pixels. It is well known that the generation of super pixel segmentations is not unique. This has motivated many researchers to use multiple super pixel segmentations for problems such as semantic segmentation or single view reconstruction. These super-pixels have not yet been combined in a principled manner, this is a difficult problem, as they may overlap, or be nested in such a way that the segmentations form a segmentation tree. Our new hierarchical random field model allows information from all of the multiple segmentations to contribute to a global energy. MAP inference in this model can be performed efficiently using powerful graph cut based move making algorithms. Our framework generalizes much of the previous work based on pixels or segments, and the resulting labelings can be viewed both as a detailed segmentation at the pixel level, or at the other extreme, as a segment selector that pieces together a solution like a jigsaw, selecting the best segments from different segmentations as pieces. We evaluate its performance on some of the most challenging data sets for object class segmentation, and show that this ability to perform inference using multiple overlapping segmentations leads to state-of-the-art results.

Humans describe images in terms of nouns and adjectives while algorithms operate on images represented as sets of pixels. Bridging this gap between how humans would like to access images versus their typical representation is the goal of image parsing, which involves assigning object and attribute labels to pixels. In this article we propose treating nouns as object labels and adjectives as visual attribute labels. This allows us to formulate the image parsing problem as one of jointly estimating per-pixel object and attribute labels from a set of training images. We propose an efficient (interactive time) solution. Using the extracted labels as handles, our system empowers a user to verbally refine the results. This enables hands-free parsing of an image into pixel-wise object/attribute labels that correspond to human semantics. Verbally selecting objects of interest enables a novel and natural interaction modality that can possibly be used to interact with new generation devices (e.g., smartphones, Google Glass, livingroom devices). We demonstrate our system on a large number of real-world images with varying complexity. To help understand the trade-offs compared to traditional mouse-based interactions, results are reported for both a large-scale quantitative evaluation and a user study.

Our goal is to detect humans and estimate their 2D pose in single images. In particular, handling cases of partial visibility where some limbs may be occluded or one person is partially occluding another. Two standard, but disparate, approaches have developed in the field: the first is the part based approach for layout type problems, involving optimising an articulated pictorial structure, the second is the pixel based approach for image labelling involving optimising a random field graph defined on the image. Our novel contribution is a formulation for pose estimation which combines these two models in a principled way in one optimisation problem and thereby inherits the advantages of both of them. Inference on this joint model finds the set of instances of persons in an image, the location of their joints, and a pixel-wise body part labelling. We achieve near or state of the art results on standard human pose data sets, and demonstrate the correct estimation for cases of self-occlusion, person overlap and image truncation.

Semantic reconstruction of a scene is important for a variety of applications such as 3D modelling, object recognition and autonomous robotic navigation. However, most object labelling methods work in the image domain and fail to capture the information present in 3D space. In this work we propose a principled way to generate object labelling in 3D. Our method builds a triangulated meshed representation of the scene from multiple depth estimates. We then define a CRF over this mesh, which is able to capture the consistency of geometric properties of the objects present in the scene. In this framework, we are able to generate object hypotheses by combining information from multiple sources: geometric properties (from the 3D mesh), and appearance properties (from images). We demonstrate the robustness of our framework in both indoor and outdoor scenes. For indoor scenes we created an augmented version of the NYU indoor scene dataset (RGBD images) with object labelled meshes for training and evaluation. For outdoor scenes, we created ground truth object labellings for the KITTY odometry dataset (stereo image sequence). We observe a significant speed-up in the inference stage by performing labelling on the mesh, and additionally achieve higher accuracies.

In this paper we propose a robust algorithm that generates an efficient and accurate dense 3D reconstruction with associated semantic labellings. Intelligent autonomous systems require accurate 3D reconstructions for applications such as navigation and localisation. Such systems also need to recognise their surroundings in order to identify and interact with objects of interest. Considerable emphasis has been given to generating a good reconstruction but less effort has gone into generating a 3D semantic model. The inputs to our algorithm are street level stereo image pairs acquired from a camera mounted on a moving vehicle. The depth-maps, generated from the stereo pairs across time, are fused into a global 3D volume online in order to accommodate arbitrary long image sequences. The street level images are automatically labelled using a Conditional Random Field (CRF) framework exploiting stereo images, and label estimates are aggregated to annotate the 3D volume. We evaluate our approach on the KITTI odometry dataset and have manually generated ground truth for object class segmentation. Our qualitative evaluation is performed on various sequences of the dataset and we also quantify our results on a representative subset.

