Vision-Language Models (VLMs) have achieved impressive performance across a wide range of multimodal tasks, yet they often exhibit inconsistent behavior when faced with semantically equivalent inputs—undermining their reliability and robustness. Recent benchmarks, such as MM-R3, highlight that even state-of-the-art VLMs can produce divergent response across semantically equivalent inputs, despite maintaining high average accuracy. Prior work addresses this issue by modifying model architectures or conducting large-scale fine-tuning on curated datasets. In contrast, we propose a simple and effective test-time consistency framework that enhances semantic consistency without supervised re-training. Our method is entirely post-hoc, model-agnostic, and applicable to any VLM with access to its weights. Given a single test point, we enforce consistent predictions via two complementary objectives: (i) a Cross-Entropy Agreement Loss that aligns predictive distributions across semantically equivalent inputs, and (ii) a Pseudo-Label Consistency Loss that draws outputs toward a self-averaged consensus. Our method is plug-and-play, and leverages information from a single test-input itself to improve consistency. Experiments on the MM-R3benchmark show that our framework yields substantial gains in consistency across state-of-the-art models, establishing a new direction for inference-time adaptation in VLMs.1

Shivam Chandhok, Wan-Cyuan Fan, Vered Shwartz, Vineeth N. Balasubramanian, Leonid Sigal. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The commonsense reasoning capabilities of vision-language models (VLMs), especially in abductive reasoning and defeasible reasoning, remain poorly understood. Most benchmarks focus on typical visual scenarios [1], [23], [42], making it difficult to discern whether model performance stems from keen perception and reasoning skills, or reliance on pure statistical recall. We argue that by focusing on atypical events in videos, clearer insights can be gained on the core capabilities of VLMs. Explaining and understanding such out-of-distribution events requires models to extend beyond basic pattern recognition and regurgitation of their prior knowledge. To this end, we introduce Black-SwanSuite, a benchmark for evaluating VLMs’ ability to reason about unexpected events through abductive and defeasible tasks. Our tasks artificially limit the amount of visual information provided to models while questioning them about hidden unexpected events, or provide new visual information that could change an existing hypothesis about the event. We curate a comprehensive benchmark suite comprising over 3,800 MCQ, 4,900 generative and 6,700 yes/no questions, spanning 1,655 videos. After extensively evaluating various state-of-the-art VLMs, including GPT-4o and Gemini 1.5 Pro, as well as open-source VLMs such as LLaVA-Video, we find significant performance gaps of up to 32% from humans on these tasks. Our findings reveal key limitations in current VLMs, emphasizing the need for enhanced model architectures and training strategies. Our data and leaderboard is available at https://blackswan.cs.ubc.ca.

Current research on generating 3D hand-object interaction motion primarily focuses on in-domain objects. Generalization to unseen objects is essential for practical applications, yet it remains both challenging and largely unexplored. In this paper, we propose LatentHOI, a novel approach designed to tackle the challenges of generalizing hand-object interaction synthesis to unseen objects. Our main insight lies in decoupling high-level temporal motion from fine-grained spatial hand-object interactions via a latent diffusion model coupled with a Grasping Variational Autoencoder (Grasp-VAE). This configuration introduces regularization by enforcing a conditional dependency between spatial grasping and temporal motion, as well as through the regularized latent space for better generalization ability. We conducted extensive experiments in an unseen-object setting on both single-hand grasping and bi-manual motion datasets, including GRAB, DexYCB†, and OakInk. Quantitative and qualitative evaluations demonstrate that our method significantly enhances the realism and physical plausibility of generated motions for unseen objects, both in single and bimanual manipulations, compared to the state-of-the-art.

Deep neural networks (DNNs) offer significant promise for improving breast cancer diagnosis in medical imaging. However, these models are highly susceptible to adversarial attacks—small, imperceptible changes that can mislead classifiers—raising critical concerns about their reliability and security. Traditional attacks rely on fixed-norm perturbations, misaligning with human perception. In contrast, diffusion-based attacks require pre-trained models, demanding substantial data when these models are unavailable, limiting practical use in data-scarce scenarios. In medical imaging, however, this is often unfeasible due to the limited availability of datasets. Building on recent advancements in learnable prompts, we propose Prompt2Perturb (P2P), a novel language-guided attack method capable of generating meaningful attack examples driven by text instructions. During the prompt learning phase, our approach leverages learnable prompts within the text encoder to create subtle, yet impactful, perturbations that remain imperceptible while guiding the model towards targeted outcomes. In contrast to current prompt learning-based approaches, our P2P stands out by directly updating text embeddings, avoiding the need for retraining diffusion models. Further, we leverage the finding that optimizing only the early diffusion steps boosts efficiency while ensuring that the generated adversarial examples incorporate subtle noise, thus preserving ultrasound image quality without introducing noticeable artifacts. We show that our method outperforms state-of-the-art attack techniques across three breast ultrasound datasets in FID and LPIPS. Moreover, the generated images are both more natural in appearance and more effective compared to existing adversarial attacks. Our Code is publicly available on GitHub.

Ali Salamatian, Amirhossein Abaskohi, Wan-Cyuan Fan, Mir Rayat Imtiaz Hossain, Leonid Sigal, Giuseppe Carenini. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

We present locality-sensitive avatar, a neural radiance field (NeRF) based network to learn human motions from monocular videos. To this end, we estimate a canonical representation between different frames of a video with a non-linear mapping from observation to canonical space, which we decompose into a skeletal rigid motion and a non-rigid counterpart. Our key contribution is to retain fine-grained details by modeling the non-rigid part with a graph neural network (GNN) that keeps the pose information local to neighboring body parts. Compared to former canonical representation based methods which solely operate on the coordinate space of a whole shape, our locality-sensitive motion modeling can reproduce both realistic shape contours and vivid fine-grained details. We evaluate on ZJU-MoCap, SynWild, ActorsHQ, MVHumanNet and various outdoor videos. The experiments reveal that with the locality sensitive deformation to canonical feature space, we are the first to achieve state-of-the-art results across novel view synthesis, novel pose animation and 3D shape reconstruction simultaneously. Our code is available at https://github.com/ChunjinSong/lsavatar.

Text watermarking aims to subtly embeds statistical signals into text by controlling the Large Language Model (LLM)’s sampling process, enabling watermark detectors to verify that the output was generated by the specified model. The robustness of these watermarking algorithms has become a key factor in evaluating their effectiveness. Current text watermarking algorithms embed watermarks in high-entropy tokens to ensure text quality. In this paper, we reveal that this seemingly benign design can be exploited by attackers, posing a significant risk to the robustness of the watermark. We introduce a generic efficient paraphrasing attack, the Self-Information Rewrite Attack (SIRA), which leverages the vulnerability by calculating the self-information of each token to identify potential pattern tokens and perform targeted attack. Our work exposes a widely prevalent vulnerability in current watermarking algorithms. The experimental results show SIRA achieves nearly 100% attack success rates on seven recent watermarking methods with only $0.88 per million tokens cost. Our approach does not require any access to the watermark algorithms or the watermarked LLM and can seamlessly transfer to any LLM as the attack model even mobile-level models. Our findings highlight the urgent need for more robust watermarking.

Advances in Large Language Models (LLMs) have sparked interest in their ability to solve Olympiad-level math problems. However, the training and evaluation of these models are constrained by the limited size and quality of available datasets, as creating large-scale data for such advanced problems requires extensive effort from human experts. In addition, current benchmarks are prone to contamination, leading to unreliable evaluations. In this paper, we present an automated pipeline that leverages the rich resources of the Art of Problem Solving (AoPS) forum, which predominantly features Olympiad-level problems and community-driven solutions. Using open-source LLMs, we develop a method to extract question-answer pairs from the forum, resulting in AoPS-Instruct, a dataset of more than 600,000 high-quality QA pairs. Our experiments demonstrate that fine-tuning LLMs on AoPS-Instruct improves their reasoning abilities across various benchmarks. Moreover, we build an automatic pipeline that introduces LiveAoPSBench, an evolving evaluation set with timestamps, derived from the latest forum data, providing a contamination-resistant benchmark for assessing LLM performance. Notably, we observe a significant decline in LLM performance over time, suggesting their success on older examples may stem from pre-training exposure rather than true reasoning ability. Our work presents a scalable approach to creating and maintaining large-scale, high-quality datasets for advanced math reasoning, offering valuable insights into the capabilities and limitations of LLMs in this domain.

Multi-modal Large Language Models (LLM) have advanced conversational abilities but struggle with providing live, interactive step-by-step guidance, a key capability for future AI assistants. Effective guidance requires not only delivering instructions but also detecting their successful execution, as well as identifying and alerting users to mistakes, all of which has to happen in real-time. This requires models that are not turn-based, but that can react asynchronously to a video stream, as well as video data showing users performing tasks including mistakes and their corrections. To this end, we introduce Qualcomm Interactive Cooking, a new benchmark and dataset built upon CaptainCook4D, which contains user mistakes during task execution. Our dataset and benchmark features densely annotated, timed instructions and feedback messages, specifically including mistake alerts precisely timestamped to their visual occurrence in the video. We evaluate state-of-the-art multi-modal LLMs on the Qualcomm Interactive Cooking benchmark and introduce LiveMamba, a streaming multi-modal LLM designed for interactive instructional guidance. This work provides the first dedicated benchmark and a strong baseline for developing and evaluating on live, situated coaching.

Video Object Segmentation (VOS) has emerged as an increasingly important problem with availability of larger datasets and more complex and realistic settings, which involve long videos with global motion (e.g., in egocentric settings), depicting small objects undergoing both rigid and non-rigid (including state) deformations. While a number of recent approaches have been explored for this task, these data characteristics still present challenges. In this work we propose a novel, clip-based DETR-style encoder-decoder architecture, which focuses on systematically analyzing and addressing aforementioned challenges. Specifically, we propose a novel transformation-aware loss thatfocuses learning on portions of the video where an object undergoes significant deformations - a form of “soft” hard examples mining. Further, we propose a multiplicative time-coded memory, beyond vanilla additive positional encoding, which helps propagate context across long videos. Finally, we incorporate these in our proposed holistic multiscale video transformer for tracking via multi-scale memory matching and decoding to ensure sensitivity and accuracy for long videos and small objects. Our model enables on-line inference with long videos in a windowed fashion, by breaking the video into clips and propagating context among them. We illustrate that short clip length and longer memory with learned time-coding are important design choices for improved performance. Collectively, these technical contributions enable our model to achieve new state-of-the-art (SoTA) performance on two complex egocentric datasets - VISOR [13] and VOST [44], while achieving comparable to SoTA results on the conventional VOS benchmark, DAVIS’ 17 [38]. Detailed ablations validate our design choices and provide insights into the importance of parameter choices and impact on performance.11Code is available at the project page.

Abstract For reconstructing high‐fidelity human 3D models from monocular videos, it is crucial to maintain consistent large‐scale body shapes along with finely matched subtle wrinkles. This paper explores how per‐frame rendering results can be factorized into a pose‐independent component and a corresponding pose‐dependent counterpart to facilitate frame consistency at multiple scales. Pose adaptive texture features are further improved by restricting the frequency bands of these two components. Pose‐independent outputs are expected to be low‐frequency, while high‐frequency information is linked to pose‐dependent factors. We implement this with a dual‐branch network. The first branch takes coordinates in the canonical space as input, while the second one additionally considers features outputted by the first branch and pose information of each frame. A final network integrates the information predicted by both branches and utilizes volume rendering to generate photo‐realistic 3D human images. Through experiments, we demonstrate that our method consistently surpasses all state‐of‐the‐art methods in preserving high‐frequency details and ensuring consistent body contours. Our code is accessible at https://github.com/ChunjinSong/facavatar .

Classical light field rendering for novel view synthesis can accurately reproduce view-dependent effects such as reflection, refraction, and translucency, but requires a dense view sampling of the scene. Methods based on geometric reconstruction need only sparse views, but cannot accurately model non-Lambertian effects. We introduce a model that combines the strengths and mitigates the limitations of these two directions. By operating on a four-dimensional representation of the light field, our model learns to represent view-dependent effects accurately. By enforcing geometric constraints during training and inference, the scene geometry is implicitly learned from a sparse set of views. Concretely, we introduce a two-stage transformer-based model that first aggregates features along epipolar lines, then aggregates features along reference views to produce the color of a target ray. Additionally, we propose modifications that allow the model to generalize to scenes without any fine-tuning. Our model outperforms the state-of-the-art on multiple forward-facing and 360$^\circ$∘ datasets, with larger margins on scenes with severe view-dependent variations.

