Recent advances in large language models (LLMs) have greatly improved their reasoning and decision-making abilities when deployed as agents. Richer reasoning, however, often comes at the cost of longer chain of thought (CoT), hampering interaction efficiency in real-world scenarios. Nevertheless, there still lacks systematic definition of LLM‑Agent efficiency, hindering targeted improvements. To this end, we introduce dual‑efficiency, comprising (i) step-level efficiency, which minimizes tokens per step, and (ii) trajectory-level efficiency, which minimizes the number of steps to complete a task. Building on this definition, we propose DEPO, a dual-efficiency preference‑based optimization method that jointly rewards succinct responses and fewer action steps. Experiments on WebShop and BabyAI show that DEPO cuts token usage by up to 60.9% and steps by up to 26.9%, while achieving up to a 29.3% improvement in task performance. DEPO also generalizes to three out-of-domain math benchmarks and retains its efficiency gains when trained on only 25% of the data.

We tackle the task of customized image editing using a text-conditioned Diffusion Model (DM). The goal is to fuse the subject in a reference image (e.g., sunglasses) with a source one (e.g., a boy), while retaining the fidelity of them both (e.g., the boy wearing the sunglasses). An intuitive approach, called LoRA fusion, first separately trains a DM LoRA for each image to encode its details. Then the two LoRAs are linearly combined by a weight to generate a fused image. Unfortunately, even through careful grid search or learning the weight, this approach still trades off the fidelity of one image against the other. We point out that the evil lies in the overlooked role of diffusion time-step in the generation process, i.e., a smaller time-step controls the generation of a more fine-grained attribute. For example, a large LoRA weight for the source may help preserve its fine-grained details (e.g., face attributes) at a small time-step, but could overpower the reference subject LoRA and lose the fidelity of its overall shape at a larger time-step. To address this deficiency, we propose TimeFusion, which learns a time-step-specific LoRA fusion weight that resolves the trade-off, i.e., generating the source and reference subject in high fidelity given their respective prompt. Then we can customize image editing using this weight and a target prompt.

Personalizing 3D scenes from a single reference image enables intuitive user-guided editing, which requires achieving both multi-view consistency across perspectives and referential consistency with the input image. However, these goals are particularly challenging due to the viewpoint bias caused by the limited perspective provided in a single image. Lacking the mechanisms to effectively expand reference information beyond the original view, existing methods of image-conditioned 3DGS personalization often suffer from this viewpoint bias and struggle to produce consistent results. Therefore, in this paper, we present Consistent Personalization for 3D Gaussian Splatting (CP-GS), a framework that progressively propagates the single-view reference appearance to novel perspectives. In particular, CP-GS integrates pre-trained image-to-3D generation and iterative LoRA fine-tuning to extract and extend the reference appearance, and finally produces faithful multi-view guidance images and the personalized 3DGS outputs through a view-consistent generation process guided by geometric cues. Extensive experiments on real-world scenes show that our CP-GS effectively mitigates the viewpoint bias, achieving high-quality image-conditioned 3DGS personalization that significantly outperforms existing methods.

3D Gaussian Splatting (3DGS) has demonstrated impressive Novel View Synthesis (NVS) results in a real-time rendering manner. During training, it relies heavily on the average magnitude of view-space positional gradients to grow Gaussians to reduce rendering loss. However, this average operation smooths the positional gradients from different viewpoints and rendering errors from different pixels, hindering the growth and optimization of many defective Gaussians. This leads to strong spurious artifacts in some areas. To address this problem, we propose Hard Gaussian Splatting, dubbed HGS, which considers multi-view significant positional gradients and rendering errors to grow hard Gaussians that fill the gaps of classical Gaussian Splatting on 3D scenes, thus achieving superior NVS results. In detail, we present positional gradient driven HGS, which leverages multi-view significant positional gradients to uncover hard Gaussians. Moreover, we propose rendering error guided HGS, which identifies noticeable pixel rendering errors and potentially over-large Gaussians to jointly mine hard Gaussians. By growing and optimizing these hard Gaussians, our method helps to resolve blurring and needle-like artifacts. Experiments on various datasets demonstrate that our method achieves state-of-the-art rendering quality while maintaining real-time efficiency, yielding LPIPS improvements of 5.1%, 19.7% and 6.3% on Mip-NeRF360, Tanks&Temples and Deep Blending, respectively.

In this paper, we aim to create physical digital twins of deformable objects under interaction. Existing methods focus more on the physical learning of current state modeling, but generalize worse to future prediction. This is because existing methods ignore the intrinsic physical properties of deformable objects, resulting in the limited physical learning in the current state modeling. To address this, we present NeuSpring, a neural spring field for the reconstruction and simulation of deformable objects from videos. Built upon spring-mass models for realistic physical simulation, our method consists of two major innovations: 1) a piecewise topology solution that efficiently models multi-region spring connection topologies using zero-order optimization, which considers the material heterogeneity of real-world objects. 2) a neural spring field that represents spring physical properties across different frames using a canonical coordinate-based neural network, which effectively leverages the spatial associativity of springs for physical learning. Experiments on real-world datasets demonstrate that our NeuSping achieves superior reconstruction and simulation performance for current state modeling and future prediction, with Chamfer distance improved by 20% and 25%, respectively.

Drag-Based Image Editing (DBIE), which allows users to manipulate images by directly dragging objects within them, has recently attracted much attention from the community. However, it faces two key challenges: (i) point-based drag is often highly ambiguous and difficult to align with user intentions; (ii) current DBIE methods primarily rely on alternating between motion supervision and point tracking, which is not only cumbersome but also fails to produce high-quality results. These limitations motivate us to explore DBIE from a new perspective---unifying it as a Latent Region Optimization (LRO) problem that aims to use region-level geometric transformations to optimize latent code to realize drag manipulation. Thus, by specifying the areas and types of geometric transformations, we can effectively address the ambiguity issue. We also propose a simple yet effective editing framework, dubbed DragNeXt. It solves LRO through Progressive Backward Self-Intervention (PBSI), simplifying the overall procedure of the alternating workflow while further enhancing quality by fully leveraging region-level structure information and progressive guidance from intermediate drag states. We validate DragNeXt on our NextBench, and extensive experiments demonstrate that our proposed method can significantly outperform existing approaches.

Image clustering is a classic problem in computer vision, which categorizes images into different groups. Recent studies utilize nouns as external semantic knowledge to improve clustering performance. However, these methods often overlook the inherent ambiguity of nouns, which can distort semantic representations and degrade clustering quality. To address this issue, we propose a hierarChical semAntic alignmEnt method for image clustering, dubbed CAE, which improves clustering performance in a training-free manner. In our approach, we incorporate two complementary types of textual semantics: caption-level descriptions, which convey fine-grained attributes of image content, and noun-level concepts, which represent high-level object categories. We first select relevant nouns from WordNet and descriptions from caption datasets to construct a semantic space aligned with image features. Then, we design a residual attention mechanism to further enhance the discriminability of this space. Finally, we combine the enhanced semantic and image features to perform clustering. Extensive experiments across 8 datasets demonstrate the effectiveness of our method, notably surpassing the state-of-the-art training-free approach with a 4.2% improvement in accuracy and a 2.9% improvement in adjusted rand index (ARI) on the ImageNet-1K dataset.

Despite recent progress, video generative models still struggle to generate delicate human actions (e.g., gymnastics), particularly when they are required to start from a user-provided reference image. In this paper, we explore the task of learning to animate images into videos that portray delicate human actions using a small number of videos — 16 or fewer — which reduces the need for extensive data collection and enhances practicality for real-world applications. Learning generalizable motion patterns that smoothly transition from user-provided reference images in such a few-shot setting is highly challenging. We propose FLASH (Few-shot Learning to Animate and Steer Humans), which enhances generalization of motion by training the model to reconstruct a video using the motion features and cross-frame correspondences extracted from another video with the same motion but different appearance. This encourages the learning of transferable motion and mitigates overfitting to the appearance in limited training data. Additionally, FLASH extends the decoder with additional layers to propagate details from the reference image to generated frames, improving transition smoothness. Human judges significantly favor FLASH, with 65.78% of 488 responses prefer FLASH over baselines. We strongly recommend watching the videos on the webpage1, as motion artifacts are hard to notice from images.

Recent studies have revealed the potential of training open-source Large Language Models (LLMs) to unleash LLMs' reasoning ability for enhancing vision-language navigation (VLN) performance, and simultaneously mitigate the domain gap between LLMs' training corpus and the VLN task. However, these approaches predominantly adopt straightforward input-output mapping paradigms, causing the mapping learning difficult and the navigational decisions unexplainable. Chain-of-Thought (CoT) training is a promising way to improve both navigational decision accuracy and interpretability, while the complexity of the navigation task makes the perfect CoT labels unavailable and may lead to overfitting through pure CoT supervised fine-tuning. To address these issues, we propose EvolveNav, a novel sElf-improving embodied reasoning paradigm that realizes adaptable and generalizable navigational reasoning for boosting LLM-based vision-language Navigation. Specifically, EvolveNav involves a two-stage training process: (1) Formalized CoT Supervised Fine-Tuning, where we train the model with curated formalized CoT labels to first activate the model's navigational reasoning capabilities, and simultaneously increase the reasoning speed; (2) Self-Reflective Post-Training, where the model is iteratively trained with its own reasoning outputs as self-enriched CoT labels to enhance the supervision diversity. A self-reflective auxiliary task is also designed to encourage the model to learn correct reasoning patterns by contrasting with wrong ones. Experimental results under both task-specific and cross-task training paradigms demonstrate the consistent superiority of EvolveNav over previous LLM-based VLN approaches on various popular benchmarks, including R2R, REVERIE, CVDN, and SOON. EvolveNav open avenues for exploring effective self-improving reasoning paradigms, enabling building agents capable of self-evolving for promoting LLM-based embodied AI research.

VQA explanation task aims to explain the decision-making process of VQA models in a way that is easily understandable to humans. Existing methods mostly use visual location or natural language explanation approaches to generate corresponding rationales. Although significant progress has been made, these frameworks are bottlenecked by the following challenges: 1) Uni-modal paradigm inevitably leads to semantic ambiguity of explanations. 2) The reasoning process cannot be faithfully responded to and suffers from logical inconsistency. 3) Human-annotated explanations are expensive and time-consuming to collect. In this paper, we introduce a new Semi-supervised VQA Multi-modal Explanation (SME) method via self-critical learning, which addresses the above challenges by leveraging both visual and textual explanations to comprehensively reveal the inference process of the model. Meanwhile, in order to improve the logical consistency between answers and rationales, we design a novel self-critical strategy to evaluate candidate explanations based on answer reward scores. More importantly, our method can benefit from a tremendous amount of samples without human-annotated explanations with semi-supervised learning. Extensive automatic measures and human evaluations all show the effectiveness of our method. Finally, the framework achieves a new state-of-the-art performance on the three VQA explanation datasets.

Visual relation detection (VRD) aims to identify relationships (or interactions) between object pairs in an image. Although recent VRD models have achieved impressive performance, they are all restricted to pre-defined relation categories, while failing to consider thesemantic ambiguitycharacteristic of visual relations. Unlike objects, the appearance of visual relations is always subtle and can be described by multiple predicate words from different perspectives, e.g., “ride” can be depicted as “race” and “sit on”, from the sports and spatial position views, respectively. To this end, we propose to model visual relations as continuous embeddings, and design diffusion models to achieve generalized VRD in a conditional generative manner, termed Diff-VRD. We model the diffusion process in a latent space and generate all possible relations in the image as an embedding sequence. During the generation, the visual and text embeddings of subject-object pairs serve as conditional signals and are injected via cross-attention. After the generation, we design a subsequent matching stage to assign the relation words to subject-object pairs by considering their semantic similarities. Benefiting from the diffusion-based generative process, our Diff-VRD is able to generate visual relations beyond the pre-defined category labels of datasets. To properly evaluate this generalized VRD task, we introduce two evaluation metrics, i.e., text-to-image retrieval and SPICE PR Curve inspired by image captioning. Extensive experiments in both human-object interaction (HOI) detection and scene graph generation (SGG) benchmarks attest to the superiority and effectiveness of Diff-VRD.

Few-Shot Class-Incremental Learning (FSCIL) requires models to be updated incrementally with limited labeled samples given in each session, differing from the traditional training paradigm and easily resulting in severe spurious relations between categories. Thus, in this paper, we propose to address FSCIL from a new perspective: enhancing FSCIL via disentangling spurious relations between categories. Accordingly, we propose a simple yet effective approach, dubbed ConTrollable Relation-disentangLed Few-Shot Class-Incremental Learning (CTRL-FSCIL). Specifically, during a base session, we propose to anchor base class embeddings in feature space and build disentangled proxies to bridge gaps between the learning processes of categories encountered in different sessions, making category relations controllable. Furthermore, during incremental learning, the parameters of the backbone network are frozen in order to relieve the negative impact of data scarcity. Meanwhile, a relation disentanglement loss is employed to guide a relation control module to disentangle spurious relations between learned categories. In this way, spurious relation issues in FSCIL can be alleviated. Extensive experiments on CIFAR-100, mini-ImageNet, and CUB-200 demonstrate the effectiveness of CTRL-FSCIL. Our code has been publicly released on github.

In this paper, we propose the LoRA of Change (LoC) framework for image editing with visual instructions, i.e., before-after image pairs. Compared to the ambiguities, insufficient specificity, and diverse interpretations of natural language, visual instructions can accurately reflect users' intent. Building on the success of LoRA in text-based image editing and generation, we dynamically learn an instruction-specific LoRA to encode the "change" in a before-after image pair, enhancing the interpretability and reusability of our model. Furthermore, generalizable models for image editing with visual instructions typically require quad data, i.e., a before-after image pair, along with query and target images. Due to the scarcity of such quad data, existing models are limited to a narrow range of visual instructions. To overcome this limitation, we introduce the LoRA Reverse optimization technique, enabling large-scale training with paired data alone. Extensive qualitative and quantitative experiments demonstrate that our model produces high-quality images that align with user intent and support a broad spectrum of real-world visual instructions.

Distribution estimation is a pivotal strategy in few-shot learning (FSL) to mitigate data scarcity by sampling from estimated distributions, utilizing statistical properties (mean and variance) transferred from related base categories. However, category-level estimation alone often fails to generate representative samples due to significant dissimilarities between base and novel categories, leading to suboptimal performance. To address this limitation, we propose Hybrid Granularity Distribution Estimation (HGDE), which integrates both coarse-grained category-level statistics and fine-grained instance-level statistics. By leveraging instance statistics from the nearest base samples, HGDE enhances the characterization of novel categories, capturing subtle features that category-level estimation overlooks. These statistics are fused through linear interpolation to form a robust distribution for novel categories, ensuring both diversity and representativeness in generated samples. Additionally, HGDE employs refined estimation techniques, such as weighted summation for mean calculation and principal component retention for covariance, to further improve accuracy. Empirical evaluations on four FSL benchmarks, including Mini-ImageNet, Tiered-ImageNet, CUB and CIFAR-FS, demonstrate that HGDE offers effective distribution estimation capabilities and leads to notable accuracy gains, with improvements of more than 1.8% in 1-shot tasks on CUB. These results highlight HGDE's ability to balance mean precision and variance diversity, making it a versatile and effective solution for FSL.

Recent advancements in multimodal large models have significantly bridged the representation gap between diverse modalities, catalyzing the evolution of video multimodal interpretation, which enhances users' understanding of video content by generating correlated modalities. However, most existing video multimodal interpretation methods primarily concentrate on global comprehension with limited user interaction. To address this, we propose a novel task, Controllable Video Segmentation and Captioning (SegCaptioning), which empowers users to provide specific prompts, such as a bounding box around an object of interest, to simultaneously generate correlated masks and captions that precisely embody user intent. An innovative framework, Scene Graph-guided Fine-grained SegCaptioning Transformer (SG-FSCFormer), is designed to integrate a Prompt-guided Temporal Graph Former to effectively capture and represent user intent through an adaptive prompt adaptor, ensuring that the generated content aligns well with the user's requirements. Furthermore, our model introduces a Fine-grained Mask-linguistic Decoder to collaboratively predict high-quality caption-mask pairs using a Multi-entity Contrastive loss, while providing fine-grained alignment between each mask and its corresponding caption tokens, thereby enhancing the user's comprehension of videos. Comprehensive experiments conducted on two benchmark datasets demonstrate that SG-FSCFormer achieves remarkable performance, effectively capturing user intent and generating precise multimodal outputs tailored to user specifications. Our code is available at https://github.com/XuZhang1211/SG-FSCFormer.

We propose a novel Transformer-based image-to-text generation model termed asACFthat adaptively clusters vision patches into object regions and language words into phrases to implicitly learn object-phrase alignments for better visual-text coherence. To achieve this, we design a novel self-attention layer that applies self-attention over the elements in a local cluster window instead of the whole sequence. The window size is softly decided by a clustering matrix that is calculated by the current input data and thus this process is adaptive. By stacking these revised self-attention layers to construct ACF, the small clusters in the lower layers can be grouped into a bigger cluster, e.g., vision/language. ACF clusters small objects/phrases into bigger ones. In this gradual clustering process, a parsing tree is generated which embeds the hierarchical knowledge of the input sequence. As a result, by using ACF to build the vision encoder and language decoder, the hierarchical object-phrase alignments are embedded and then transferred from vision to language domains in two popular image-to-text tasks: Image captioning and Visual Question Answering. The experiment results demonstrate the effectiveness of ACF, which outperforms most SOTA captioning and VQA models and achieves comparable scores compared with some large-scale pre-trained models.

Controllable image semantic understanding tasks, such as captioning or segmentation, necessitate users to input a prompt (e.g., text or bounding boxes) to predict a unique outcome, presenting challenges such as high-cost prompt input or limited information output. This paper introduces a new task ``Image Collaborative Segmentation and Captioning'' (SegCaptioning), which aims to translate a straightforward prompt, like a bounding box around an object, into diverse semantic interpretations represented by (caption, masks) pairs, allowing flexible result selection by users. This task poses significant challenges, including accurately capturing a user's intention from a minimal prompt while simultaneously predicting multiple semantically aligned caption words and masks. Technically, we propose a novel Scene Graph Guided Diffusion Model that leverages structured scene graph features for correlated mask-caption prediction. Initially, we introduce a Prompt-Centric Scene Graph Adaptor to map a user's prompt to a scene graph, effectively capturing his intention. Subsequently, we employ a diffusion process incorporating a Scene Graph Guided Bimodal Transformer to predict correlated caption-mask pairs by uncovering intricate correlations between them. To ensure accurate alignment, we design a Multi-Entities Contrastive Learning loss to explicitly align visual and textual entities by considering inter-modal similarity, resulting in well-aligned caption-mask pairs. Extensive experiments conducted on two datasets demonstrate that SGDiff achieves superior performance in SegCaptioning, yielding promising results for both captioning and segmentation tasks with minimal prompt input.

Recent endeavors in Multimodal Large Language Models (MLLMs) aim to unify visual comprehension and generation by combining LLM and diffusion models, the state-of-the-art in each task, respectively. Existing approaches rely on spatial visual tokens, where image patches are encoded and arranged according to a spatial order (e.g., raster scan). However, we show that spatial tokens lack the recursive structure inherent to languages, hence form an impossible language for LLM to master. In this paper, we build a proper visual language by leveraging diffusion timesteps to learn discrete, recursive visual tokens. Our proposed tokens recursively compensate for the progressive attribute loss in noisy images as timesteps increase, enabling the diffusion model to reconstruct the original image at any timestep. This approach allows us to effectively integrate the strengths of LLMs in autoregressive reasoning and diffusion models in precise image generation, achieving seamless multimodal comprehension and generation within a unified framework. Extensive experiments show that we achieve a new SOTA for multimodal comprehension and generation simultaneously compared with other MLLMs. Project Page: https://DDT-LLaMA.github.io/.

Training diffusion models is always a computation-intensive task. In this paper, we introduce a novel speed-up method for diffusion model training, called SpeeD, which is based on a closer look at time steps. Our key findings are: i) Time steps can be empirically divided into acceleration, deceleration, and convergence areas based on the process increment. ii) These time steps are imbalanced, with many concentrated in the convergence area. iii) The concentrated steps provide limited benefits for diffusion training. To address this, we design an asymmetric sampling strategy that reduces the frequency of steps from the convergence area while increasing the sampling probability in other areas. Additionally, we propose a weighting strategy to emphasize the importance of time steps with rapid-change process increments. As a plug-and-play and architecture-agnostic approach, SpeeD consistently achieves 3 × acceleration across various diffusion architectures, datasets, and tasks. Notably, due to its simple design, our approach significantly reduces the cost of diffusion model training with minimal overhead. Our research enables more researchers to train diffusion models at a lower cost.

The latest emerged 4D Panoptic Scene Graph (4D-PSG) provides an advanced-ever representation for comprehensively modeling the dynamic 4D visual real world. Unfortunately, current pioneering 4D-PSG research can primarily suffer from data scarcity issues severely, as well as the resulting out-of-vocabulary problems; also, the pipeline nature of the benchmark generation method can lead to suboptimal performance. To address these challenges, this paper investigates a novel framework for 4D-PSG generation that leverages rich 2D visual scene annotations to enhance 4D scene learning. First, we introduce a 4D Large Language Model (4D-LLM) integrated with a 3D mask decoder for end-to-end generation of 4D-PSG. A chained SG inference mechanism is further designed to exploit LLMs’ open-vocabulary capabilities to infer accurate and comprehensive object and relation labels iteratively. Most importantly, we propose a 2D-to-4D visual scene transfer learning framework, where a spatial-temporal scene transcending strategy effectively transfers dimension-invariant features from abundant 2D SG annotations to 4D scenes, effectively compensating for data scarcity in 4D-PSG. Extensive experiments on the benchmark data demonstrate that we strikingly outperform baseline models by a large margin, highlighting the effectiveness of our method. The project page is https://sqwu.top/PSG-4D-LLM/.

Instruction-based image editing aims to modify specific image elements with natural language instructions. However, current models in this domain often struggle to execute complex user instructions accurately, as they are trained on low-quality data with limited editing types. We present AnyEdit, a comprehensive multi-modal instruction editing dataset, comprising 2.5 million high-quality editing pairs spanning over 20 editing types and five domains. We ensure the diversity and quality of the AnyEdit collection through three aspects: initial data diversity, adaptive editing process, and automated selection of editing results. Using the dataset, we further train a novel AnyEdit Stable Diffusion with task-aware routing and learnable task embedding for unified image editing. Comprehensive experiments on three benchmark datasets show that AnyEdit consistently boosts the performance of diffusion-based editing models. This presents prospects for developing instruction-driven image editing models that support human creativity.

Recently, large efforts have been made to design efficient linear-complexity visual Transformers. However, current linear attention models are generally unsuitable to be deployed in resource-constrained mobile devices, due to suffering from either few efficiency gains or significant accuracy drops. In this paper, we propose a new deCoupled duAl-interactive lineaR attEntion (CARE) mechanism, revealing that features’ decoupling and interaction can fully unleash the power of linear attention. We first propose an asymmetrical feature decoupling strategy that asymmetrically decouples the learning process for local inductive bias and long-range dependencies, thereby preserving sufficient local and global information while effectively enhancing the efficiency of models. Then, a dynamic memory unit is employed to maintain critical information along the network pipeline. Moreover, we design a dual interaction module to effectively facilitate interaction between local inductive bias and long-range information as well as among features at different layers. By adopting a decoupled learning way and fully exploiting complementarity across features, our method can achieve both high efficiency and accuracy. Extensive experiments on ImageNet-1K, COCO, and ADE20K datasets demonstrate the effectiveness of our approach, e.g., achieving 78.4/82.1% top-1 accuracy on ImagegNet-1K at the cost of only 0.7/1.9 GMACs. Codes will be released on github.

While image-text foundation models have succeeded across diverse downstream tasks, they still face challenges in the presence of spurious correlations between the input and label. To address this issue, we propose a simple three-step approach-Project-Probe-Aggregate (PPA)-that enables parameter-efficient fine-tuning for foundation models without relying on group annotations. Building upon the failure-based debiasing scheme, our method, PPA, improves its two key components: minority samples identification and the robust training algorithm. Specifically, we first train biased classifiers by projecting image features onto the nullspace of class proxies from text encoders. Next, we infer group labels using the biased classifier and probe group targets with prior correction. Finally, we aggregate group weights of each class to produce the debiased classifier Our theoretical analysis shows that our PPA enhances minority group identification and is Bayes optimal for minimizing the balanced group error, mitigating spurious correlations. Extensive experimental results confirm the effectiveness of our PPA: it outperforms the state-of-the-art by an average worst-group accuracy while requiring less than 0.01% tunable parameters without training group labels.

Recent advancements in multimodal large language models (MLLMs) have demonstrated exceptional performance in multimodal perception and understanding. However, leading open-source MLLMs exhibit significant limitations in complex and structured reasoning, particularly in tasks requiring deep reasoning for decision-making and problem-solving. In this work, we present Corvid, an MLLM with enhanced chain-of-thought (CoT) reasoning capabilities. Architecturally, Corvid incorporates a hybrid vision encoder for informative visual representation and a meticulously designed connector (GateMixer) to facilitate cross-modal alignment. To enhance Corvid's CoT reasoning capabilities, we introduce MCoT-Instruct-287K, a high-quality multimodal CoT instruction-following dataset, refined and standardized from diverse public reasoning sources. Leveraging this dataset, we fine-tune Corvid with a two-stage CoT-formatted training approach to progressively enhance its step-by-step reasoning abilities. Furthermore, we propose an effective inference-time scaling strategy that enables Corvid to mitigate over-reasoning and under-reasoning through self-verification. Extensive experiments demonstrate that Corvid outperforms existing o1-like MLLMs and state-of-the-art MLLMs with similar parameter scales, with notable strengths in mathematical reasoning and science problem-solving. Project page: https://mm-vl.github.io/corvid.

Triangle meshes are fundamental to 3D applications, enabling efficient modification and rasterization while maintaining compatibility with standard rendering pipelines. However, current automatic mesh generation methods typically rely on intermediate representations that lack the continuous surface quality inherent to meshes. Converting these representations into meshes produces dense, suboptimal outputs. Although recent autoregressive approaches demonstrate promise in directly modeling mesh vertices and faces, they are constrained by the limitation in face count, scalability, and structural fidelity. To address these challenges, we propose Nautilus, a locality-aware autoencoder for artist-like mesh generation that leverages the local properties of manifold meshes to achieve structural fidelity and efficient representation. Our approach introduces a novel tokenization algorithm that preserves face proximity relationships and compresses sequence length through locally shared vertices and edges, enabling the generation of meshes with an unprecedented scale of up to 5,000 faces. Furthermore, we develop a Dual-stream Point Conditioner that provides multi-scale geometric guidance, ensuring global consistency and local structural fidelity by capturing fine-grained geometric features. Extensive experiments demonstrate that Nautilus significantly outperforms state-of-the-art methods in both fidelity and scalability. The project page is at https://nautilusmeshgen.github.io.