Many methods have been proposed to recover the intrinsic scene properties such as shape, reflectance and illumination from a single image. However, most of these models have been applied on laboratory datasets. In this work we explore the synergy effects between intrinsic scene properties recovered from an image, and the objects and attributes present in the scene. We cast the problem in a joint energy minimization framework; thus our model is able to encode the strong correlations between intrinsic properties (reflectance, shape, illumination), objects (table, tv-monitor), and materials (wooden, plastic) in a given scene. We tested our approach on the NYU and Pascal datasets, and observe both qualitative and quantitative improvements in the overall accuracy.

Current state-of-the-art action classification methods extract feature representations from the entire video clip in which the action unfolds, however this representation may include irrelevant scene context and movements which are shared amongst multiple action classes. For example, a waving action may be performed whilst walking, however if the walking movement and scene context appear in other action classes, then they should not be included in a waving movement classifier. In this work, we propose an action classification framework in which more discriminative action subvolumes are learned in a weakly supervised setting, owing to the difficulty of manually labelling massive video datasets. The learned models are used to simultaneously classify video clips and to localise actions to a given space-time subvolume. Each subvolume is cast as a bag-offeatures (BoF) instance in a multiple-instance-learning framework, which in turn is used to learn its class membership. We demonstrate quantitatively that even with single fixedsized subvolumes, the classification performance of our proposed algorithm is superior to the state-of-the-art BoF baseline on the majority of performance measures, and shows promise for space-time action localisation on the most challenging video datasets.

Semantic image segmentation is a problem of simultaneous segmentation and recognition of an input image into regions and their associated categorical labels, such as person, car or cow. A popular way to achieve this goal is to assign a label to every pixel in the input image and impose simple structural constraints on the output label space. Efficient approximation algorithms for solving this labelling problem such as α-expansion have, at best, linear runtime complexity with respect to the number of labels, making them practical only when working in a specific domain that has few classes-of-interest. However when working in a more general setting where the number of classes could easily reach tens of thousands, sub-linear complexity is desired. In this paper we propose meeting this requirement by performing cascaded inference that wraps around the α-expansion algorithm. The cascade both divides the large label set into smaller more manageable ones by way of a hierarchy, and dynamically subdivides the image into smaller and smaller regions during inference. We test our method on the SUN09 dataset with 107 accurately hand labelled classes.

Recently, Krahenbuhl and Koltun proposed an efficient inference method for densely connected pairwise random fields using the mean-field approximation for a Conditional Random Field (CRF). However, they restrict their pairwise weights to take the form of a weighted combination of Gaussian kernels where each Gaussian component is allowed to take only zero mean, and can only be rescaled by a single value for each label pair. Further, their method is sensitive to initialization. In this paper, we propose methods to alleviate these issues. First, we propose a hierarchical mean-field approach where labelling from the coarser level is propagated to the finer level for better initialisation. Further, we use SIFT-flow based label transfer to provide a good initial condition at the coarsest level. Second, we allow our approach to take general Gaussian pairwise weights, where we learn the mean, the co-variance matrix, and the mixing co-efficient for every mixture component. We propose a variation of Expectation Maximization (EM) for piecewise learning of the parameters of the mixture model determined by the maximum likelihood function. Finally, we demonstrate the efficiency and accuracy offered by our method for object class segmentation problems on two challenging datasets: PascalVOC-10 segmentation and CamVid datasets. We show that we are able to achieve state of the art performance on the CamVid dataset, and an almost 3% improvement on the PascalVOC10 dataset compared to baseline graph-cut and mean-field methods, while also reducing the inference time by almost a factor of 3 compared to graph-cuts based methods.