The emergence of attention-based transformer models has led to their extensive use in various tasks, due to their superior generalization and transfer properties. Recent research has demonstrated that such models, when prompted appropriately, are excellent for few-shot inference. However, such techniques are under-explored for dense prediction tasks like semantic segmentation. In this work, we examine the effectiveness of prompting a transformer-decoder with learned visual prompts for the generalized few-shot segmentation (GFSS) task. Our goal is to achieve strong performance not only on novel categories with limited examples, but also to retain performance on base cat-egories. We propose an approach to learn visual prompts with limited examples. These learned visual prompts are used to prompt a multiscale transformer decoder to facilitate accurate dense predictions. Additionally, we introduce a unidirectional causal attention mechanism between the novel prompts, learned with limited examples, and the base prompts, learned with abundant data. This mecha-nism enriches the novel prompts without deteriorating the base class performance. Overall, this form of prompting helps us achieve state-of-the-art performance for GFSS on two different benchmark datasets: COCO-20i and Pascal-Si, without the need for test-time optimization (or transduction). Furthermore, test-time optimization leveraging unlabelled test data can be used to improve the prompts, which we refer to as transductive prompt tuning.

From image-text pairs, large-scale vision-language models (VLMs) learn to implicitly associate image regions with words, which prove effective for tasks like visual question answering. However, leveraging the learned association for open-vocabulary semantic segmentation remains a challenge. In this paper, we propose a simple, yet extremely effective, training-free technique, Plug-and-Play Open- Vocabulary Semantic Segmentation (PnP-OVSS) for this task. PnP-OVSS leverages a VLM with direct text-to-image cross-attention and an image-text matching loss. To balance between over-segmentation and under-segmentation, we introduce Salience Dropout; by iteratively dropping patches that the model is most attentive to, we are able to better resolve the entire extent of the segmentation mask. PnP-OVSS does not require any neural net-work training and performs hyperparameter tuning without the need for any segmentation annotations, even for a validation set. PnP-OVSS demonstrates substantial improvements over comparable baselines (+29.4% mIoU on Pascal VOC, +13.2% mIoU on Pascal Context, +14.0% mIoU on MS COCO, +2.4% mIoU on COCO Stuff) and even outper-forms most baselines that conduct additional network training on top of pretrained VLMs. Our codebase is at https://github.com/letitiabanana/PnP-OVSS.

The quality of the prompts provided to text-to-image diffusion models determines how faithful the generated content is to the user's intent, often requiring 'prompt engineering ‘. To harness visual concepts from target images without prompt engineering, current approaches largely rely on embedding inversion by optimizing and then mapping them to pseudo-tokens. However, working with such highdimensional vector representations is challenging because they lack semantics and interpretability, and only allow simple vector operations when using them. Instead, this work focuses on inverting the diffusion model to obtain interpretable language prompts directly. The challenge of doing this lies in the fact that the resulting optimization problem is fundamentally discrete and the space of prompts is exponentially large; this makes using standard optimization techniques, such as stochastic gradient descent, difficult. To this end, we utilize a delayed projection scheme to optimize for prompts representative of the vocabulary space in the model. Further, we leverage the findings that different timesteps of the diffusion process cater to different levels of detail in an image. The later, noisy, timesteps of the forward diffusion process correspond to the semantic information, and therefore, prompt inversion in this range provides tokens representative of the image semantics. We show that our approach can identify semantically interpretable and meaningful prompts for a target image which can be used to synthesize diverse images with similar content. We further illustrate the application of the optimized prompts in evolutionary image generation and concept removal.

Predicting video saliency is crucial for improving sports video processing efficiency, thereby providing an enriched viewing experience for a wide-ranging audience. However, there is a long-term absence of well-established eye-tracking database and learning-based approach, particularly tailored for sports videos. In this paper, we establish a large-scale eye-tracking database dubbed audio-visual sports (AVS). AVS consists of 1,000 high-quality sports videos with eye fixations from 60 participants. Through the data analysis on AVS, we observe that human attention patterns exhibit significant variations based on the specific scene context of the sports. Motivated by this, we propose a sport-aware audiovisual saliency model, which can adaptively learn the scene context in a hyper manner. Specifically, a new audio-visual fusion (AVF) block is developed to effectively fuse features from the visual and audio backbone. After that, a hyper network is introduced to learn sport-aware priors, which are then adopted to guide the self-adaptive saliency predictor for predicting saliency map. Experimental results demonstrate that our approach outperforms other state-of-the-art saliency prediction models over the only two sports video eye-tracking databases.

Video understanding has witnessed significant progress with recent video foundation models demonstrating strong performance owing to self-supervised pre-training objectives; Masked Autoencoders (MAE) being the design of choice. Nevertheless, the majority of prior works that leverage MAE pre-training have focused on relatively short video representations (16 / 32 frames in length) largely due to hardware memory and compute limitations that scale poorly with video length due to the dense memory-intensive self-attention decoding. One natural strategy to address these challenges is to subsample tokens to reconstruct during decoding (or decoder masking). In this work, we propose an effective strategy for prioritizing tokens which allows training on longer video sequences (128 frames) and gets better performance than, more typical, random and uniform masking strategies. The core of our approach is an adaptive decoder masking strategy that prioritizes the most important tokens and uses quantized tokens as reconstruction objectives. Our adaptive strategy leverages a powerful MAGVIT-based tokenizer that jointly learns the tokens and their priority. We validate our design choices through exhaustive ablations and observe improved performance of the resulting long-video (128 frames) encoders over short-video (32 frames) counterparts. With our long-video masked autoencoder (LVMAE) strategy, we surpass state-of-the-art on Diving48 by 3.9 points and EPIC-Kitchens-100 verb classification by 2.5 points while relying on a simple core architecture and video-only pre-training (unlike some of the prior works that require millions of labeled video-text pairs or specialized encoders).

Recent advances in pixel-level tasks (e.g. segmentation) illustrate the benefit of of long-range interactions between aggregated region-based representations that can enhance local features. However, such aggregated representations, often in the form of attention, fail to model the underlying semantics of the scene (e.g. individual objects and, by extension, their interactions). In this work, we address the issue by proposing a component that learns to project image features into latent representations and reason between them using a transformer encoder to generate contextualized and scene-consistent representations which are fused with original image features. Our design encourages the latent regions to represent semantic concepts by ensuring that the activated regions are spatially disjoint and the union of such regions corresponds to a connected object segment. The proposed semantic global reasoning (SGR) component is end-to-end trainable and can be easily added to a wide variety of backbones (CNN or transformer-based) and segmentation heads (per-pixel or mask classification) to consistently improve the segmentation results on different datasets. In addition, our latent tokens are semantically interpretable and diverse and provide a rich set of features that can be transferred to downstream tasks like object detection and segmentation, with improved performance. Furthermore, we also proposed metrics to quantify the semantics of latent tokens at both class & instance level.

The goal of scene graph generation is to predict a graph from an input image, where nodes correspond to identified and localized objects and edges to their corresponding interaction predicates. Existing methods are trained in a fully supervised manner and focus on message passing mechanisms, loss functions, and/or bias mitigation. In this work we introduce a simple-yet-effective self-supervised relational alignment regularization designed to improve the scene graph generation performance. The proposed alignment is general and can be combined with any existing scene graph generation framework, where it is trained alongside the original model’s objective. The alignment is achieved through distillation, where an auxiliary relation prediction branch, that mirrors and shares parameters with the supervised counterpart, is designed. In the auxiliary branch, relational input features are partially masked prior to message passing and predicate prediction. The predictions for masked relations are then aligned with the supervised counterparts after the message passing. We illustrate the effectiveness of this self-supervised relational alignment in conjunction with two scene graph generation architectures, SGTR [25] and Neural Motifs [53], and show that in both cases we achieve significantly improved performance.

Existing dense or paragraph video captioning approaches rely on holistic representations of videos, possibly coupled with learned object/action representations, to condition hierarchical language decoders. However, they fundamentally lack the commonsense knowledge of the world required to reason about progression of events, causality, and even the function of certain objects within a scene. To address this limitation we propose a novel video captioning Transformer-based model, that takes into account both implicit (visuo-lingual and purely linguistic) and explicit (knowledge-base) commonsense knowledge. We show that these forms of knowledge, in isolation and in combination, enhance the quality of produced captions. Further, inspired by imitation learning, we propose a new task of instruction generation, where the goal is to produce a set of linguistic instructions from a video demonstration of its performance. We formalize the task using the ALFRED dataset generated using an AI2-THOR environment. While instruction generation is conceptually similar to paragraph captioning, it differs in the fact that it exhibits stronger object persistence, as well as spatially-aware and causal sentence structure. We show that our commonsense knowledge enhanced approach produces significant improvements on this task (up to 57% in METEOR and 8.5% in CIDEr), as well as the state-of-the-art result on more traditional video captioning in the ActivityNet Captions dataset.

Salient object ranking (SOR) aims to segment salient objects in an image and simultaneously predict their saliency rankings, according to the shifted human attention over different objects. The existing SOR approaches mainly focus on object-based attention, e.g., the semantic and appearance of object. However, we find that the scene context plays a vital role in SOR, in which the saliency ranking of the same object varies a lot at different scenes. In this paper, we thus make the first attempt towards explicitly learning scene context for SOR. Specifically, we establish a large-scale SOR dataset of 24,373 images with rich context annotations, i.e., scene graphs, segmentation, and saliency rankings. Inspired by the data analysis on our dataset, we propose a novel graph hypernetwork, named HyperSOR, for context-aware SOR. In HyperSOR, an initial graph module is developed to segment objects and construct an initial graph by considering both geometry and semantic information. Then, a scene graph generation module with multi-path graph attention mechanism is designed to learn semantic relationships among objects based on the initial graph. Finally, a saliency ranking prediction module dynamically adopts the learned scene context through a novel graph hypernetwork, for inferring the saliency rankings. Experimental results show that our HyperSOR can significantly improve the performance of SOR.

People normally watch 360 ° videos through a head-mounted display, inside which only the content of viewports can be seen. Therefore, viewport proposal, referring to detecting potential viewport candidates, plays an important role in many 360 ° video processing tasks. In this paper, we advance the viewport proposal by further aligning the predicted viewports across frames for individual subject. This provides a better methodology and a deeper perspective to learn the human perceptual behaviours on 360 ° videos. Specifically, we first analyze three 360 ° video datasets and obtain several findings on human consistency, objectness and motion of viewports. Inspired by these findings, we propose a bi-directional transformer approach, named BiT, for 360 ° video viewport proposal and alignment. Specifically, BiT is composed of a multi-level residual module, a bi-directional encoder-decoder module and a spherical matching module. This way, the viewports can be well proposed and aligned via considering multi-level, bi-directional and non-local information. Moreover, the aligned viewports by BiT are used to refine the viewports and improve viewport proposal accuracy in return. Finally, we validate that our BiT approach is superior on viewport proposal, compared with the state-of-the-art approaches. Besides, the aligned viewports from BiT is verified to be effective in multiple applications, such as saliency prediction, trajectory prediction and perceptual video compression.

With the rapid development of virtual reality, 360° images have gained increasing popularity. Their wide field of view necessitates high resolution to ensure image quality. This, however, makes it harder to acquire, store and even process such 360° images. To alleviate this issue, we propose the first attempt at 360° image rescaling, which refers to downscaling a 360° image to a visually valid lowresolution (LR) counterpart and then upscaling to a highresolution (HR) 360° image given the LR variant. Specifically, we first analyze two 360° image datasets and observe several findings that characterize how 360° images typically change along their latitudes. Inspired by these findings, we propose a novel deformable invertible neural network (INN), named DINN360, for latitude-aware 360° image rescaling. In DINN360, a deformable INN is designed to downscale the LR image, and project the high-frequency (HF) component to the latent space by adaptively handling various deformations occurring at different latitude regions. Given the downscaled LR image, the high-quality HR image is then reconstructed in a conditional latitude-aware manner by recovering the structure-related HF component from the latent space. Extensive experiments over four public datasets show that our DINN360 method performs considerably better than other state-of-the-art methods for 2 x, 4 x and 8 x 360° image rescaling.