With the increasing attention to pre-trained vision-language models (VLMs), \eg, CLIP, substantial efforts have been devoted to many downstream tasks, especially in test-time adaptation (TTA). However, previous works focus on learning prototypes only in the textual modality while overlooking the ambiguous semantics in class names. These ambiguities lead to textual prototypes that are insufficient to capture visual concepts, resulting in limited performance. To address this issue, we introduce \textbf{ProtoMM}, a training-free framework that constructs multimodal prototypes to adapt VLMs during the test time. By viewing the prototype as a discrete distribution over the textual descriptions and visual particles, ProtoMM has the ability to combine the multimodal features for comprehensive prototype learning. More importantly, the visual particles are dynamically updated as the testing stream flows. This allows our multimodal prototypes to continually learn from the data, enhancing their generalizability in unseen scenarios. In addition, we quantify the importance of the prototypes and test images by formulating their semantic distance as an optimal transport problem. Extensive experiments on 15 zero-shot benchmarks demonstrate the effectiveness of our method, achieving a 1.03\% average accuracy improvement over state-of-the-art methods on ImageNet and its variant datasets.

Diffusion models (DMs) have achieved state-of-the-art generative performance but suffer from high sampling latency due to their sequential denoising nature. Existing solver-based acceleration methods often face image quality degradation under a low-latency budget. In this paper, we propose the Ensemble Parallel Direction solver (dubbed as \ours), a novel ODE solver that mitigates truncation errors by incorporating multiple parallel gradient evaluations in each ODE step. Importantly, since the additional gradient computations are independent, they can be fully parallelized, preserving low-latency sampling. Our method optimizes a small set of learnable parameters in a distillation fashion, ensuring minimal training overhead. In addition, our method can serve as a plugin to improve existing ODE samplers. Extensive experiments on various image synthesis benchmarks demonstrate the effectiveness of our \ours~in achieving high-quality and low-latency sampling. For example, at the same latency level of 5 NFE, EPD achieves an FID of 4.47 on CIFAR-10, 7.97 on FFHQ, 8.17 on ImageNet, and 8.26 on LSUN Bedroom, surpassing existing learning-based solvers by a significant margin. Codes are available in https://github.com/BeierZhu/EPD.

Multimodal Large Language Models (MLLMs) have demonstrated exceptional capabilities in processing vision-language tasks. One of the crux of MLLMs lies in vision tokenization, which involves efficiently transforming input visual signals into feature representations that are most beneficial for LLMs. However, existing vision tokenizers, essential for semantic alignment between vision and language, remain problematic. Existing methods aggressively fragment visual input, corrupting the visual semantic integrity. To address this, this paper proposes a novel dynamic Semantic-Equivalent Vision Tokenizer (SeTok), which groups visual features into semantic units via a dynamic clustering algorithm, flexibly determining the number of tokens based on image complexity. The resulting vision tokens effectively preserve semantic integrity and capture both low-frequency and high-frequency visual features. The proposed MLLM (Setokim) equipped with SeTok significantly demonstrates superior performance across various tasks, as evidenced by our experimental results.

The Multimodal Large Language Model (MLLM) is currently experiencing rapid growth, driven by the advanced capabilities of language-based LLMs. Unlike their specialist predecessors, existing MLLMs are evolving towards a Multimodal Generalist paradigm. Initially limited to understanding multiple modalities, these models have advanced to not only comprehend but also generate across modalities. Their capabilities have expanded from coarse-grained to fine-grained multimodal understanding and from supporting singular modalities to accommodating a wide array of or even arbitrary modalities. To assess the capabilities of various MLLMs, a diverse array of benchmark test sets has been proposed. This leads to a critical question: Can we simply assume that higher performance across tasks indicates a stronger MLLM capability, bringing us closer to human-level AI? We argue that the answer is not as straightforward as it seems. In this project, we introduce an evaluation framework to delineate the capabilities and behaviors of current multimodal generalists. This framework, named General-Level, establishes 5-scale levels of MLLM performance and generality, offering a methodology to compare MLLMs and gauge the progress of existing systems towards more robust multimodal generalists and, ultimately, towards AGI (Artificial General Intelligence). Central to our framework is the use of Synergy as the evaluative criterion, categorizing capabilities based on whether MLLMs preserve synergy across comprehension and generation, as well as across multimodal interactions. To evaluate the comprehensive abilities of various generalists, we present a massive multimodal benchmark, General-Bench, which encompasses a broader spectrum of skills, modalities, formats, and capabilities, including over 700 tasks and 325,800 instances. The evaluation results that involve over 100 existing state-of-the-art MLLMs uncover the capability rankings of generalists, highlighting the challenges in reaching genuine AI. We expect this project to pave the way for future research on next-generation multimodal foundation models, providing a robust infrastructure to accelerate the realization of AGI. Project Page: https://generalist.top/, Leaderboard: https://generalist.top/leaderboard/, Benchmark: https://huggingface.co/General-Level/.

With the advance of diffusion models, today's video generation has achieved impressive quality. To extend the generation length and facilitate real-world applications, a majority of video diffusion models (VDMs) generate videos in an autoregressive manner, i.e., generating subsequent clips conditioned on the last frame(s) of the previous clip. However, existing autoregressive VDMs are highly inefficient and redundant: The model must re-compute all the conditional frames that are overlapped between adjacent clips. This issue is exacerbated when the conditional frames are extended autoregressively to provide the model with long-term context. In such cases, the computational demands increase significantly (i.e., with a quadratic complexity w.r.t. the autoregression step). In this paper, we propose **Ca2-VDM**, an efficient autoregressive VDM with **Ca**usal generation and **Ca**che sharing. For **causal generation**, it introduces unidirectional feature computation, which ensures that the cache of conditional frames can be precomputed in previous autoregression steps and reused in every subsequent step, eliminating redundant computations. For **cache sharing**, it shares the cache across all denoising steps to avoid the huge cache storage cost. Extensive experiments demonstrated that our Ca2-VDM achieves state-of-the-art quantitative and qualitative video generation results and significantly improves the generation speed. Code is available: https://github.com/Dawn-LX/CausalCache-VDM

Large vision-language models (LVLMs) have significantly advanced numerous fields. In this work, we explore how to harness their potential to address 3D scene understanding tasks, using 3D question answering (3D-QA) as a representative example. Due to the limited training data in 3D, we do not train LVLMs but infer in a zero-shot manner. Specifically, we sample 2D views from a 3D point cloud and feed them into 2D models to answer a given question. When the 2D model is chosen, e.g., LLAVA-OV, the quality of sampled views matters the most. We propose cdViews, a novel approach to automatically selecting critical and diverse Views for 3D-QA. cdViews consists of two key components: viewSelector prioritizing critical views based on their potential to provide answer-specific information, and viewNMS enhancing diversity by removing redundant views based on spatial overlap. We evaluate cdViews on the widely-used ScanQA and SQA benchmarks, demonstrating that it achieves state-of-the-art performance in 3D-QA while relying solely on 2D models without fine-tuning. These findings support our belief that 2D LVLMs are currently the most effective alternative (of the resource-intensive 3D LVLMs) for addressing 3D tasks.

Recent advancements in image generation have enabled the creation of high-quality images from diverse conditions, such as text and images. However, existing methods struggle to balance multiple conditions effectively, typically favoring certain modalities over others. To address this challenge, we introduce EMMA, a novel multi-modal image generation model that integrates both text and additional modalities, which we refer to as "text + X", to guide the generation of images. EMMA incorporates an innovative Multi-modal Feature Connector design, which effectively integrates textual and supplementary modal information through a special attention mechanism. Additionally, we propose a method for the assembly of existing modules to produce images conditioned on multiple modalities at the same time, eliminating the need for additional training. This modular nature also facilitates easy adaptation to different existing frameworks, making EMMA a flexible and effective tool for producing personalized and context-aware images. Extensive experiments demonstrate the effectiveness of EMMA in maintaining high fidelity and detail in generated images, showcasing its potential as a robust solution for advanced multi-modal conditional image generation tasks.

Multi-modal large language models have demonstrated impressive performances on most vision-language tasks. However, the model generally lacks the understanding capabilities for specific domain data, particularly when it comes to interpreting chart figures. This is mainly due to the lack of relevant multi-modal instruction tuning datasets. In this article, we create a high-quality instruction-tuning dataset leveraging GPT-4. We develop a multi-step data generation process in which different steps are responsible for generating tabular data, creating chart figures, and designing instruction tuning data separately. Our method’s flexibility enables us to generate diverse, high-quality instruction-tuning data consistently and efficiently while maintaining a low resource expenditure. Additionally, it allows us to incorporate a wider variety of chart and task types not yet featured in existing datasets. Next, we introduce ChartLlama, a multi-modal large language model that we’ve trained using our created dataset. ChartLlama outperforms all prior methods in ChartQA, Chart-to-text, and Chart-extraction evaluation benchmarks. Additionally, ChartLlama significantly improves upon the baseline in our specially compiled chart dataset, which includes new chart and task types. The results of ChartLlama confirm the value and huge potential of our proposed data generation method in enhancing chart comprehension.

This workshop aims to explore the potential of large generative models to revolutionize the way we interact with multimodal information. A Large Language Model (LLM) represents a sophisticated form of artificial intelligence engineered to comprehend and produce natural language text, exemplified by technologies such as GPT, LLaMA, Flan-T5, ChatGLM, and Qwen, etc. These models undergo training on extensive text datasets, exhibiting commendable attributes including robust language generation, zeroshot transfer capabilities, and In-Context Learning (ICL). With the surge in multimodal content-encompassing images, videos, audio, and 3D models-over the recent period, Large MultiModal Models (LMMs) have seen significant enhancements. These improvements enable the augmentation of conventional LLMs to accommodate multimodal inputs or outputs, as seen in BLIP, Flamingo, KOSMOS, LLaVA, Gemini, GPT-4, etc. Concurrently, certain research initiatives have delved into generating specific modalities, with Kosmos2 and MiniGPT-5 focusing on image generation, and SpeechGPT on speech production. There are also endeavors to integrate LLMs with external tools to achieve a near 'any-to-any' multimodal comprehension and generation capacity, illustrated by projects like Visual-ChatGPT, ViperGPT, MMREACT, HuggingGPT, and AudioGPT. Collectively, these models, spanning not only text and image generation but also other modalities, are referred to as large generative models. This workshop will provide an opportunity for researchers, practitioners, and industry professionals to explore the latest trends and best practices in the field of multimodal applications of large generative models. We also remark that the submissions are not limited to the use of such models. The workshop will also focus on exploring the challenges and opportunities of integrating large language models with other AI technologies such as computer vision and speech recognition. Additionally, the workshop will provide a platform for participants to present their research, share their experiences, and discuss potential collaborations.

This paper presents a pioneering exploration of reinforcement learning (RL) via group relative policy optimization for unified multimodal large language models (ULMs), aimed at simultaneously reinforcing generation and understanding capabilities. Through systematic pilot studies, we uncover the significant potential of ULMs to enable the synergistic co-evolution of dual capabilities within a shared policy optimization framework. Building on this insight, we introduce \textbf{CoRL}, a \textbf{Co}-\textbf{R}einforcement \textbf{L}earning framework comprising a unified RL stage for joint optimization and a refined RL stage for task-specific enhancement. With the proposed CoRL, our resulting model, \textbf{ULM-R1}, achieves average improvements of 7\% on three text-to-image generation datasets and 23\% on nine multimodal understanding benchmarks. These results demonstrate the effectiveness of CoRL and highlight the substantial benefits of reinforcement learning in facilitating cross-task synergy and optimization for ULMs. Code is available at \url{https://github.com/mm-vl/ULM-R1}.

With the continuous development of large language models and reasoning chain technologies, the potential of deep reasoning based on reinforcement learning has shown remarkable promise in multi-task scenarios. However, existing unified models have yet to achieve end-to-end integration in image generation and understanding tasks, limiting the model’s self-reflection ability and the realization of cross-modal reasoning chains. To address this, we propose Vinic, a novel framework designed to enable interleaved image generation and understanding through deep reasoning capabilities. We leverage a small amount of multimodal chain-of-thought (MCoT) data for cold-start and employ reinforcement learning to guide the integration of image generation and understanding tasks. Additionally, we introduce a momentum-based reward function, which dynamically adjusts the reward distribution by considering historical improvements, ensuring the stability of the model across multiple generations. Experimental results demonstrate that integrating MCoT can achieve a +22% improvement over the base model on Geneval, effectively enhancing both image generation quality and instruction alignment capabilities.

Reinforcement learning (RL) has become an indispensable post-training step for unlocking the full potential of Large Language Models (LLMs). Its core motivation is to incentivize the model’s inference trajectory via a reward model, effectively balancing the exploration–exploitation trade-off in scenarios where collecting exhaustive input–output ground-truth pairs is infeasible. This motivation naturally extends to visual generation, where perfect alignment between an image and a textual prompt is inherently ambiguous and often unattainable. However, existing visual generative models are not yet ready for RL due to the following two fundamental drawbacks that undermine the foundations of RL: 1) For diffusion-based models, the actual generation trajectories of sampled images cannot be reliably rewarded, as diffusion inversion is notoriously difficult. 2) For autoregressive (AR) models, we show that the widely used spatial visual tokens do not satisfy the Bellman equation and thus violate the policy improvement theorem of RL. To this end, we propose to use Selftok (Self-consistency Tokenizer), which represents each image as a sequential 1D stream of discrete, autoregressive tokens. Together with language, we train a pure AR vision-language model (VLM) for visual generation. Impressively, without using any text-image training pairs, a simple policy gradient algorithm applied to Selftok tokens significantly boosts visual generation performance, surpassing existing models by a large margin. Implementation details are provided in the Appendix.

Vision-language models (VLMs) such as CLIP demonstrate strong generalization in zero-shot classification but remain highly vulnerable to adversarial perturbations. Existing methods primarily focus on adversarial fine-tuning or prompt optimization, they often overlook the gaps in CLIP’s encoded features, which is shown as the text and image features lie far apart from each other. This misalignment is significantly amplified under adversarial perturbations, leading to severe degradation in classification performance. To address this problem, we propose **C**r**O**ss-moda**L**ity **A**lignment, dubbed **COLA**, an optimal transport-based framework that explicitly addresses adversarial misalignment by restoring both global image-text alignment and local structural consistency in the feature space. (1) COLA first projects adversarial image embeddings onto a subspace spanned by class text features, effectively filtering out non-semantic distortions while preserving discriminative information. (2) It then models images and texts as discrete distributions over multiple augmented views and refines their alignment via OT, with the subspace projection seamlessly integrated into the cost computation. This design ensures stable cross-modal alignment even under adversarial conditions. COLA is training-free and compatible with existing fine-tuned models. Extensive evaluations across 14 zero-shot classification benchmarks demonstrate the effectiveness of COLA, especially with an average improvement of 6.7% on ImageNet and its variants under PGD adversarial attacks, while maintaining high accuracy on clean samples.

Large Language Models (LLMs) have evolved into Multimodal Large Language Models (MLLMs), significantly enhancing their capabilities by integrating visual information and other types, thus aligning more closely with the nature of human intelligence, which processes a variety of data forms beyond just text. Despite advancements, the undesirable generation of these models remains a critical concern, particularly due to vulnerabilities exposed by text-based jailbreak attacks, which have represented a significant threat by challenging existing safety protocols. Motivated by the unique security risks posed by the integration of new and old modalities for MLLMs, we propose a unified multimodal universal jailbreak attack framework that leverages iterative image-text interactions and transfer-based strategy to generate a universal adversarial suffix and image. Our work not only highlights the interaction of image-text modalities can be used as a critical vulnerability but also validates that multimodal universal jailbreak attacks can bring higher-quality undesirable generations across different MLLMs. We evaluate the undesirable context generation of MLLMs like LLaVA, Yi-VL, MiniGPT4, MiniGPT-v2, and InstructBLIP, and reveal significant multimodal safety alignment issues, highlighting the inadequacy of current safety mechanisms against sophisticated multimodal attacks. This study underscores the urgent need for robust safety measures in MLLMs, advocating for a comprehensive review and enhancement of security protocols to mitigate potential risks associated with multimodal capabilities.

The integration of visual and textual data in Vision-Language Pre-training (VLP) models is crucial for enhancing vision-language understanding. However, the adversarial robustness of these models, especially in the alignment of image-text features, has not yet been sufficiently explored. In this paper, we introduce a novel gradient-based multimodal adversarial attack method, underpinned by contrastive learning, to improve the transferability of multimodal adversarial samples in VLP models. This method concurrently generates adversarial texts and images within imperceptive perturbation, employing both image-text and intra-modal contrastive loss. We evaluate the effectiveness of our approach on image-text retrieval and visual entailment tasks, using publicly available datasets in a black-box setting. Extensive experiments indicate a significant advancement over existing single-modal transfer-based adversarial attack methods and current multimodal adversarial attack approaches.

Learning correspondences aims to find correct correspondences (inliers) from the initial correspondence set with an uneven correspondence distribution and a low inlier rate, which can be regarded as graph data. Recent advances usually use graph neural networks (GNNs) to build a single type of graph or simply stack local graphs into the global one to complete the task. But they ignore the complementary relationship between different types of graphs, which can effectively capture potential relationships among sparse correspondences. To address this problem, we propose MGNet to effectively combine multiple complementary graphs. To obtain information integrating implicit and explicit local graphs, we construct local graphs from implicit and explicit aspects and combine them effectively, which is used to build a global graph. Moreover, we propose Graph Soft Degree Attention (GSDA) to make full use of all sparse correspondence information at once in the global graph, which can capture and amplify discriminative features. Extensive experiments demonstrate that MGNet outperforms state-of-the-art methods in different visual tasks. The code is provided in https://github.com/DAILUANYUAN/MGNet-2024AAAI.

Semi-supervised 3D object detection is a promising yet under-explored direction to reduce data annotation costs, especially for cluttered indoor scenes. A few prior works, such as SESS and 3DIoUMatch, attempt to solve this task by utilizing a teacher model to generate pseudo-labels for unlabeled samples. However, the availability of unlabeled samples in the 3D domain is relatively limited compared to its 2D counterpart due to the greater effort required to collect 3D data. Moreover, the loose consistency regularization in SESS and restricted pseudo-label selection strategy in 3DIoUMatch lead to either low-quality supervision or a limited amount of pseudo labels. To address these issues, we present a novel Dual-Perspective Knowledge Enrichment approach named DPKE for semi-supervised 3D object detection. Our DPKE enriches the knowledge of limited training data, particularly unlabeled data, from two perspectives: data-perspective and feature-perspective. Specifically, from the data-perspective, we propose a class-probabilistic data augmentation method that augments the input data with additional instances based on the varying distribution of class probabilities. Our DPKE achieves feature-perspective knowledge enrichment by designing a geometry-aware feature matching method that regularizes feature-level similarity between object proposals from the student and teacher models. Extensive experiments on the two benchmark datasets demonstrate that our DPKE achieves superior performance over existing state-of-the-art approaches under various label ratio conditions. The source code and models will be made available to the public.

Text-to-video (T2V) synthesis has gained increasing attention in the community, in which the recently emerged diffusion models (DMs) have promisingly shown stronger performance than the past approaches. While existing state-of-the-art DMs are competent to achieve high-resolution video generation, they may largely suffer from key limitations (e.g., action occurrence disorders, crude video motions) with respect to the intricate temporal dynamics modeling, one of the crux of video synthesis. In this work, we investigate strengthening the awareness of video dynamics for DMs, for high-quality T2V generation. Inspired by human intuition, we design an innovative dynamic scene manager (dubbed as Dysen) module, which includes (step-1) extracting from input text the key actions with proper time-order arrangement, (step-2) transforming the action schedules into the dynamic scene graph (DSG) representations, and (step-3) enriching the scenes in the DSG with sufficient and reasonable details. Taking advantage of the existing powerful LLMs (e.g., ChatGPT) via in-context learning, Dysen realizes (nearly) human-level temporal dynamics understanding. Finally, the resulting video DSG with rich action scene details is encoded as fine-grained spatio-temporal features, integrated into the backbone T2V DM for video generating. Experiments on popular T2V datasets suggest that our Dysen- VDM consistently outperforms prior arts with significant margins, especially in scenarios with complex actions. Codes at http://haofei.vip/Dysen-VDM/.

In this paper, we present an empirical study on image recognition unfairness, i.e., extreme class accuracy disparity on balanced data like ImageNet. We demonstrate that classes are not equal and unfairness is prevalent for image classification models across various datasets, network architectures, and model capacities. Moreover, several intriguing properties of fairness are identified. First, the unfairness lies in problematic representation rather than classifier bias distinguished from long-tailed recognition. Second, with the proposed concept of Model Prediction Bias, we investigate the origins of problematic representation during training optimization. Our findings reveal that models tend to exhibit greater prediction biases for classes that are more challenging to recognize. It means that more other classes will be confused with harder classes. Then the False Positives (FPs) will dominate the learning in optimization, thus leading to their poor accuracy. Further, we conclude that data augmentation and representation learning algorithms improve overall performance by promoting fairness to some degree in image classification.

Despite advancements in text-to-image generation (T2I), prior methods often face text-image misalignment problems such as relation confusion in generated images. Existing solutions involve cross-attention manipulation for better compositional understanding or integrating large language models for improved layout planning. However, the inherent alignment capabilities of T2I models are still inadequate. By reviewing the link between generative and discriminative modeling, we posit that T2I models' discriminative abilities may reflect their text-image alignment proficiency during generation. In this light, we advocate bolstering the discriminative abilities of T2I models to achieve more precise text-to-image alignment for generation. We present a discriminative adapter built on T2I models to probe their discriminative abilities on two representative tasks and leverage discriminative fine-tuning to improve their text-image alignment. As a bonus of the discriminative adapter, a self-correction mechanism can leverage discriminative gradients to better align generated images to text prompts during inference. Comprehensive evaluations across three benchmark datasets, including both in-distribution and out-of-distribution scenarios, demonstrate our method's superior generation performance. Meanwhile, it achieves state-of-the-art discriminative performance on the two discriminative tasks compared to other generative models. The code is available at https://dpt-t2i.github.io/.

We study text-based image editing (TBIE) of a single image by counterfactual inference because it is an elegant formulation to precisely address the requirement: the edited image should retain the fidelity of the original one. Through the lens of the formulation, we find that the crux of TBIE is that existing techniques hardly achieve a good tradeoff between editability and fidelity, mainly due to the overfitting of the single-image fine-tuning. To this end, we propose a Doubly Abductive Counterfactual inference framework (DAC). We first parameterize an exogenous variable as a UNet LoRA, whose abduction can encode all the image details. Second, we abduct another exogenous variable parameterized by a text encoder LoRA, which recovers the lost editability caused by the overfitted first abduction. Thanks to the second abduction, which exclusively encodes the visual transition from post-edit to pre-edit, its inversion-subtracting the LoRA-effectively reverts pre-edit back to post-edit, thereby accomplishing the edit. Through extensive experiments, our DAC achieves a good tradeoff between editability and fidelity. Thus, we can support a wide spectrum of user editing intents, including addition, removal, manipulation, replacement, style transfer, and facial change, which are extensively validated in both qualitative and quantitative evaluations. Codes are in https://github.com/xuesong39/DAC.

Generative AI has made significant strides in computer vision, particularly in text-driven image/video synthesis (T2I/T2V). Despite the notable advancements, it remains challenging in human-centric content synthesis such as realistic dance generation. Current methodologies, primarily tailored for human motion transfer, encounter difficulties when confronted with real-world dance scenarios (e.g., social media dance), which require to generalize across a wide spectrum of poses and intricate human details. In this paper, we depart from the traditional paradigm of human motion transfer and emphasize two additional critical attributes for the synthesis of human dance content in social media contexts: (i) Generalizability: the model should be able to generalize beyond generic human viewpoints as well as unseen human subjects, backgrounds, and poses; (ii) Compositionality: it should allow for the seamless composition of seen/unseen subjects, backgrounds, and poses from different sources. To address these challenges, we introduce Disco, which includes a novel model architecture with disentangled control to improve the compositionality of dance synthesis, and an effective human attribute pre-training for better generalizability to unseen humans. Extensive qualitative and quantitative results demonstrate that DISCO can generate high-quality human dance images and videos with diverse appearances and flexible motions. Code is available at https://disco-dance.github.io/.

Score distillation sampling (SDS) and its variants have greatly boosted the development of text-to-3D generation, but are vulnerable to geometry collapse and poor textures yet. To solve this issue, we first deeply analyze the SDS and find that its distillation sampling process indeed corresponds to the trajectory sampling of a stochastic differential equation (SDE): SDS samples along an SDE trajectory to yield a less noisy sample which then serves as a guidance to optimize a 3D model. However, the randomness in SDE sampling often leads to a diverse and unpredictable sample which is not always less noisy, and thus is not a consistently correct guidance, explaining the vulnerability of SDS. Since for any SDE, there always exists an ordinary differential equation (ODE) whose trajectory sampling can deterministically and consistently converge to the desired target point as the SDE, we propose a novel and effective “Consistent3D” method that explores the ODE deterministic sampling prior for text-to-3D generation. Specifically, at each training iteration, given a rendered image by a 3D model, we first estimate its desired 3D score function by a pre-trained 2D diffusion model, and build an ODE for trajectory sampling. Next, we design a consistency distillation sampling loss which samples along the ODE trajectory to generate two adjacent samples and uses the less noisy sample to guide another more noisy one for distilling the deterministic prior into the 3D model. Experimental results show the efficacy of our Consistent3D in generating high-fidelity and diverse 3D objects and large-scale scenes, as shown in Fig. 1. The codes are available at https://github.com/sail-sg/Consistent3D.

Score distillation sampling (SDS) has been widely adopted to overcome the absence of unseen views in reconstructing 3D objects from a single image. It leverages pretrained 2D diffusion models as teacher to guide the reconstruction of student 3D models. Despite their remarkable success, SDS-based methods often encounter geometric artifacts and texture saturation. We find out the crux is the overlooked indiscriminate treatment of diffusion time-steps during optimization: it unreasonably treats the student-teacher knowledge distillation to be equal at all time-steps and thus entangles coarse-grained and fine-grained modeling. Therefore, we propose the Diffusion Time-step Curriculum one-image-to-3D pipeline (DTC123), which involves both the teacher and student models collaborating with the time-step curriculum in a coarse-to-fine manner. Extensive experiments on NeRF4, RealFusion15, GSO and Level50 benchmark demonstrate that DTC123 can produce multiview consistent, high-quality, and diverse 3D assets. Codes and more generation demos will be released in https://github.com/yxymessi/DTC123.