Efficient keypoint-based object detection methods are used in many real-time computer vision applications. These approaches often model an object as a collection of keypoints and associated descriptors, and detection then involves first constructing a set of correspondences between object and image keypoints via descriptor matching, and subsequently using these correspondences as input to a robust geometric estimation algorithm such as RANSAC to find the transformation of the object in the image. In such approaches, the object model is generally constructed offline, and does not adapt to a given environment at runtime. Furthermore, the feature matching and transformation estimation stages are treated entirely separately. In this paper, we introduce a new approach to address these problems by combining the overall pipeline of correspondence generation and transformation estimation into a single structured output learning framework. Following the recent trend of using efficient binary descriptors for feature matching, we also introduce an approach to approximate the learned object model as a collection of binary basis functions which can be evaluated very efficiently at runtime. Experiments on challenging video sequences show that our algorithm significantly improves over state-of-the-art descriptor matching techniques using a range of descriptors, as well as recent online learning based approaches.

A large number of problems in computer vision can be modeled as energy minimization problems in a markov random field (MRF) framework. Many methods have been developed over the years for efficient inference, especially in pairwise MRFs. In general there is a trade-off between the complexity/efficiency of the algorithm and its convergence properties, with certain problems requiring more complex inference to handle general pairwise potentials. Graphcuts based α-expansion performs well on certain classes of energies, and sequential tree reweighted message passing (TRWS) and loopy belief propagation (LBP) can be used for non-submodular cases. These methods though suffer from poor convergence and often oscillate between solutions. In this paper, we propose a tiered move making algorithm which is an iterative method. Each move to the next configuration is based on the current labeling and an optimal tiered move, where each tiered move requires one application of the dynamic programming based tiered labeling method introduced in Felzenszwalb et. al. [2]. The algorithm converges to a local minimum for any general pairwise potential, and we give a theoretical analysis of the properties of the algorithm, characterizing the situations in which we can expect good performance. We evaluate the algorithm on many benchmark labeling problems such as stereo, image segmentation, image stitching and image denoising, as well as random energy minimization. Our method consistently gets better energy values than α-expansion, LBP, quadratic pseudo-boolean optimization (QPBO), and is competitive with TRWS.

Linear SVMs are efficient in both training and testing, however the data in real applications is rarely linearly separable. Non-linear kernel SVMs are too computationally intensive for applications with large-scale data sets. Recently locally linear classifiers have gained popularity due to their efficiency whilst remaining competitive with kernel methods. The vanilla nearest neighbor algorithm is one of the simplest locally linear classifiers, but it lacks robustness due to the noise often present in real-world data. In this paper, we introduce a novel local classifier, Parametric Nearest Neighbor (P-NN) and its extension Ensemble of P-NN (EP-NN). We parameterize the nearest neighbor algorithm based on the minimum weighted squared Euclidean distances between the data points and the prototypes, where a prototype is represented by a locally linear combination of some data points. Meanwhile, our method attempts to jointly learn both the prototypes and the classifier parameters discriminatively via max-margin. This makes our classifiers suitable to approximate the classification decision boundaries locally based on nonlinear functions. During testing, the computational complexity of both classifiers is linear in the product of the dimension of data and the number of prototypes. Our classification results on MNIST, USPS, LETTER, and Chars 74K are comparable and in some cases are better than many other methods such as the state-of-the-art locally linear classifiers.

This paper describes a method for producing a semantic map from multi-view street-level imagery. We define a semantic map as an overhead, or bird's eye view of a region with associated semantic object labels, such as car, road and pavement. We formulate the problem using two conditional random fields. The first is used to model the semantic image segmentation of the street view imagery treating each image independently. The outputs of this stage are then aggregated over many images to form the input for our semantic map that is a second random field defined over a ground plane. Each image is related by a simple, yet effective, geometrical function that back projects a region from the street view image into the overhead ground plane map. We introduce, and make publicly available, a new dataset created from real world data. Our qualitative evaluation is performed on this data consisting of a 14.8 km track, and we also quantify our results on a representative subset.