There has been a recent explosion of impressive generative models that can produce high quality images (or videos) conditioned on text descriptions. However, all such approaches rely on conditional sentences that contain un-ambiguous descriptions of scenes and main actors in them. Therefore employing such models for more complex task of story visualization, where naturally references and co-references exist, and one requires to reason about when to maintain consistency of actors and backgrounds across frames/scenes, and when not to, based on story progression, remains a challenge. In this work, we address the aforementioned challenges and propose a novel autoregressive diffusion-based framework with a visual memory module that implicitly captures the actor and background context across the generated frames. Sentence-conditioned soft attention over the memories enables effective reference resolution and learns to maintain scene and actor consistency when needed. To validate the effectiveness of our approach, we extend the MUGEN dataset [19] and introduce additional characters, backgrounds and referencing in multi-sentence storylines. Our experiments for story generation on the MUGEN, the PororoSV [30] and the FlintstonesSV [16] dataset show that our method not only outperforms prior state-of-the-art in generating frames with high visual quality, which are consistent with the story, but also models appropriate correspondences between the characters and the background.

We propose a method to decompose a video into a background and a set of foreground layers, where the background captures stationary elements while the foreground layers capture moving objects along with their associated effects (e.g. shadows and reflections). Our approach is designed for unconstrained monocular videos, with an arbitrary camera and object motion. Prior work that tackles this problem assumes that the video can be mapped onto a fixed 2D canvas, severely limiting the possible space of camera motion. Instead, our method applies recent progress in monocular camera pose and depth estimation to create a full, RGBD video layer for the background, along with a video layer for each foreground object. To solve the underconstrained decomposition problem, we propose a new loss formulation based on multi-view consistency. We test our method on challenging videos with complex camera motion and show significant qualitative improvement over current approaches.

Correlation based stereo matching has achieved outstanding performance, which pursues cost volume between two feature maps. Unfortunately, current methods with a fixed model do not work uniformly well across various datasets, greatly limiting their real-world applicability. To tackle this issue, this paper proposes a new perspective to dynamically calculate correlation for robust stereo matching. A novel Uncertainty Guided Adaptive Correlation (UGAC) module is introduced to robustly adapt the same model for different scenarios. Specifically, a variance-based uncertainty estimation is employed to adaptively adjust the sampling area during warping operation. Additionally, we improve the traditional non-parametric warping with learnable parameters, such that the position-specific weights can be learned. We show that by empowering the recurrent network with the UGAC module, stereo matching can be exploited more robustly and effectively. Extensive experiments demonstrate that our method achieves state-of-the-art performance over the ETH3D, KITTI, and Middlebury datasets when employing the same fixed model over these datasets without any retraining procedure. To target real-time applications, we further design a lightweight model based on UGAC, which also outperforms other methods over KITTI benchmarks with only 0.6 M parameters.

In the existing coding standards, rate control (RC) plays a critical role in optimally allocating bit-rates to each coding unit, for improving rate-distortion performance under the limited bandwidth. However, the existing RC methods are mainly based on traditional distortion metrics, which fail to take the advantage of the emerging DNN based image quality assessment (IQA) metrics. In this paper, we set up the first attempt to achieve IQA score based RC for image compression. Specifically, a novel visualization based score-distortion (VSD) model and ρ-slope model are proposed to explicitly establish the relationship between IQA score and bit-rates. Then, by solving optimal rate-distortion optimization based on the IQA score, we propose a novel RC method for the HEVC standard. The experimental results show that, given the target bit-rates, the proposed RC method can accurately control the bit-rates and generate the compressed images with higher IQA score and better perceptual quality. More importantly, the proposed RC method is evaluated to be effective over two DNN based IQA metrics and four image datasets, exhibiting the potential in practical use. The code is available at https://github.com/Ffangqy/IQA-RC.

Intelligent systems possess a crucial characteristic of breaking complicated problems into smaller reusable components or parts and adjusting to new tasks using these part representations. However, current part-learners encounter difficulties in dealing with incidental correlations resulting from the limited observations of objects that may appear only in specific arrangements or with specific backgrounds. These incidental correlations may have a detrimental impact on the generalization and interpretability of learned part representations. This study asserts that part-based representations could be more interpretable and generalize better with limited data, employing two innovative regularization methods. The first regularization separates foreground and background information's generative process via a unique mixture-of-parts formulation. Structural constraints are imposed on the parts using a weakly-supervised loss, guaranteeing that the mixture-of-parts for foreground and background entails soft, object-agnostic masks. The second regularization assumes the form of a distillation loss, ensuring the invariance of the learned parts to the incidental background correlations. Furthermore, we incorporate sparse and orthogonal constraints to facilitate learning high-quality part representations. By reducing the impact of incidental background correlations on the learned parts, we exhibit state-of-the-art (SoTA) performance on few-shot learning tasks on benchmark datasets, including MiniImagenet, TieredImageNet, and FC100. We also demonstrate that the part-based representations acquired through our approach generalize better than existing techniques, even under domain shifts of the background and common data corruption on the ImageNet-9 dataset.

There has been a growing interest in solving Visual Question Answering (VQA) tasks that require the model to reason beyond the content present in the image. In this work, we focus on questions that require commonsense reasoning. In contrast to previous methods which inject knowledge from static knowledge bases, we investigate the incorporation of contextualized knowledge using Commonsense Transformer (COMET), an existing knowledge model trained on human-curated knowledge bases. We propose a method to generate, select, and encode external commonsense knowledge alongside visual and textual cues in a new pre-trained Vision-Language-Commonsense transformer model, VLC-BERT. Through our evaluation on the knowledge-intensive OK-VQA and A-OKVQA datasets, we show that VLC-BERT is capable of outperforming existing models that utilize static knowledge bases. Furthermore, through a detailed analysis, we explain which questions benefit, and which don’t, from contextualized commonsense knowledge from COMET. Code: https://github.com/aditya10/VLC-BERT

Abstract Artist generated clip‐art images typically consist of a small number of distinct, uniformly colored regions with clear boundaries. Legacy artist created images are often stored in low‐resolution (100x100px or less) anti‐aliased raster form. Compared to anti‐aliasing free rasterization, anti‐aliasing blurs inter‐region boundaries and obscures the artist's intended region topology and color palette; at the same time, it better preserves subpixel details. Recovering the underlying artist‐intended images from their low‐resolution anti‐aliased rasterizations can facilitate resolution independent rendering, lossless vectorization, and other image processing applications. Unfortunately, while human observers can mentally deblur these low‐resolution images and reconstruct region topology, color and subpixel details, existing algorithms applicable to this task fail to produce outputs consistent with human expectations when presented with such images. We recover these viewer perceived blur‐free images at subpixel resolution, producing outputs where each input pixel is replaced by four corresponding (sub)pixels. Performing this task requires computing the size of the output image color palette, generating the palette itself, and associating each pixel in the output with one of the colors in the palette. We obtain these desired output components by leveraging a combination of perceptual and domain priors, and real world data. We use readily available data to train a network that predicts, for each anti‐aliased image, a low‐blur approximation of the blur‐free double‐resolution outputs we seek. The images obtained at this stage are perceptually closer to the desired outputs but typically still have hundreds of redundant differently colored regions with fuzzy boundaries. We convert these low‐blur intermediate images into blur‐free outputs consistent with viewer expectations using a discrete partitioning procedure guided by the characteristic properties of clip‐art images, observations about the antialiasing process, and human perception of anti‐aliased clip‐art. This step dramatically reduces the size of the output color palettes, and the region counts bringing them in line with viewer expectations and enabling the image processing applications we target. We demonstrate the utility of our method by using our outputs for a number of image processing tasks, and validate it via extensive comparisons to prior art. In our comparative study, participants preferred our deblurred outputs over those produced by the best‐performing alternative by a ratio of 75 to 8.5.

Classical light field rendering for novel view synthesis can accurately reproduce view-dependent effects such as reflection, refraction, and translucency, but requires a dense view sampling of the scene. Methods based on geometric reconstruction need only sparse views, but cannot accurately model non-Lambertian effects. We introduce a model that combines the strengths and mitigates the limitations of these two directions. By operating on a four-dimensional representation of the light field, our model learns to represent view-dependent effects accurately. By enforcing geometric constraints during training and inference, the scene geometry is implicitly learned from a sparse set of views. Concretely, we introduce a two-stage transformer-based model that first aggregates features along epipolar lines, then aggregates features along reference views to produce the color of a target ray. Our model outperforms the state-of-the-art on multiple forward-facing and 360° datasets, with larger margins on scenes with severe view-dependent variations. Code and results can be found at light-field-neural-rendering.github. io.

The task of scene graph generation entails identifying object entities and their corresponding interaction predicates in a given image (or video). Due to the combinatorially large solution space, existing approaches to scene graph generation assume certain factorization of the joint distribution to make the estimation feasible (e.g., assuming that objects are conditionally independent of predicate predictions). However, this fixed factorization is not ideal under all scenarios (e.g., for images where an object entailed in interaction is small and not discernible on its own). In this work, we propose a novel framework for scene graph generation that addresses this limitation, as well as introduces dynamic conditioning on the image, using message passing in a Markov Random Field. This is implemented as an iterative refinement procedure wherein each modification is conditioned on the graph generated in the previous iteration. This conditioning across refinement steps allows joint reasoning over entities and relations. This framework is realized via a novel and end-to-end trainable transformer-based architecture. In addition, the proposed framework can improve existing approach performance. Through extensive experiments on Visual Genome and Action Genome benchmark datasets we show improved performance on the scene graph generation.

In this paper, we propose a novel task for saliency-guided image translation, with the goal of image-to-image translation conditioned on the user specified saliency map. To address this problem, we develop a novel Generative Adversarial Network (GAN)-based model, called SalG-GAN. Given the original image and target saliency map, SalG-GAN can generate a translated image that satisfies the target saliency map. In SalG-GAN, a disentangled representation framework is proposed to encourage the model to learn diverse translations for the same target saliency condition. A saliency-based attention module is introduced as a special attention mechanism for facilitating the developed structures of saliency-guided generator, saliency cue encoder and saliency-guided global and local discriminators. Furthermore, we build a synthetic dataset and a real-world dataset with labeled visual attention for training and evaluating our SalG-GAN. The experimental results over both datasets verify the effectiveness of our model for saliency-guided image translation.

Methods for object detection and segmentation rely on large scale instance-level annotations for training, which are difficult and time-consuming to collect. Efforts to alleviate this look at varying degrees and quality of supervision. Weakly-supervised approaches draw on image-level labels to build detectors/segmentors, while zero/few-shot methods assume abundant instance-level data for a set of base classes, and none to a few examples for novel classes. This taxonomy has largely siloed algorithmic designs. In this work, we aim to bridge this divide by proposing an intuitive and unified semi-supervised model that is applicable to a range of supervision: from zero to a few instance-level samples per novel class. For base classes, our model learns a mapping from weakly-supervised to fully-supervised detectors/segmentors. By learning and leveraging visual and lingual similarities between the novel and base classes, we transfer those mappings to obtain detectors/segmentors for novel classes: refining them with a few novel class instance-level annotated samples, if available. The overall model is end-to-end trainable and highly flexible1. Through extensive experiments on MS- COCO [32] and Pascal VOC [14] benchmark datasets we show improved performance in a variety of settings.

Traditional scene graph generation methods are trained using cross-entropy losses that treat objects and relationships as independent entities. Such a formulation, however, ignores the structure in the output space, in an inherently structured prediction problem. In this work, we introduce a novel energy-based learning framework for generating scene graphs. The proposed formulation allows for efficiently incorporating the structure of scene graphs in the output space. This additional constraint in the learning framework acts as an inductive bias and allows models to learn efficiently from a small number of labels. We use the proposed energy-based framework†to train existing stateof-the-art models and obtain a significant performance improvement, of up to 21% and 27%, on the Visual Genome [9] and GQA [5] benchmark datasets, respectively. Furthermore, we showcase the learning efficiency of the proposed framework by demonstrating superior performance in the zero- and few-shot settings where data is scarce.

Scene graph generation has emerged as an important problem in computer vision. While scene graphs provide a grounded representation of objects, their locations and relations in an image, they do so only at the granularity of proposal bounding boxes. In this work, we propose the first, to our knowledge, framework for pixel-level segmentation-grounded scene graph generation. Our framework is agnostic to the underlying scene graph generation method and address the lack of segmentation annotations in target scene graph datasets (e.g., Visual Genome [24]) through transfer and multi-task learning from, and with, an auxiliary dataset (e.g., MS COCO [29]). Specifically, each target object being detected is endowed with a segmentation mask, which is expressed as a lingual-similarity weighted linear combination over categories that have annotations present in an auxiliary dataset. These inferred masks, along with a Gaussian masking mechanism which grounds the relations at a pixel-level within the image, allow for improved relation prediction. The entire framework is end-to-end trainable and is learned in a multi-task manner. Code is available at github.com/ubc-vision/segmentation-sg.