Even when using large multi-modal foundation models, few-shot learning is still challenging-if there is no proper inductive bias, it is nearly impossible to keep the nuanced class attributes while removing the visually prominent attributes that spuriously correlate with class labels. To this end, we find an inductive bias that the time-steps of a Diffusion Model (DM) can isolate the nuanced class attributes, i.e., as the forward diffusion adds noise to an image at each time-step, nuanced attributes are usually lost at an earlier time-step than the spurious attributes that are visually prominent. Building on this, we propose Time-step Few-shot (TiF) learner. We train class-specific low-rank adapters for a text-conditioned DM to make up for the lost attributes, such that images can be accurately reconstructed from their noisy ones given a prompt. Hence, at a small time-step, the adapter and prompt are essentially a parameterization of only the nuanced class attributes. For a test image, we can use the parameterization to only extract the nuanced class attributes for classification. TiF learner significantly outperforms OpenCLIP and its adapters on a variety of fine-grained and customized few-shot learning tasks. Codes are in https://github.com/yue-zhongqi/tif.

Facial Attribute Classification (FAC) holds substantial promise in widespread applications. However, FAC models trained by traditional methodologies can be unfair by exhibiting accuracy inconsistencies across varied data sub-populations. This unfairness is largely attributed to bias in data, where some spurious attributes (e.g., Male) statistically correlate with the target attribute (e.g., Smiling). Most of existing fairness-aware methods rely on the labels of spurious attributes, which may be unavailable in practice. This work proposes a novel, generation-based two-stage framework to train a fair FAC model on biased data without additional annotation. Initially, we identify the potential spurious attributes based on generative models. Notably, it enhances interpretability by explicitly showing the spurious attributes in image space. Following this, for each image, we first edit the spurious attributes with a random degree sampled from a uniform distribution, while keeping target attribute unchanged. Then we train a fair FAC model by fostering model invariance to these augmentation. Extensive experiments on three common datasets demonstrate the effectiveness of our method in promoting fairness in FAC without compromising accuracy. Codes are in https://github.com/heqianpei/DiGA.

Recent advancements in Multimodal Large Language Models (MLLMs) have been utilizing Visual Prompt Generators (VPGs) to convert visual features into tokens that LLMs can recognize. This is achieved by training the VPGs on millions of image-caption pairs, where the VPG-generated tokens of images are fed into a frozen LLM to generate the corresponding captions. However, this image-captioning based training objective inherently biases the VPG to concentrate solely on the primary visual contents sufficient for caption generation, often neglecting other visual details. This shortcoming results in MLLMs’ underperformance in comprehending demonstrative instructions consisting of multiple, interleaved, and multimodal instructions that demonstrate the required context to complete a task. To address this issue, we introduce a generic and lightweight Visual Prompt Generator Complete module (VPG-C), which can infer and complete the missing details essential for comprehending demonstrative instructions. Further, we propose a synthetic discriminative training strategy to fine-tune VPG-C, eliminating the need for supervised demonstrative instructions. As for evaluation, we build DEMON, a comprehensive benchmark for demonstrative instruction understanding. Synthetically trained with the proposed strategy, VPG-C achieves significantly stronger zero-shot performance across all tasks of DEMON. Further evaluation on the MME and OwlEval benchmarks also demonstrate the superiority of VPG-C. The code and models are available at https://github.com/DCDmllm/Cheetah.

Representation learning is all about discovering the hidden modular attributes that generate the data faithfully. We explore the potential of Denoising Diffusion Probabilistic Model (DM) in unsupervised learning of the modular attributes. We build a theoretical framework that connects the diffusion time-steps and the hidden attributes, which serves as an effective inductive bias for unsupervised learning. Specifically, the forward diffusion process incrementally adds Gaussian noise to samples at each time-step, which essentially collapses different samples into similar ones by losing attributes, e.g., fine-grained attributes such as texture are lost with less noise added (i.e., early time-steps), while coarse-grained ones such as shape are lost by adding more noise (i.e., late time-steps). To disentangle the modular attributes, at each time-step t, we learn a t-specific feature to compensate for the newly lost attribute, and the set of all {1,...,t}-specific features, corresponding to the cumulative set of lost attributes, are trained to make up for the reconstruction error of a pre-trained DM at time-step t. On CelebA, FFHQ, and Bedroom datasets, the learned feature significantly improves attribute classification and enables faithful counterfactual generation, e.g., interpolating only one specified attribute between two images, validating the disentanglement quality. Codes are in https://github.com/yue-zhongqi/diti.

Existing research of video understanding still struggles to achieve in-depth comprehension and reasoning in complex videos, primarily due to the under-exploration of two key bottlenecks: fine-grained spatial-temporal perceptive understanding and cognitive-level video scene comprehension. This paper bridges the gap by presenting a novel solution. We first introduce a novel video Multimodal Large Language Model (MLLM), MotionEpic, which achieves fine-grained pixel-level spatial-temporal video grounding by integrating video spatial-temporal scene graph (STSG) representation. Building upon MotionEpic, we then develop a Video-of-Thought (VoT) reasoning framework. VoT inherits the Chain-of-Thought (CoT) core, breaking down a complex task into simpler and manageable sub-problems, and addressing them step-by-step from a low-level pixel perception to high-level cognitive interpretation. Extensive experiments across various complex video QA benchmarks demonstrate that our overall framework strikingly boosts existing state-of-the-art. To our knowledge, this is the first attempt at successfully implementing the CoT technique for achieving human-level video reasoning, where we show great potential in extending it to a wider range of video understanding scenarios. Systems and codes will be open later.

We tackle the common challenge of inter-concept visual confusion in compositional concept generation using text-guided diffusion models (TGDMs). It becomes even more pronounced in the generation of customized concepts, due to the scarcity of user-provided concept visual examples. By revisiting the two major stages leading to the success of TGDMs---1) contrastive image-language pre-training (CLIP) for text encoder that encodes visual semantics, and 2) training TGDM that decodes the textual embeddings into pixels---we point that existing customized generation methods only focus on fine-tuning the second stage while overlooking the first one. To this end, we propose a simple yet effective solution called CLIF: contrastive image-language fine-tuning. Specifically, given a few samples of customized concepts, we obtain non-confusing textual embeddings of a concept by fine-tuning CLIP via contrasting a concept and the over-segmented visual regions of other concepts. Experimental results demonstrate the effectiveness of CLIF in preventing the confusion of multi-customized concept generation. Project page: https://clif-official.github.io/clif.

For multimodal LLMs, the synergy of visual comprehension (textual output) and generation (visual output) presents an ongoing challenge. This is due to a conflicting objective: for comprehension, an MLLM needs to abstract the visuals; for generation, it needs to preserve the visuals as much as possible. Thus, the objective is a dilemma for visual-tokens. To resolve the conflict, we propose encoding images into morph-tokens to serve a dual purpose: for comprehension, they act as visual prompts instructing MLLM to generate texts; for generation, they take on a different, non-conflicting role as complete visual-tokens for image reconstruction, where the missing visual cues are recovered by the MLLM. Extensive experiments show that morph-tokens can achieve a new SOTA for multimodal comprehension and generation simultaneously. Our project is available at https://github.com/DCDmllm/MorphTokens.

Artificial intelligence (AI) encompasses knowledge acquisition and real-world grounding across various modalities, including language, visual, auditory, and sensory data. Multimodal large language models (MLLMs) have thus recently garnered growing interest in both academia and industry, showing an unprecedented trend to achieve human-level AI. This tutorial aims to deliver a comprehensive review of cutting-edge research in MLLMs, focusing on three key areas: MLLM architecture design, instructional learning, and multimodal reasoning of MLLMs. We will explore technical advancements, synthesize key challenges, and discuss potential avenues for future research. All the resources and materials will be made available online. https://mllm2024.github.io/ACM-MM2024

Vision-language models such as CLIP are capable of mapping the different modality data into a unified feature space, enabling zero/few-shot inference by measuring the similarity of given images and texts. However, most existing methods overlook modality gaps in CLIP's encoded features, which is shown as the text and image features lie far apart from each other, resulting in limited classification performance. To tackle this issue, we introduce a method called Selective Vision-Language Subspace Projection (SSP), which incorporates local image features and utilizes them as a bridge to enhance the alignment between image-text pairs. Specifically, our SSP framework comprises two parallel modules: a vision projector and a language projector. Both projectors utilize local image features to span the respective subspaces for image and texts, thereby projecting the image and text features into their respective subspaces to achieve alignment. Moreover, our approach entails only training-free matrix calculations and can be seamlessly integrated into advanced CLIP-based few-shot learning frameworks. Extensive experiments on 11 datasets have demonstrated SSP's superior text-image alignment capabilities, outperforming the state-of-the-art alignment methods. The code is available at https://github.com/zhuhsingyuu/SSP

Recent developments of vision large language models (LLMs) have seen remarkable progress, yet still encounter challenges towards multimodal generalists, such as coarse-grained instance-level understanding, lack of unified support for both images and videos, and insufficient coverage across various vision tasks. In this paper we present Vitron, a universal pixel-level vision LLM designed for comprehensive understanding, generating, segmenting, and editing of both static images and dynamic videos. Building on top of an LLM backbone, Vitron incorporates encoders for images, videos, and pixel-level regional visuals within its frontend modules, while employing state-of-the-art visual specialists as its backend, via which Vitron supports a spectrum of vision end tasks, spanning visual comprehension to visual generation, from low level to high level. To ensure an effective and precise message passing from LLM to backend modules for function invocation, we propose a novel hybrid method by simultaneously integrating discrete textual instructions and continuous signal embeddings. Further, we design various pixel-level spatiotemporal vision-language alignment learning for Vitron to reach the best fine-grained visual capability. Finally, a cross-task synergy module is advised to learn to maximize the task-invariant fine-grained visual features, enhancing the synergy between different visual tasks. Demonstrated over 12 visual tasks and evaluated across 22 datasets, Vitron showcases its extensive capabilities in the four main vision task clusters. Overall, this work illuminates the great potential of developing a more unified multimodal generalist.

In recent times, Vision-Language Models (VLMs) have been trained under two predominant paradigms. Generative training has enabled Multimodal Large Language Models (MLLMs) to tackle various complex tasks, yet issues such as hallucinations and weak object discrimination persist. Discriminative training, exemplified by models like CLIP, excels in zero-shot image-text classification and retrieval, yet struggles with complex scenarios requiring fine-grained semantic differentiation. This paper addresses these challenges by proposing a unified approach that integrates the strengths of both paradigms. Considering interleaved image-text sequences as the general format of input samples, we introduce a structure-induced training strategy that imposes semantic relationships between input samples and the MLLM’s hidden state. This approach enhances the MLLM’s ability to capture global semantics and distinguish fine-grained semantics. By leveraging dynamic sequence alignment within the Dynamic Time Warping framework and integrating a novel kernel for fine-grained semantic differentiation, our method effectively balances generative and discriminative tasks. Extensive experiments demonstrate the effectiveness of our approach, achieving state-of-the-art results in multiple generative tasks, especially those requiring cognitive and discrimination abilities. Additionally, our method surpasses discriminative benchmarks in interleaved and fine-grained retrieval tasks. By employing a retrieval-augmented generation strategy, our approach further enhances performance in some generative tasks within one model, offering a promising direction for future research in vision-language modeling.

In this paper, we delve deeper into the Kullback–Leibler (KL) Divergence loss and mathematically prove that it is equivalent to the Decoupled Kullback-Leibler (DKL) Divergence loss that consists of 1) a weighted Mean Square Error ($\mathbf{w}$MSE) loss and 2) a Cross-Entropy loss incorporating soft labels. Thanks to the decomposed formulation of DKL loss, we have identified two areas for improvement. Firstly, we address the limitation of KL/DKL in scenarios like knowledge distillation by breaking its asymmetric optimization property. This modification ensures that the $\mathbf{w}$MSE component is always effective during training, providing extra constructive cues. Secondly, we introduce class-wise global information into KL/DKL to mitigate bias from individual samples. With these two enhancements, we derive the Improved Kullback–Leibler (IKL) Divergence loss and evaluate its effectiveness by conducting experiments on CIFAR-10/100 and ImageNet datasets, focusing on adversarial training, and knowledge distillation tasks. The proposed approach achieves new state-of-the-art adversarial robustness on the public leaderboard --- \textit{RobustBench} and competitive performance on knowledge distillation, demonstrating the substantial practical merits. Our code is available at https://github.com/jiequancui/DKL.

We propose a novel method, \textbf{TwinAct}, to tackle the challenge of decoupling actions and actors in order to customize the text-guided diffusion models (TGDMs) for few-shot action image generation. TwinAct addresses the limitations of existing methods that struggle to decouple actions from other semantics (e.g., the actor's appearance) due to the lack of an effective inductive bias with few exemplar images. Our approach introduces a common action space, which is a textual embedding space focused solely on actions, enabling precise customization without actor-related details. Specifically, TwinAct involves three key steps: 1) Building common action space based on a set of representative action phrases; 2) Imitating the customized action within the action space; and 3) Generating highly adaptable customized action images in diverse contexts with action similarity loss. To comprehensively evaluate TwinAct, we construct a novel benchmark, which provides sample images with various forms of actions. Extensive experiments demonstrate TwinAct's superiority in generating accurate, context-independent customized actions while maintaining the identity consistency of different subjects, including animals, humans, and even customized actors.

The swift advancement in Multimodal LLMs (MLLMs) also presents significant challenges for effective knowledge editing. Current methods, including intrinsic knowledge editing and external knowledge resorting, each possess strengths and weaknesses, struggling to balance the desired properties of reliability, generality, and locality when applied to MLLMs. In this paper, we propose \textbf{UniKE}, a novel multimodal editing method that establishes a unified perspective and paradigm for intrinsic knowledge editing and external knowledge resorting. Both types of knowledge are conceptualized as vectorized key-value memories, with the corresponding editing processes resembling the assimilation and accommodation phases of human cognition, conducted at the same semantic levels. Within such a unified framework, we further promote knowledge collaboration by disentangling the knowledge representations into the semantic and truthfulness spaces. Extensive experiments validate the effectiveness of our method, which ensures that the post-edit MLLM simultaneously maintains excellent reliability, generality, and locality. The code for UniKE is available at https://github.com/beepkh/UniKE.

As Archimedes famously said, ``Give me a lever long enough and a fulcrum on which to place it, and I shall move the world'', in this study, we propose to use a tiny Language Model (LM), \eg, a Transformer with 67M parameters, to lever much larger Vision-Language Models (LVLMs) with 9B parameters. Specifically, we use this tiny \textbf{Lever-LM} to configure effective in-context demonstration (ICD) sequences to improve the In-Context Learinng (ICL) performance of LVLMs. Previous studies show that diverse ICD configurations like the selection and ordering of the demonstrations heavily affect the ICL performance, highlighting the significance of configuring effective ICD sequences. Motivated by this and by re-considering the the process of configuring ICD sequence, we find this is a mirror process of human sentence composition and further assume that effective ICD configurations may contain internal statistical patterns that can be captured by Lever-LM. Then a dataset with effective ICD sequences is constructed to train Lever-LM. After training, given novel queries, new ICD sequences are configured by the trained Lever-LM to solve vision-language tasks through ICL. Experiments show that these ICD sequences can improve the ICL performance of two LVLMs compared with some strong baselines in Visual Question Answering and Image Captioning, validating that Lever-LM can really capture the statistical patterns for levering LVLMs. The code is available at \url{https://anonymous.4open.science/r/Lever-LM-604A/}.

Recent 3D large reconstruction models (LRMs) can generate high-quality 3D content in sub-seconds by integrating multi-view diffusion models with scalable multi-view reconstructors. Current works further leverage 3D Gaussian Splatting as 3D representation for improved visual quality and rendering efficiency. However, we observe that existing Gaussian reconstruction models often suffer from multi-view inconsistency and blurred textures. We attribute this to the compromise of multi-view information propagation in favor of adopting powerful yet computationally intensive architectures (\eg, Transformers). To address this issue, we introduce MVGamba, a general and lightweight Gaussian reconstruction model featuring a multi-view Gaussian reconstructor based on the RNN-like State Space Model (SSM). Our Gaussian reconstructor propagates causal context containing multi-view information for cross-view self-refinement while generating a long sequence of Gaussians for fine-detail modeling with linear complexity. With off-the-shelf multi-view diffusion models integrated, MVGamba unifies 3D generation tasks from a single image, sparse images, or text prompts. Extensive experiments demonstrate that MVGamba outperforms state-of-the-art baselines in all 3D content generation scenarios with approximately only $0.1\times$ of the model size. The codes are available at \url{https://github.com/SkyworkAI/MVGamba}.

When fine-tuning zero-shot models like CLIP, our desideratum is for the fine-tuned model to excel in both in-distribution (ID) and out-of-distribution (OOD). Recently, ensemble-based models (ESM) have been shown to offer significant robustness improvement, while preserving high ID accuracy. However, our study finds that ESMs do not solve the ID-OOD trade-offs: they achieve peak performance for ID and OOD accuracy at different mixing coefficients. When optimized for OOD accuracy, the ensemble model exhibits a noticeable decline in ID accuracy, and vice versa. In contrast, we propose a sample-wise ensembling technique that can simultaneously attain the best ID and OOD accuracy without the trade-offs. Specifically, we construct a Zero-Shot Failure (ZSF) set containing training samples incorrectly predicted by the zero-shot model. For each test sample, we calculate its distance to the ZSF set and assign a higher weight to the fine-tuned model in the ensemble if the distance is small. We term our method Variance Reduction Fine-tuning (VRF), as it effectively reduces the variance in ensemble predictions, thereby decreasing residual error. On ImageNet and five derived distribution shifts, our VRF further improves the OOD accuracy by 1.5 - 2.0 pp over the ensemble baselines while maintaining or increasing ID accuracy. VRF achieves similar large robustness gains on (0.9 - 3.1 pp) on other distribution shifts 19 benchmarks. Codes are available in https://github.com/BeierZhu/VRF.

Vision-language models, such as CLIP, have shown impressive generalization capacities when using appropriate text descriptions. While optimizing prompts on downstream labeled data has proven effective in improving performance, these methods entail labor costs for annotations and are limited by their quality. Additionally, since CLIP is pre-trained on highly imbalanced Web-scale data, it suffers from inherent label bias that leads to suboptimal performance. To tackle the above challenges, we propose a label-**F**ree p**ro**mpt distribution **l**earning and b**i**as **c**orrection framework, dubbed as **Frolic**, which boosts zero-shot performance without the need for labeled data. Specifically, our Frolic learns distributions over prompt prototypes to capture diverse visual representations and adaptively fuses these with the original CLIP through confidence matching. This fused model is further enhanced by correcting label bias via a label-free logit adjustment. Notably, our method is not only training-free but also circumvents the necessity for hyper-parameter tuning. Extensive experimental results across 16 datasets demonstrate the efficacy of our approach, particularly outperforming the state-of-the-art by an average of $2.6\%$ on 10 datasets with CLIP ViT-B/16 and achieving an average margin of $1.5\%$ on ImageNet and its five distribution shifts with CLIP ViT-B/16. Codes are available in [https://github.com/zhuhsingyuu/Frolic](https://github.com/zhuhsingyuu/Frolic).

Nearly all existing scene graph generation (SGG) models have overlooked the ground-truth annotation qualities of mainstream SGG datasets, i.e., they assume: 1) all the manually annotated positive samples are equally correct; 2) all the un-annotated negative samples are absolutely background. In this paper, we argue that neither of the assumptions applies to SGG: there are numerous “noisy” ground-truth predicate labels that break these two assumptions and harm the training of unbiased SGG models. To this end, we propose a novelNoIsy labelCorrEction andSampleTrainingstrategy for SGG:NICEST, which rules out these noisy label issues by generating high-quality samples and designing an effective training strategy. Specifically, it consists of: 1)NICE: it detects noisy samples and then reassigns higher-quality soft predicate labels to them. To achieve this goal, NICE contains three main steps: negative Noisy Sample Detection (Neg-NSD), positive NSD (Pos-NSD), and Noisy Sample Correction (NSC). Firstly, in Neg-NSD, it is treated as an out-of-distribution detection problem, and the pseudo labels are assigned to all detected noisy negative samples. Then, in Pos-NSD, we use a density-based clustering algorithm to detect noisy positive samples. Lastly, in NSC, we use weighted KNN to reassign more robust soft predicate labels rather than hard labels to all noisy positive samples. 2)NIST: it is a multi-teacher knowledge distillation based training strategy, which enables the model to learn unbiased fusion knowledge. A dynamic trade-off weighting strategy in NIST is designed to penalize the bias of different teachers. Due to the model-agnostic nature of both NICE and NIST, NICEST can be seamlessly incorporated into any SGG architecture to boost its performance on different predicate categories. In addition, to better assess the generalization ability of SGG models, we propose a new benchmark,VG-OOD, by reorganizing the prevalent VG dataset. This reorganization deliberately makes the predicate distributions between the training and test sets as different as possible for each subject-object category pair. This new benchmark helps disentangle the influence of subject-object category biases. Extensive ablations and results on different backbones and tasks have attested to the effectiveness and generalization ability of each component of NICEST.

Large pre-trained vision-language models, such as CLIP [1], have demonstrated remarkable performance in few shot image classification. To facilitate the rapid adaptation of CLIP in downstream tasks with limited visual samples, two primary frameworks have been proposed. The first framework centers on the image encoder and introduces a trainable visual classifier after the backbone to generate logits for each object class. Nevertheless, this framework heavily depends on limited visual features extracted by the pre-trained visual encoder, which can result in over-fitting issues. The second framework aims to optimize the text encoder by using trainable soft language prompts and computing logits for each class based on the similarity between image features and optimized prompt features. However, this framework encounters the issue of imperfect alignment between the representations extracted by the image and text encoders, making it difficult to fine-tune the language prompts using visual samples. This paper proposes a Multi- Modal Prototype Regularization (MMPR) method for CLIP based few-shot fine-tuning for image classification. MMPR can address the challenges of effectively utilizing both image and text features. MMPR fine-tunes a classifier and regularizes its weights using both image-based (ImgPR) and text-based (TexPR) prototypes. ImgPR represents the mean of image representations within the same class, derived from the image encoder, to distill specific visual distribution knowledge for classifier adaptation. TexPR represents the hand-crafted prompt associated with the class, derived from the text encoder, to incorporate general encyclopedic knowledge and mitigate visual over-fitting. MMPR significantly leverages both image and text information without increasing computational complexity during the inference stage compared to existing methods. Experimental results on various challenging public benchmarks demonstrate the superiority of the proposed MMPR method over state-of-the-art methods.

The present machine learning schema typically uses a one-pass model inference (e.g., forward propagation) to make predictions in the testing phase. It is inherently different from human students who double-check the answer during examinations especially when the confidence is low. To bridge this gap, we propose a learning to double-check (L2D) framework, which formulates double check as a learnable procedure with two core operations: recognizing unreliable predictions and revising predictions. To judge the correctness of a prediction, we resort to counterfactual faithfulness in causal theory and design a contrastive faithfulness measure. In particular, L2D generates counterfactual features by imagining: "what would the sample features be if its label was the predicted class" and judges the prediction by the faithfulness of the counterfactual features. Furthermore, we design a simple and effective revision module to revise the original model prediction according to the faithfulness. We apply the L2D framework to three classification models and conduct experiments on two public datasets for image classification, validating the effectiveness of L2D in prediction correctness judgment and revision.

We present a new paradigm for fine-tuning large-scale vision-language pre-trained models on downstream task, dubbed Prompt Regularization (ProReg). Different from traditional fine-tuning which easily overfits to the downstream task data, ProReg uses the prediction by prompting the pretrained model to regularize the fine-tuning. The motivation is: by prompting the large model “a photo of a [CLASS]”, the fill-in answer is only dependent on the pretraining encyclopedic knowledge while independent of the task data distribution, which is usually biased. Specifically, given a training sample prediction during fine-tuning, we first calculate its Kullback-Leibler loss of the prompt prediction and Cross-Entropy loss of the ground-truth label, and then combine them with a proposed sample-wise adaptive trade- off weight, which automatically adjusts the transfer between the pretrained and downstream domains. On various out-of-distribution benchmarks, we show the consistently strong performance of ProReg compared with conventional fine-tuning, zero-shot prompt, prompt tuning, and other state-of-the-art methods.

We study the out-of-distribution generalization of active learning that adaptively selects samples for annotation in learning the decision boundary of classification.Our empirical study finds that increasingly annotating seen samples may hardly benefit the generalization.To address the problem, we propose Counterfactual Active Learning (CounterAL) that empowers active learning with counterfactual thinking to bridge the seen samples with unseen cases.In addition to annotating factual samples, Coun-terAL requires annotators to answer counterfactual questions to construct counterfactual samples for training.To achieve CounterAL, we design a new acquisition strategy that selects the informative factual-counterfactual pairs for annotation; and a new training strategy that pushes the model update to focus on the discrepancy between factual and counterfactual samples.We evaluate CounterAL on multiple public datasets of sentiment analysis and natural language inference.The experiment results show that CounterAL requires fewer acquisition rounds and outperforms existing active learning methods by a large margin in OOD tests with comparable IID performance.* Wenjie Wang is the corresponding author.This work is

Weakly Supervised Video Anomaly Detection (WSVAD) is challenging because the binary anomaly label is only given on the video level, but the output requires snippet-level predictions. So, Multiple Instance Learning (MIL) is prevailing in WSVAD. However, MIL is notoriously known to suffer from many false alarms because the snippet-level detector is easily biased towards the abnormal snippets with simple context, confused by the normality with the same bias, and missing the anomaly with a different pattern. To this end, we propose a new MIL framework: Unbiased MIL (UMIL), to learn unbiased anomaly features that improve WSVAD. At each MIL training iteration, we use the current detector to divide the samples into two groups with different context biases: the most confident abnormal/normal snippets and the rest ambiguous ones. Then, by seeking the invariant features across the two sample groups, we can remove the variant context biases. Extensive experiments on benchmarks UCF-Crime and TAD demonstrate the effectiveness of our UMIL. Our code is provided at https://github.com/ktr-hubrt/UMIL.