In this work, we propose a method for instance based human segmentation in images and videos, extending the recent detector-based conditional random field model of Ladicky et.al. Instance based human segmentation involves pixel level labeling of an image, partitioning it into distinct human instances and background. To achieve our goal, we add three new components to their framework. First, we include human partsbased detection potentials to take advantage of the structure present in human instances. Further, in order to generate a consistent segmentation from different human parts, we incorporate shape prior information, which biases the segmentation to characteristic overall human shapes. Also, we enhance the representative power of the energy function by adopting exemplar instance based matching terms, which helps our method to adapt easily to different human sizes and poses. Finally, we extensively evaluate our proposed method on the Buffy dataset with our new segmented ground truth images, and show a substantial improvement over existing CRF methods. These new annotations will be made available for future use as well.

This paper extends classical object pose and relative camera motion estimation algorithms for imaging sensors sampling the scene through light-paths. Many algorithms in multi-view geometry assume that every pixel observes light traveling in a single line in space. We wish to relax this assumption and address various theoretical and practical issues in modeling camera rays as piece-wise linear-paths. Such paths consisting of finitely many linear segments are typical of any simple camera configuration with reflective and refractive elements. Our main contribution is to propose efficient algorithms that can work with the complete light-path without knowing the correspondence between their individual segments and the scene points. Second, we investigate light-paths containing infinitely many and small piece-wise linear segments that can be modeled using simple parametric curves such as conics. We show compelling simulations and real experiments, involving catadioptric configurations and mirages, to validate our study.

Object recognition has made great strides recently. However, the best methods, such as those based on kernel-SVMs are highly computationally intensive. The problem of how to accelerate the evaluation process without decreasing accuracy is thus of current interest. In this paper, we deal with this problem by using the idea of ranking. We propose a cascaded architecture which using the ranking SVM generates an ordered set of proposals for windows containing object instances. The top ranking windows may then be fed to a more complex detector. Our experiments demonstrate that our approach is robust, achieving higher overlap-recall values using fewer output proposals than the state-of-the-art. Our use of simple gradient features and linear convolution indicates that our method is also faster than the state-of-the-art.

Adaptive tracking-by-detection methods are widely used in computer vision for tracking arbitrary objects. Current approaches treat the tracking problem as a classification task and use online learning techniques to update the object model. However, for these updates to happen one needs to convert the estimated object position into a set of labelled training examples, and it is not clear how best to perform this intermediate step. Furthermore, the objective for the classifier (label prediction) is not explicitly coupled to the objective for the tracker (accurate estimation of object position). In this paper, we present a framework for adaptive visual object tracking based on structured output prediction. By explicitly allowing the output space to express the needs of the tracker, we are able to avoid the need for an intermediate classification step. Our method uses a kernelized structured output support vector machine (SVM), which is learned online to provide adaptive tracking. To allow for real-time application, we introduce a budgeting mechanism which prevents the unbounded growth in the number of support vectors which would otherwise occur during tracking. Experimentally, we show that our algorithm is able to outperform state-of-the-art trackers on various benchmark videos. Additionally, we show that we can easily incorporate additional features and kernels into our framework, which results in increased performance.

Local Coordinate Coding (LCC) [18] is a method for modeling functions of data lying on non-linear manifolds. It provides a set of anchor points which form a local coordinate system, such that each data point on the manifold can be approximated by a linear combination of its anchor points, and the linear weights become the local coordinate coding. In this paper we propose encoding data using orthogonal anchor planes, rather than anchor points. Our method needs only a few orthogonal anchor planes for coding, and it can linearize any (\alpha,\beta,p)-Lipschitz smooth nonlinear function with a fixed expected value of the upper-bound approximation error on any high dimensional data. In practice, the orthogonal coordinate system can be easily learned by minimizing this upper bound using singular value decomposition (SVD). We apply our method to model the coordinates locally in linear SVMs for classification tasks, and our experiment on MNIST shows that using only 50 anchor planes our method achieves 1.72% error rate, while LCC achieves 1.90% error rate using 4096 anchor points.

This work is supported by EPSRC research grants, HMGCC, TUBITAK researcher exchange grant, the IST Programme of the European Community, under the PASCAL2 Network of Excellence, IST-2007-216886.