Learning how to localize and separate individual object sounds in the audio channel of the video is a difficult task. Current state-of-the-art methods predict audio masks from artificially mixed spectrograms, known as Mix-and-Separate framework. We propose an audio-visual co-segmentation, where the network learns both what individual objects look and sound like, from videos labeled with only object labels. Unlike other recent visually-guided audio source separation frameworks, our architecture can be learned in an end-to-end manner and requires no additional supervision or bounding box proposals. Specifically, we introduce weakly-supervised object segmentation in the context of sound separation. We also formulate spectrogram mask prediction using a set of learned mask bases, which combine using coefficients conditioned on the output of object segmentation — a design that facilitates separation. Extensive experiments on the MUSIC dataset show that our proposed approach outperforms state-of-the-art methods on visually guided sound source separation and sound denoising.

As an important step towards visual reasoning, visual grounding (e.g., phrase localization, referring expression comprehension / segmentation) has been widely explored. Previous approaches to referring expression comprehension (REC) or segmentation (RES) either suffer from limited performance, due to a two-stage setup, or require the designing of complex task-specific one-stage architectures. In this paper, we propose a simple one-stage multi-task framework for visual grounding tasks. Specifically, we leverage a transformer architecture, where two modalities are fused in a visual-lingual encoder. In the decoder, the model learns to generate contextualized lingual queries which are then decoded and used to directly regress the bounding box and produce a segmentation mask for the corresponding referred regions. With this simple but highly contextualized model, we outperform state-of-the-art methods by a large margin on both REC and RES tasks. We also show that a simple pre-training schedule (on an external dataset) further improves the performance. Extensive experiments and ablations illustrate that our model benefits greatly from contextualized information and multi-task training.

The recent success of transformer models in language, such as BERT, has motivated the use of such architectures for multi-modal feature learning and tasks. However, most multi-modal variants (e.g., ViLBERT) have limited themselves to visual-linguistic data. Relatively few have explored its use in audio-visual modalities, and none, to our knowledge, illustrate them in the context of granular audio-visual detection or segmentation tasks such as sound source separation and localization. In this work, we introduce TriBERT -- a transformer-based architecture, inspired by ViLBERT, which enables contextual feature learning across three modalities: vision, pose, and audio, with the use of flexible co-attention. The use of pose keypoints is inspired by recent works that illustrate that such representations can significantly boost performance in many audio-visual scenarios where often one or more persons are responsible for the sound explicitly (e.g., talking) or implicitly (e.g., sound produced as a function of human manipulating an object). From a technical perspective, as part of the TriBERT architecture, we introduce a learned visual tokenization scheme based on spatial attention and leverage weak-supervision to allow granular cross-modal interactions for visual and pose modalities. Further, we supplement learning with sound-source separation loss formulated across all three streams. We pre-train our model on the large MUSIC21 dataset and demonstrate improved performance in audio-visual sound source separation on that dataset as well as other datasets through fine-tuning. In addition, we show that the learned TriBERT representations are generic and significantly improve performance on other audio-visual tasks such as cross-modal audio-visual-pose retrieval by as much as 66.7% in top-1 accuracy.

Unsupervised multi-object scene decomposition is a fast-emerging problem in representation learning. Despite significant progress in static scenes, such models are unable to leverage important dynamic cues present in videos. We propose PROVIDE, a novel unsupervised framework for PRObabilistic VIdeo DEcomposition based on a temporal extension of iterative inference. PROVIDE is powerful enough to jointly model complex individual multi-object representations and explicit temporal dependencies between latent variables across frames. This is achieved by leveraging 2D-LSTM, temporally conditioned inference and generation within the iterative amortized inference for posterior refinement. Our method improves the overall quality of decompositions, encodes information about the objects’ dynamics, and can be used to predict trajectories of each object separately. Additionally, we show that our model has a high accuracy even without color information. We demonstrate the decomposition capabilities of our model and show that it outperforms the state-of-the-art on several benchmark datasets, one of which was curated for this work and will be made publicly available.

Despite significant progress, controlled generation of complex images with interacting people remains difficult. Existing layout generation methods fall short of synthesizing realistic person instances; while pose-guided generation approaches focus on a single person and assume simple or known backgrounds. To tackle these limitations, we propose a new problem, Persons in Context Synthesis, which aims to synthesize diverse person instance(s) in consistent contexts, with user control over both. The context is specified by the bounding box object layout which lacks shape information, while pose of the person(s) by keypoints which are sparsely annotated. To handle the stark difference in input structures, we proposed two separate neural branches to attentively composite the respective (context/person) inputs into shared "compositional structural space", which encodes shape, location and appearance information for both context and person structures in a disentangled manner. This structural space is then decoded to the image space using multi-level feature modulation strategy, and learned in a self supervised manner from image collections and their corresponding inputs. Extensive experiments on two large-scale datasets (COCO-Stuff [6] and Visual Genome [19]) demonstrate that our framework outperforms state-of-theart methods w.r.t. synthesis quality.

Large–scale visual sensor networks have become progressively an essential part of our daily lives underpinning many technological, financial, and social advancements today, with applications in smart cities, traffic monitoring, environmental pollution control, public safety, and crime prevention.

Few-shot learning is a fundamental task in computer vision that carries the promise of alleviating the need for exhaustively labeled data. Most few-shot learning approaches to date have focused on progressively more complex neural feature extractors and classifier adaptation strategies, and the refinement of the task definition itself. In this paper, we explore the hypothesis that a simple class-covariance-based distance metric, namely the Mahalanobis distance, adopted into a state of the art few-shot learning approach (CNAPS) can, in and of itself, lead to a significant performance improvement. We also discover that it is possible to learn adaptive feature extractors that allow useful estimation of the high dimensional feature covariances required by this metric from surprisingly few samples. The result of our work is a new ``Simple CNAPS'' architecture which has up to 9.2% fewer trainable parameters than CNAPS and performs up to 6.1% better than state of the art on the standard few-shot image classification benchmark dataset.

Reconstruction of a 3D shape from a single 2D image is a classical computer vision problem, whose difficulty stems from the inherent ambiguity of recovering occluded or only partially observed surfaces. Recent methods address this challenge through the use of largely unstructured neural networks that effectively distill conditional mapping and priors over 3D shape. In this work, we induce structure and geometric constraints by leveraging three core observations: (1) the surface of most everyday objects is often almost entirely exposed from pairs of typical opposite views; (2) everyday objects often exhibit global reflective symmetries which can be accurately predicted from single views; (3) opposite orthographic views of a 3D shape share consistent silhouettes. Following these observations, we first predict orthographic 2.5D visible surface maps (depth, normal and silhouette) from perspective 2D images, and detect global reflective symmetries in this data; second, we predict the back facing depth and normal maps using as input the front maps and, when available, the symmetric reflections of these maps; and finally, we reconstruct a 3D mesh from the union of these maps using a surface reconstruction method best suited for this data. Our experiments demonstrate that our framework outperforms state-of-the art approaches for 3D shape reconstructions from 2D and 2.5D data in terms of input fidelity and details preservation. Specifically, we achieve 12% better performance on average in ShapeNet benchmark dataset, and up to 19% for certain classes of objects (e.g., chairs and vessels).

Despite significant progress in object categorization, in recent years, a number of important challenges remain; mainly, the ability to learn from limited labeled data and to recognize object classes within large, potentially open, set of labels. Zero-shot learning is one way of addressing these challenges, but it has only been shown to work with limited sized class vocabularies and typically requires separation between supervised and unsupervised classes, allowing former to inform the latter but not vice versa. We propose the notion of vocabulary-informed learning to alleviate the above mentioned challenges and address problems of supervised, zero-shot, generalized zero-shot and open set recognition using a unified framework. Specifically, we propose a weighted maximum margin framework for semantic manifold-based recognition that incorporates distance constraints from (both supervised and unsupervised) vocabulary atoms. Distance constraints ensure that labeled samples are projected closer to their correct prototypes, in the embedding space, than to others. We illustrate that resulting model shows improvements in supervised, zero-shot, generalized zero-shot, and large open set recognition, with up to 310K class vocabulary on Animal with Attributes and ImageNet datasets.

Psychological studies have shown that personality traits are associated with book preferences. However, past findings based on questionnaires are limited to conventional book genres and do not capture niche content (e.g., family drama) and reading behaviors (e.g., backburners). For a more comprehensive measure of book content, this study harnesses a massive archive of content labels, also known as `tags', created by users of a book review website, Goodreads.com. Combined with data on preferences and personality scores collected from Facebook users, the tag labels achieve high accuracy in personality prediction by psychological standards. Additionally, we group tags into broader genres to check their validity against past findings. Our results are robust across both tag-level and genre-level analyses and are consistent with existing literature. Moreover, user-generated tag labels reveal unexpected insights, such as cultural differences, book reading behaviors, and other non-content factors affecting preferences. To our knowledge, this is currently the largest study that explores the relationship between personality and book content preferences.

Raster clip-art images, which consist of distinctly colored regions separated by sharp boundaries typically allow for a clear mental vector interpretation. Converting these images into vector format can facilitate compact lossless storage and enable numerous processing operations. Despite recent progress, existing vectorization methods that target such data frequently produce vectorizations that fail to meet viewer expectations. We present PolyFit , a new clip-art vectorization method that produces vectorizations well aligned with human preferences. Since segmentation of such inputs into regions had been addressed successfully, we specifically focus on fitting piecewise smooth vector curves to the raster input region boundaries, a task prior methods are particularly prone to fail on. While perceptual studies suggest the criteria humans are likely to use during mental boundary vectorization, they provide no guidance as to the exact interaction between them; learning these interactions directly is problematic due to the large size of the solution space. To obtain the desired solution, we first approximate the raster region boundaries with coarse intermediate polygons leveraging a combination of perceptual cues with observations from studies of human preferences. We then use these intermediate polygons as auxiliary inputs for computing piecewise smooth vectorizations of raster inputs. We define a finite set of potential polygon to curve primitive maps, and learn the mapping from the polygons to their best fitting primitive configurations from human annotations, arriving at a compact set of local raster and polygon properties whose combinations reliably predict human-expected primitive choices. We use these primitives to obtain a final globally consistent spline vectorization. Extensive comparative user studies show that our method outperforms state-of-the-art approaches on a wide range of data, where our results are preferred three times as often as those of the closest competitor across multiple types of inputs with various resolutions.

We address the problem of text-based activity retrieval in video. Given a sentence describing an activity, our task is to retrieve matching clips from an untrimmed video. To capture the inherent structures present in both text and video, we introduce a multilevel model that integrates vision and language features earlier and more tightly than prior work. First, we inject text features early on when generating clip proposals, to help eliminate unlikely clips and thus speed up processing and boost performance. Second, to learn a fine-grained similarity metric for retrieval, we use visual features to modulate the processing of query sentences at the word level in a recurrent neural network. A multi-task loss is also employed by adding query re-generation as an auxiliary task. Our approach significantly outperforms prior work on two challenging benchmarks: Charades-STA and ActivityNet Captions.

We propose an end-to-end approach for phrase grounding in images. Unlike prior methods that typically attempt to ground each phrase independently by building an image-text embedding, our architecture formulates grounding of multiple phrases as a sequential and contextual process. Specifically, we encode region proposals and all phrases into two stacks of LSTM cells, along with so-far grounded phrase-region pairs. These LSTM stacks collectively capture context for grounding of the next phrase. The resulting architecture, which we call SeqGROUND, supports many-to-many matching by allowing an image region to be matched to multiple phrases and vice versa. We show competitive performance on the Flickr30K benchmark dataset and, through ablation studies, validate the efficacy of sequential grounding as well as individual design choices in our model architecture.

We propose a novel probabilistic generative model for action sequences. The model is termed the Action Point Process VAE (APP-VAE), a variational auto-encoder that can capture the distribution over the times and categories of action sequences. Modeling the variety of possible action sequences is a challenge, which we show can be addressed via the APP-VAE's use of latent representations and non-linear functions to parameterize distributions over which event is likely to occur next in a sequence and at what time. We empirically validate the efficacy of APP-VAE for modeling action sequences on the MultiTHUMOS and Breakfast datasets.