Semantic Scene Completion (SSC) transforms an image of single-view depth and/or RGB 2D pixels into 3D voxels, each of whose semantic labels are predicted. SSC is a well-known ill-posed problem as the prediction model has to “imagine” what is behind the visible surface, which is usually represented by Truncated Signed Distance Function (TSDF). Due to the sensory imperfection of the depth camera, most existing methods based on the noisy TSDF estimated from depth values suffer from 1) incomplete volumetric predictions and 2) confused semantic labels. To this end, we use the ground-truth 3D voxels to generate a perfect visible surface, called TSDF-CAD, and then train a “cleaner” SSC model. As the model is noise-free, it is expected to focus more on the “imagination” of unseen voxels. Then, we propose to distill the intermediate “cleaner” knowledge into another model with noisy TSDF input. In particular, we use the 3D occupancy feature and the semantic relations of the “cleaner self” to supervise the counterparts of the “noisy self” to respectively address the above two incorrect predictions. Experimental results validate that our method improves the noisy counterparts with 3.1% IoU and 2.2% mIoU for measuring scene completion and SSC, and also achieves new state-of-the-art accuracy on the popular NYU dataset. The code is available at https://github.com/fereenwong/CleanerS.

Novel Class Discovery (NCD) aims to discover unknown classes without any annotation, by exploiting the transferable knowledge already learned from a base set of known classes. Existing works hold an impractical assumption that the novel class distribution prior is uniform, yet neglect the imbalanced nature of real-world data. In this paper, we relax this assumption by proposing a new challenging task: distribution-agnostic NCD, which allows data drawn from arbitrary unknown class distributions and thus renders existing methods useless or even harmful. We tackle this challenge by proposing a new method, dubbed “Boot-strapping Your Own Prior (BYOP)”, which iteratively estimates the class prior based on the model prediction it-self. At each iteration, we devise a dynamic temperature technique that better estimates the class prior by encouraging sharper predictions for less-confident samples. Thus, BYOP obtains more accurate pseudo-labels for the novel samples, which are beneficial for the next training iteration. Extensive experiments show that existing methods suffer from imbalanced class distributions, while BYOp11Code: https://github.com/muliyangm/BYOP. out-performs them by clear margins, demonstrating its effectiveness across various distribution scenarios.

In video action recognition, shortcut static features can interfere with the learning of motion features, resulting in poor out-of-distribution (OOD) generalization. The video background is clearly a source of static bias, but the video foreground, such as the clothing of the actor, can also provide static bias. In this paper, we empirically verify the existence of foreground static bias by creating test videos with conflicting signals from the static and moving portions of the video. To tackle this issue, we propose a simple yet effective technique, StillMix, to learn robust action representations. Specifically, StillMix identifies bias-inducing video frames using a 2D reference network and mixes them with videos for training, serving as effective bias suppression even when we cannot explicitly extract the source of bias within each video frame or enumerate types of bias. Finally, to precisely evaluate static bias, we synthesize two new benchmarks, SCUBA for static cues in the background, and SCUFO for static cues in the foreground. With extensive experiments, we demonstrate that StillMix mitigates both types of static bias and improves video representations for downstream applications. Code is available at https://github.com/lihaoxin05/StillMix.

Fair face recognition is all about learning invariant feature that generalizes to unseen faces in any demographic group. Unfortunately, face datasets inevitably capture the imbalanced demographic attributes that are ubiquitous in real-world observations, and the model learns biased feature that generalizes poorly in the minority group. We point out that the bias arises due to the confounding demographic attributes, which mislead the model to capture the spurious demographic-specific feature. The confounding effect can only be removed by causal intervention, which requires the confounder annotations. However, such annotations can be prohibitively expensive due to the diversity of the demographic attributes. To tackle this, we propose to generate diverse data partitions iteratively in an unsupervised fashion. Each data partition acts as a self-annotated confounder, enabling our Invariant Feature Regularization (Inv-Reg) to deconfound. Inv-Reg is orthogonal to existing methods, and combining Inv-Reg with two strong baselines (Arcface and CIFP) leads to new state-of-the-art that improves face recognition on a variety of demographic groups. Code is available at https://github.com/milliema/InvReg.

This study explores the concept of equivariance in vision-language foundation models (VLMs), focusing specifically on the multimodal similarity function that is not only the major training objective but also the core delivery to support downstream tasks. Unlike the existing image-text similarity objective which only categorizes matched pairs as similar and unmatched pairs as dissimilar, equivariance also requires similarity to vary faithfully according to the semantic changes. This allows VLMs to generalize better to nuanced and unseen multimodal compositions. However, modeling equivariance is challenging as the ground truth of semantic change is difficult to collect. For example, given an image-text pair about a dog, it is unclear to what extent the similarity changes when the pixel is changed from dog to cat? To this end, we propose EqSim, a regularization loss that can be efficiently calculated from any two matched training pairs and easily pluggable into existing image-text retrieval fine-tuning. Meanwhile, to further diagnose the equivariance of VLMs, we present a new challenging benchmark EqBen. Compared to the existing evaluation sets, EqBen is the first to focus on "visual-minimal change". Extensive experiments show the lack of equivariance in current VLMs1and validate the effectiveness of EqSim2.

We show that classifiers trained with random region proposals achieve state-of-the-art Open-world Object Detection (OWOD): they can not only maintain the accuracy of the known objects (w/training labels), but also considerably improve the recall of unknown ones (w/o training labels). Specifically, we propose RandBox, a Fast R-CNN based architecture trained on random proposals at each training iteration, surpassing existing Faster R-CNN and Transformer based OWOD. Its effectiveness stems from the following two benefits introduced by randomness. First, as the randomization is independent of the distribution of the limited known objects, the random proposals become the instrumental variable that prevents the training from being confounded by the known objects. Second, the unbiased training encourages more proposal explorations by using our proposed matching score that does not penalize the random proposals whose prediction scores do not match the known objects. On two benchmarks: Pascal-VOC/MS-COCO and LVIS, Rand-Box significantly outperforms the previous state-of-the-art in all metrics. We also detail the ablations on randomization and loss designs. Codes are available at https://github.com/scuwyh2000/RandBox.

In a video, an object usually appears as the trajectory, i.e., it spans over a few spatial but longer temporal patches, that contains abundant spatiotemporal contexts. However, modern Video-Language BERTs (VDL-BERTs) neglect this trajectory characteristic that they usually follow image-language BERTs (IL-BERTs) to deploy the patch-to-word (P2W) attention that may over-exploit trivial spatial contexts and neglect significant temporal contexts. To amend this, we propose a novel TW-BERT to learn Trajectory-Word alignment by a newly designed trajectory-to-word (T2W) attention for solving video-language tasks. Moreover, previous VDL-BERTs usually uniformly sample a few frames into the model while different trajectories have diverse graininess, i.e., some trajectories span longer frames and some span shorter, and using a few frames will lose certain useful temporal contexts. However, simply sampling more frames will also make pre-training infeasible due to the largely increased training burdens. To alleviate the problem, during the fine-tuning stage, we insert a novel Hierarchical Frame-Selector (HFS) module into the video encoder. HFS gradually selects the suitable frames conditioned on the text context for the later cross-modal encoder to learn better trajectory-word alignments. By the proposed T2W attention and HFS, our TW-BERT achieves SOTA performances on text-to-video retrieval tasks, and comparable performances on video question-answering tasks with some VDL-BERTs trained on much more data. The code will be available in the supplementary material.

We tackle the data scarcity challenge in few-shot point cloud recognition of 3D objects by using a joint prediction from a conventional 3D model and a well-trained 2D model. Surprisingly, such an ensemble, though seems trivial, has hardly been shown effective in recent 2D-3D models. We find out the crux is the less effective training for the "joint hard samples", which have high confidence prediction on different wrong labels, implying that the 2D and 3D models do not collaborate well. To this end, our proposed invariant training strategy, called INVJOINT, does not only emphasize the training more on the hard samples, but also seeks the invariance between the conflicting 2D and 3D ambiguous predictions. INVJOINT can learn more collaborative 2D and 3D representations for better ensemble. Extensive experiments on 3D shape classification with widely adopted ModelNet10/40, ScanObjectNN and Toys4K, and shape retrieval with ShapeNet-Core validate the superiority of our INVJOINT. Codes will be publicly Available1.

Thanks to the large pre-trained vision-language models (VLMs) like CLIP [37], we can craft a zero-shot classifier by discrete prompt design, e.g., the confidence score of an image being "[CLASS]" can be obtained by using the VLM provided similarity between the image and the prompt sentence "a photo of a [CLASS]". Furthermore, prompting shows great potential for fast adaptation of VLMs to downstream tasks if we fine-tune the soft prompts with few samples. However, we find a common failure that improper fine-tuning or learning with extremely few-shot samples may even under-perform the zero-shot prediction. Existing methods still address this problem by using traditional anti-overfitting techniques such as early stopping and data augmentation, which lack a principled solution specific to prompting. In this paper, we present Prompt-aligned Gradient, dubbed ProGrad to prevent prompt tuning from forgetting the general knowledge learned from VLMs. In particular, ProGrad only updates the prompt whose gradient is aligned (or non-conflicting) to the general knowledge, which is represented as the optimization direction offered by the pre-defined prompt predictions. Extensive experiments under the few-shot learning, domain generalization, base-to-new generalization and cross-dataset transfer settings demonstrate the stronger few-shot generalization ability of ProGrad over state-of-the-art prompt tuning methods.

Motivated by the semi-supervised learning that uses the unlabeled data and pseudo annotations to improve the image classification, this paper proposes a new semi-supervised few-shot segmentation (FSS) framework of which the training process uses not only the annotated images, but also the unlabeled images, e.g. images from other available datasets, to enhance the training of the FSS model. Furthermore, in the test phase, more support images and pseudo-annotations can also be generated by the proposed framework to enrich the support set of novel classes and therefore benefit the inference. However, unlabeled images are not a free lunch. The noisy intra-class samples and inter-class samples existed in the unlabeled images as well as the interferences of the bad quality of pseudo annotations make it difficult to utilize the correct images and pseudo annotations for a certain class. To this end, we further propose a ranking algorithm consisting of an inter-class confidence term and an intra-class confidence term to efficiently utilize the pseudo annotations of the class with high quality. Extensive experiments on COCO-20idataset demonstrate that the proposed semi-supervised FSS framework is superior to many state-of-the-art methods.

Advancements in prompt tuning of vision-language models have underscored their potential in enhancing open-world visual concept comprehension. However, prior works only primarily focus on single-mode (only one prompt for each modality) and holistic level (image or sentence) semantic alignment, which fails to capture the sample diversity, leading to sub-optimal prompt discovery. To address the limitation, we propose a multi-mode token-level tuning framework that leverages the optimal transportation to learn and align a set of prompt tokens across modalities. Specifically, we rely on two essential factors: 1) multi-mode prompts discovery, which guarantees diverse semantic representations, and 2) token-level alignment, which helps explore fine-grained similarity. Consequently, the similarity can be calculated as a hierarchical transportation problem between the modality-specific sets. Extensive experiments on popular image recognition benchmarks show the superior generalization and few-shot abilities of our approach. The qualitative analysis demonstrates that the learned prompt tokens have the ability to capture diverse visual concepts.

In this work, we investigate the task of text-to-image (T2I) synthesis under the abstract-to-intricate setting, i.e., generating intricate visual content from simple abstract text prompts. Inspired by human imagination intuition, we propose a novel scene-graph hallucination (SGH) mechanism for effective abstract-to-intricate T2I synthesis. SGH carries out scene hallucination by expanding the initial scene graph (SG) of the input prompt with more feasible specific scene structures, in which the structured semantic representation of SG ensures high controllability of the intrinsic scene imagination. To approach the T2I synthesis, we deliberately build an SG-based hallucination diffusion system. First, we implement the SGH module based on the discrete diffusion technique, which evolves the SG structure by iteratively adding new scene elements. Then, we utilize another continuous-state diffusion model as the T2I synthesizer, where the overt image-generating process is navigated by the underlying semantic scene structure induced from the SGH module. On the benchmark COCO dataset, our system outperforms the existing best-performing T2I model by a significant margin, especially improving on the abstract-to-intricate T2I generation. Further in-depth analyses reveal how our methods advance.

Domain Adaptation (DA) is always challenged by the spurious correlation between the domain-invariant features (e.g., class identity) and the domain-specific ones (e.g., environment) that does not generalize to the target domain. Unfortunately, even enriched with additional unsupervised target domains, existing Unsupervised DA (UDA) methods still suffer from it. This is because the source domain supervision only considers the target domain samples as auxiliary data (e.g., by pseudo-labeling), yet the inherent distribution in the target domain---where the valuable de-correlation clues hide---is disregarded. We propose to make the U in UDA matter by giving equal status to the two domains. Specifically, we learn an invariant classifier whose prediction is simultaneously consistent with the labels in the source domain and clusters in the target domain, hence the spurious correlation inconsistent in the target domain is removed. We dub our approach "Invariant CONsistency learning" (ICON). Extensive experiments show that ICON achieves the state-of-the-art performance on the classic UDA benchmarks: Office-Home and VisDA-2017, and outperforms all the conventional methods on the challenging WILDS 2.0 benchmark. Codes are in https://github.com/yue-zhongqi/ICON.

Foundation models like CLIP allow zero-shot transfer on various tasks without additional training data. Yet, the zero-shot performance is less competitive than a fully supervised one. Thus, to enhance the performance, fine-tuning and ensembling are also commonly adopted to better fit the downstream tasks. However, we argue that such prior work has overlooked the inherent biases in foundation models. Due to the highly imbalanced Web-scale training set, these foundation models are inevitably skewed toward frequent semantics, and thus the subsequent fine-tuning or ensembling is still biased. In this study, we systematically examine the biases in foundation models and demonstrate the efficacy of our proposed Generalized Logit Adjustment (GLA) method. Note that bias estimation in foundation models is challenging, as most pre-train data cannot be explicitly assessed like in traditional long-tailed classification tasks. To this end, GLA has an optimization-based bias estimation approach for debiasing foundation models. As our work resolves a fundamental flaw in the pre-training, the proposed GLA demonstrates significant improvements across a diverse range of tasks: it achieves 1.5 pp accuracy gains on ImageNet, an large average improvement (1.4-4.6 pp) on 11 few-shot datasets, 2.4 pp gains on long-tailed classification. Codes are in https://github.com/BeierZhu/GLA.

Today's VQA models still tend to capture superficial linguistic correlations in the training set and fail to generalize to the test set with different QA distributions. To reduce these language biases, recent VQA works introduce an auxiliary question-only model to regularize the training of targeted VQA model, and achieve dominating performance on diagnostic benchmarks for out-of-distribution testing. However, due to the complex model design, ensemble-based methods are unable to equip themselves with two indispensable characteristics of an ideal VQA model: 1) Visual-explainable: The model should rely on the right visual regions when making decisions. 2) Question-sensitive: The model should be sensitive to the linguistic variations in questions. To this end, we propose a novel model-agnostic Counterfactual Samples Synthesizing and Training (CSST) strategy. After training with CSST, VQA models are forced to focus on all critical objects and words, which significantly improves both visual-explainable and question-sensitive abilities. Specifically, CSST is composed of two parts: Counterfactual Samples Synthesizing (CSS) and Counterfactual Samples Training (CST). CSS generates counterfactual samples by carefully masking critical objects in images or words in questions and assigning pseudo ground-truth answers. CST not only trains the VQA models with both complementary samples to predict respective ground-truth answers, but also urges the VQA models to further distinguish the original samples and superficially similar counterfactual ones. To facilitate the CST training, we propose two variants of supervised contrastive loss for VQA, and design an effective positive and negative sample selection mechanism based on CSS. Extensive experiments have shown the effectiveness of CSST. Particularly, by building on top of model LMH+SAR (Clark et al. 2019), (Si et al. 2021), we achieve record-breaking performance on all out-of-distribution benchmarks (e.g., VQA-CP v2, VQA-CP v1, and GQA-OOD).

Dataset bias in vision-language tasks is becoming one of the main problems which hinders the progress of our community. Existing solutions lack a principled analysis about why modern image captioners easily collapse into dataset bias. In this paper, we present a novel perspective: Deconfounded Image Captioning (DIC), to find out the answer of this question, then retrospect modern neural image captioners, and finally propose a DIC framework: DICv1.0 to alleviate the negative effects brought by dataset bias. DIC is based on causal inference, whose two principles: the backdoor and front-door adjustments, help us review previous studies and design new effective models. In particular, we showcase that DICv1.0 can strengthen two prevailing captioning models and can achieve a single-model 131.1 CIDEr-D and 128.4 c40 CIDEr-D on Karpathy split and online split of the challenging MS COCO dataset, respectively. Interestingly, DICv1.0 is a natural derivation from our causal retrospect, which opens promising directions for image captioning.

Graph Convolutional Network (GCN) is an emerging technique that performs learning and reasoning on graph data. It operates feature learning on the graph structure, through aggregating the features of the neighbor nodes to obtain the embedding of each target node. Owing to the strong representation power, recent research shows that GCN achieves state-of-the-art performance on several tasks such as recommendation and linked document classification. Despite its effectiveness, we argue that existing designs of GCN forgo modeling cross features, making GCN less effective for tasks or data where cross features are important. Although neural network can approximate any continuous function, including the multiplication operator for modeling feature crosses, it can be rather inefficient to do so (i.e., wasting many parameters at the risk of overfitting) if there is no explicit design. To this end, we design a new operator named Cross-feature Graph Convolution, which explicitly models the arbitrary-order cross features with complexity linear to feature dimension and order size. We term our proposed architecture as Cross-GCN, and conduct experiments on three graphs to validate its effectiveness. Extensive analysis validates the utility of explicitly modeling cross features in GCN, especially for feature learning at lower layers.

Deep learning models often fit undesired dataset bias in training. In this paper, we formulate the bias usingcausal inference, which helps us uncover the ever-elusive causalities among the key factors in training, and thus pursue the desired causal effect without the bias. We start from revisiting the process of building a visual recognition system, and then propose a structural causal model (SCM) for the key variables involved in dataset collection and recognition model: object, common sense, bias, context, and label prediction. Based on the SCM, one can observe that there are “good” and “bad” biases. Intuitively, in the image where a car is driving on a high way in a desert, the “good” bias denoting the common-sense context is the highway, and the “bad” bias accounting for the noisy context factor is the desert. We tackle this problem with a novel causal interventional training (CIT) approach, where we control theobservedcontext in each object class. We offer theoretical justifications forCITand validate it with extensive classification experiments on CIFAR-10, CIFAR-100 and ImageNet,e.g., surpassing the standard deep neural networks ResNet-34 and ResNet-50, respectively, by 0.95% and 0.70% accuracies on the ImageNet. Our code is open-sourced on the GitHubhttps://github.com/qinwei-hfut/CIT.

The prevailing framework for matching multimodal inputs is based on a two-stage process: (1) detecting proposals with an object detector and (2) matching text queries with proposals. Existing two-stage solutions mostly focus on the matching step. In this article, we argue that these methods overlook an obvious mismatch between the roles of proposals in the two stages: they generate proposals solely based on the detection confidence (i.e., query-agnostic), hoping that the proposals contain all instances mentioned in the text query (i.e., query-aware). Due to this mismatch, chances are that proposals relevant to the text query are suppressed during the filtering process, which in turn bounds the matching performance. To this end, we propose VL-NMS, which is the first method to yield query-aware proposals at the first stage. VL-NMS regards all mentioned instances as critical objects and introduces a lightweight module to predict a score for aligning each proposal with a critical object. These scores can guide the NMS operation to filter out proposals irrelevant to the text query, increasing the recall of critical objects, and resulting in a significantly improved matching performance. Since VL-NMS is agnostic to the matching step, it can be easily integrated into any state-of-the-art two-stage matching method. We validate the effectiveness of VL-NMS on three multimodal matching tasks, namely referring expression grounding, phrase grounding, and image-text matching. Extensive ablation studies on several baselines and benchmarks consistently demonstrate the superiority of VL-NMS.

We focus on the confounding bias between language and location in the visual grounding pipeline, where we find that the bias is the major visual reasoning bottleneck. For example, the grounding process is usually a trivial languagelocation association without visual reasoning, e.g., grounding any language query containing sheep to the nearly central regions, due to that most queries about sheep have ground-truth locations at the image center. First, we frame the visual grounding pipeline into a causal graph, which shows the causalities among image, query, target location and underlying confounder. Through the causal graph, we know how to break the grounding bottleneck: deconfounded visual grounding. Second, to tackle the challenge that the confounder is unobserved in general, we propose a confounder-agnostic approach called: Referring Expression Deconfounder (RED), to remove the confounding bias. Third, we implement RED as a simple language attention, which can be applied in any grounding method. On popular benchmarks, RED improves various state-of-the-art grounding methods by a significant margin. Code is available at: https://github.com/JianqiangH/Deconfounded_VG.

We address the overlooked unbiasedness in existing long-tailed classification methods: we find that their overall improvement is mostly attributed to the biased preference of "tail" over "head", as the test distribution is assumed to be balanced; however, when the test is as imbalanced as the long-tailed training data---let the test respect Zipf's law of nature---the "tail" bias is no longer beneficial overall because it hurts the "head" majorities. In this paper, we propose Cross-Domain Empirical Risk Minimization (xERM) for training an unbiased test-agnostic model to achieve strong performances on both test distributions, which empirically demonstrates that xERM fundamentally improves the classification by learning better feature representation rather than the "head vs. tail" game. Based on causality, we further theoretically explain why xERM achieves unbiasedness: the bias caused by the domain selection is removed by adjusting the empirical risks on the imbalanced domain and the balanced but unseen domain.

Shulin Cao, Jiaxin Shi, Liangming Pan, Lunyiu Nie, Yutong Xiang, Lei Hou, Juanzi Li, Bin He, Hanwang Zhang. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

Neural discrete reasoning (NDR) has shown remarkable progress in combining deep models with discrete reasoning.However, we find that existing NDR solution suffers from large performance drop on hypothetical questions, e.g., "what the annualized rate of return would be if the revenue in 2020 was doubled".The key to hypothetical question answering (HQA) is counterfactual thinking, which is a natural ability of human reasoning but difficult for deep models.In this work, we devise a Learning to Imagine (L2I) module, which can be seamlessly incorporated into NDR models to perform the imagination of unseen counterfactual.In particular, we formulate counterfactual thinking into two steps: 1) identifying the fact to intervene, and 2) deriving the counterfactual from the fact and assumption, which are designed as neural networks.Based on TAT-QA, we construct a very challenging HQA dataset with 8,283 hypothetical questions.We apply the proposed L2I to TAGOP, the state-of-theart solution on TAT-QA, validating the rationality and effectiveness of our approach.

Extracting class activation maps (CAM) is arguably the most standard step of generating pseudo masks for weakly-supervised semantic segmentation (WSSS). Yet, we find that the crux of the unsatisfactory pseudo masks is the binary crossentropy loss (BCE) widely used in CAM. Specifically, due to the sum-over-class pooling nature of BCE, each pixel in CAM may be responsive to multiple classes co-occurring in the same receptive field. As a result, given a class, its hot CAM pixels may wrongly invade the area belonging to other classes, or the non-hot ones may be actually a part of the class. To this end, we introduce an embarrassingly simple yet surprisingly effective method: Reactivating the converged CAM with BCE by using softmax crossentropy loss (SCE), dubbed ReCAM. Given an image, we use CAM to extract the feature pixels of each single class, and use them with the class label to learn another fully-connected layer (after the backbone) with SCE. Once converged, we extract ReCAM in the same way as in CAM. Thanks to the contrastive nature of SCE, the pixel response is disentangled into different classes and hence less mask ambiguity is expected. The evaluation on both PASCAL VOC and MS COCO shows that ReCAM not only generates high-quality masks, but also supports plug-and-play in any CAM variant with little overhead. Our code is public at https://github.com/zhaozhengChenIReCAM.

Deep learning models have achieved great success in many fields, yet they are vulnerable to adversarial examples. This paper follows a causal perspective to look into the adversarial vulnerability and proposes Causal Intervention by Semantic Smoothing (CISS), a novel framework towards robustness against natural language attacks. Instead of merely fitting observational data, CISS learns causal effects p(y|do(x)) by smoothing in the latent semantic space to make robust predictions, which scales to deep architectures and avoids tedious construction of noise customized for specific attacks. CISS is provably robust against word substitution attacks, as well as empirically robust even when perturbations are strengthened by unknown attack algorithms. For example, on YELP, CISS surpasses the runner-up by 6.8% in terms of certified robustness against word substitutions, and achieves 80.7% empirical robustness when syntactic attacks are integrated.

Knowledge distillation (KD) is essentially a process of transferring a teacher model's behavior, e.g., network response, to a student model. The network response serves as additional supervision to formulate the machine domain, which uses the data collected from the human domain as a transfer set. Traditional KD methods hold an underlying assumption that the data collected in both human domain and machine domain are both independent and identically distributed (IID). We point out that this naive assumption is unrealistic and there is indeed a transfer gap between the two domains. Although the gap offers the student model external knowledge from the machine domain, the imbalanced teacher knowledge would make us incorrectly estimate how much to transfer from teacher to student per sample on the non-IID transfer set. To tackle this challenge, we propose Inverse Probability Weighting Distillation (IPWD) that estimates the propensity of a training sample belonging to the machine domain, and assigns its inverse amount to compensate for under-represented samples. Experiments on CIFAR-100 and ImageNet demonstrate the effectiveness of \ours~for both two-stage distillation and one-stage self-distillation.

Current deep learning-based image manipulation localization methods achieve impressive performance when rich spatial features and information are fully utilized. However, most of them suffer from the irrelevance of semantic awareness when identifying various manipulation categories. This leads to false alarms on recognizing forged regions. In this paper, we propose a Progressively-Refined Neural Network (PR-Net), to localize the tampered regions progressively under a coarse-to-fine workflow. Specifically, PR-Net is composed of a Feature Extractor (FE) that captures feature intrinsic correlations and a Mask Generation Module (MGM) with three refining generators. The FE takes a CNN to extract the image features and introduces an attention mechanism Convolution Block Attention Module (CBAM) to suppress the image content and guide the extractor in exploring the inconsistencies between the manipulated and authentic regions. The MGM comprises three generators where the Coarse Mask RR-Generator generates a localization result roughly, the Candidate Mask RR-Generator generates a possible tampered region according to the rough localization measure, and the Fine Mask RR-Generator produces the final prediction of manipulated regions. We also utilize the Rotated Residual (RR) structure to suppress the image content during the generative process. The extensive experimental results on four benchmark data sets (NIST16, COVER, CASIA v1.0, and In-The-Wild) demonstrate the superior performance of PR-Net compared with the state-of-the-art methods in localizing the manipulated regions.