We introduce a novel bilinear boosting algorithm, which extends the multi-class boosting framework of JointBoost to optimize a bilinear objective function. This allows style parameters to be introduced to aid classification, where style is any factor which the classes vary with systematically, modeled by a vector quantity. The algorithm allows learning to take place across different styles. We apply this Style Parameterized Boosting framework (StyP-Boost) to two object class segmentation tasks: road surface segmentation and general scene parsing. In the former the style parameters represent global surface appearance, and in the latter the probability of belonging to a scene-class. We show how our framework improves on 1) learning without style, and 2) learning independent classifiers within each style. Further, we achieve state-of-the-art results on the Corel database for scene parsing.

Conditional Random Field models have proved effective for several low-level computer vision problems. Inference in these models involves solving a combinatorial optimization problem, with methods such as graph cuts, belief propagation. Although several methods have been proposed to learn the model parameters from training data, they suffer from various drawbacks. Learning these parameters involves computing the partition function, which is intractable. To overcome this, state-of-the-art structured learning methods frame the problem as one of large margin estimation. Iterative solutions have been proposed to solve the resulting convex optimization problem. Each iteration involves solving an inference problem over all the labels, which limits the efficiency of these structured methods. In this paper we present an efficient large margin piece-wise learning method which is widely applicable. We show how the resulting optimization problem can be reduced to an equivalent convex problem with a small number of constraints, and solve it using an efficient scheme. Our method is both memory and computationally efficient. We show results on publicly available standard datasets.

Graph cuts have proven useful for image segmentation and for volumetric reconstruction in multiple view stereo. However, solutions are biased: the cost function tends to favour either a short boundary (in 2D) or a boundary with a small area (in 3D). This bias can be avoided by instead minimising the cut ratio, which normalises the cost by a measure of the boundary size. This paper uses ideas from discrete differential geometry to develop a linear programming formulation for finding a minimum ratio cut in arbitrary dimension, which allows constraints on the solution to be specified in a natural manner, and which admits an efficient and globally optimal solution. Results are shown for 2D segmentation and for 3D volumetric reconstruction.

In this paper, we present novel techniques that improve the computational and memory efficiency of algorithms for solving multilabel energy functions arising from discrete mrfs or crfs. These methods are motivated by the observations that the performance of minimization algorithms depends on: 1) the initialization used for the primal and dual variables and 2) the number of primal variables involved in the energy function. Our first method (dynamic alpha-expansion) works by "recycling" results from previous problem instances. The second method simplifies the energy function by "reducing" the number of unknown variables present in the problem. Further, we show that it can also be used to generate a good initialization for the dynamic alpha-expansion algorithm by "reusing" dual variables. We test the performance of our methods on energy functions encountered in the problems of stereo matching and color and object-based segmentation. Experimental results show that our methods achieve a substantial improvement in the performance of alpha-expansion, as well as other popular algorithms such as sequential tree-reweighted message passing and max-product belief propagation. We also demonstrate the applicability of our schemes for certain higher order energy functions, such as the one described in [1], for interactive texture-based image and video segmentation. In most cases, we achieve a 10-15 times speed-up in the computation time. Our modified alpha-expansion algorithm provides similar performance to Fast-PD, but is conceptually much simpler. Both alpha-expansion and Fast-PD can be made orders of magnitude faster when used in conjunction with the "reduce" scheme proposed in this paper.

We present a probabilistic method for segmenting instances of a particular object category within an image. Our approach overcomes the deficiencies of previous segmentation techniques based on traditional grid conditional random fields (CRF), namely that 1) they require the user to provide seed pixels for the foreground and the background and 2) they provide a poor prior for specific shapes due to the small neighborhood size of grid CRF. Specifically, we automatically obtain the pose of the object in a given image instead of relying on manual interaction. Furthermore, we employ a probabilistic model which includes shape potentials for the object to incorporate top-down information that is global across the image, in addition to the grid clique potentials which provide the bottom-up information used in previous approaches. The shape potentials are provided by the pose of the object obtained using an object category model. We represent articulated object categories using a novel layered pictorial structures model. Nonarticulated object categories are modeled using a set of exemplars. These object category models have the advantage that they can handle large intraclass shape, appearance, and spatial variation. We develop an efficient method, OBJCUT, to obtain segmentations using our probabilistic framework. Novel aspects of this method include: 1) efficient algorithms for sampling the object category models of our choice and 2) the observation that a sampling-based approximation of the expected log-likelihood of the model can be increased by a single graph cut. Results are presented on several articulated (e.g., animals) and nonarticulated (e.g., fruits) object categories. We provide a favorable comparison of our method with the state of the art in object category specific image segmentation, specifically the methods of Leibe and Schiele and Schoenemann and Cremers.