Despite significant recent progress on generative models, controlled generation of images depicting multiple and complex object layouts is still a difficult problem. Among the core challenges are the diversity of appearance a given object may possess and, as a result, exponential set of images consistent with a specified layout. To address these challenges, we propose a novel approach for layout-based image generation; we call it Layout2Im. Given the coarse spatial layout (bounding boxes + object categories), our model can generate a set of realistic images which have the correct objects in the desired locations. The representation of each object is disentangled into a specified/certain part (category) and an unspecified/uncertain part (appearance). The category is encoded using a word embedding and the appearance is distilled into a low-dimensional vector sampled from a normal distribution. Individual object representations are composed together using convolutional LSTM, to obtain an encoding of the complete layout, and then decoded to an image. Several loss terms are introduced to encourage accurate and diverse generation. The proposed Layout2Im model significantly outperforms the previous state of the art, boosting the best reported inception score by 24.66% and 28.57% on the very challenging COCO-Stuff and Visual Genome datasets, respectively. Extensive experiments also demonstrate our method's ability to generate complex and diverse images with multiple objects.

In this paper we present an end-to-end framework for grounding of phrases in images. In contrast to previous works, our model, which we call G3raphGround, uses graphs to formulate more complex, non-sequential dependencies among proposal image regions and phrases. We capture intra-modal dependencies using a separate graph neural network for each modality (visual and lingual), and then use conditional message-passing in another graph neural network to fuse their outputs and capture cross-modal relationships. This final representation results in grounding decisions. The framework supports many-to-many matching and is able to ground single phrase to multiple image regions and vice versa. We validate our design choices through a series of ablation studies and illustrate state-of-the-art performance on Flickr30k and ReferIt Game benchmark datasets.

Recently there is an increasing interest in scene generation within the research community. However, models used for generating scene layouts from textual description largely ignore plausible visual variations within the structure dictated by the text. We propose LayoutVAE, a variational autoencoder based framework for generating stochastic scene layouts. LayoutVAE is a versatile modeling framework that allows for generating full image layouts given a label set, or per label layouts for an existing image given a new label. In addition, it is also capable of detecting unusual layouts, potentially providing a way to evaluate layout generation problem. Extensive experiments on MNIST-Layouts and challenging COCO 2017 Panoptic dataset verifies the effectiveness of our proposed framework.

Visual attention mechanisms have proven to be integrally important constituent components of many modern deep neural architectures. They provide an efficient and effective way to utilize visual information selectively, which has shown to be especially valuable in multi-modal learning tasks. However, all prior attention frameworks lack the ability to explicitly model structural dependencies among attention variables, making it difficult to predict consistent attention masks. In this paper we develop a novel structured spatial attention mechanism which is end-to-end trainable and can be integrated with any feed-forward convolutional neural network. This proposed AttentionRNN layer explicitly enforces structure over the spatial attention variables by sequentially predicting attention values in the spatial mask in a bi-directional raster-scan and inverse raster-scan order. As a result, each attention value depends not only on local image or contextual information, but also on the previously predicted attention values. Our experiments show consistent quantitative and qualitative improvements on a variety of recognition tasks and datasets; including image categorization, question answering and image generation.

Multi-modal learning, particularly among imaging and linguistic modalities, has made amazing strides in many high-level fundamental visual understanding problems, ranging from language grounding to dense event captioning. However, much of the research has been limited to approaches that either do not take audio corresponding to video into account at all, or those that model the audio-visual correlations in service of sound or sound source localization. In this paper, we present the evidence, that audio signals can carry surprising amount of information when it comes to high-level visual-lingual tasks. Specifically, we focus on the problem of weakly-supervised dense event captioning in videos and show that audio on its own can nearly rival performance of a state-of-the-art visual model and, combined with video, can improve on the state-of-the-art performance. Extensive experiments on the ActivityNet Captions dataset show that our proposed multi-modal approach outperforms state-of-the-art unimodal methods, as well as validate specific feature representation and architecture design choices.

Understanding images beyond salient actions involves reasoning about scene context, objects, and the roles they play in the captured event. Situation recognition has recently been introduced as the task of jointly reasoning about the verbs (actions) and a set of semantic-role and entity (noun) pairs in the form of action frames. Labeling an image with an action frame requires an assignment of values (nouns) to the roles based on the observed image content. Among the inherent challenges are the rich conditional structured dependencies between the output role assignments and the overall semantic sparsity. In this paper, we propose a novel mixture-kernel attention graph neural network (GNN) architecture designed to address these challenges. Our GNN enables dynamic graph structure during training and inference, through the use of a graph attention mechanism, and context-aware interactions between role pairs. We illustrate the efficacy of our model and design choices by conducting experiments on imSitu benchmark dataset, with accuracy improvements of up to 10% over the state-of-the-art.

We explore the power of spatial context as a self-supervisory signal for learning visual representations. In particular, we propose spatial context networks that learn to predict a representation of one image patch from another image patch, within the same image, conditioned on their real-valued relative spatial offset. Unlike auto-encoders, that aim to encode and reconstruct original image patches, our network aims to encode and reconstruct intermediate representations of the spatially offset patches. As such, the network learns a spatially conditioned contextual representation. By testing performance with various patch selection mechanisms we show that focusing on object-centric patches is important, and that using object proposal as a patch selection mechanism leads to the highest improvement in performance. Further, unlike auto-encoders, context encoders [21], or other forms of unsupervised feature learning, we illustrate that contextual supervision (with pre-trained model initialization) can improve on existing pre-trained model performance. We build our spatial context networks on top of standard VGG_19 and CNN_M architectures and, among other things, show that we can achieve improvements (with no additional explicit supervision) over the original ImageNet pre-trained VGG_19 and CNN_M models in object categorization and detection on VOC2007.

Dense video captioning is a fine-grained video understanding task that involves two sub-problems: localizing distinct events in a long video stream, and generating captions for the localized events. We propose the Joint Event Detection and Description Network (JEDDi-Net), which solves the dense video captioning task in an end-to-end fashion. Our model continuously encodes the input video stream with three-dimensional convolutional layers, proposes variable-length temporal events based on pooled features, and generates their captions. Proposal features are extracted within each proposal segment through 3D Segment-of-Interest pooling from shared video feature encoding. In order to explicitly model temporal relationships between visual events and their captions in a single video, we also propose a two-level hierarchical captioning module that keeps track of context. On the large-scale ActivityNet Captions dataset, JEDDi-Net demonstrates improved results as measured by standard metrics. We also present the first dense captioning results on the TACoS-MultiLevel dataset.

The ability to quickly recognize and learn new visual concepts from limited samples enables humans to quickly adapt to new tasks and environments. This ability is enabled by semantic association of novel concepts with those that have already been learned and stored in memory. Computers can start to ascertain similar abilities by utilizing a semantic concept space. A concept space is a high-dimensional semantic space in which similar abstract concepts appear close and dissimilar ones far apart. In this paper, we propose a novel approach to one-shot learning that builds on this core idea. Our approach learns to map a novel sample instance to a concept, relates that concept to the existing ones in the concept space and, using these relationships, generates new instances, by interpolating among the concepts, to help learning. Instead of synthesizing new image instance, we propose to directly synthesize instance features by leveraging semantics using a novel auto-encoder network we call dual TriNet. The encoder part of the TriNet learns to map multi-layer visual features from CNN to a semantic vector. In semantic space, we search for related concepts, which are then projected back into the image feature spaces by the decoder portion of the TriNet. Two strategies in the semantic space are explored. Notably, this seemingly simple strategy results in complex augmented feature distributions in the image feature space, leading to substantially better performance. The codes and models are released in the github: https://github.com/tankche1/ Semantic-Feature-Augmentation-in-Few-shot-Learning.

The alignment of heterogeneous sequential data (video to text) is an important and challenging problem. Standard techniques for this task, including Dynamic Time Warping (DTW) and Conditional Random Fields (CRFs), suffer from inherent drawbacks. Mainly, the Markov assumption implies that, given the immediate past, future alignment decisions are independent of further history. The separation between similarity computation and alignment decision also prevents end-to-end training. In this paper, we propose an end-to-end neural architecture where alignment actions are implemented as moving data between stacks of Long Short-term Memory (LSTM) blocks. This flexible architecture supports a large variety of alignment tasks, including one-to-one, one-to-many, skipping unmatched elements, and (with extensions) non-monotonic alignment. Extensive experiments on semi-synthetic and real datasets show that our algorithm outperforms state-of-the-art baselines.

We propose an end-to-end network for visual illustration of a sequence of sentences forming a story. At the core of our model is the ability to model the inter-related nature of the sentences within a story, as well as the ability to learn coherence to support reference resolution. The framework takes the form of an encoder-decoder architecture, where sentences are encoded using a hierarchical two-level sentence-story GRU, combined with an encoding of coherence, and sequentially decoded using a predicted feature representation into a consistent illustrative image sequence. We optimize all parameters of our network in an end-to-end fashion with respect to order embedding loss, encoding entailment between images and sentences. Experiments on the VIST storytelling dataset [9] highlight the importance of our algorithmic choices and efficacy of our overall model.

Despite being virtually ubiquitous, sequence-to-sequence models are challenged by their lack of diversity and inability to be externally controlled. In this paper, we speculate that a fundamental shortcoming of sequence generation models is that the decoding is done strictly from left-to-right, meaning that outputs values generated earlier have a profound effect on those generated later. To address this issue, we propose a novel middle-out decoder architecture that begins from an initial middle-word and simultaneously expands the sequence in both directions. To facilitate information flow and maintain consistent decoding, we introduce a dual self-attention mechanism that allows us to model complex dependencies between the outputs. We illustrate the performance of our model on the task of video captioning, as well as a synthetic sequence de-noising task. Our middle-out decoder achieves significant improvements on de-noising and competitive performance in the task of video captioning, while quantifiably improving the caption diversity. Furthermore, we perform a qualitative analysis that demonstrates our ability to effectively control the generation process of our decoder.

Abstract We present a method for the automatic creation of fictional storybooks based on personal photographs. Unlike previous attempts that summarize such collections by picking salient or diverse photos, or creating personal literal narratives, we focus on the creation of fictional stories. This provides new value to users, as well as an engaging way for people (especially children) to experience their own photographs. We use a graph model to represent an artist‐generated story, where each node is a ‘frame’, akin to frames in comics or storyboards. A node is described by story elements, comprising actors, location, supporting objects and time. The edges in the graph encode connections between these elements and provide the discourse of the story. Based on this construction, we develop a constraint satisfaction algorithm for one‐to‐one assignment of nodes to photographs. Once each node is assigned to a photograph, a visual depiction of the story can be generated in different styles using various templates. We show results of several fictional visual stories created from different personal photo sets and in different styles.

Emotion is a key element in user-generated video. However, it is difficult to understand emotions conveyed in such videos due to the complex and unstructured nature of user-generated content and the sparsity of video frames expressing emotion. In this paper, for the first time, we propose a technique for transferring knowledge from heterogeneous external sources, including image and textual data, to facilitate three related tasks in understanding video emotion: emotion recognition, emotion attribution and emotion-oriented summarization. Specifically, our framework (1) learns a video encoding from an auxiliary emotional image dataset in order to improve supervised video emotion recognition, and (2) transfers knowledge from an auxiliary textual corpora for zero-shot recognition of emotion classes unseen during training. The proposed technique for knowledge transfer facilitates novel applications of emotion attribution and emotion-oriented summarization. A comprehensive set of experiments on multiple datasets demonstrate the effectiveness of our framework.

We propose a weakly-supervised approach that takes image-sentence pairs as input and learns to visually ground (i.e., localize) arbitrary linguistic phrases, in the form of spatial attention masks. Specifically, the model is trained with images and their associated image-level captions, without any explicit region-to-phrase correspondence annotations. To this end, we introduce an end-to-end model which learns visual groundings of phrases with two types of carefully designed loss functions. In addition to the standard discriminative loss, which enforces that attended image regions and phrases are consistently encoded, we propose a novel structural loss which makes use of the parse tree structures induced by the sentences. In particular, we ensure complementarity among the attention masks that correspond to sibling noun phrases, and compositionality of attention masks among the children and parent phrases, as defined by the sentence parse tree. We validate the effectiveness of our approach on the Microsoft COCO and Visual Genome datasets.