Grounding natural language in images, such as localizing "the black dog on the left of the tree", is one of the core problems in artificial intelligence, as it needs to comprehend the fine-grained language compositions. However, existing solutions merely rely on the association between the holistic language features and visual features, while neglect the nature of composite reasoning implied in the language. In this paper, we propose a natural language grounding model that can automatically compose a binary tree structure for parsing the language and then perform visual reasoning along the tree in a bottom-up fashion. We call our model RvG-Tree: Recursive Grounding Tree, which is inspired by the intuition that any language expression can be recursively decomposed into two constituent parts, and the grounding confidence score can be recursively accumulated by calculating their grounding scores returned by the two sub-trees.RvG-Tree can be trained end-to-end by using the Straight-Through Gumbel-Softmax estimator that allows the gradients from the continuous score functions passing through the discrete tree construction. Experiments on several benchmarks show that our model achieves the state-of-the-art performance with more explainable reasoning.

We propose scene graph auto-encoder (SGAE) that incorporates the language inductive bias into the encoder-decoder image captioning framework for more human-like captions. Intuitively, we humans use the inductive bias to compose collocations and contextual inferences in discourse. For example, when we see the relation “a person on a bike”, it is natural to replace “on” with “ride” and infer “a person riding a bike on a road” even when the “road” is not evident. Therefore, exploiting such bias as a language prior is expected to help the conventional encoder-decoder models reason as we humans and generate more descriptive captions. Specifically, we use the scene graph—a directed graph ($\mathcal {G}$) where an object node is connected by adjective nodes and relationship nodes—to represent the complex structural layout of both image ($\mathcal {I}$) and sentence ($\mathcal {S}$). In the language domain, we use SGAE to learn a dictionary set ($\mathcal {D}$) that helps reconstruct sentences in the$\mathcal {S}\rightarrow \mathcal {G}_{\mathcal {S}} \rightarrow \mathcal {D} \rightarrow \mathcal {S}$auto-encoding pipeline, where$\mathcal {D}$encodes the desired language prior and the decoder learns to caption from such a prior; in the vision-language domain, we share$\mathcal {D}$in the$\mathcal {I}\rightarrow \mathcal {G}_{\mathcal {I}} \rightarrow \mathcal {D} \rightarrow \mathcal {S}$pipeline and distill the knowledge of the language decoder of the auto-encoder to that of the encoder-decoder based image captioner to transfer the language inductive bias. In this way, the shared$\mathcal {D}$provides hidden embeddings about descriptive collocations to the encoder-decoder and the distillation strategy teaches the encoder-decoder to transform these embeddings to human-like captions as the auto-encoder. Thanks to the scene graph representation, the shared dictionary set, and the Knowledge Distillation strategy, the inductive bias is transferred across domains in principle. We validate the effectiveness of SGAE on the challenging MS-COCO image captioning benchmark, where our SGAE-based single-model achieves a new state-of-the-art 129.6 CIDEr-D on the Karpathy split, and a competitive 126.6 CIDEr-D (c40) on the official server, which is even comparable to other ensemble models. Furthermore, we validate the transferability of SGAE on two more challenging settings: transferring inductive bias from other language corpora and unpaired image captioning. Once again, the results of both settings confirm the superiority of SGAE. The code is released inhttps://github.com/yangxuntu/SGAE.

With the maturity of visual detection techniques, we are more ambitious in describing visual content with open-vocabulary, fine-grained and free-form language, i.e., the task of image captioning. In particular, we are interested in generating longer, richer and more fine-grained sentences and paragraphs as image descriptions. Image captioning can be translated to the task of sequential language prediction given visual content, where the output sequence forms natural language description with plausible grammar. However, existing image captioning methods focus only on language policy while not visual policy, and thus fail to capture visual context that are crucial for compositional reasoning such as object relationships (e.g., "man riding horse") and visual comparisons (e.g., "small(er) cat"). This issue is especially severe when generating longer sequences such as a paragraph. To fill the gap, we propose a Context-Aware Visual Policy network (CAVP) for fine-grained image-to-language generation: image sentence captioning and image paragraph captioning. During captioning, CAVP explicitly considers the previous visual attentions as context, and decides whether the context is used for the current word/sentence generation given the current visual attention. Compared against traditional visual attention mechanism that only fixes a single visual region at each step, CAVP can attend to complex visual compositions over time. The whole image captioning model-CAVP and its subsequent language policy network-can be efficiently optimized end-to-end by using an actor-critic policy gradient method. We have demonstrated the effectiveness of CAVP by state-of-the-art performances on MS-COCO and Stanford captioning datasets, using various metrics and sensible visualizations of qualitative visual context.

Computer Vision (CV) and Natural Language Processing (NLP) are two most fundamental disciplines under a broad area of artificial intelligence (AI). CV is regarded as a field of research that explores the techniques to teach computers to see and understand digital content such as images and videos. NLP is a branch of linguistics that enables computers to process, interpret, and even generate human language. With the rise and development of deep learning over the past decade, there has been a steady momentum of innovation and breakthroughs that convincingly push the limits and improve the state-of-the-art of both vision and language modeling. An interesting observation is that the research in the two areas starts to interact, with a significant growth in both the volume of publications and extensive applications. Meanwhile, many previous experiences have shown that this can naturally build up the circle of human intelligence.

The cross-domain image captioning, which is trained on a source domain and generalized to other domains, usually faces the large domain shift problem. Although prior work has attempted to leverage both paired source and unpaired target data to minimize this shift, the performance is still unsatisfactory. One main reason lies in the large discrepancy in language expression between two domains, where diverse language styles are adopted to describe an image from different views, resulting in different semantic descriptions for an image. To tackle this problem, this paper proposes a Style-based Cross-domain Image Captioner (SCIC) which incorporates the discriminative style information into the encoder-decoder framework, and interprets an image as a special sentence according to external style instructions. Technically, we design a novel "Instruction-based LSTM", which adds the instruct gate to collect a style instruction, and then outputs a specified format according to that instruction. Two objectives are designed to train I-LSTM: 1) generating correct image descriptions and 2) generating correct styles, thus the model is expected to accurately capture the semantic meanings of an image by the special caption as well as understand the syntactic structure of the caption. We use MS-COCO as the source domain, and Oxford-102, CUB-200, Flickr30k as the target domains. Experimental results demonstrate that our model consistently outperforms the previous methods, and the style information incorporating with I-LSTM significantly improves the performance, with 5% CIDEr improvements at least on all datasets.

Scene graphs connect individual objects with visual relationships. They serve as a comprehensive scene representation for downstream multimodal tasks. However, by exploring recent progress in Scene Graph Generation (SGG), we find that the performance of recent works is highly limited by the pairwise relationship modeling by naive feature concatenation. Such pairwise features lack sufficient object interaction due to the mis-aligned object parts, resulting in non-discriminative pairwise features for visual relationship prediction. For example, naive concatenated pairwise feature usually make the model fail to discriminate betweenridingandfeedingfor object pairpersonandhorse. To this end, we design a meta-architecture— learning-to-align — for dynamic object feature concatenation. We call our model:Align R-CNN. Specifically, we introduce a novel attention-based multiple region alignment module that can be jointly optimized with SGG. Experiments on the large-scale SGG benchmark Visual Genome show that the proposedAlign R-CNNcan replace the naive feature concatenation and thus boost all the existing SGG methods.

The prevailing framework for solving referring expression grounding is based on a two-stage process: 1) detecting proposals with an object detector and 2) grounding the referent to one of the proposals. Existing two-stage solutions mostly focus on the grounding step, which aims to align the expressions with the proposals. In this paper, we argue that these methods overlook an obvious mismatch between the roles of proposals in the two stages: they generate proposals solely based on the detection confidence (i.e., expression-agnostic), hoping that the proposals contain all right instances in the expression (i.e., expression-aware). Due to this mismatch, current two-stage methods suffer from a severe performance drop between detected and ground-truth proposals. To this end, we propose Ref-NMS, which is the first method to yield expression-aware proposals at the first stage. Ref-NMS regards all nouns in the expression as critical objects, and introduces a lightweight module to predict a score for aligning each box with a critical object. These scores can guide the NMS operation to filter out the boxes irrelevant to the expression, increasing the recall of critical objects, resulting in a significantly improved grounding performance. Since Ref- NMS is agnostic to the grounding step, it can be easily integrated into any state-of-the-art two-stage method. Extensive ablation studies on several backbones, benchmarks, and tasks consistently demonstrate the superiority of Ref-NMS. Codes are available at: https://github.com/ChopinSharp/ref-nms.

Yixin Cao, Xiang Ji, Xin Lv, Juanzi Li, Yonggang Wen, Hanwang Zhang. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

We propose a causal framework to explain the catastrophic forgetting in Class-Incremental Learning (CIL) and then derive a novel distillation method that is orthogonal to the existing anti-forgetting techniques, such as data replay and feature/label distillation. We first 1) place CIL into the framework, 2) answer why the forgetting happens: the causal effect of the old data is lost in new training, and then 3) explain how the existing techniques mitigate it: they bring the causal effect back. Based on the causal framework, we propose to distill the Colliding Effect between the old and the new data, which is fundamentally equivalent to the causal effect of data replay, but without any cost of replay storage. Thanks to the causal effect analysis, we can further capture the Incremental Momentum Effect of the data stream, removing which can help to retain the old effect overwhelmed by the new data effect, and thus alleviate the forgetting of the old class in testing. Extensive experiments on three CIL benchmarks: CIFAR-100, ImageNet-Sub&Full, show that the proposed causal effect distillation can improve various state-of-the-art CIL methods by a large margin (0.72%–9.06%).1

Weakly-supervised Temporal Action Localization (WTAL) aims to detect the action segments with only video-level action labels in training. The key challenge is how to distinguish the action of interest segments from the background, which is unlabelled even on the video-level. While previous works treat the background as "curses", we consider it as "blessings". Specifically, we first use causal analysis to point out that the common localization errors are due to the unobserved confounder that resides ubiquitously in visual recognition. Then, we propose a Temporal Smoothing PCA-based (TS-PCA) deconfounder, which exploits the unlabelled background to model an observed substitute for the unobserved confounder, to remove the confounding effect. Note that the proposed deconfounder is model-agnostic and non-intrusive, and hence can be applied in any WTAL method without model re-designs. Through extensive experiments on four state-of-the-art WTAL methods, we show that the deconfounder can improve all of them on the public datasets: THUMOS-14 and ActivityNet-1.31.

VQA models may tend to rely on language bias as a shortcut and thus fail to sufficiently learn the multi-modal knowledge from both vision and language. Recent debiasing methods proposed to exclude the language prior during inference. However, they fail to disentangle the "good" language context and "bad" language bias from the whole. In this paper, we investigate how to mitigate language bias in VQA. Motivated by causal effects, we proposed a novel counterfactual inference framework, which enables us to capture the language bias as the direct causal effect of questions on answers and reduce the language bias by subtracting the direct language effect from the total causal effect. Experiments demonstrate that our proposed counterfactual inference framework 1) is general to various VQA backbones and fusion strategies, 2) achieves competitive performance on the language-bias sensitive VQA-CP dataset while performs robustly on the balanced VQA v2 dataset without any augmented data. The code is available at https://github.com/yuleiniu/cfvqa.

We present a novel attention mechanism: Causal Attention (CATT), to remove the ever-elusive confounding effect in existing attention-based vision-language models. This effect causes harmful bias that misleads the attention module to focus on the spurious correlations in training data, damaging the model generalization. As the confounder is unobserved in general, we use the front-door adjustment to realize the causal intervention, which does not require any knowledge on the confounder. Specifically, CATT is implemented as a combination of 1) In-Sample Attention (IS-ATT) and 2) Cross-Sample Attention (CS-ATT), where the latter forcibly brings other samples into every IS-ATT, mimicking the causal intervention. CATT abides by the QK-V convention and hence can replace any attention module such as top-down attention and self-attention in Transformers. CATT improves various popular attention-based vision-language models by considerable margins. In particular, we show that CATT has great potential in large-scale pre-training, e.g., it can promote the lighter LXMERT [57], which uses fewer data and less computational power, comparable to the heavier UNITER [14]. Code is published in https://github.com/yangxuntu/lxmertcatt.

We present a novel counterfactual framework for both Zero-Shot Learning (ZSL) and Open-Set Recognition (OSR), whose common challenge is generalizing to the unseen-classes by only training on the seen-classes. Our idea stems from the observation that the generated samples for unseen-classes are often out of the true distribution, which causes severe recognition rate imbalance between the seen-class (high) and unseen-class (low). We show that the key reason is that the generation is not Counterfactual Faithful, and thus we propose a faithful one, whose generation is from the sample-specific counterfactual question: What would the sample look like, if we set its class attribute to a certain class, while keeping its sample attribute unchanged? Thanks to the faithfulness, we can apply the Consistency Rule to perform unseen/seen binary classification, by asking: Would its counterfactual still look like itself? If "yes", the sample is from a certain class, and "no" otherwise. Through extensive experiments on ZSL and OSR, we demonstrate that our framework effectively mitigates the seen/unseen imbalance and hence significantly improves the overall performance. Note that this framework is orthogonal to existing methods, thus, it can serve as a new baseline to evaluate how ZSL/OSR models generalize. Codes are available at https://github.com/yue-zhongqi/gcm-cf.

Multi-hop Question Answering (QA) is a challenging task because it requires precise reasoning with entity relations at every step towards the answer.The relations can be represented in terms of labels in knowledge graph (e.g., spouse) or text in text corpus (e.g., they have been married for 26 years).Existing models usually infer the answer by predicting the sequential relation path or aggregating the hidden graph features.The former is hard to optimize, and the latter lacks interpretability.In this paper, we propose Trans-ferNet, an effective and transparent model for multi-hop QA, which supports both label and text relations in a unified framework.Trans-ferNet jumps across entities at multiple steps.At each step, it attends to different parts of the question, computes activated scores for relations, and then transfer the previous entity scores along activated relations in a differentiable way.We carry out extensive experiments on three datasets and demonstrate that TransferNet surpasses the state-of-the-art models by a large margin.In particular, on MetaQA, it achieves 100% accuracy in 2-hop and 3-hop questions.By qualitative analysis, we show that TransferNet has transparent and interpretable intermediate results.

Attention module does not always help deep models learn causal features that are robust in any confounding context, e.g., a foreground object feature is invariant to different backgrounds. This is because the confounders trick the attention to capture spurious correlations that benefit the prediction when the training and testing data are IID (identical & independent distribution); while harm the prediction when the data are OOD (out-of-distribution). The sole fundamental solution to learn causal attention is by causal intervention, which requires additional annotations of the confounders, e.g., a "dog" model is learned within "grass+dog" and "road+dog" respectively, so the "grass" and "road" contexts will no longer confound the "dog" recognition. However, such annotation is not only prohibitively expensive, but also inherently problematic, as the confounders are elusive in nature. In this paper, we propose a causal attention module (CaaM) that self-annotates the confounders in unsupervised fashion. In particular, multiple CaaMs can be stacked and integrated in conventional attention CNN and self-attention Vision Transformer. In OOD settings, deep models with CaaM outperform those without it significantly; even in IID settings, the attention localization is also improved by CaaM, showing a great potential in applications that require robust visual saliency. Codes are available at https://github.com/Wangt-CN/CaaM.

We propose an Auto-Parsing Network (APN) to discover and exploit the input data’s hidden tree structures for improving the effectiveness of the Transformer-based vision-language systems. Specifically, we impose a Probabilistic Graphical Model (PGM) parameterized by the attention operations on each self-attention layer to incorporate sparse assumption. We use this PGM to softly segment an input sequence into a few clusters where each cluster can be treated as the parent of the inside entities. By stacking these PGM constrained self-attention layers, the clusters in a lower layer compose into a new sequence, and the PGM in a higher layer will further segment this sequence. Iteratively, a sparse tree can be implicitly parsed, and this tree’s hierarchical knowledge is incorporated into the transformed embeddings, which can be used for solving the target vision-language tasks. Specifically, we showcase that our APN can strengthen Transformer based networks in two major vision-language tasks: Captioning and Visual Question Answering. Also, a PGM probability-based parsing algorithm is developed by which we can discover what the hidden structure of input is during the inference.

Existing Unsupervised Domain Adaptation (UDA) literature adopts the covariate shift and conditional shift assumptions, which essentially encourage models to learn common features across domains. However, due to the lack of supervision in the target domain, they suffer from the semantic loss: the feature will inevitably lose non-discriminative semantics in source domain, which is however discriminative in target domain. We use a causal view—transportability theory [40]—to identify that such loss is in fact a confounding effect, which can only be removed by causal intervention. However, the theoretical solution provided by transportability is far from practical for UDA, because it requires the stratification and representation of the unobserved confounder that is the cause of the domain gap. To this end, we propose a practical solution: Transporting Causal Mechanisms (TCM), to identify the confounder stratum and representations by using the domain-invariant disentangled causal mechanisms, which are discovered in an unsupervised fashion. Our TCM is both theoretically and empirically grounded. Extensive experiments show that TCM achieves state-of-the-art performance on three challenging UDA benchmarks: ImageCLEF-DA, Office-Home, and VisDA-2017. Codes are available at https://github.com/yue-zhongqi/tcm.

In this paper, we seek reasons for the two major failure cases in Semantic Segmentation (SS): 1) missing small objects or minor object parts, and 2) mislabeling minor parts of large objects as wrong classes. We have an interesting finding that Failure-1 is due to the underuse of detailed features and Failure-2 is due to the underuse of visual contexts. To help the model learn a better trade-off, we introduce several Self-Regulation (SR) losses for training SS neural networks. By "self", we mean that the losses are from the model per se without using any additional data or supervision. By applying the SR losses, the deep layer features are regulated by the shallow ones to preserve more details; meanwhile, shallow layer classification logits are regulated by the deep ones to capture more semantics. We conduct extensive experiments on both weakly and fully supervised SS tasks, and the results show that our approach consistently surpasses the baselines. We also validate that SR losses are easy to implement in various state-of-the-art SS models, e.g., SPGNet [7] and OCRNet [62], incurring little computational overhead during training and none for testing1.

The fine-tuning of pre-trained language models has a great success in many NLP fields. Yet, it is strikingly vulnerable to adversarial examples, e.g., word substitution attacks using only synonyms can easily fool a BERT-based sentiment analysis model. In this paper, we demonstrate that adversarial training, the prevalent defense technique, does not directly fit a conventional fine-tuning scenario, because it suffers severely from catastrophic forgetting: failing to retain the generic and robust linguistic features that have already been captured by the pre-trained model. In this light, we propose Robust Informative Fine-Tuning (RIFT), a novel adversarial fine-tuning method from an information-theoretical perspective. In particular, RIFT encourages an objective model to retain the features learned from the pre-trained model throughout the entire fine-tuning process, whereas a conventional one only uses the pre-trained weights for initialization. Experimental results show that RIFT consistently outperforms the state-of-the-arts on two popular NLP tasks: sentiment analysis and natural language inference, under different attacks across various pre-trained language models.

Question answering (QA) models are well-known to exploit data bias, e.g., the language prior in visual QA and the position bias in reading comprehension. Recent debiasing methods achieve good out-of-distribution (OOD) generalizability with a considerable sacrifice of the in-distribution (ID) performance. Therefore, they are only applicable in domains where the test distribution is known in advance. In this paper, we present a novel debiasing method called Introspective Distillation (IntroD) to make the best of both worlds for QA. Our key technical contribution is to blend the inductive bias of OOD and ID by introspecting whether a training sample fits in the factual ID world or the counterfactual OOD one. Experiments on visual QA datasets VQA v2, VQA-CP, and reading comprehension dataset SQuAD demonstrate that our proposed IntroD maintains the competitive OOD performance compared to other debiasing methods, while sacrificing little or even achieving better ID performance compared to the non-debiasing ones.

A good visual representation is an inference map from observations (images) to features (vectors) that faithfully reflects the hidden modularized generative factors (semantics). In this paper, we formulate the notion of "good" representation from a group-theoretic view using Higgins' definition of disentangled representation, and show that existing Self-Supervised Learning (SSL) only disentangles simple augmentation features such as rotation and colorization, thus unable to modularize the remaining semantics. To break the limitation, we propose an iterative SSL algorithm: Iterative Partition-based Invariant Risk Minimization (IP-IRM), which successfully grounds the abstract semantics and the group acting on them into concrete contrastive learning. At each iteration, IP-IRM first partitions the training samples into two subsets that correspond to an entangled group element. Then, it minimizes a subset-invariant contrastive loss, where the invariance guarantees to disentangle the group element. We prove that IP-IRM converges to a fully disentangled representation and show its effectiveness on various benchmarks. Codes are available at https://github.com/Wangt-CN/IP-IRM.

Recommendation is a prevalent and critical service in information systems. To provide personalized suggestions to users, industry players embrace machine learning, more specifically, building predictive models based on the click behavior data. This is known as the Click-Through Rate (CTR) prediction, which has become the gold standard for building personalized recommendation service. However, we argue that there is a significant gap between clicks and user satisfaction --- it is common that a user is "cheated" to click an item by the attractive title/cover of the item. This will severely hurt user's trust on the system if the user finds the actual content of the clicked item disappointing. What's even worse, optimizing CTR models on such flawed data will result in the Matthew Effect, making the seemingly attractive but actually low-quality items be more frequently recommended.

We focus on grounding (i.e., localizing or linking) referring expressions in images, e.g., "largest elephant standing behind baby elephant". This is a general yet challenging vision-language task since it does not only require the localization of objects, but also the multimodal comprehension of context - visual attributes (e.g., "largest", "baby") and relationships (e.g., "behind") that help to distinguish the referent from other objects, especially those of the same category. Due to the exponential complexity involved in modeling the context associated with multiple image regions, existing work oversimplifies this task to pairwise region modeling by multiple instance learning. In this paper, we propose a variational Bayesian method, called Variational Context, to solve the problem of complex context modeling in referring expression grounding. Specifically, our framework exploits the reciprocal relation between the referent and context, i.e., either of them influences estimation of the posterior distribution of the other, and thereby the search space of context can be greatly reduced. In addition to reciprocity, our framework considers the semantic information of context, i.e., the referring expression can be reproduced based on the estimated context. We also extend the model to unsupervised setting where no annotation for the referent is available. Extensive experiments on various benchmarks show consistent improvement over state-of-the-art methods in both supervised and unsupervised settings.

Vision and language understanding is one of the most fundamental and difficult tasks in Multimedia Intelligence. Simultaneously Visual Question Answering (VQA) is even more challenging since it requires complex reasoning steps to the correct answer. To achieve this, Neural Module Network (NMN) and its variants rely on parsing the natural language question into a module layout (i.e., a problem-solving program). In particular, this process follows a feedforward encoder-decoder pipeline: the encoder embeds the question into a static vector and the decoder generates the layout. However, we argue that such conventional encoder-decoder neglects the dynamic nature of question comprehension (i.e., we should attend to different words from step to step) and per-module intermediate results (i.e., we should discard module performing badly) in the reasoning steps. In this paper, we present a novel NMN, called Self-Adaptive Neural Module Transformer (SANMT), which adaptively adjusts both of the question feature encoding and the layout decoding by considering intermediate Q&A results. Specifically, we encode the intermediate results with the given question features by a novel transformer module to generate dynamic question feature embedding which evolves over reasoning steps. Besides, the transformer utilizes the intermediate results from each reasoning step to guide subsequent layout arrangement. Extensive experimental evaluations demonstrate the superiority of the proposed SANMT over NMN and its variants on four challenging benchmarks, including CLEVR, CLEVR-CoGenT, VQAv1.0, and VQAv2.0 (on average the relative improvement over NMN are 1.5, 2.3, 0.7 and 0.5 points with respect to accuracy).

introduction Share on Introduction to the Special Issue on Fine-grained Visual Computing Editors: Shaohua Wan Zhongnan University of Economics and Law Zhongnan University of Economics and LawView Profile , Zan Gao Qilu University of Technology Qilu University of TechnologyView Profile , Hanwang Zhang Nanyang Technological University Nanyang Technological UniversityView Profile , Xiaojun Chang Monash University Monash UniversityView Profile Authors Info & Claims ACM Transactions on Multimedia Computing, Communications, and ApplicationsVolume 17Issue 1sJanuary 2021 Article No.: 11pp 1–3https://doi.org/10.1145/3447532Online:31 March 2021Publication History 0citation70DownloadsMetricsTotal Citations0Total Downloads70Last 12 Months24Last 6 weeks1 Get Citation AlertsNew Citation Alert added!This alert has been successfully added and will be sent to:You will be notified whenever a record that you have chosen has been cited.To manage your alert preferences, click on the button below.Manage my Alerts New Citation Alert!Please log in to your account Save to BinderSave to BinderCreate a New BinderNameCancelCreateExport CitationPublisher SiteGet Access

We formulate a practical yet challenging problem: General Partial Label Learning (GPLL). Compared to the traditional Partial Label Learning (PLL) problem, GPLL relaxes the supervision assumption from instance-level — a label set partially labels an instance — to group-level: 1) a label set partially labels a group of instances, where the within-group instance-label link annotations are missing, and 2) cross-group links are allowed — instances in a group may be partially linked to the label set from another group. Such ambiguous group-level supervision is more practical in real-world scenarios as additional annotation on the instance-level is no longer required, e.g., face-naming in videos where the group consists of faces in a frame, labeled by a name set in the corresponding caption. In this paper, we propose a novel graph convolutional network (GCN) called Dual Bipartite Graph Autoencoder (DB-GAE) to tackle the label ambiguity challenge of GPLL. First, we exploit the cross-group correlations to represent the instance groups as dual bipartite graphs: within-group and cross-group, which reciprocally complements each other to resolve the linking ambiguities. Second, we design a GCN autoencoder to encode and decode them, where the decodings are considered as the refined results. It is worth noting that DB-GAE is self-supervised and transductive, as it only uses the group-level supervision without a separate offline training stage. Extensive experiments on two real-world datasets demonstrate that DB-GAE significantly outperforms the best baseline over absolute 0.159 F1-score and 24.8% accuracy. We further offer analysis on various levels of label ambiguities.

Despite Visual Question Answering (VQA) has realized impressive progress over the last few years, today's VQA models tend to capture superficial linguistic correlations in the train set and fail to generalize to the test set with different QA distributions. To reduce the language biases, several recent works introduce an auxiliary question-only model to regularize the training of targeted VQA model, and achieve dominating performance on VQA-CP. However, since the complexity of design, current methods are unable to equip the ensemble-based models with two indispensable characteristics of an ideal VQA model: 1) visual-explainable: the model should rely on the right visual regions when making decisions. 2) question-sensitive: the model should be sensitive to the linguistic variations in question. To this end, we propose a model-agnostic Counterfactual Samples Synthesizing (CSS) training scheme. The CSS generates numerous counterfactual training samples by masking critical objects in images or words in questions, and assigning different ground-truth answers. After training with the complementary samples (ie, the original and generated samples), the VQA models are forced to focus on all critical objects and words, which significantly improves both visual-explainable and question-sensitive abilities. In return, the performance of these models is further boosted. Extensive ablations have shown the effectiveness of CSS. Particularly, by building on top of the model LMH, we achieve a record-breaking performance of 58.95% on VQA-CP v2, with 6.5% gains.