In this paper we present a framework for pixel-wise object segmentation of road scenes that combines motion and appearance features. It is designed to handle street-level imagery such as that on Google Street View and Microsoft Bing Maps. We formulate the problem in a CRF framework in order to probabilistically model the label likelihoods and the a priori knowledge. An extended set of appearance-based features is used, which consists of textons, colour, location and HOG descriptors. A novel boosting approach is then applied to combine the motion and appearance-based features. We also incorporate higher order potentials in our CRF model, which produce segmentations with precise object boundaries. We evaluate our method both quantitatively and qualitatively on the challenging Cambridge-driving Labeled Video dataset. Our approach shows an overall recognition accuracy of 84 % compared to the state-of-the-art accuracy of 69%.

Supervised learning of a parts-based model can be formulated as an optimization problem with a large (exponential in the number of parts) set of constraints. We show how this seemingly difficult problem can be solved by (i) reducing it to an equivalent convex problem with a small, polynomial number of constraints (taking advantage of the fact that the model is tree-structured and the potentials have a special form); and (ii) obtaining the globally optimal model using an efficient dual decomposition strategy. Each component of the dual decomposition is solved by a modified version of the highly optimized SVM-Light algorithm. To demonstrate the effectiveness of our approach, we learn human upper body models using two challenging, publicly available datasets. Our model accounts for the articulation of humans as well as the occlusion of parts. We compare our method with a baseline iterative strategy as well as a state of the art algorithm and show significant efficiency improvements.

Most methods for object class segmentation are formulated as a labelling problem over a single choice of quantisation of an image space - pixels, segments or group of segments. It is well known that each quantisation has its fair share of pros and cons; and the existence of a common optimal quantisation level suitable for all object categories is highly unlikely. Motivated by this observation, we propose a hierarchical random field model, that allows integration of features computed at different levels of the quantisation hierarchy. MAP inference in this model can be performed efficiently using powerful graph cut based move making algorithms. Our framework generalises much of the previous work based on pixels or segments. We evaluate its efficiency on some of the most challenging data-sets for object class segmentation, and show it obtains state-of-the-art results.

In this paper, we extend the class of energy functions for which the optimal \alpha-expansion and \alpha \beta-swap moves can be computed in polynomial time. Specifically, we introduce a novel family of higher order clique potentials, and show that the expansion and swap moves for any energy function composed of these potentials can be found by minimizing a submodular function. We also show that for a subset of these potentials, the optimal move can be found by solving an st-mincut problem. We refer to this subset as the {\cal P};n Potts model. Our results enable the use of powerful \alpha-expansion and \alpha \beta-swap move making algorithms for minimization of energy functions involving higher order cliques. Such functions have the capability of modeling the rich statistics of natural scenes and can be used for many applications in Computer Vision. We demonstrate their use in one such application, i.e., the texture-based image or video-segmentation problem.

Second-order priors on the smoothness of 3D surfaces are a better model of typical scenes than first-order priors. However, stereo reconstruction using global inference algorithms, such as graph cuts, has not been able to incorporate second-order priors because the triple cliques needed to express them yield intractable (nonsubmodular) optimization problems. This paper shows that inference with triple cliques can be effectively performed. Our optimization strategy is a development of recent extensions to alpha -- expansion, based on the "QPBO" algorithm. The strategy is to repeatedly merge proposal depth maps using a novel extension of QPBO. Proposal depth maps can come from any source, for example, frontoparallel planes as in alpha-expansion, or indeed any existing stereo algorithm, with arbitrary parameter settings.