Deep neural networks (DNNs) and probabilistic graphical models (PGMs) are the two main tools for statistical modeling. While DNNs provide the ability to model rich and complex relationships between input and output variables, PGMs provide the ability to encode dependencies among the output variables themselves. End-to-end training methods for models with structured graphical dependencies on top of neural predictions have recently emerged as a principled way of combining these two paradigms. While these models have proven to be powerful in discriminative settings with discrete outputs, extensions to structured continuous spaces, as well as performing efficient inference in these spaces, are lacking. We propose non-parametric structured output networks (NSON), a modular approach that cleanly separates a non-parametric, structured posterior representation from a discriminative inference scheme but allows joint end-to-end training of both components. Our experiments evaluate the ability of NSONs to capture structured posterior densities (modeling) and to compute complex statistics of those densities (inference). We compare our model to output spaces of varying expressiveness and popular variational and sampling-based inference algorithms.

Visual dialog is a task of answering a series of inter-dependent questions given an input image, and often requires to resolve visual references among the questions. This problem is different from visual question answering (VQA), which relies on spatial attention ({\em a.k.a. visual grounding}) estimated from an image and question pair. We propose a novel attention mechanism that exploits visual attentions in the past to resolve the current reference in the visual dialog scenario. The proposed model is equipped with an associative attention memory storing a sequence of previous (attention, key) pairs. From this memory, the model retrieves previous attention, taking into account recency, that is most relevant for the current question, in order to resolve potentially ambiguous reference(s). The model then merges the retrieved attention with the tentative one to obtain the final attention for the current question; specifically, we use dynamic parameter prediction to combine the two attentions conditioned on the question. Through extensive experiments on a new synthetic visual dialog dataset, we show that our model significantly outperforms the state-of-the-art (by ~16 % points) in the situation where the visual reference resolution plays an important role. Moreover, the proposed model presents superior performance (~2 % points improvement) in the Visual Dialog dataset, despite having significantly fewer parameters than the baselines.

Most approaches for scene parsing, recognition or retrieval use detectors that are either (i) independently trained or (ii) jointly trained for conjunctions of object-object or object-attribute phrases. We posit that neither of these two extremes is uniformly optimal, in terms of performance, across all categories and conjunctions. The choice of whether one should train an independent or composite detector should be made for each possible conjunction separately, and depends on the statistics of the dataset as well. For example, person holding phone may be more accurately modeled using a single composite detector, while tall person may be more accurately modeled as combination of two detectors. We extensively study this issue in the context of multiple problems and datasets. Further, for e ciency, we propose a predictor that is based on a number of category speci c features (e.g., sample size, entropy, etc.) for whether independent or joint composite detector may be more accurate for a given conjunction. We show that our prediction and selection mechanism generalizes and leads to improved performance on a number of large-scale datasets and vision tasks.

We propose a novel learning framework for object categorization with interactive semantic feedback. In this framework, a discriminative categorization model improves through human-guided iterative semantic feedbacks. Specifically, the model identifies the most helpful relational semantic queries to discriminatively refine the model. The user feedback on whether the relationship is semantically valid or not is incorporated back into the model, in the form of regularization, and the process iterates. We validate the proposed model in a few-shot multi-class classification scenario, where we measure classification performance on a set of ‘target’ classes, with few training instances, by leveraging and transferring knowledge from ‘anchor’ classes, that contain larger set of labeled instances.

Viewpoint estimation, especially in case of multiple object classes, remains an important and challenging problem. First, objects under different views undergo extreme appearance variations, often making within-class variance larger than between-class variance. Second, obtaining precise ground truth for real-world images, necessary for training supervised viewpoint estimation models, is extremely difficult and time consuming. As a result, annotated data is often available only for a limited number of classes. Hence it is desirable to share viewpoint information across classes. Additional complexity arises from unaligned pose labels between classes, i.e. a side view of a car might look more like a frontal view of a toaster, than its side view. To address these problems, we propose a metric learning approach for joint class prediction and pose estimation. Our approach allows to circumvent the problem of viewpoint alignment across multiple classes, and does not require dense viewpoint labels. Moreover, we show, that the learned metric generalizes to new classes, for which the pose labels are not available, and therefore makes it possible to use only partially annotated training sets, relying on the intrinsic similarities in the viewpoint manifolds. We evaluate our approach on two challenging multi-class datasets, 3DObjects and PASCAL3D+.

Researchers often summarize their work in the form of posters. Posters provide a coherent and efficient way to convey core ideas from scientific papers. Generating a good scientific poster, however, is a complex and time consuming cognitive task, since such posters need to be readable, informative, and visually aesthetic. In this paper, for the first time, we study the challenging problem of learning to generate posters from scientific papers. To this end, a data-driven framework, that utilizes graphical models, is proposed. Specifically, given content to display, the key elements of a good poster, including panel layout and attributes of each panel, are learned and inferred from data. Then, given inferred layout and attributes, composition of graphical elements within each panel is synthesized. To learn and validate our model, we collect and make public a Poster-Paper dataset, which consists of scientific papers and corresponding posters with exhaustively labelled panels and attributes. Qualitative and quantitative results indicate the effectiveness of our approach.

Despite significant progress in object categorization, in recent years, a number of important challenges remain, mainly, ability to learn from limited labeled data and ability to recognize object classes within large, potentially open, set of labels. Zero-shot learning is one way of addressing these challenges, but it has only been shown to work with limited sized class vocabularies and typically requires separation between supervised and unsupervised classes, allowing former to inform the latter but not vice versa. We propose the notion of semi-supervised vocabulary-informed learning to alleviate the above mentioned challenges and address problems of supervised, zero-shot and open set recognition using a unified framework. Specifically, we propose a maximum margin framework for semantic manifold-based recognition that incorporates distance constraints from (both supervised and unsupervised) vocabulary atoms, ensuring that labeled samples are projected closest to their correct prototypes, in the embedding space, than to others. We show that resulting model shows improvements in supervised, zero-shot, and large open set recognition, with up to 310K class vocabulary on AwA and ImageNet datasets.

In this work we improve training of temporal deep models to better learn activity progression for activity detection and early detection tasks. Conventionally, when training a Recurrent Neural Network, specifically a Long Short Term Memory (LSTM) model, the training loss only considers classification error. However, we argue that the detection score of the correct activity category, or the detection score margin between the correct and incorrect categories, should be monotonically non-decreasing as the model observes more of the activity. We design novel ranking losses that directly penalize the model on violation of such monotonicities, which are used together with classification loss in training of LSTM models. Evaluation on ActivityNet shows significant benefits of the proposed ranking losses in both activity detection and early detection tasks.

Large-scale action recognition and video categorization are important problems in computer vision. To address these problems, we propose a novel object-and scene-based semantic fusion network and representation. Our semantic fusion network combines three streams of information using a three-layer neural network: (i) frame-based low-level CNN features, (ii) object features from a state-of-the-art large-scale CNN object-detector trained to recognize 20K classes, and (iii) scene features from a state-of-the-art CNN scene-detector trained to recognize 205 scenes. The trained network achieves improvements in supervised activity and video categorization in two complex large-scale datasets - ActivityNet and FCVID, respectively. Further, by examining and back propagating information through the fusion network, semantic relationships (correlations) between video classes and objects/scenes can be discovered. These video class-object/video class-scene relationships can in turn be used as semantic representation for the video classes themselves. We illustrate effectiveness of this semantic representation through experiments on zero-shot action/video classification and clustering.

Despite growing research interest, emotion understanding for user-generated videos remains a challenging problem. Major obstacles include the diversity and complexity of video content, as well as the sparsity of expressed emotions. For the first time, we systematically study large-scale video emotion recognition by transferring deep feature encodings. In addition to the traditional, supervised recognition, we study the problem of zero-shot emotion recognition, where emotions in the test set are unseen during training. To cope with this task, we utilize knowledge transferred from auxiliary image and text corpora. A novel auxiliary Image Transfer Encoding (ITE) process is proposed to efficiently encode and generate video representation. We also thoroughly investigate different configurations of convolutional neural networks. Comprehensive experiments on multiple datasets demonstrate the effectiveness of our framework.

We propose a method to rank and retrieve image sequences from a natural language text query, consisting of multiple sentences or paragraphs. One of the method's key applications is to visualize visitors' text-only reviews on TRIPADVISOR or YELP, by automatically retrieving the most illustrative image sequences. While most previous work has dealt with the relations between a natural language sentence and an image or a video, our work extends to the relations between paragraphs and image sequences. Our approach leverages the vast user-generated resource of blog posts and photo streams on the Web. We use blog posts as text-image parallel training data that co-locate informative text with representative images that are carefully selected by users. We exploit large-scale photo streams to augment the image samples for retrieval. We design a latent structural SVM framework to learn the semantic relevance relations between text and image sequences. We present both quantitative and qualitative results on the newly created DISNEYLAND dataset.

We propose an approach that utilizes large collections of photo streams and blog posts, two of the most prevalent sources of data on the Web, for joint story-based summarization and exploration. Blogs consist of sequences of images and associated text; they portray events and experiences with concise sentences and representative images. We leverage blogs to help achieve story-based semantic summarization of collections of photo streams. In the opposite direction, blog posts can be enhanced with sets of photo streams by showing interpolations between consecutive images in the blogs. We formulate the problem of joint alignment from blogs to photo streams and photo stream summarization in a unified latent ranking SVM framework. We alternate between solving the two coupled latent SVM problems, by first fixing the summarization and solving for the alignment from blog images to photo streams and vice versa. On a newly collected large-scale Disneyland dataset of 10K blogs (120K associated images) and 6K photo streams (540K images), we demonstrate that blog posts and photo streams are mutually beneficial for summarization, exploration, semantic knowledge transfer, and photo interpolation.

Over the last several years it has been shown that image-based object detectors are sensitive to the training data and often fail to generalize to examples that fall outside the original training sample domain (e.g., videos). A number of domain adaptation (DA) techniques have been proposed to address this problem. DA approaches are designed to adapt a fixed complexity model to the new (e.g., video) domain. We posit that unlabeled data should not only allow adaptation, but also improve (or at least maintain) performance on the original and other domains by dynamically adjusting model complexity and parameters. We call this notion domain expansion. To this end, we develop a new scalable and accurate incremental object detection algorithm, based on several extensions of large-margin embedding (LME). Our detection model consists of an embedding space and multiple class prototypes in that embedding space, that represent object classes; distance to those prototypes allows us to reason about multi-class detection. By incrementally detecting object instances in video and adding confident detections into the model, we are able to dynamically adjust the complexity of the detector over time by instantiating new prototypes to span all domains the model has seen. We test performance of our approach by expanding an object detector trained on ImageNet to detect objects in egocentric videos of Activity Daily Living (ADL) dataset and challenging videos from YouTube Objects (YTO) dataset.

Human actions are, inherently, structured patterns of body movements. We explore ensembles of hierarchical spatio-temporal trees, discovered directly from training data, to model these structures for action recognition. The hierarchical spatio-temporal trees provide a robust mid-level representation for actions. However, discovery of frequent and discriminative tree structures is challenging due to the exponential search space, particularly if one allows partial matching. We address this by first building a concise action vocabulary via discriminative clustering. Using the action vocabulary we then utilize tree mining with subsequent tree clustering and ranking to select a compact set of highly discriminative tree patterns. We show that these tree patterns, alone, or in combination with shorter patterns (action words and pairwise patterns) achieve state-of-the-art performance on two challenging datasets: UCF Sports and HighFive. Moreover, trees learned on HighFive are used in recognizing two action classes in a different dataset, Hollywood3D, demonstrating the potential for cross-dataset generality of the trees our approach discovers.

Egocentric videos are a valuable source of information as a daily log of our lives. However, large fraction of egocentric video content is typically irrelevant and boring to re-watch. It is an agonizing task, for example, to manually search for the moment when your daughter first met Mickey Mouse from hours-long egocentric videos taken at Disneyland. Although many summarization methods have been successfully proposed to create concise representations of videos, in practice, the value of the subshots to users may change according to their immediate preference/mood, thus summaries with fixed criteria may not fully satisfy users' various search intents. To address this, we propose a storyline representation that expresses an egocentric video as a set of jointly inferred, through MRF inference, story elements comprising of actors, locations, supporting objects and events, depicted on a timeline. We construct such a storyline with very limited annotation data (a list of map locations and weak knowledge of what events may be possible at each location), by bootstrapping the process with data obtained through focused Web image and video searches. Our representation promotes story-based search with queries in the form of AND-OR graphs, which span any subset of story elements and their spatio-temporal composition. We show effectiveness of our approach on a set of unconstrained YouTube egocentric videos of visits to Disneyland.