Visual dialog is a challenging task that requires the comprehension of the semantic dependencies among implicit visual and textual contexts. This task can refer to the relation inference in a graphical model with sparse contexts and unknown graph structure (relation descriptor), and how to model the underlying context-aware relation inference is critical. To this end, we propose a novel Context-Aware Graph (CAG) neural network. Each node in the graph corresponds to a joint semantic feature, including both object-based (visual) and history-related (textual) context representations. The graph structure (relations in dialog) is iteratively updated using an adaptive top-K message passing mechanism. Specifically, in every message passing step, each node selects the most K relevant nodes, and only receives messages from them. Then, after the update, we impose graph attention on all the nodes to get the final graph embedding and infer the answer. In CAG, each node has dynamic relations in the graph (different related K neighbor nodes), and only the most relevant nodes are attributive to the context-aware relational graph inference. Experimental results on VisDial v0.9 and v1.0 datasets show that CAG outperforms comparative methods. Visualization results further validate the interpretability of our method.

Filter pruning has been widely applied to neural network compression and acceleration. Existing methods usually utilize pre-defined pruning criteria, such as Lp-norm, to prune unimportant filters. There are two major limitations to these methods. First, existing methods fail to consider the variety of filter distribution across layers. To extract features of the coarse level to the fine level, the filters of different layers have various distributions. Therefore, it is not suitable to utilize the same pruning criteria to different functional layers. Second, prevailing layer-by-layer pruning methods process each layer independently and sequentially, failing to consider that all the layers in the network collaboratively make the final prediction. In this paper, we propose Learning Filter Pruning Criteria (LFPC) to solve the above problems. Specifically, we develop a differentiable pruning criteria sampler. This sampler is learnable and optimized by the validation loss of the pruned network obtained from the sampled criteria. In this way, we could adaptively select the appropriate pruning criteria for different functional layers. Besides, when evaluating the sampled criteria, LFPC comprehensively consider the contribution of all the layers at the same time. Experiments validate our approach on three image classification benchmarks. Notably, on ILSVRC-2012, our LFPC reduces more than 60% FLOPs on ResNet-50 with only 0.83% top-5 accuracy loss.

Real-world visual recognition requires handling the extreme sample imbalance in large-scale long-tailed data. We propose a “divide&conquer” strategy for the challenging LVIS task: divide the whole data into balanced parts and then apply incremental learning to conquer each one. This derives a novel learning paradigm: class-incremental few-shot learning, which is especially effective for the challenge evolving over time: 1) the class imbalance among the old class knowledge review and 2) the few-shot data in new-class learning. We call our approach Learning to Segment the Tail (LST). In particular, we design an instance-level balanced replay scheme, which is a memory-efficient approximation to balance the instance-level samples from the old-class images. We also propose to use a meta-module for new-class learning, where the module parameters are shared across incremental phases, gaining the learning-to-learn knowledge incrementally, from the data-rich head to the data-poor tail. We empirically show that: at the expense of a little sacrifice of head-class forgetting, we can gain a significant 8.3% AP improvement for the tail classes with less than 10 instances, achieving an overall 2.0% AP boost for the whole 1,230 classes.

This paper unravels the design tricks adopted by us, the champion team MReaL-BDAI, for Visual Dialog Challenge 2019: two causal principles for improving Visual Dialog (VisDial). By "improving", we mean that they can promote almost every existing VisDial model to the state-of-the-art performance on the leader-board. Such a major improvement is only due to our careful inspection on the causality behind the model and data, finding that the community has overlooked two causalities in VisDial. Intuitively, Principle 1 suggests: we should remove the direct input of the dialog history to the answer model, otherwise a harmful shortcut bias will be introduced; Principle 2 says: there is an unobserved confounder for history, question, and answer, leading to spurious correlations from training data. In particular, to remove the confounder suggested in Principle 2, we propose several causal intervention algorithms, which make the training fundamentally different from the traditional likelihood estimation. Note that the two principles are model-agnostic, so they are applicable in any VisDial model.

Today's scene graph generation (SGG) task is still far from practical, mainly due to the severe training bias, e.g., collapsing diverse "human walk on / sit on / lay on beach" into "human on beach". Given such SGG, the down-stream tasks such as VQA can hardly infer better scene structures than merely a bag of objects. However, debiasing in SGG is not trivial because traditional debiasing methods cannot distinguish between the good and bad bias, e.g., good context prior (e.g., "person read book" rather than "eat") and bad long-tailed bias (e.g., "near" dominating "behind / in front of"). In this paper, we present a novel SGG framework based on causal inference but not the conventional likelihood. We first build a causal graph for SGG, and perform traditional biased training with the graph. Then, we propose to draw the counterfactual causality from the trained graph to infer the effect from the bad bias, which should be removed. In particular, we use Total Direct Effect (TDE) as the proposed final predicate score for unbiased SGG. Note that our framework is agnostic to any SGG model and thus can be widely applied in the community who seeks unbiased predictions. By using the proposed Scene Graph Diagnosis toolkit on the SGG benchmark Visual Genome and several prevailing models, we observed significant improvements over the previous state-of-the-art methods.

We present a novel unsupervised feature representation learning method, Visual Commonsense Region-based Convolutional Neural Network (VC R-CNN), to serve as an improved visual region encoder for high-level tasks such as captioning and VQA. Given a set of detected object regions in an image (e.g., using Faster R-CNN), like any other unsupervised feature learning methods (e.g., word2vec), the proxy training objective of VC R-CNN is to predict the contextual objects of a region. However, they are fundamentally different: the prediction of VC R-CNN is by using causal intervention: P(Y|do(X)), while others are by using the conventional likelihood: P(Y|X). This is also the core reason why VC R-CNN can learn “sense-making” knowledge like chair can be sat - while not just “common” co-occurrences such as chair is likely to exist if table is observed. We extensively apply VC R-CNN features in prevailing models of three popular tasks: Image Captioning, VQA, and VCR, and observe consistent performance boosts across them, achieving many new state-of-the-arts.

Visual attention not only improves the performance of image captioners, but also serves as a visual interpretation to qualitatively measure the caption rationality and model transparency. Specifically, we expect that a captioner can fix its attentive gaze on the correct objects while generating the corresponding words. This ability is also known as grounded image captioning. However, the grounding accuracy of existing captioners is far from satisfactory.To improve the grounding accuracy while retaining the captioning quality, it is expensive to collect the word-region alignment as strong supervision.To this end, we propose a Part-of-Speech (POS) enhanced image-text matching model (SCAN[24]): POS-SCAN, as the effective knowledge distillation for more grounded image captioning. The benefits are two-fold: 1) given a sentence and an image, POS-SCAN can ground the objects more accurately than SCAN; 2) POS-SCAN serves as a word-region alignment regularization for the captioner's visual attention module. By showing benchmark experimental results, we demonstrate that conventional image captioners equipped with POS-SCAN can significantly improve the grounding accuracy without strong supervision. Last but not the least, we explore the indispensable Self-Critical Sequence Training (SCST[46]) in the context of grounded image captioning and show that the image-text matching score can serve as a reward for more grounded captioning1.

Occluded faces-as a common scene in real life-have a significant negative impact on most face recognition systems. Existing methods try to remove the occlusions by a single-stage generative adversarial network (GAN), which is unaware of the occlusion and thus has difficulties in generalizing to a large variety of occlusion types, e.g., different objects at various positions. To this end, we propose the two-stage Occlusion-Aware GAN (OA-GAN), where the first GAN is for disentangling the occlusions, which will be served as the additional input of the second GAN for synthesizing the final de-occluded faces. In this way, our two-stage model can handle diverse occlusions in the wild and is naturally more explainable because of its awareness of the occluded objects. Extensive experiments on both synthetic and real-world datasets validate the superiority of the two-stage OAGAN design. Furthermore, by applying the generated de-occluded faces to facial expression recognition (FER) systems, we find that our two-stage de-occlusion process significantly increases the accuracy of FER under occlusion.

When we humans tell a long paragraph about an image, we usually first implicitly compose a mental "script'' and then comply with it to generate the paragraph. Inspired by this, we render the modern encoder-decoder based image paragraph captioning model such ability by proposing Hierarchical Scene Graph Encoder-Decoder (HSGED) for generating coherent and distinctive paragraphs. In particular, we use the image scene graph as the "script" to incorporate rich semantic knowledge and, more importantly, the hierarchical constraints into the model. Specifically, we design a sentence scene graph RNN (SSG-RNN) to generate sub-graph level topics, which constrain the word scene graph RNN (WSG-RNN) to generate the corresponding sentences. We propose irredundant attention in SSG-RNN to improve the possibility of abstracting topics from rarely described sub-graphs and inheriting attention in WSG-RNN to generate more grounded sentences with the abstracted topics, both of which give rise to more distinctive paragraphs. An efficient sentence-level loss is also proposed for encouraging the sequence of generated sentences to be similar to that of the ground-truth paragraphs. We validate HSGED on Stanford image paragraph dataset and show that it not only achieves a new state-of-the-art 36.02 CIDEr-D, but also generates more coherent and distinctive paragraphs under various metrics.

As the class size grows, maintaining a balanced dataset across many classes is challenging because the data are long-tailed in nature; it is even impossible when the sample-of-interest co-exists with each other in one collectable unit, e.g., multiple visual instances in one image. Therefore, long-tailed classification is the key to deep learning at scale. However, existing methods are mainly based on re-weighting/re-sampling heuristics that lack a fundamental theory. In this paper, we establish a causal inference framework, which not only unravels the whys of previous methods, but also derives a new principled solution. Specifically, our theory shows that the SGD momentum is essentially a confounder in long-tailed classification. On one hand, it has a harmful causal effect that misleads the tail prediction biased towards the head. On the other hand, its induced mediation also benefits the representation learning and head prediction. Our framework elegantly disentangles the paradoxical effects of the momentum, by pursuing the direct causal effect caused by an input sample. In particular, we use causal intervention in training, and counterfactual reasoning in inference, to remove the bad'' while keep thegood''. We achieve new state-of-the-arts on three long-tailed visual recognition benchmarks: Long-tailed CIFAR-10/-100, ImageNet-LT for image classification and LVIS for instance segmentation.

We uncover an ever-overlooked deficiency in the prevailing Few-Shot Learning (FSL) methods: the pre-trained knowledge is indeed a confounder that limits the performance. This finding is rooted from our causal assumption: a Structural Causal Model (SCM) for the causalities among the pre-trained knowledge, sample features, and labels. Thanks to it, we propose a novel FSL paradigm: Interventional Few-Shot Learning (IFSL). Specifically, we develop three effective IFSL algorithmic implementations based on the backdoor adjustment, which is essentially a causal intervention towards the SCM of many-shot learning: the upper-bound of FSL in a causal view. It is worth noting that the contribution of IFSL is orthogonal to existing fine-tuning and meta-learning based FSL methods, hence IFSL can improve all of them, achieving a new 1-/5-shot state-of-the-art on miniImageNet, tieredImageNet, and cross-domain CUB. Code is released at https://github.com/yue-zhongqi/ifsl.

We present a causal inference framework to improve Weakly-Supervised Semantic Segmentation (WSSS). Specifically, we aim to generate better pixel-level pseudo-masks by using only image-level labels -- the most crucial step in WSSS. We attribute the cause of the ambiguous boundaries of pseudo-masks to the confounding context, e.g., the correct image-level classification of "horse" and "person" may be not only due to the recognition of each instance, but also their co-occurrence context, making the model inspection (e.g., CAM) hard to distinguish between the boundaries. Inspired by this, we propose a structural causal model to analyze the causalities among images, contexts, and class labels. Based on it, we develop a new method: Context Adjustment (CONTA), to remove the confounding bias in image-level classification and thus provide better pseudo-masks as ground-truth for the subsequent segmentation model. On PASCAL VOC 2012 and MS-COCO, we show that CONTA boosts various popular WSSS methods to new state-of-the-arts.

In this paper, we introduce a stochastic dynamics video infilling (SDVI) framework to generate frames between long intervals in a video. Our task differs from video interpolation which aims to produce transitional frames for a short interval between every two frames and increase the temporal resolution. Our task, namely video infilling, however, aims to infill long intervals with plausible frame sequences. Our framework models the infilling as a constrained stochastic generation process and sequentially samples dynamics from the inferred distribution. SDVI consists of two parts: (1) a bi-directional constraint propagation module to guarantee the spatial-temporal coherence among frames, (2) a stochastic sampling process to generate dynamics from the inferred distributions. Experimental results show that SDVI can generate clear frame sequences with varying contents. Moreover, motions in the generated sequence are realistic and able to transfer smoothly from the given start frame to the terminal frame.

Many achievements have been made on learning to hash for uni-modal and cross-modal retrieval. However, it is still an unsolved problem that how to directly and efficiently learn discriminative discrete hash codes for the multimedia retrieval, where both query and database samples are represented with heterogeneous multi-modal features. With this motivation, we propose a Fast Discrete Collaborative Multi-modal Hashing (FDCMH) method in this paper. We first propose an efficient collaborative multi-modal mapping that first transforms heterogeneous multi-modal features into the unified factors to exploit the complementarity of multi-modal features and preserve the semantic correlations in multiple modalities with linear computation and space complexity. Such shared factors also bridge the heterogeneous modality gap and remove the inter-modality redundancy. Further, we develop an asymmetric hashing learning module to simultaneously correlate the learned hash codes with low-level data distribution and high-level semantics. In particular, this design could avoid the challenging symmetric semantic matrix factorization and O(n2) memory cost (n is the number of training samples). It can support both computation and memory efficient discrete hash optimization. Experiments on several public multimedia retrieval datasets demonstrate the superiority of the proposed approach compared with state-of-the-art hashing techniques, in terms of both model learning efficiency and retrieval accuracy.

Image captioning is one of the most challenging tasks in AI because it requires an understanding of both complex visuals and natural language. Because image captioning is essentially a sequential prediction task, recent advances in image captioning have used reinforcement learning (RL) to better explore the dynamics of word-by-word generation. However, the existing RL-based image captioning methods rely primarily on a single policy network and reward function-an approach that is not well matched to the multi-level (word and sentence) and multi-modal (vision and language) nature of the task. To solve this problem, we propose a novel multi-level policy and reward RL framework for image captioning that can be easily integrated with RNN-based captioning models, language metrics, or visual-semantic functions for optimization. Specifically, the proposed framework includes two modules: 1) a multi-level policy network that jointly updates the word- and sentence-level policies for word generation; and 2) a multi-level reward function that collaboratively leverages both a vision-language reward and a language-language reward to guide the policy. Furthermore, we propose a guidance term to bridge the policy and the reward for RL optimization. The extensive experiments on the MSCOCO and Flickr30k datasets and the analyses show that the proposed framework achieves competitive performances on a variety of evaluation metrics. In addition, we conduct ablation studies on multiple variants of the proposed framework and explore several representative image captioning models and metrics for the word-level policy network and the language-language reward function to evaluate the generalization ability of the proposed framework.

We propose DeepChannel, a robust, data-efficient, and interpretable neural model for extractive document summarization. Given any document-summary pair, we estimate a salience score, which is modeled using an attention-based deep neural network, to represent the salience degree of the summary for yielding the document. We devise a contrastive training strategy to learn the salience estimation network, and then use the learned salience score as a guide and iteratively extract the most salient sentences from the document as our generated summary. In experiments, our model not only achieves state-of-the-art ROUGE scores on CNN/Daily Mail dataset, but also shows strong robustness in the out-of-domain test on DUC2007 test set. Moreover, our model reaches a ROUGE-1 F-1 score of 39.41 on CNN/Daily Mail test set with merely 1/100 training set, demonstrating a tremendous data efficiency.

Visual dialog is a challenging vision-language task, which requires the agent to answer multi-round questions about an image. It typically needs to address two major problems: (1) How to answer visually-grounded questions, which is the core challenge in visual question answering (VQA); (2) How to infer the co-reference between questions and the dialog history. An example of visual co-reference is: pronouns (e.g., “they”) in the question (e.g., “Are they on or off?”) are linked with nouns (e.g., “lamps”) appearing in the dialog history (e.g., “How many lamps are there?”) and the object grounded in the image. In this work, to resolve the visual co-reference for visual dialog, we propose a novel attention mechanism called Recursive Visual Attention (RvA). Specifically, our dialog agent browses the dialog history until the agent has sufficient confidence in the visual co-reference resolution, and refines the visual attention recursively. The quantitative and qualitative experimental results on the large-scale VisDial v0.9 and v1.0 datasets demonstrate that the proposed RvA not only outperforms the state-of-the-art methods, but also achieves reasonable recursion and interpretable attention maps without additional annotations. The code is available at https://github.com/yuleiniu/rva.

We aim to dismantle the prevalent black-box neural architectures used in complex visual reasoning tasks, into the proposed eXplainable and eXplicit Neural Modules (XNMs), which advance beyond existing neural module networks towards using scene graphs - objects as nodes and the pairwise relationships as edges - for explainable and explicit reasoning with structured knowledge. XNMs allow us to pay more attention to teach machines how to "think'', regardless of what they "look''. As we will show in the paper, by using scene graphs as an inductive bias, 1) we can design XNMs in a concise and flexible fashion, i.e., XNMs merely consist of 4 meta-types, which significantly reduce the number of parameters by 10 to 100 times, and 2) we can explicitly trace the reasoning-flow in terms of graph attentions. XNMs are so generic that they support a wide range of scene graph implementations with various qualities. For example, when the graphs are detected perfectly, XNMs achieve 100% accuracy on both CLEVR and CLEVR CoGenT, establishing an empirical performance upper-bound for visual reasoning; when the graphs are noisily detected from real-world images, XNMs are still robust to achieve a competitive 67.5% accuracy on VQAv2.0, surpassing the popular bag-of-objects attention models without graph structures.

We propose to compose dynamic tree structures that place the objects in an image into a visual context, helping visual reasoning tasks such as scene graph generation and visual Q&A. Our visual context tree model, dubbed VCTree, has two key advantages over existing structured object representations including chains and fully-connected graphs: 1) The efficient and expressive binary tree encodes the inherent parallel/hierarchical relationships among objects, e.g., ``clothes'' and ``pants'' are usually co-occur and belong to ``person''; 2) the dynamic structure varies from image to image and task to task, allowing more content-/task-specific message passing among objects. To construct a VCTree, we design a score function that calculates the task-dependent validity between each object pair, and the tree is the binary version of the maximum spanning tree from the score matrix. Then, visual contexts are encoded by bidirectional TreeLSTM and decoded by task-specific models. We develop a hybrid learning procedure which integrates end-task supervised learning and the tree structure reinforcement learning, where the former's evaluation result serves as a self-critic for the latter's structure exploration. Experimental results on two benchmarks, which require reasoning over contexts: Visual Genome for scene graph generation and VQA2.0 for visual Q&A, show that VCTree outperforms state-of-the-art results while discovering interpretable visual context structures.

We propose Scene Graph Auto-Encoder (SGAE) that incorporates the language inductive bias into the encoder-decoder image captioning framework for more human-like captions. Intuitively, we humans use the inductive bias to compose collocations and contextual inference in discourse. For example, when we see the relation "person on bike'', it is natural to replace "on'' with "ride'' and infer "person riding bike on a road'' even the "road'' is not evident. Therefore, exploiting such bias as a language prior is expected to help the conventional encoder-decoder models less likely to overfit to the dataset bias and focus on reasoning. Specifically, we use the scene graph - a directed graph (G) where an object node is connected by adjective nodes and relationship nodes - to represent the complex structural layout of both image (I) and sentence (S). In the textual domain, we use SGAE to learn a dictionary (D) that helps to reconstruct sentences in the S → G → D → S pipeline, where D encodes the desired language prior; in the vision-language domain, we use the shared D to guide the encoder-decoder in the I → G → D → S pipeline. Thanks to the scene graph representation and shared dictionary, the inductive bias is transferred across domains in principle. We validate the effectiveness of SGAE on the challenging MS-COCO image captioning benchmark, e.g., our SGAE-based single-model achieves a new state-of-the-art 127.8 CIDEr-D on the Karpathy split, and a competitive 125.5 CIDEr-D (c40) on the official server even compared to other ensemble models. Code has been made available at: https://github.com/yangxuntu/SGAE.

Scene graphs --- objects as nodes and visual relationships as edges --- describe the whereabouts and interactions of objects in an image for comprehensive scene understanding. To generate coherent scene graphs, almost all existing methods exploit the fruitful visual context by modeling message passing among objects. For example, ``person'' on ``bike'' can help to determine the relationship ``ride'', which in turn contributes to the confidence of the two objects. However, we argue that the visual context is not properly learned by using the prevailing cross-entropy based supervised learning paradigm, which is not sensitive to graph inconsistency: errors at the hub or non-hub nodes should not be penalized equally. To this end, we propose a Counterfactual critic Multi-Agent Training (CMAT) approach. CMAT is a multi-agent policy gradient method that frames objects into cooperative agents, and then directly maximizes a graph-level metric as the reward. In particular, to assign the reward properly to each agent, CMAT uses a counterfactual baseline that disentangles the agent-specific reward by fixing the predictions of other agents. Extensive validations on the challenging Visual Genome benchmark show that CMAT achieves a state-of-the-art performance by significant gains under various settings and metrics.

Visual grounding, a task to ground (i.e., localize) natural language in images, essentially requires composite visual reasoning. However, existing methods over-simplify the composite nature of language into a monolithic sentence embedding or a coarse composition of subject-predicate-object triplet. In this paper, we propose to ground natural language in an intuitive, explainable, and composite fashion as it should be. In particular, we develop a novel modular network called Neural Module Tree network (NMTree) that regularizes the visual grounding along the dependency parsing tree of the sentence, where each node is a neural module that calculates visual attention according to its linguistic feature, and the grounding score is accumulated in a bottom-up direction where as needed. NMTree disentangles the visual grounding from the composite reasoning, allowing the former to only focus on primitive and easy-to-generalize patterns. To reduce the impact of parsing errors, we train the modules and their assembly end-to-end by using the Gumbel-Softmax approximation and its straight-through gradient estimator, accounting for the discrete nature of module assembly. Overall, the proposed NMTree consistently outperforms the state-of-the-arts on several benchmarks. Qualitative results show explainable grounding score calculation in great detail.

We do not speak word by word from scratch; our brain quickly structures a pattern like STH DO STH AT SOMEPLACE and then fills in the detailed descriptions. To render existing encoder-decoder image captioners such humanlike reasoning, we propose a novel framework: learning to Collocate Neural Modules (CNM), to generate the “inner pattern” connecting visual encoder and language decoder. Unlike the widely-used neural module networks in visual Q&A, where the language (i.e., question) is fully observable, CNM for captioning is more challenging as the language is being generated and thus is partially observable. To this end, we make the following technical contributions for CNM training: 1) compact module design - one for function words and three for visual content words (e.g., noun, adjective, and verb), 2) soft module fusion and multistep module execution, robustifying the visual reasoning in partial observation, 3) a linguistic loss for module controller being faithful to part-of-speech collocations (e.g., adjective is before noun). Extensive experiments on the challenging MS-COCO image captioning benchmark validate the effectiveness of our CNM image captioner. In particular, CNM achieves a new state-of-the-art 127.9 CIDErD on Karpathy split and a single-model 126.0 c40 on the official server. CNM is also robust to few training samples, e.g., by training only one sentence per image, CNM can halve the performance loss compared to a strong baseline.

We study the multi-round response generation in visual dialog, where a response is generated according to a visually grounded conversational history. Given a triplet: an image, Q&A history, and current question, all the prevailing methods follow a codec (i.e., encoder-decoder) fashion in a supervised learning paradigm: a multimodal encoder encodes the triplet into a feature vector, which is then fed into the decoder for the current answer generation, supervised by the ground-truth. However, this conventional supervised learning does NOT take into account the impact of imperfect history, violating the conversational nature of visual dialog and thus making the codec more inclined to learn history bias but not contextual reasoning. To this end, inspired by the actor-critic policy gradient in reinforcement learning, we propose a novel training paradigm called History Advantage Sequence Training (HAST). Specifically, we intentionally impose wrong answers in the history, obtaining an adverse critic, and see how the historic error impacts the codec's future behavior by History Advantage - a quantity obtained by subtracting the adverse critic from the gold reward of ground-truth history. Moreover, to make the codec more sensitive to the history, we propose a novel attention network called History-Aware Co-Attention Network (HACAN) which can be effectively trained by using HAST. Experimental results on three benchmarks: VisDial v0.9&v1.0 and GuessWhat?!, show that the proposed HAST strategy consistently outperforms the state-of-the-art supervised counterparts.

Food recognition for user-uploaded images is crucial in visual diet tracking, an emerging application linking multimedia and healthcare domains. However, it is challenging due to the various visual appearances of food images. This is caused by different conditions when taking the photos, such as angles, distances, light conditions, food containers, and background scenes. To alleviate such a semantic gap, this paper presents a cross-modal alignment and transfer network (ATNet), which is motivated by the paradigm of learning using privileged information (LUPI). It additionally utilizes the ingredients in food images as an "intelligent teacher" in the training stage to facilitate cross-modal information passing. Specifically, ATNet first uses a pair of synchronized autoencoders to build the base image and ingredient channels for information flow. Subsequently, the information passing is enabled through a two-stage cross-modal interaction. The first stage of interaction adopts a two-step method, called partial heterogeneous transfer, to 1) alleviate the intrinsic heterogeneity between images and ingredients and 2) align them in a shared space to make their carried information about food classes interact. In the second stage, ATNet learns to map the visual embeddings of images to the ingredient channel for food recognition from the view of "teacher''. This leads a refined recognition by a multi-view fusion. Experiments on two real-world datasets show that ATNet can be incorporated with any state-of-the-art CNN models to consistently improve their performance.

Semantic image inpainting - a task to speculate and fill in large missing areas of a natural image, has shown exciting progress with the introduction of generative adversarial networks (GANs). But due to lack of sufficient understanding of semantic and spatial context, existing methods easily generate blurred boundary and distorted structure, which are inconsistent with the surrounding area. In this paper, we propose a new end-to-end framework named Single-shot Densely Connected Generative Network (SSDCGN), which generates visually realistic and semantically distinct pixels for the missing content by a battery of symmetric encoder-decoder groups. To maximize semantic extraction and realize precise spatial context localization, we involve a deeper densely skip connection in our network. Extensive experiments on Paris StreetView and ImageNet datasets show the superiority of our method.