In this paper we show how online images can be automatically exploited for scene visualization and reconstruction starting from a mere visual query provided by the user. A visual query is used to retrieve images of a landmark place using a visual search engine. These images are used to reconstruct robust 3–D features and camera poses in projective space. Novel views are then rendered corresponding to a virtual camera flying smoothly through the projective space by triangulation of the projected points in the output view. We introduce a method to fuse the rendered novel views from all input images at each virtual view point by computing their intrinsic image and illuminations. This approach allows us to remove the occlusions and maintain consistent and controlled illumination throughout the rendered sequence. We demonstrate the performance of our prototype system on two landmark structures. 1

In this paper, we present novel techniques that improve the computational and memory efficiency of algorithms for solving multi-label energy functions arising from discrete MRFs or CRFs. These methods are motivated by the observations that the performance of minimization algorithms depends on: (a) the initialization used for the primal and dual variables; and (b) the number of primal variables involved in the energy function. Our first method (dynamic alpha-expansion) works by dasiarecyclingpsila results from previous problem instances. The second method simplifies the energy function by dasiareducingpsila the number of unknown variables, and can also be used to generate a good initialization for the dynamic alpha-expansion algorithm by dasiareusingpsila dual variables. We test the performance of our methods on energy functions encountered in the problems of stereo matching, and colour and object based segmentation. Experimental results show that our methods achieve a substantial improvement in the performance of alpha-expansion, as well as other popular algorithms such as sequential tree-reweighted message passing, and max-product belief propagation. In most cases we achieve a 10-15 times speed-up in the computation time. Our modified alpha-expansion algorithm provides similar performance to Fast-PD. However, it is much simpler and can be made orders of magnitude faster by using the initialization schemes proposed in the paper.

This paper proposes a novel framework for labelling problems which is able to combine multiple segmentations in a principled manner. Our method is based on higher order conditional random fields and uses potentials defined on sets of pixels (image segments) generated using unsupervised segmentation algorithms. These potentials enforce label consistency in image regions and can be seen as a strict generalization of the commonly used pairwise contrast sensitive smoothness potentials. The higher order potential functions used in our framework take the form of the Robust Pnmodel. This enables the use of powerful graph cut based move making algorithms for performing inference in the framework [14]. We test our method on the problem of multi-class object segmentation by augmenting the conventional CRF used for object segmentation with higher order potentials defined on image regions. Experiments on challenging data sets show that integration of higher order potentials quantitatively and qualitatively improves results leading to much better definition of object boundaries. We believe that this method can be used to yield similar improvements for many other labelling problems.

This paper addresses the problem of exactly inferring the maximum a posteriori solutions of discrete multi-label MRFs or CRFs with higher order cliques. We present a framework to transform special classes of multi-label higher order functions to submodular second order Boolean functions (referred to as Fs2), which can be minimized exactly using graph cuts and we characterize those classes. The basic idea is to use two or more Boolean variables to encode the states of a single multi-label variable. There are many ways in which this can be done and much interesting research lies in finding ways which are optimal or minimal in some sense. We study the space of possible encodings and find the ones that can transform the most general class of functions to Fs2. Our main contributions are two-fold. First, we extend the subclass of submodular energy functions that can be minimized exactly using graph cuts. Second, we show how higher order potentials can be used to improve single view 3D reconstruction results. We believe that our work on exact minimization of higher order energy functions will lead to similar improvements in solutions of other labelling problems.

This paper addresses human pose recognition from video sequences by formulating it as a classification problem. Unlike much previous work we do not make any assumptions on the availability of clean segmentation. The first step of this work consists in a novel method of aligning the training images using 3D Mocap data. Next we define classes by discretizing a 2D manifold whose two dimensions are camera viewpoint and actions. Our main contribution is a pose detection algorithm based on random forests. A bottom-up approach is followed to build a decision tree by recursively clustering and merging the classes at each level. For each node of the decision tree we build a list of potentially discriminative features using the alignment of training images; in this paper we consider Histograms of Orientated Gradient (HOG). We finally grow an ensemble of trees by randomly sampling one of the selected HOG blocks at each node. Our proposed approach gives promising results with both fixed and moving cameras.