Learning from synthetic data has many important and practical applications, An example of application is photo-sketch recognition. Using synthetic data is challenging due to the differences in feature distributions between synthetic and real data, a phenomenon we term synthetic gap. In this paper, we investigate and formalize a general framework -- Stacked Multichannel Autoencoder (SMCAE) that enables bridging the synthetic gap and learning from synthetic data more efficiently. In particular, we show that our SMCAE can not only transform and use synthetic data on the challenging face-sketch recognition task, but that it can also help simulate real images, which can be used for training classifiers for recognition. Preliminary experiments validate the effectiveness of the framework.

We present an approach for generating pictorial storylines from large collections of online photo streams shared by visitors to theme parks (e.g. Disneyland), along with publicly available information such as visitor's maps. The story graph visualizes various events and activities recurring across visitors' photo sets, in the form of hierarchically branching narrative structure associated with attractions and districts in theme parks. We first estimate story elements of each photo stream, including the detection of faces and supporting objects, and attraction-based localization. We then create spatio-temporal story graphs via an inference of sparse time-varying directed graphs. Through quantitative evaluation and crowdsourcing-based user studies via Amazon Mechanical Turk, we show that the story graphs serve as a more convenient mid-level data structure to perform photo-based recommendation tasks than other alternatives. We also present storybook-like demo examples regarding exploration, recommendation, and temporal analysis, which may be most beneficial uses of the story graphs to visitors.

Family photo collections often contain richer semantics than arbitrary images of people because families contain a handful of specific individuals who can be associated with certain social roles (e.g. father, mother, or child). As a result, family photo collections have unique challenges and opportunities for face recognition compared to random groups of photos containing people. We address the problem of unsupervised family member discovery: given a collection of family photos, we infer the size of the family, as well as the visual appearance and social role of each family member. As a result, we are able to recognize the same individual across many different photos. We propose an unsupervised EM-style joint inference algorithm with a probabilistic CRF that models identity and role assignments for all detected faces, along with associated pair wise relationships between them. Our experiments illustrate how joint inference of both identity and role (across all photos simultaneously) outperforms independent estimates of each. Joint inference also improves the ability to recognize the same individual across many different photos.

We propose the problem of automated photo album creation from an unordered image collection. The problem is difficult as it involves a number of complex perceptual tasks that facilitate selection and ordering of photos to create a compelling visual narrative. To help solve this problem, we collect (and will make available) a new benchmark dataset based on Flickr images. Flickr Album Dataset and provides a variety of annotations useful for the task, including manually created albums of various lengths. We analyze the problem and provide experimental evidence, through user studies, that both selection and ordering of photos within an album is important for human observers. To capture and learn rules of album composition, we propose a discriminative structured model capable of encoding simple preferences for contextual layout of the scene (e.g., spatial layout of faces, global scene context, and presence/absence of attributes) and ordering between photos (e.g., exclusion principles or correlations). The parameters of the model are learned using a structured SVM framework. Once learned, the model allows automatic composition of photo albums from unordered and untagged collections of images. We quantitatively evaluate the results obtained using our model against manually created albums and baselines on a dataset of 63 personal photo collections from 5 different topics.

The goal of cross-domain matching (CDM) is to find correspondences between two sets of objects in different domains in an unsupervised way. CDM has various interesting applications, including photo album summarization where photos are automatically aligned into a designed frame expressed in the Cartesian coordinate system, and temporal alignment which aligns sequences such as videos that are potentially expressed using different features. In this paper, we propose an information-theoretic CDM framework based on squared-loss mutual information (SMI). The proposed approach can directly handle non-linearly related objects/sequences with different dimensions, with the ability that hyper-parameters can be objectively optimized by cross-validation. We apply the proposed method to several real-world problems including image matching, unpaired voice conversion, photo album summarization, cross-feature video and cross-domain video-to-mocap alignment, and Kinect-based action recognition, and experimentally demonstrate that the proposed method is a promising alternative to state-of-the-art CDM methods.

We present a perceptual control space for simulation of cloth that works with any physical simulator, treating it as a black box. The perceptual control space provides intuitive, art-directable control over the simulation behavior based on a learned mapping from common descriptors for cloth ( e.g. , flowiness, softness) to the parameters of the simulation. To learn the mapping, we perform a series of perceptual experiments in which the simulation parameters are varied and participants assess the values of the common terms of the cloth on a scale. A multi-dimensional sub-space regression is performed on the results to build a perceptual generative model over the simulator parameters. We evaluate the perceptual control space by demonstrating that the generative model does in fact create simulated clothing that is rated by participants as having the expected properties. We also show that this perceptual control space generalizes to garments and motions not in the original experiments.

In this paper, we address the problem of jointly summarizing large sets of Flickr images and YouTube videos. Starting from the intuition that the characteristics of the two media types are different yet complementary, we develop a fast and easily-parallelizable approach for creating not only high-quality video summaries but also novel structural summaries of online images as storyline graphs. The storyline graphs can illustrate various events or activities associated with the topic in a form of a branching network. The video summarization is achieved by diversity ranking on the similarity graphs between images and video frames. The reconstruction of storyline graphs is formulated as the inference of sparse time-varying directed graphs from a set of photo streams with assistance of videos. For evaluation, we collect the datasets of 20 outdoor activities, consisting of 2.7M Flickr images and 16K YouTube videos. Due to the large-scale nature of our problem, we evaluate our algorithm via crowdsourcing using Amazon Mechanical Turk. In our experiments, we demonstrate that the proposed joint summarization approach outperforms other baselines and our own methods using videos or images only.

The goal of supervised feature selection is to find a subset of input features that are responsible for predicting output values. The least absolute shrinkage and selection operator (Lasso) allows computationally efficient feature selection based on linear dependency between input features and output values. In this letter, we consider a feature-wise kernelized Lasso for capturing nonlinear input-output dependency. We first show that with particular choices of kernel functions, nonredundant features with strong statistical dependence on output values can be found in terms of kernel-based independence measures such as the Hilbert-Schmidt independence criterion. We then show that the globally optimal solution can be efficiently computed; this makes the approach scalable to high-dimensional problems. The effectiveness of the proposed method is demonstrated through feature selection experiments for classification and regression with thousands of features.

Discriminative, or (structured) prediction, methods have proved effective for variety of problems in computer vision; a notable example is 3D monocular pose estimation. All methods to date, however, relied on an assumption that training (source) and test (target) data come from the same underlying joint distribution. In many real cases, including standard data sets, this assumption is flawed. In the presence of training set bias, the learning results in a biased model whose performance degrades on the (target) test set. Under the assumption of covariate shift, we propose an unsupervised domain adaptation approach to address this problem. The approach takes the form of training instance reweighting, where the weights are assigned based on the ratio of training and test marginals evaluated at the samples. Learning with the resulting weighted training samples alleviates the bias in the learned models. We show the efficacy of our approach by proposing weighted variants of kernel regression (KR) and twin Gaussian processes (TGP). We show that our weighted variants outperform their unweighted counterparts and improve on the state-of-the-art performance in the public (HumanEva) data set.

In this paper, we develop a new model for recognizing human actions. An action is modeled as a very sparse sequence of temporally local discriminative key frames - collections of partial key-poses of the actor(s), depicting key states in the action sequence. We cast the learning of key frames in a max-margin discriminative framework, where we treat key frames as latent variables. This allows us to (jointly) learn a set of most discriminative key frames while also learning the local temporal context between them. Key frames are encoded using a spatially-localizable pose let-like representation with HoG and BoW components learned from weak annotations, we rely on structured SVM formulation to align our components and mine for hard negatives to boost localization performance. This results in a model that supports spatio-temporal localization and is insensitive to dropped frames or partial observations. We show classification performance that is competitive with the state of the art on the benchmark UT-Interaction dataset and illustrate that our model outperforms prior methods in an on-line streaming setting.

The appearance of an object changes profoundly with pose, camera view and interactions of the object with other objects in the scene. This makes it challenging to learn detectors based on an object-level label (e.g., "car"). We postulate that having a richer set of labelings (at different levels of granularity) for an object, including finer-grained subcategories, consistent in appearance and view, and higher order composites - contextual groupings of objects consistent in their spatial layout and appearance, can significantly alleviate these problems. However, obtaining such a rich set of annotations, including annotation of an exponentially growing set of object groupings, is simply not feasible. We propose a weakly-supervised framework for object detection where we discover subcategories and the composites automatically with only traditional object-level category labels as input. To this end, we first propose an exemplar-SVM-based clustering approach, with latent SVM refinement, that discovers a variable length set of discriminative subcategories for each object class. We then develop a structured model for object detection that captures interactions among object subcategories and automatically discovers semantically meaningful and discriminatively relevant visual composites. We show that this model produces state-of-the-art performance on UIUC phrase object detection benchmark.

We organized a Grand Challenge and Workshop on Multi-Modal Gesture Recognition.

We propose a new weakly-supervised structured learning approach for recognition and spatio-temporal localization of actions in video. As part of the proposed approach we develop a generalization of the Max-Path search algorithm, which allows us to efficiently search over a structured space of multiple spatio-temporal paths, while also allowing to incorporate context information into the model. Instead of using spatial annotations, in the form of bounding boxes, to guide the latent model during training, we utilize human gaze data in the form of a weak supervisory signal. This is achieved by incorporating gaze, along with the classification, into the structured loss within the latent SVM learning framework. Experiments on a challenging benchmark dataset, UCF-Sports, show that our model is more accurate, in terms of classification, and achieves state-of-the-art results in localization. In addition, we show how our model can produce top-down saliency maps conditioned on the classification label and localized latent paths.

We propose a simulation-based dynamical motion prior for tracking human motion from video in presence of physical ground-person interactions. Most tracking approaches to date have focused on efficient inference algorithms and/or learning of prior kinematic motion models; however, few can explicitly account for the physical plausibility of recovered motion. Here, we aim to recover physically plausible motion of a single articulated human subject. Toward this end, we propose a full-body 3D physical simulation-based prior that explicitly incorporates a model of human dynamics into the Bayesian filtering framework. We consider the motion of the subject to be generated by a feedback “control loop” in which Newtonian physics approximates the rigid-body motion dynamics of the human and the environment through the application and integration of interaction forces, motor forces, and gravity. Interaction forces prevent physically impossible hypotheses, enable more appropriate reactions to the environment (e.g., ground contacts), and are produced from detected human-environment collisions. Motor forces actuate the body, ensure that proposed pose transitions are physically feasible, and are generated using a motion controller. For efficient inference in the resulting high-dimensional state space, we utilize an exemplar-based control strategy that reduces the effective search space of motor forces. As a result, we are able to recover physically plausible motion of human subjects from monocular and multiview video. We show, both quantitatively and qualitatively, that our approach performs favorably with respect to Bayesian filtering methods with standard motion priors.

We present a hierarchical model for human activity recognition in entire multi-person scenes. Our model describes human behaviour at multiple levels of detail, ranging from low-level actions through to high-level events. We also include a model of social roles, the expected behaviours of certain people, or groups of people, in a scene. The hierarchical model includes these varied representations, and various forms of interactions between people present in a scene. The model is trained in a discriminative max-margin framework. Experimental results demonstrate that this model can improve performance at all considered levels of detail, on two challenging datasets.

We formulate the problem of 3D human pose estimation and tracking as one of inference in a graphical model. Unlike traditional kinematic tree representations, our model of the body is a collection of loosely-connected body-parts. In particular, we model the body using an undirected graphical model in which nodes correspond to parts and edges to kinematic, penetration, and temporal constraints imposed by the joints and the world. These constraints are encoded using pair-wise statistical distributions, that are learned from motion-capture training data. Human pose and motion estimation is formulated as inference in this graphical model and is solved using Particle Message Passing ( PaMPas ). PaMPas is a form of non-parametric belief propagation that uses a variation of particle filtering that can be applied over a general graphical model with loops. The loose-limbed model and decentralized graph structure allow us to incorporate information from “bottom-up” visual cues, such as limb and head detectors, into the inference process. These detectors enable automatic initialization and aid recovery from transient tracking failures. We illustrate the method by automatically tracking people in multi-view imagery using a set of calibrated cameras and present quantitative evaluation using the HumanEva dataset.

Latent variable models, such as the GPLVM and related methods, help mitigate overfitting when learning from small or moderately sized training sets. Nevertheless, existing methods suffer from several problems: 1) complexity, 2) the lack of explicit mappings to and from the latent space, 3) an inability to cope with multimodality, and 4) the lack of a well-defined density over the latent space. We propose an LVM called the Kernel Information Embedding (KIE) that defines a coherent joint density over the input and a learned latent space. Learning is quadratic, and it works well on small data sets. We also introduce a generalization, the shared KIE (sKIE), that allows us to model multiple input spaces (e.g., image features and poses) using a single, shared latent representation. KIE and sKIE permit missing data during inference and partially labeled data during learning. We show that with data sets too large to learn a coherent global model, one can use the sKIE to learn local online models. We use sKIE for human pose inference.