Video Question & Answering (VideoQA), a task to answer questions in videos, involves rich spatio-temporal content (e.g., appearance and motion) and requires multi-hop reasoning process. However, existing methods usually deal with appearance and motion separately and fail to synchronize the attentions on appearance and motion features, neglecting two key properties of video QA: (1) appearance and motion features are usually concomitant and complementary to each other at time slice level. Some questions rely on joint representations of both kinds of features at some point in the video; (2) appearance and motion have different importance in multi-step reasoning. In this paper, we propose a novel Question- Aware Tube-Switch Network (TSN) for video question answering which contains (1) a Mix module to synchronously combine the appearance and motion representation at time slice level, achieving fine-grained temporal alignment and correspondence between appearance and motion at every time slice and (2) a Switch mod- ule to adaptively choose appearance or motion tube as primary at each reasoning step, guiding the multi-hop reasoning process. To end-to-end train TSN, we utilize the Gumbel-Softmax strategy to account for the discrete tube-switch process. Extensive experimental results on two benchmarks: MSVD-QA and MSRVTT-QA, have demonstrated that the proposed TSN consistently outperforms state-of-the-art on all metrics.

Recently, a great progress in automatic image captioning has been achieved by using semantic concepts detected from the image. However, we argue that existing concepts-to-caption framework, in which the concept detector is trained using the image-caption pairs to minimize the vocabulary discrepancy, suffers from the deficiency of insufficient concepts. The reasons are two-fold: 1) the extreme imbalance between the number of occurrence positive and negative samples of the concept and 2) the incomplete labeling in training captions caused by the biased annotation and usage of synonyms. In this paper, we propose a method, termed online positive recall and missing concepts mining, to overcome those problems. Our method adaptively re-weights the loss of different samples according to their predictions for online positive recall and uses a two-stage optimization strategy for missing concepts mining. In this way, more semantic concepts can be detected and a high accuracy will be expected. On the caption generation stage, we explore an element-wise selection process to automatically choose the most suitable concepts at each time step. Thus, our method can generate more precise and detailed caption to describe the image. We conduct extensive experiments on the MSCOCO image captioning data set and the MSCOCO online test server, which shows that our method achieves superior image captioning performance compared with other competitive methods.

Recently, much advance has been made in image captioning, and an encoder-decoder framework has achieved outstanding performance for this task. In this paper, we propose an extension of the encoder-decoder framework by adding a component called guiding network. The guiding network models the attribute properties of input images, and its output is leveraged to compose the input of the decoder at each time step. The guiding network can be plugged into the current encoder-decoder framework and trained in an end-to-end manner. Hence, the guiding vector can be adaptively learned according to the signal from the decoder, making itself to embed information from both image and language. Additionally, discriminative supervision can be employed to further improve the quality of guidance. The advantages of our proposed approach are verified by experiments carried out on the MS COCO dataset.

We propose a novel framework called Semantics-Preserving Adversarial Embedding Network (SP-AEN) for zero-shot visual recognition (ZSL), where test images and their classes are both unseen during training. SP-AEN aims to tackle the inherent problem - semantic loss - in the prevailing family of embedding-based ZSL, where some semantics would be discarded during training if they are non-discriminative for training classes, but could become critical for recognizing test classes. Specifically, SP-AEN prevents the semantic loss by introducing an independent visual-to-semantic space embedder which disentangles the semantic space into two subspaces for the two arguably conflicting objectives: classification and reconstruction. Through adversarial learning of the two subspaces, SP-AEN can transfer the semantics from the reconstructive subspace to the discriminative one, accomplishing the improved zero-shot recognition of unseen classes. Comparing with prior works, SP-AEN can not only improve classification but also generate photo-realistic images, demonstrating the effectiveness of semantic preservation. On four popular benchmarks: CUB, AWA, SUN and aPY, SP-AEN considerably outperforms other state-of-the-art methods by an absolute performance difference of 12.2%, 9.3%, 4.0% and 3.6% in terms of harmonic mean values [62].

We focus on grounding (i.e., localizing or linking) referring expressions in images, e.g., "largest elephant standing behind baby elephant". This is a general yet challenging vision-language task since it does not only require the localization of objects, but also the multimodal comprehension of context - visual attributes (e.g., "largest", "baby") and relationships (e.g., "behind") that help to distinguish the referent from other objects, especially those of the same category. Due to the exponential complexity involved in modeling the context associated with multiple image regions, existing work oversimplifies this task to pairwise region modeling by multiple instance learning. In this paper, we propose a variational Bayesian method, called Variational Context, to solve the problem of complex context modeling in referring expression grounding. Our model exploits the reciprocal relation between the referent and context, i.e., either of them influences estimation of the posterior distribution of the other, and thereby the search space of context can be greatly reduced. We also extend the model to unsupervised setting where no annotation for the referent is available. Extensive experiments on various benchmarks show consistent improvement over state-of-the-art methods in both supervised and unsupervised settings. The code is available at https://github.com/yuleiniu/vc/.

User and item features of side information are crucial for accurate recommendation. However, the large number of feature dimensions, e.g., usually larger than 107, results in expensive storage and computational cost. This prohibits fast recommendation especially on mobile applications where the computational resource is very limited. In this paper, we develop a generic feature-based recommendation model, called Discrete Factorization Machine (DFM), for fast and accurate recommendation. DFM binarizes the real-valued model parameters (e.g., float32) of every feature embedding into binary codes (e.g., boolean), and thus supports efficient storage and fast user-item score computation. To avoid the severe quantization loss of the binarization, we propose a convergent updating rule that resolves the challenging discrete optimization of DFM. Through extensive experiments on two real-world datasets, we show that 1) DFM consistently outperforms state-of-the-art binarized recommendation models, and 2) DFM shows very competitive performance compared to its real-valued version (FM), demonstrating the minimized quantization loss.

Image captioning is one of the most challenging hallmark of AI, due to its complexity in visual and natural language understanding. As it is essentially a sequential prediction task, recent advances in image captioning use Reinforcement Learning (RL) to better explore the dynamics of word-by-word generation. However, existing RL-based image captioning methods mainly rely on a single policy network and reward function that does not well fit the multi-level (word and sentence) and multi-modal (vision and language) nature of the task. To this end, we propose a novel multi-level policy and reward RL framework for image captioning. It contains two modules: 1) Multi-Level Policy Network that can adaptively fuse the word-level policy and the sentence-level policy for the word generation; and 2) Multi-Level Reward Function that collaboratively leverages both vision-language reward and language-language reward to guide the policy. Further, we propose a guidance term to bridge the policy and the reward for RL optimization. Extensive experiments and analysis on MSCOCO and Flickr30k show that the proposed framework can achieve competing performances with respect to different evaluation metrics.

Recommendation systems play a vital role in online information systems and have become a major monetization tool for user-oriented platforms. In recent years, there has been increasing research interest in recommendation technologies in the information retrieval and data mining community, and significant progress has been made owing to the fast development of deep learning. However, in the multimedia community, there has been relatively less attention paid to the development of multimedia recommendation technologies. In this tutorial, we summarize existing research efforts on multimedia recommendation. We first provide an overview on fundamental techniques and recent advances on personalized recommendation for general items. We then summarize existing developments on recommendation technologies for multimedia content. Lastly, we present insight into the challenges and future directions in this emerging and promising area.

For decades, we are interested in detecting objects and classifying them into a fixed vocabulary of lexicon. With the maturity of these low-level vision solutions, we are hunger for a higher-level representation of the visual data, so as to extract visual knowledge rather than merely bags of visual entities, allowing machines to reason about human-level decision-making and even manipulate the visual data at the pixel-level. In this tutorial, we will introduce a various of machine learning techniques for modeling visual relationships (e.g., subject-predicate-object triplet detection) and contextual generative models (e.g., generating photo-realistic images using conditional generative adversarial networks). In particular, we plan to start from fundamental theories on object detection, relationship detection, generative adversarial networks, to more advanced topics on referring expression visual grounding, pose guided person image generation, and context based image inpainting.

Many vision-language tasks can be reduced to the problem of sequence prediction for natural language output. In particular, recent advances in image captioning use deep reinforcement learning (RL) to alleviate the "exposure bias'' during training: ground-truth subsequence is exposed in every step prediction, which introduces bias in test when only predicted subsequence is seen. However, existing RL-based image captioning methods only focus on the language policy while not the visual policy (eg, visual attention), and thus fail to capture the visual context that are crucial for compositional reasoning such as visual relationships (\eg, "man riding horse'') and comparisons (eg. "smaller cat"). To fill the gap, we propose a Context-Aware Visual Policy network (CAVP) for sequence-level image captioning. At every time step, CAVP explicitly accounts for the previous visual attentions as the context, and then decides whether the context is helpful for the current word generation given the current visual attention. Compared against traditional visual attention that only fixes a single image region at every step, CAVP can attend to complex visual compositions over time. The whole image captioning model --- CAVP and its subsequent language policy network --- can be efficiently optimized end-to-end by using an actor-critic policy gradient method with respect to any caption evaluation metric. We demonstrate the effectiveness of CAVP by state-of-the-art performances on MS-COCO offline split and online server, using various metrics and sensible visualizations of qualitative visual context. The code is available at \urlhttps://github.com/daqingliu/CAVP

Deep neural networks suffer from over-fitting and catastrophic forgetting when trained with small data. One natural remedy for this problem is data augmentation, which has been recently shown to be effective. However, previous works either assume that intra-class variances can always be generalized to new classes, or employ naive generation methods to hallucinate finite examples without modeling their latent distributions. In this work, we propose Covariance-Preserving Adversarial Augmentation Networks to overcome existing limits of low-shot learning. Specifically, a novel Generative Adversarial Network is designed to model the latent distribution of each novel class given its related base counterparts. Since direct estimation on novel classes can be inductively biased, we explicitly preserve covariance information as the ``variability'' of base examples during the generation process. Empirical results show that our model can generate realistic yet diverse examples, leading to substantial improvements on the ImageNet benchmark over the state of the art.

Existing video hash functions are built on three isolated stages: frame pooling, relaxed learning, and binarization, which have not adequately explored the temporal order of video frames in a joint binary optimization model, resulting in severe information loss. In this paper, we propose a novel unsupervised video hashing framework dubbed self-supervised video hashing (SSVH), which is able to capture the temporal nature of videos in an end-to-end learning to hash fashion. We specifically address two central problems: 1) how to design an encoder-decoder architecture to generate binary codes for videos and 2) how to equip the binary codes with the ability of accurate video retrieval. We design a hierarchical binary auto-encoder to model the temporal dependencies in videos with multiple granularities, and embed the videos into binary codes with less computations than the stacked architecture. Then, we encourage the binary codes to simultaneously reconstruct the visual content and neighborhood structure of the videos. Experiments on two real-world data sets show that our SSVH method can significantly outperform the state-of-the-art methods and achieve the current best performance on the task of unsupervised video retrieval.

Embedding network data into a low-dimensional vector space has shown promising performance for many real-world applications, such as node classification and entity retrieval. However, most existing methods focused only on leveraging network structure. For social networks, besides the network structure, there also exists rich information about social actors, such as user profiles of friendship networks and textual content of citation networks. These rich attribute information of social actors reveal the homophily effect, exerting huge impacts on the formation of social networks. In this paper, we explore the rich evidence source of attributes in social networks to improve network embedding. We propose a generic Attributed Social Network Embedding framework (ASNE), which learns representations for social actors (i.e., nodes) by preserving both the structural proximity and attribute proximity. While the structural proximity captures the global network structure, the attribute proximity accounts for the homophily effect. To justify our proposal, we conduct extensive experiments on four real-world social networks. Compared to the state-of-the-art network embedding approaches, ASNE can learn more informative representations, achieving substantial gains on the tasks of link prediction and node classification. Specifically, ASNE significantly outperforms node2vec with an 8.2 percent relative improvement on the link prediction task, and a 12.7 percent gain on the node classification task.

Visual attention has been successfully applied in structural prediction tasks such as visual captioning and question answering. Existing visual attention models are generally spatial, i.e., the attention is modeled as spatial probabilities that re-weight the last conv-layer feature map of a CNN encoding an input image. However, we argue that such spatial attention does not necessarily conform to the attention mechanism - a dynamic feature extractor that combines contextual fixations over time, as CNN features are naturally spatial, channel-wise and multi-layer. In this paper, we introduce a novel convolutional neural network dubbed SCA-CNN that incorporates Spatial and Channel-wise Attentions in a CNN. In the task of image captioning, SCA-CNN dynamically modulates the sentence generation context in multi-layer feature maps, encoding where (i.e., attentive spatial locations at multiple layers) and what (i.e., attentive channels) the visual attention is. We evaluate the proposed SCA-CNN architecture on three benchmark image captioning datasets: Flickr8K, Flickr30K, and MSCOCO. It is consistently observed that SCA-CNN significantly outperforms state-of-the-art visual attention-based image captioning methods.

Visual relations, such as person ride bike and bike next to car, offer a comprehensive scene understanding of an image, and have already shown their great utility in connecting computer vision and natural language. However, due to the challenging combinatorial complexity of modeling subject-predicate-object relation triplets, very little work has been done to localize and predict visual relations. Inspired by the recent advances in relational representation learning of knowledge bases and convolutional object detection networks, we propose a Visual Translation Embedding network (VTransE) for visual relation detection. VTransE places objects in a low-dimensional relation space where a relation can be modeled as a simple vector translation, i.e., subject + predicate ≈ object. We propose a novel feature extraction layer that enables object-relation knowledge transfer in a fully-convolutional fashion that supports training and inference in a single forward/backward pass. To the best of our knowledge, VTransE is the first end-toend relation detection network. We demonstrate the effectiveness of VTransE over other state-of-the-art methods on two large-scale datasets: Visual Relationship and Visual Genome. Note that even though VTransE is a purely visual model, it is still competitive to the Lu's multi-modal model with language priors [27].

We propose a novel method for video object proposal to generate sequences of bounding boxes for each object candidate in videos, namely object trajectory proposals. Unlike the image-based methods that produce object proposals independently in each video frame, our method generates temporally consistent proposals in the form of object trajectories, which is crucial for subsequent analysis of object appearance and motion characteristics. Given a video, we extract motion seeds through estimating the outliers of global motion, and generate moving object trajectory proposals from the seeds. By ignoring the motion outliers, we consistently sample bounding boxes with pruning to form static object trajectory proposals. Finally, we rank both the moving and static object trajectory proposals under a unified scoring mechanism. The experimental results show that our method can effectively generate object trajectory proposals and outperform the state-of-the-art methods.

We present VidedWhisfer, a novel approach for unsupervised video representation learning, in which video sequence is treated as a self-supervision entity based on the observation that the sequence encodes video temporal dynamics (e.g., object movement and event evolution). Specifically, for each video sequence, we use a pre-learned visual dictionary to generate a sequence of high-level semantics, dubbed “whisper”, which encodes both visual contents at the frame level and visual dynamics at the sequence level. VidedWhisfer is driven by a novel “sequence-to-whisper” learning strategy. Naturally, an end-to-end sequence-to-sequence learning model using RNN is modeled and trained to predict the whisper sequence. We propose two ways to generate video representation from the model. Through extensive experiments we demonstrate that video representation learned by VidedWhisfer is effective to boost fundamental video-related applications such as video retrieval and classification.

Factorization Machines (FMs) are a supervised learning approach that enhances the linear regression model by incorporating the second-order feature interactions. Despite effectiveness, FM can be hindered by its modelling of all feature interactions with the same weight, as not all feature interactions are equally useful and predictive. For example, the interactions with useless features may even introduce noises and adversely degrade the performance. In this work, we improve FM by discriminating the importance of different feature interactions. We propose a novel model named Attentional Factorization Machine (AFM), which learns the importance of each feature interaction from data via a neural attention network. Extensive experiments on two real-world datasets demonstrate the effectiveness of AFM. Empirically, it is shown on regression task AFM betters FM with a 8.6% relative improvement, and consistently outperforms the state-of-the-art deep learning methods Wide&Deep [Cheng et al., 2016] and DeepCross [Shan et al., 2016] with a much simpler structure and fewer model parameters. Our implementation of AFM is publicly available at: https://github.com/hexiangnan/attentional_factorization_machine

Different from traditional long videos, micro-videos are much shorter and usually recorded at a specific place with mobile devices. To better understand the semantics of a micro-video and facilitate downstream applications, it is crucial to estimate the venue where the micro-video is recorded, for example, in a concert or on a beach. However, according to our statistics over two million micro-videos, only $1.22%$ of them were labeled with location information. For the remaining large number of micro-videos without location information, we have to rely on their content to estimate their venue categories. This is a highly challenging task, as micro-videos are naturally multi-modal (with textual, visual and, acoustic content), and more importantly, the quality of each modality varies greatly for different micro-videos.

As a bridge to connect vision and language, visual relations between objects in the form of relation triplet $łangle subject,predicate,object\rangle$, such as "person-touch-dog'' and "cat-above-sofa'', provide a more comprehensive visual content understanding beyond objects. In this paper, we propose a novel vision task named Video Visual Relation Detection (VidVRD) to perform visual relation detection in videos instead of still images (ImgVRD). As compared to still images, videos provide a more natural set of features for detecting visual relations, such as the dynamic relations like "A-follow-B'' and "A-towards-B'', and temporally changing relations like "A-chase-B'' followed by "A-hold-B''. However, VidVRD is technically more challenging than ImgVRD due to the difficulties in accurate object tracking and diverse relation appearances in video domain. To this end, we propose a VidVRD method, which consists of object tracklet proposal, short-term relation prediction and greedy relational association. Moreover, we contribute the first dataset for VidVRD evaluation, which contains 1,000 videos with manually labeled visual relations, to validate our proposed method. On this dataset, our method achieves the best performance in comparison with the state-of-the-art baselines.

Recently image question answering (ImageQA) has gained lots of attention in the research community. However, as its natural extension, video question answering (VideoQA) is less explored. Although both tasks look similar, VideoQA is more challenging mainly because of the complexity and diversity of videos. As such, simply extending the ImageQA methods to videos is insufficient and suboptimal. Particularly, working with the video needs to model its inherent temporal structure and analyze the diverse information it contains. In this paper, we consider exploiting the appearance and motion information resided in the video with a novel attention mechanism. More specifically, we propose an end-to-end model which gradually refines its attention over the appearance and motion features of the video using the question as guidance. The question is processed word by word until the model generates the final optimized attention. The weighted representation of the video, as well as other contextual information, are used to generate the answer. Extensive experiments show the advantages of our model compared to other baseline models. We also demonstrate the effectiveness of our model by analyzing the refined attention weights during the question answering procedure.

In this paper, we focus on improving Event Extraction (EE) by incorporating visual knowledge with words and phrases from text documents. We first discover visual patterns from large-scale text-image pairs in a weakly-supervised manner and then propose a multimodal event extraction algorithm where the event extractor is jointly trained with textual features and visual patterns. Extensive experimental results on benchmark data sets demonstrate that the proposed multimodal EE method can achieve significantly better performance on event extraction: absolute 7.1% F-score gain on event trigger labeling and 8.5% F-score gain on event argument labeling.

Multimedia content is dominating today's Web information. The nature of multimedia user-item interactions is 1/0 binary implicit feedback (e.g., photo likes, video views, song downloads, etc.), which can be collected at a larger scale with a much lower cost than explicit feedback (e.g., product ratings). However, the majority of existing collaborative filtering (CF) systems are not well-designed for multimedia recommendation, since they ignore the implicitness in users' interactions with multimedia content. We argue that, in multimedia recommendation, there exists item- and component-level implicitness which blurs the underlying users' preferences. The item-level implicitness means that users' preferences on items (e.g. photos, videos, songs, etc.) are unknown, while the component-level implicitness means that inside each item users' preferences on different components (e.g. regions in an image, frames of a video, etc.) are unknown. For example, a 'view'' on a video does not provide any specific information about how the user likes the video (i.e.item-level) and which parts of the video the user is interested in (i.e.component-level). In this paper, we introduce a novel attention mechanism in CF to address the challenging item- and component-level implicit feedback in multimedia recommendation, dubbed Attentive Collaborative Filtering (ACF). Specifically, our attention model is a neural network that consists of two attention modules: the component-level attention module, starting from any content feature extraction network (e.g. CNN for images/videos), which learns to select informative components of multimedia items, and the item-level attention module, which learns to score the item preferences. ACF can be seamlessly incorporated into classic CF models with implicit feedback, such as BPR and SVD++, and efficiently trained using SGD. Through extensive experiments on two real-world multimedia Web services: Vine and Pinterest, we show that ACF significantly outperforms state-of-the-art CF methods.

In recent years, deep neural networks have yielded immense success on speech recognition, computer vision and natural language processing. However, the exploration of deep neural networks on recommender systems has received relatively less scrutiny. In this work, we strive to develop techniques based on neural networks to tackle the key problem in recommendation --- collaborative filtering --- on the basis of implicit feedback.

Social media websites have become important information sharing platforms. The rapid development of social media platforms has led to increasingly large-scale social media data, which has shown remarkable societal and marketing values. There are needs to extract important events in live social media streams. However, microblogs event classification is challenging due to two facts, i.e., the short/conversational nature and the incompatible meanings between the text and the corresponding image in social posts, and the rapidly evolving contents. In this article, we propose to conduct event classification via deep learning and social tracking. First, we introduce a Multi-modal Multi-instance Deep Network (M 2 DN) for microblogs classification, which is able to handle the weakly labeled microblogs data oriented from the incompatible meanings inside microblogs. Besides predicting each microblogs as predefined events, we propose to employ social tracking to extract social-related auxiliary information to enrich the testing samples. We extract a set of candidate-relevant microblogs in a short time window by using social connections, such as related users and geographical locations. All these selected microblogs and the testing data are formulated in a Markov Random Field model. The inference on the Markov Random Field is conducted to update the classification results of the testing microblogs. This method is evaluated on the Brand-Social-Net dataset for classification of 20 events. Experimental results and comparison with the state of the arts show that the proposed method can achieve better performance for the event classification task.

Sentence auto-completion is an important feature that saves users many keystrokes in typing the entire sentence by providing suggestions as they type. Despite its value, the existing sentence auto-completion methods, such as query completion models, can hardly be applied to solving the object completion problem in sentences with the form of (subject, verb, object), due to the complex natural language description and the data deficiency problem. Towards this goal, we treat an SVO sentence as a three-element triple (subject, sentence pattern, object), and cast the sentence object completion problem as an element inference problem. These elements in all triples are encoded into a unified low-dimensional embedding space by our proposed TRANSFER model, which leverages the external knowledge base to strengthen the representation learning performance. With such representations, we can provide reliable candidates for the desired missing element by a linear model. Extensive experiments on a real-world dataset have well-validated our model. Meanwhile, we have successfully applied our proposed model to factoid question answering systems for answer candidate selection, which further demonstrates the applicability of the TRANSFER model.

Recent progress in using long short-term memory (LSTM) for image captioning has motivated the exploration of their applications for video captioning. By taking a video as a sequence of features, an LSTM model is trained on video-sentence pairs and learns to associate a video to a sentence. However, most existing methods compress an entire video shot or frame into a static representation, without considering attention mechanism which allows for selecting salient features. Furthermore, existing approaches usually model the translating error, but ignore the correlations between sentence semantics and visual content. To tackle these issues, we propose a novel end-to-end framework named aLSTMs, an attention-based LSTM model with semantic consistency, to transfer videos to natural sentences. This framework integrates attention mechanism with LSTM to capture salient structures of video, and explores the correlation between multimodal representations (i.e., words and visual content) for generating sentences with rich semantic content. Specifically, we first propose an attention mechanism that uses the dynamic weighted sum of local two-dimensional convolutional neural network representations. Then, an LSTM decoder takes these visual features at time t and the word-embedding feature at time t-1 to generate important words. Finally, we use multimodal embedding to map the visual and sentence features into a joint space to guarantee the semantic consistence of the sentence description and the video visual content. Experiments on the benchmark datasets demonstrate that our method using single feature can achieve competitive or even better results than the state-of-the-art baselines for video captioning in both BLEU and METEOR.

In sharp contrast to the traditional category/subcategory level image retrieval, product image search aims to find the images containing the exact same product. This is a challenging problem because in addition to being robust under different imaging conditions such as varying viewpoints and illumination changes, the features should also be able to distinguish the specific product among many similar products. Consequently, it is important to utilize a large dataset, containing many product classes, to learn a strongly discriminative representation. Building such a dataset requires laborious manual annotation. Toward learning fine-grained, robust, discriminative features for product image search, we present a novel paradigm that can construct the required dataset without any human annotation. Unlike other fine-grained recognition works that rely on high-quality annotated datasets and are very narrowly focused on a specific object category, our method handles multiple object classes and requires minimum human effort. First, an ImageNet pretrained model is used to generate product clusters. As the original features from ImageNet are not discriminative, the clusters generated by this unsupervised procedure contain much noise. We alleviate noise by explicitly modeling noise distribution and automatically detecting errors during learning. The proposed paradigm is general, requires minimum human efforts, and is applicable to any deep learning task where fine-grained discriminative features are desired. Extensive experiments on the ALISC dataset have demonstrated that our approach is sound and effective, surpassing the baseline GoogleNet model by 15.09%.

We present VideoWhisper, a novel approach for unsupervised video representation learning. Based on the observation that the frame sequence encodes the temporal dynamics of a video (e.g., object movement and event evolution), we treat the frame sequential order as a self-supervision to learn video representations. Unlike other unsupervised video feature learning methods based on frame-level feature reconstruction that is sensitive to visual variance, VideoWhisper is driven by a novel video “sequence-to-whisper” learning strategy. Specifically, for each video sequence, we use a prelearned visual dictionary to generate a sequence of high-level semantics, dubbed “whisper,” which can be considered as the language describing the video dynamics. In this way, we model VideoWhisper as an end-to-end sequence-to-sequence learning model using attention-based recurrent neural networks. This model is trained to predict the whisper sequence and hence it is able to learn the temporal structure of videos. We propose two ways to generate video representation from the model. Through extensive experiments on two real-world video datasets, we demonstrate that video representation learned by V ideoWhisper is effective to boost fundamental multimedia applications such as video retrieval and event classification.

We address the problem of image hashing by learning binary codes from large and weakly supervised photo collections. Due to the explosive growth of user generated media on the Web, this problem is becoming critical for large-scale visual applications like image retrieval. While most existing hashing methods fail to address this challenge well, our method shows promising improvement due to the following two key advantages.First, we formulate a novel hashing objective that can effectively mine implicit weak supervision by collaborative filtering. Second, we propose a discrete hashing algorithm, offered with efficient optimization, to overcome the inferior optimizations in obtaining binary codes from real-valued solutions. In this way, our method can be considered as a weakly-supervised discrete hashing framework which jointly learns image semantics and their corresponding binary codes. Through training on one million weakly annotated images, our experimental results demonstrate that image retrieval using the proposed hashing method outperforms the other state-of-the-art ones on image and video benchmarks.