Second-order priors on the smoothness of 3D surfaces are a better model of typical scenes than first-order priors. However, stereo reconstruction using global inference algorithms, such as graph-cuts, has not been able to incorporate second-order priors because the triple cliques needed to express them yield intractable (non-submodular) optimization problems. This paper shows that inference with triple cliques can be effectively optimized. Our optimization strategy is a development of recent extensions to a-expansion, based on the "QPBO" algorithm [5, 14, 26]. The strategy is to repeatedly merge proposal depth maps using a novel extension of QPBO. Proposal depth maps can come from any source, for example fronto-parallel planes as in a-expansion, or indeed any existing stereo algorithm, with arbitrary parameter settings. Experimental results demonstrate the usefulness of the second-order prior and the efficacy of our optimization framework. An implementation of our stereo framework is available online [34].

We consider the problem of optimizing multilabel MRFs, which is in general NP-hard and ubiquitous in low-level computer vision. One approach for its solution is to formulate it as an integer linear programming and relax the integrality constraints. The approach we consider in this paper is to first convert the multi-label MRF into an equivalent binary-label MRF and then to relax it. The resulting relaxation can be efficiently solved using a maximum flow algorithm. Its solution provides us with a partially optimal labelling of the binary variables. This partial labelling is then easily transferred to the multi-label problem. We study the theoretical properties of the new relaxation and compare it with the standard one. Specifically, we compare tightness, and characterize a subclass of problems where the two relaxations coincide. We propose several combined algorithms based on the technique and demonstrate their performance on challenging computer vision problems.

The problem of obtaining the maximum a posteriori (MAP) estimate of a discrete random field is of fundamental importance in many areas of Computer Science. In this work, we build on the tree reweighted message passing (TRW) framework of (Kolmogorov, 2006; Wainwright et al., 2005). TRW iteratively optimizes the Lagrangian dual of a linear programming relaxation for MAP estimation. We show how the dual formulation of TRW can be extended to include cycle inequalities (Barahona & Mahjoub, 1986) and some recently proposed second order cone (SOC) constraints (Kumar et al., 2007). We propose efficient iterative algorithms for solving the resulting duals. Similar to the method described in (Kolmogorov, 2006), these algorithms are guaranteed to converge. We test our approach on a large set of synthetic data, as well as real data. Our experiments show that the additional constraints (i.e. cycle inequalities and SOC constraints) provide better results in cases where the TRW framework fails (namely MAP estimation for non-submodular energy functions).

We consider the problem of obtaining the approximate maximum a posteriori estimate of a discrete random field characterized by pairwise potentials that form a truncated convex model. For this problem, we propose an improved st-mincut based move making algorithm. Unlike previous move making approaches, which either provide a loose bound or no bound on the quality of the solution (in terms of the corresponding Gibbs energy), our algorithm achieves the same guarantees as the standard linear programming (LP) relaxation. Compared to previous approaches based on the LP relaxation, e.g. interior-point algorithms or tree-reweighted message passing (TRW), our method is faster as it uses only the efficient st-mincut algorithm in its design. Furthermore, it directly provides us with a primal solution (unlike TRW and other related methods which attempt to solve the dual of the LP). We demonstrate the effectiveness of the proposed approach on both synthetic and standard real data problems. Our analysis also opens up an interesting question regarding the relationship between move making algorithms (such as $\alpha$-expansion and the algorithms presented in this paper) and the randomized rounding schemes used with convex relaxations. We believe that further explorations in this direction would help design efficient algorithms for more complex relaxations.

We show that application of modern multiview stereo techniques to the newview synthesis (NVS) problem introduces a number of non-trivial complexities. By simultaneously solving for the colour and depth of the new-view pixels we can eliminate the visual artefacts that conventional NVS-via-stereo suffers. The global occlusion reasoning which has led to considerable improvements in recent stereo algorithms can easily be included in the new algorithm, using a recently improved graph-cut-based optimizer for general multi-label conditional random fields (CRFs). However, the CRF priors that are important to success in stereo cannot be easily applied if the reconstruction is to be computed in the reference frame of the novel view. We address this problem by extending recent work on the fast optimization of texture priors in NVS to model the image edge structure, yielding a synthesis of the two approaches which yields good results on difficult image sequences. 1