Marker-less motion capture is a challenging problem, particularly when only monocular video is available. We estimate human motion from monocular video by recovering three-dimensional controllers capable of implicitly simulating the observed human behavior and replaying this behavior in other environments and under physical perturbations. Our approach employs a state-space biped controller with a balance feedback mechanism that encodes control as a sequence of simple control tasks. Transitions among these tasks are triggered on time and on proprioceptive events ( e.g ., contact). Inference takes the form of optimal control where we optimize a high-dimensional vector of control parameters and the structure of the controller based on an objective function that compares the resulting simulated motion with input observations. We illustrate our approach by automatically estimating controllers for a variety of motions directly from monocular video. We show that the estimation of controller structure through incremental optimization and refinement leads to controllers that are more stable and that better approximate the reference motion. We demonstrate our approach by capturing sequences of walking, jumping, and gymnastics.

To deal with multi-modality in human pose estimation, mixture models or local models are introduced. However, problems with over-fitting and generalization are caused by our necessarily limited data, and the regression parameters need to be determined without resorting to slow and processor-hungry techniques, such as cross validation. To compensate these problems, we have developed a semi-parametric regression model in latent space with variational inference. Our method performed competitively in comparison to other current methods.

We present a type of Temporal Restricted Boltzmann Machine that defines a probability distribution over an output sequence conditional on an input sequence. It shares the desirable properties of RBMs: efficient exact inference, an exponentially more expressive latent state than HMMs, and the ability to model nonlinear structure and dynamics. We apply our model to a challenging real-world graphics problem: facial expression transfer. Our results demonstrate improved performance over several baselines modeling high-dimensional 2D and 3D data.

Motion capture technology generally requires that recordings be performed in a laboratory or closed stage setting with controlled lighting. This restriction precludes the capture of motions that require an outdoor setting or the traversal of large areas. In this paper, we present the theory and practice of using body-mounted cameras to reconstruct the motion of a subject. Outward-looking cameras are attached to the limbs of the subject, and the joint angles and root pose are estimated through non-linear optimization. The optimization objective function incorporates terms for image matching error and temporal continuity of motion. Structure-from-motion is used to estimate the skeleton structure and to provide initialization for the non-linear optimization procedure. Global motion is estimated and drift is controlled by matching the captured set of videos to reference imagery. We show results in settings where capture would be difficult or impossible with traditional motion capture systems, including walking outside and swinging on monkey bars. The quality of the motion reconstruction is evaluated by comparing our results against motion capture data produced by a commercially available optical system.

We introduce a new class of probabilistic latent variable model called the Implicit Mixture of Conditional Restricted Boltzmann Machines (imCRBM) for use in human pose tracking. Key properties of the imCRBM are as follows: (1) learning is linear in the number of training exemplars so it can be learned from large datasets; (2) it learns coherent models of multiple activities; (3) it automatically discovers atomic “movemes” and (4) it can infer transitions between activities, even when such transitions are not present in the training set. We describe the model and how it is learned and we demonstrate its use in the context of Bayesian filtering for multi-view and monocular pose tracking. The model handles difficult scenarios including multiple activities and transitions among activities. We report state-of-the-art results on the HumanEva dataset.

We present a technique for learning clothing models that enables the simultaneous animation of thousands of detailed garments in real-time. This surprisingly simple conditional model learns and preserves the key dynamic properties of a cloth motion along with folding details. Our approach requires no a priori physical model, but rather treats training data as a "black box." We show that the models learned with our method are stable over large time-steps and can approximately resolve cloth-body collisions. We also show that within a class of methods, no simpler model covers the full range of cloth dynamics captured by ours. Our method bridges the current gap between skinning and physical simulation, combining benefits of speed from the former with dynamic effects from the latter. We demonstrate our approach on a variety of apparel worn by male and female human characters performing a varied set of motions typically used in video games ( e.g. , walking, running, jumping, etc. ).

Latent variable models (LVM), like the shared-GPLVM and the spectral latent variable model, help mitigate over-fitting when learning discriminative methods from small or moderately sized training sets. Nevertheless, existing methods suffer from several problems: (1) complexity; (2) the lack of explicit mappings to and from the latent space; (3) an inability to cope with multi-modality; and (4) the lack of a well-defined density over the latent space. We propose a LVM called the shared kernel information embedding (sKIE). It defines a coherent density over a latent space and multiple input/output spaces (e.g., image features and poses), and it is easy to condition on a latent state, or on combinations of the input/output states. Learning is quadratic, and it works well on small datasets. With datasets too large to learn a coherent global model, one can use sKIE to learn local online models. sKIE permits missing data during inference, and partially labelled data during learning. We use sKIE for human pose inference.

Motion and interaction with the environment are fundamentally intertwined. Few people-tracking algorithms exploit such interactions, and those that do assume that surface geometry and dynamics are given. This paper concerns the converse problem, i.e., the inference of contact and environment properties from motion. For 3D human motion, with a 12-segment articulated body model, we show how one can estimate the forces acting on the body in terms of internal forces (joint torques), gravity, and the parameters of a contact model (e.g., the geometry and dynamics of a spring-based model). This is tested on motion capture data and video-based tracking data, with walking, jogging, cartwheels, and jumping.

Human motion tracking is an important problem in computer vision. Most prior approaches have concentrated on efficient inference algorithms and prior motion models; however, few can explicitly account for physical plausibility of recovered motion. The primary purpose of this work is to enforce physical plausibility in the tracking of a single articulated human subject. Towards this end, we propose a full-body 3D physical simulation-based prior that explicitly incorporates motion control and dynamics into the Bayesian filtering framework. We consider the humanpsilas motion to be generated by a ldquocontrol looprdquo. In this control loop, Newtonian physics approximates the rigid-body motion dynamics of the human and the environment through the application and integration of forces. Collisions generate interaction forces to prevent physically impossible hypotheses. This allows us to properly model human motion dynamics, ground contact and environment interactions. For efficient inference in the resulting high-dimensional state space, we introduce exemplar-based control strategy to reduce the effective search space. As a result we are able to recover the physically-plausible kinematic and dynamic state of the body from monocular and multi-view imagery. We show, both quantitatively and qualitatively, that our approach performs favorably with respect to standard Bayesian filtering methods.

Much of the research on video-based human motion capture assumes the body shape is known a priori and is represented coarsely (e.g. using cylinders or superquadrics to model limbs). These body models stand in sharp contrast to the richly detailed 3D body models used by the graphics community. Here we propose a method for recovering such models directly from images. Specifically, we represent the body using a recently proposed triangulated mesh model called SCAPE which employs a low-dimensional, but detailed, parametric model of shape and pose-dependent deformations that is learned from a database of range scans of human bodies. Previous work showed that the parameters of the SCAPE model could be estimated from marker-based motion capture data. Here we go further to estimate the parameters directly from image data. We define a cost function between image observations and a hypothesized mesh and formulate the problem as optimization over the body shape and pose parameters using stochastic search. Our results show that such rich generative models enable the automatic recovery of detailed human shape and pose from images.

Strong lighting is common in natural scenes yet is often viewed as a nuisance for object pose estimation and tracking. In human shape and pose estimation, cast shadows can be confused with foreground structure while self shadowing and shading variation on the body cause the appearance of the person to change with pose. Rather than attempt to minimize the effects of lighting and shadows, we show that strong lighting in a scene actually makes pose and shape estimation more robust. Additionally, by recovering multiple body poses we are able to automatically estimate the lighting in the scene and the albedo of the body. Our approach makes use of a detailed 3D body model, the parameters of which are directly recovered from image data. We provide a thorough exploration of human pose estimation under strong lighting conditions and show: 1. the estimation of the light source from cast shadows; 2. the estimation of the light source and the albedo of the body from multiple body poses; 3. that a point light and cast shadows on the ground plane can be treated as an additional "shadow camera" that improves pose and shape recovery, particularly in monocular scenes. Additionally we introduce the notion of albedo constancy which employs lighting normalized image data for matching. Our experiments with multiple subjects show that rather than causing problems, strong lighting improves human pose and shape estimation.

Estimation of three-dimensional articulated human pose and motion from images is a central problem in computer vision. Much of the previous work has been limited by the use of crude generative models of humans represented as articu- lated collections of simple parts such as cylinders. Automatic initialization of such models has proved difficult and most approaches assume that the size and shape of the body parts are known a priori. In this paper we propose a method for automatically recovering a detailed parametric model of non-rigid body shape and pose from monocular imagery. Specifically, we represent the body using a param- eterized triangulated mesh model that is learned from a database of human range scans. We demonstrate a discriminative method to directly recover the model pa- rameters from monocular images using a conditional mixture of kernel regressors. This predicted pose and shape are used to initialize a generative model for more detailed pose and shape estimation. The resulting approach allows fully automatic pose and shape recovery from monocular and multi-camera imagery. Experimen- tal results show that our method is capable of robustly recovering articulated pose, shape and biometric measurements (e.g. height, weight, etc.) in both calibrated and uncalibrated camera environments.

Part-based tree-structured models have been widely used for 2D articulated human pose-estimation. These approaches admit efficient inference algorithms while capturing the important kinematic constraints of the human body as a graphical model. These methods often fail however when multiple body parts fit the same image region resulting in global pose estimates that poorly explain the overall image evidence. Attempts to solve this problem have focused on the use of strong prior models that are limited to learned activities such as walking. We argue that the problem actually lies with the image observations and not with the prior. In particular, image evidence for each body part is estimated independently of other parts without regard to self-occlusion. To address this we introduce occlusion-sensitive local likelihoods that approximate the global image likelihood using per-pixel hidden binary variables that encode the occlusion relationships between parts. This occlusion reasoning introduces interactions between non-adjacent body parts creating loops in the underlying graphical model. We deal with this using an extension of an approximate belief propagation algorithm (PAMPAS). The algorithm recovers the real-valued 2D pose of the body in the presence of occlusions, does not require strong priors over body pose and does a quantitatively better job of explaining image evidence than previous methods.

A novel approach for real-time skin segmentation in video sequences is described. The approach enables reliable skin segmentation despite wide variation in illumination during tracking. An explicit second order Markov model is used to predict evolution of the skin-color (HSV) histogram over time. Histograms are dynamically updated based on feedback from the current segmentation and predictions of the Markov model. The evolution of the skin-color distribution at each frame is parameterized by translation, scaling, and rotation in color space. Consequent changes in geometric parameterization of the distribution are propagated by warping and resampling the histogram. The parameters of the discrete-time dynamic Markov model are estimated using Maximum Likelihood Estimation and also evolve over time. The accuracy of the new dynamic skin color segmentation algorithm is compared to that obtained via a static color model. Segmentation accuracy is evaluated using labeled ground-truth video sequences taken from staged experiments and popular movies. An overall increase in segmentation accuracy of up to 24 percent is observed in 17 out of 21 test sequences. In all but one case, the skin-color classification rates for our system were higher, with background classification rates comparable to those of the static segmentation.

The detection and pose estimation of people in images and video is made challenging by the variability of human appearance, the complexity of natural scenes, and the high dimensionality of articulated body mod- els. To cope with these problems we represent the 3D human body as a graphical model in which the relationships between the body parts are represented by conditional probability distributions. We formulate the pose estimation problem as one of probabilistic inference over a graphi- cal model where the random variables correspond to the individual limb parameters (position and orientation). Because the limbs are described by 6-dimensional vectors encoding pose in 3-space, discretization is im- practical and the random variables in our model must be continuous- valued. To approximate belief propagation in such a graph we exploit a recently introduced generalization of the particle filter. This framework facilitates the automatic initialization of the body-model from low level cues and is robust to occlusion of body parts and scene clutter.

A model approach for real time skin segmentation in video sequences is described. The approach enables reliable skin segmentation despite wide variation in illumination during tracking. An explicit second order Markov model is used to predict evolution of the skin color (HSV) histogram over time. Histograms are dynamically updated based on feedback from the current segmentation and based on predictions of the Markov model. The evolution of the skin color distribution at each frame is parameterized by translation, scaling and rotation in color space. Consequent changes in geometric parameterization of the distribution are propagated by warping and re-sampling the histogram. The parameters of the discrete time dynamic Markov model are estimated using maximum likelihood estimation, and also evolve over time. Quantitative evaluation of the method was conducted on labeled ground-truth video sequences taken from popular movies.