We focus on learning open-vocabulary visual classifiers, which scale up to a large portion of natural language vocabulary (e.g., over tens of thousands of classes). In particular, the training data are large-scale weakly labeled Web images since it is difficult to acquire sufficient well-labeled data at this category scale. In this paper, we propose a novel online learning paradigm towards this challenging task. Different from traditional N-way independent classifiers that generally fail to handle the extremely sparse and inter-related labels, our classifiers learn from continuous label embeddings discovered by collaboratively decomposing the sparse image-label matrix. Leveraging on the structure of the proposed collaborative learning formulation, we develop an efficient online algorithm that can jointly learn the label embeddings and visual classifiers. The algorithm can learn over 30,000 classes of 1,000 training images within 1 second on a standard GPU. Extensively experimental results on four benchmarks demonstrate the effectiveness of our method.

Micro-videos, a new form of user generated contents (UGCs), are gaining increasing enthusiasm. Popular micro-videos have enormous commercial potential in many ways, such as online marketing and brand tracking. In fact, the popularity prediction of traditional UGCs including tweets, web images, and long videos, has achieved good theoretical underpinnings and great practical success. However, little research has thus far been conducted to predict the popularity of the bite-sized videos. This task is non-trivial due to three reasons: 1) micro-videos are short in duration and of low quality; 2) they can be described by multiple heterogeneous channels, spanning from social, visual, acoustic to textual modalities; and 3) there are no available benchmark dataset and discriminant features that are suitable for this task. Towards this end, we present a transductive multi-modal learning model. The proposed model is designed to find the optimal latent common space, unifying and preserving information from different modalities, whereby micro-videos can be better represented. This latent space can be used to alleviate the information insufficiency problem caused by the brief nature of micro-videos. In addition, we built a benchmark dataset and extracted a rich set of popularity-oriented features to characterize the popular micro-videos. Extensive experiments have demonstrated the effectiveness of the proposed model. As a side contribution, we have released the dataset, codes and parameters to facilitate other researchers.

Video browsing describes an interactive process where the user wants to find a target shot in a long video. Therefore, it is crucial for a video browsing system to be fast and accurate with minimum user effort. Previously, video browsing systems suffered from either lots of user efforts or heavy computation cost, which makes them less effective. We propose an RNN-based \textit{intention-aware} system to overcome these limitations. At each interactive round, the user only needs to select one of the displayed shots that is most visually similar to her mental target and then the user's choice will further tailor the search to the target. The search model update merely requires vector inner products which means the system is highly responsive.

Video browsing describes an interactive process where users want to find a target shot in a long video. Therefore, it is crucial for a video browsing system to be fast and accurate with minimum user effort. In sharp contrast to traditional Relevance Feedback (RF), we propose a novel paradigm for fast video browsing dubbed Mental Visual Indexing (MVI). At each interactive round, the user only needs to select one of the displayed shots that is most visually similar to her mental target and then the user's choice will further tailor the search to the target. The search model update given a user feedback only requires vector inner products, which makes MVI highly responsive. MVI is underpinned by a sequence model in terms of Recurrent Neural Network (RNN), which is trained by automatically generated shot sequences from a rigorous Bayesian framework, which simulates user feedback process. Experimental results on three 3-hour movies conducted by real users demonstrate the effectiveness of the proposed approach.

We focus on hashing videos into short binary codes for efficient Content-based Video Retrieval (CBVR), which is a fundamental technique that supports access to the ever-growing abundance of videos on the Web. Existing video hash functions are built on three isolated stages: frame pooling, relaxed learning, and binarization, which have not adequately explored the temporal order of video frames in a joint binary optimization model, resulting in severe information loss. In this paper, we propose a novel unsupervised video hashing framework called Self-Supervised Temporal Hashing (SSTH) that is able to capture the temporal nature of videos in an end-to-end learning-to-hash fashion. Specifically, the hash function of SSTH is an encoder RNN equipped with the proposed Binary LSTM (BLSTM) that generates binary codes for videos. The hash function is learned in a self-supervised fashion, where a decoder RNN is proposed to reconstruct the original video frames in both forward and reverse orders. For binary code optimization, we develop a backpropagation rule that tackles the non-differentiability of BLSTM. This rule allows efficient deep network training without suffering from the binarization loss. Through extensive CBVR experiments on two real-world consumer video datasets of Youtube and Flickr, we show that SSTH consistently outperforms state-of-the-art video hashing methods, eg., in terms of [email protected], SSTH using only 128 bits can still outperform others using 256 bits by at least 9% to 15% on both datasets.

This paper contributes improvements on both the effectiveness and efficiency of Matrix Factorization (MF) methods for implicit feedback. We highlight two critical issues of existing works. First, due to the large space of unobserved feedback, most existing works resort to assign a uniform weight to the missing data to reduce computational complexity. However, such a uniform assumption is invalid in real-world settings. Second, most methods are also designed in an offline setting and fail to keep up with the dynamic nature of online data. We address the above two issues in learning MF models from implicit feedback. We first propose to weight the missing data based on item popularity, which is more effective and flexible than the uniform-weight assumption. However, such a non-uniform weighting poses efficiency challenge in learning the model. To address this, we specifically design a new learning algorithm based on the element-wise Alternating Least Squares (eALS) technique, for efficiently optimizing a MF model with variably-weighted missing data. We exploit this efficiency to then seamlessly devise an incremental update strategy that instantly refreshes a MF model given new feedback. Through comprehensive experiments on two public datasets in both offline and online protocols, we show that our implemented, open-source (https://github.com/hexiangnan/sigir16-eals) eALS consistently outperforms state-of-the-art implicit MF methods.

We address the efficiency problem of Collaborative Filtering (CF) by hashing users and items as latent vectors in the form of binary codes, so that user-item affinity can be efficiently calculated in a Hamming space. However, existing hashing methods for CF employ binary code learning procedures that most suffer from the challenging discrete constraints. Hence, those methods generally adopt a two-stage learning scheme composed of relaxed optimization via discarding the discrete constraints, followed by binary quantization. We argue that such a scheme will result in a large quantization loss, which especially compromises the performance of large-scale CF that resorts to longer binary codes. In this paper, we propose a principled CF hashing framework called Discrete Collaborative Filtering (DCF), which directly tackles the challenging discrete optimization that should have been treated adequately in hashing. The formulation of DCF has two advantages: 1) the Hamming similarity induced loss that preserves the intrinsic user-item similarity, and 2) the balanced and uncorrelated code constraints that yield compact yet informative binary codes. We devise a computationally efficient algorithm with a rigorous convergence proof of DCF. Through extensive experiments on several real-world benchmarks, we show that DCF consistently outperforms state-of-the-art CF hashing techniques, e.g, though using only 8 bits, DCF is even significantly better than other methods using 128 bits.

We present a deep learning strategy to fuse multiple semantic cues for complex event recognition. In particular, we tackle the recognition task by answering how to jointly analyze human actions (who is doing what), objects (what), and scenes (where). First, each type of semantic features (e.g., human action trajectories) is fed into a corresponding multi-layer feature abstraction pathway, followed by a fusion layer connecting all the different pathways. Second, the correlations of how the semantic cues interacting with each other are learned in an unsupervised cross-modality autoencoder fashion. Finally, by fine-tuning a large-margin objective deployed on this deep architecture, we are able to answer the question on how the semantic cues of who, what, and where compose a complex event. As compared with the traditional feature fusion methods (e.g., various early or late strategies), our method jointly learns the essential higher level features that are most effective for fusion and recognition. We perform extensive experiments on two real-world complex event video benchmarks, MED'11 and CCV, and demonstrate that our method outperforms the best published results by 21% and 11%, respectively, on an event recognition task.

Along with the long-time evolution of popular social networks, e.g. Facebook, social media analysis research inevitably arrived at the era of considering face/user recognition with large age gaps. However, related research with adequate subjects and large age gaps is surprisingly rare. In this work, we first collect a so-called cross-age face (CAFE) dataset, ranging from child, to young, to adult, to old groups. Then, we propose a novel framework, called deep aging face verification (DAFV), for this challenging task. DAFV includes two modules: aging pattern synthesis and aging face verification. The aging pattern synthesis module synthesizes the faces of all age groups for the input face of an arbitrary age, and the core structure is a deep aging-aware denoising auto-encoder ( a2-DAE) with multiple outputs. The aging face verification module then takes the synthesized aging patterns of a face pair as the input, and each pair of synthesized images of the same age group is fed into a parallel CNN; finally, all parallel CNN outputs are fused to provide similar/dissimilar prediction. For DAFV, the training of the aging face verification module easily suffers from the overfitting results from the aging pattern synthesis module, and we propose to use the cross- validation strategy to produce error-aware outputs for the synthesis module. Extensive experiments on the CAFE dataset well demonstrate the superiority of the proposed DAFV framework over other solutions for aging face verification.

Feature representation for visual content is the key to the progress of many fundamental applications such as annotation and cross-modal retrieval. Although recent advances in deep feature learning offer a promising route towards these tasks, they are limited in application domains where high-quality and large-scale training data are expensive to obtain. In this article, we propose a novel deep feature learning paradigm based on social collective intelligence, which can be acquired from the inexhaustible social multimedia content on the Web, in particular, largely social images and tags. Differing from existing feature learning approaches that rely on high-quality image-label supervision, our weak supervision is acquired by mining the visual-semantic embeddings from noisy, sparse, and diverse social image collections. The resultant image-word embedding space can be used to (1) fine-tune deep visual models for low-level feature extractions and (2) seek sparse representations as high-level cross-modal features for both image and text. We offer an easy-to-use implementation for the proposed paradigm, which is fast and compatible with any state-of-the-art deep architectures. Extensive experiments on several benchmarks demonstrate that the cross-modal features learned by our paradigm significantly outperforms others in various applications such as content-based retrieval, classification, and image captioning.

Good representations of data do help in many machine learning tasks such as recommendation. It is often a great challenge for traditional recommender systems to learn representative features of both users and images in large social networks, in particular, social curation networks, which are characterized as the extremely sparse links between users and images, and the extremely diverse visual contents of images. To address the challenges, we propose a novel deep model which learns the unified feature representations for both users and images. This is done by transforming the heterogeneous user-image networks into homogeneous low-dimensional representations, which facilitate a recommender to trivially recommend images to users by feature similarity. We also develop a fast online algorithm that can be easily scaled up to large networks in an asynchronously parallel way. We conduct extensive experiments on a representative subset of Pinterest, containing 1,456,540 images and 1,000,000 users. Results of image recommendation experiments demonstrate that our feature learning approach significantly outperforms other state-of-the-art recommendation methods.

With the explosive increasing of web image data, image annotation has become a critical research issue for image semantic index and search. In this work, we propose a novel model, termed as multi-view semi-supervised learning (MVSSL), for robust image annotation task. Specifically, we exploit both labeled images and unlabeled images to uncover the intrinsic data structural information. Meanwhile, to comprehensively describe an individual datum, we take advantage of the correlated and complemental information derived from multiple facets of image data (i.e., multiple views or features). We devise a robust pair-wise constraint on outcomes of different views to achieve annotation consistency. Furthermore, we integrate a robust classifier learning component via l2,1 loss, which can provide effective noise identification power during the learning process. Finally, we devise an efficient iterative algorithm to solve the optimization problem in MVSSL. We conduct extensive experiments on the NUS-WIDE dataset, and the results illustrate that our proposed approach is promising for large scale web image annotation task.

In recent years, tremendous research endeavours have been dedicated to seeking effective visual representations for facilitating various multimedia applications, such as visual annotation and retrieval. Nonetheless, existing approaches can hardly achieve satisfactory performance due to the scarcity of fully exploring semantic properties of visual codes. In this paper, we present a novel visual coding approach, termed as hierarchical semantic visual coding (HSVC), to effectively encode visual objects (e.g., image and video) in a semantic hierarchy. Specifically, we first construct a semantic-enriched dictionary hierarchy, which is comprised of dictionaries corresponding to all concepts in a semantic hierarchy as well as their hierarchical semantic correlation. Moreover, we devise an on-line semantic coding model, which simultaneously 1) exploits the rich hierarchical semantic prior knowledge in the learned dictionary, 2) reflects semantic sparse property of visual codes, and 3) explores semantic relationships among concepts in the semantic hierarchy. To this end, we propose to integrate concept-level group sparsity constraint and semantic correlation matrix into a unified regularization term. We design an effective algorithm to optimize the proposed model, and a rigorous mathematical analysis has been provided to guarantee that the algorithm converges to a global optima. Extensive experiments on various multimedia datasets have been conducted to illustrate the superiority of our proposed approach as compared to state-of-the-art methods.

Feature representation for multimedia content is the key to the progress of many fundamental multimedia tasks. Although recent advances in deep feature learning offer a promising route towards these tasks, they are limited in application to domains where high-quality and large-scale training data are hard to obtain. In this paper, we propose a novel deep feature learning paradigm based on large, noisy and social image-tag collections, which can be acquired from the inexhaustible social multimedia content on the Web. Instead of learning features from high-quality image-label supervision, we propose to learn from the image-word semantic relations, in a way of seeking a unified image-word embedding space, where the pairwise feature similarities preserve the semantic relations in the original image-word pairs. We offer an easy-to-use implementation for the proposed paradigm, which is fast and compatible for integrating into any state-of-the-art deep architectures. Experiments on NUSWIDE benchmark demonstrate that the features learned by our method significantly outperforms other state-of-the-art ones.

Microblogging services have revolutionized the way people exchange information. Confronted with the ever-increasing numbers of social events and the corresponding microblogs with multimedia contents, it is desirable to provide visualized summaries to help users to quickly grasp the essence of these social events for better understanding. While existing approaches mostly focus only on text-based summary, microblog summarization with multiple media types (e.g., text, image, and video) is scarcely explored. In this paper, we propose a multimedia social event summarization framework to automatically generate visualized summaries from the microblog stream of multiple media types. Specifically, the proposed framework comprises three stages, as follows. 1) A noise removal approach is first devised to eliminate potentially noisy images. An effective spectral filtering model is exploited to estimate the probability that an image is relevant to a given event. 2) A novel cross-media probabilistic model, termed Cross-Media-LDA (CMLDA), is proposed to jointly discover subevents from microblogs of multiple media types. The intrinsic correlations among these different media types are well explored and exploited for reinforcing the cross-media subevent discovery process. 3) Finally, based on the cross-media knowledge of all the discovered subevents, a multimedia microblog summary generation process is designed to jointly identify both representative textual and visual samples, which are further aggregated to form a holistic visualized summary. We conduct extensive experiments on two real-world microblog datasets to demonstrate the superiority of the proposed framework as compared to the state-of-the-art approaches.

The task of recognizing events from video has attracted a lot of attention in recent years. However, due to the complex nature of user-defined events, the use of purely audio- visual content analysis without domain knowledge has been found to be grossly inadequate. In this paper, we propose to construct a semantic-visual knowledge base to encode the rich event-centric concepts and their relationships from the well- established lexical databases, including FrameNet, as well as the concept-specific visual knowledge from ImageNet. Based on this semantic-visual knowledge bases, we design an effective system for video event recognition. Specifically, in order to narrow the semantic gap between the high-level complex events and low-level visual representations, we utilize the event-centric semantic concepts encoded in the knowledge base as the intermediate-level event representation, which offers both human-perceivable and machine-interpretable semantic clues for event recognition. In addition, in order to leverage the abundant ImageNet images, we propose a robust transfer learning model to learn the noise- resistant concept classifiers for videos. Extensive experiments on various real-world video datasets demonstrate the superiority of our proposed system as compared to the state-of-the-art approaches.

Image tagging, also known as image annotation and image conception detection, has been extensively studied in the literature. However, most existing approaches can hardly achieve satisfactory performance owing to the deficiency and unreliability of the manually-labeled training data. In this paper, we propose a new image tagging scheme, termed social assisted media tagging (SAMT), which leverages the abundant user-generated images and the associated tags as the "social assistance" to learn the classifiers. We focus on addressing the following major challenges: (a) the noisy tags associated to the web images; and (b) the desirable robustness of the tagging model. We present a joint image tagging framework which simultaneously refines the erroneous tags of the web images as well as learns the reliable image classifiers. In particular, we devise a novel tag refinement module for identifying and eliminating the noisy tags by substantially exploring and preserving the low-rank nature of the tag matrix and the structured sparse property of the tag errors. We develop a robust image tagging module based on the l2,p-norm for learning the reliable image classifiers. The correlation of the two modules is well explored within the joint framework to reinforce each other. Extensive experiments on two real-world social image databases illustrate the superiority of the proposed approach as compared to the existing methods.

Social Curation Service (SCS) is a new type of emerging social media platform, where users can select, organize and keep track of multimedia contents they like. In this paper, we take advantage of this great opportunity and target at the very starting point in social media: user profiling, which supports fundamental applications such as personalized search and recommendation. As compared to other profiling methods in conventional Social Network Services (SNS), our work benefits from the two distinguishable characteristics of SCS: a) organized multimedia user-generated contents, and b) content-centric social network. Based on these two characteristics, we are able to deploy the state-of-the-art multimedia analysis techniques to establish content-based user profiles by extracting user preferences and their social relations. First, we automatically construct a content-based user preference ontology and learn the ontological models to generate comprehensive user profiles. In particular, we propose a new deep learning strategy called multi-task convolutional neural network (mtCNN) to learn profile models and profile-related visual features simultaneously. Second, we propose to model the multi-level social relations offered by SCS to refine the user profiles in a low-rank recovery framework. To the best of our knowledge, our work is the first that explores how social curation can help in content-based social media technologies, taking user profiling as an example. Extensive experiments on 1,293 users and 1.5 million images collected from Pinterest in fashion domain demonstrate that recommendation methods based on the proposed user profiles are considerably more effective than other state-of-the-art recommendation strategies.

Photo aesthetic quality evaluation is a challenging task in multimedia and computer vision fields. Conventional approaches suffer from the following three drawbacks: 1) the deemphasized role of semantic content that is many times more important than low-level visual features in photo aesthetics; 2) the difficulty to optimally fuse low-level and high-level visual cues in photo aesthetics evaluation; and 3) the absence of a sequential viewing path in the existing models, as humans perceive visually salient regions sequentially when viewing a photo.

Higher-level semantics such as visual attributes are crucial for fundamental multimedia applications. We present a novel attribute discovery approach that can automatically identify, model and name attributes from an arbitrary set of image and text pairs that can be easily gathered on the Web. Different from conventional attribute discovery methods, our approach does not rely on any pre-defined vocabularies and human labeling. Therefore, we are able to build a large visual knowledge base without any human efforts. The discovery is based on a novel deep architecture, named Independent Component Multimodal Autoencoder (ICMAE), that can continually learn shared higher-level representations across the visual and textual modalities. With the help of the resultant representations encoding strong visual and semantic evidences, we propose to (a) identify attributes and their corresponding high-quality training images, (b) iteratively model them with maximum compactness and comprehensiveness, and (c) name the attribute models with human understandable words. To date, the proposed system has discovered 1,898 attributes over 1.3 million pairs of image and text. Extensive experiments on various real-world multimedia datasets demonstrate the quality and effectiveness of the discovered attributes, facilitating multimedia applications such as image annotation and retrieval as compared to the state-of-the-art approaches.

Nonnegative matrix factorization (NMF) has received considerable attention in image processing, computer vision, and patter recognition. An important variant of NMF is nonnegative graph embedding (NGE), which encodes the statistical or geometric information of data in the process of matrix factorization. The NGE offers a general framework for unsupervised/supervised settings. However, NGE-like algorithms often suffer from noisy data, unreliable graphs, and noisy labels, which are commonly encountered in real-world applications. To address these issues, in this paper, we first propose a robust nonnegative graph embedding (RNGE) framework, where the joint sparsity in both graph embedding and data reconstruction endues robustness to undesirable noises. Next, we present a robust seminonnegative graph embedding (RsNGE) framework, which only constrains the coefficient matrix to be nonnegative while places no constraint on the base matrix. This extends the applicable range of RNGE to data which are not nonnegative and endows more discriminative power of the learnt base matrix. The RNGE/RsNGE provides a general formulation such that all the algorithms unified within the graph embedding framework can be easily extended to obtain their robust nonnegative/seminonnegative solutions. Further, we develop elegant multiplicative updating solutions that can solve RNGE/RsNGE efficiently and offer a rigorous convergence analysis. We conduct extensive experiments on four real-world data sets and compare the proposed RNGE/RsNGE to other representative NMF variants and data factorization methods. The experimental results demonstrate the robustness and effectiveness of the proposed approaches.

This article presents a novel attribute-augmented semantic hierarchy (A 2 SH) and demonstrates its effectiveness in bridging both the semantic and intention gaps in content-based image retrieval (CBIR). A 2 SH organizes semantic concepts into multiple semantic levels and augments each concept with a set of related attributes. The attributes are used to describe the multiple facets of the concept and act as the intermediate bridge connecting the concept and low-level visual content. An hierarchical semantic similarity function is learned to characterize the semantic similarities among images for retrieval. To better capture user search intent, a hybrid feedback mechanism is developed, which collects hybrid feedback on attributes and images. This feedback is then used to refine the search results based on A 2 SH. We use A 2 SH as a basis to develop a unified content-based image retrieval system. We conduct extensive experiments on a large-scale dataset of over one million Web images. Experimental results show that the proposed A 2 SH can characterize the semantic affinities among images accurately and can shape user search intent quickly, leading to more accurate search results as compared to state-of-the-art CBIR solutions.

This paper presents a novel Attribute-augmented Semantic Hierarchy (A2 SH) and demonstrates its effectiveness in bridging both the semantic and intention gaps in Content-based Image Retrieval (CBIR). A2 SH organizes the semantic concepts into multiple semantic levels and augments each concept with a set of related attributes, which describe the multiple facets of the concept and act as the intermediate bridge connecting the concept and low-level visual content. A hierarchical semantic similarity function is learnt to characterize the semantic similarities among images for retrieval. To better capture user search intent, a hybrid feedback mechanism is developed, which collects hybrid feedbacks on attributes and images. These feedbacks are then used to refine the search results based on A2 SH. We develop a content-based image retrieval system based on the proposed A2 SH. We conduct extensive experiments on a large-scale data set of over one million Web images. Experimental results show that the proposed A2 SH can characterize the semantic affinities among images accurately and can shape user search intent precisely and quickly, leading to more accurate search results as compared to state-of-the-art CBIR solutions.

Human group activities detection in multi-camera CCTV surveillance videos is a pressing demand on smart surveillance. Previous works on this topic are mainly based on camera topology inference that is hard to apply to real-world unconstrained surveillance videos. In this paper, we propose a new approach for multi-camera group activities detection. Our approach simultaneously exploits intra-camera and inter-camera contexts without topology inference. Specifically, a discriminative graphical model with hidden variables is developed. The intra-camera and inter-camera contexts are characterized by the structure of hidden variables. By automatically optimizing the structure, the contexts are effectively explored. Furthermore, we propose a new spatiotemporal feature, named vigilant area (VA), to characterize the quantity and appearance of the motion in an area. This feature is effective for group activity representation and is easy to extract from a dynamic and crowded scene. We evaluate the proposed VA feature and discriminative graphical model extensively on two real-world multi-camera surveillance video data sets, including a public corpus consisting of 2.5 h of videos and a 468-h video collection, which, to the best of our knowledge, is the largest video collection ever used in human activity detection. The experimental results demonstrate the effectiveness of our approach.

Non-negative data factorization has been widely used recently. However, existing techniques, such as Non-negative Graph Embedding (NGE), often suffer from noisy data, unreliable graphs, and noisy labels, which are commonly encountered in real-world applications. To address these issues, in this paper, we propose a Robust Non-negative Graph Embedding (RNGE) framework. The joint sparsity in both graph embedding and reconstruction endues the robustness of RNGE. We develop an elegant multiplicative updating solution that can solve RNGE efficiently and prove the convergence rigourously. RNGE is robust to unreliable graphs, as well as both sample and label noises in training data. Moreover, RNGE provides a general formulation such that all the algorithms unified with the graph embedding framework can be easily extended to obtain their robust non-negative solutions. We conduct extensive experiments on four real-world datasets and compared the proposed RNGE to NGE and other representative non-negative data factorization and subspace learning methods. The experimental results demonstrate the effectiveness and robustness of RNGE.

Query suggestion is an effective solution to help users deliver their search intent. While many query suggestion approaches have been proposed for test-based image retrieval with query-by-keywords, query suggestion for content-based image retrieval (CBIR) with query-by-example (QBE) has been seldom studied. QBE usually suffers from the "intention gap" problem, especially when the user fails to get an appropriate query image to express his search intention precisely. In this paper, we propose a novel query suggestion scheme named Visual Query Attributes Suggestion (VQAS) for image search with QBE. Given a query image, informative attributes are suggested to the user as complements to the query. These attributes reflect the visual properties and key components of the query. By selecting some suggested attributes, the user can provide more precise search intent which is not captured by the query image. The evaluation results on two real-world image datasets show the effectiveness of VQAS in terms of retrieval performance and the quality of query suggestions.

This demonstration presents a new interactive Content Based Image Retrieval (CBIR) system, termed Attribute Feedback (AF). Unlike traditional relevance feedback purely founded on low-level features, AF system shapes user's search intents more precisely and quickly by collecting feedbacks on intermediate-level semantic attribute. At each interaction iteration, the AF system first determines the most informative binary attributes for feedbacks and then augments the binary attribute feedbacks by a new type of attributes, "affinity attributes", each of which is learnt offline to describe the distance/similarity between user's envisioned image(s) and a retrieved image with respect to the corresponding affinity attribute. Based on the feedbacks on binary and affinity attributes, the images in corpus are further re-ranked towards better fitting user's search intents. The experimental results on two real-world image datasets have demonstrated the superiority of the AF system over other state-of-the-art relevance feedback based CBIR approaches.

This work presents a new interactive Content Based Image Retrieval (CBIR) scheme, termed Attribute Feedback (AF). Unlike traditional relevance feedback purely founded on low-level visual features, the Attribute Feedback system shapes users' information needs more precisely and quickly by collecting feedbacks on intermediate level semantic attributes. At each interactive iteration, AF first determines the most informative binary attributes for feedbacks, preferring the attributes that frequently (rarely) appear in current search results but are unlikely (likely) to be users' interest. The binary attribute feedbacks are then augmented by a new type of attributes, "affinity attributes", each of which is off-line learnt to describe the distance between user's envisioned image(s) and a retrieved image with respect to the corresponding affinity attribute. Based on the feedbacks on binary and affinity attributes, the images in corpus are further re-ranked towards better fitting the users' information needs. Extensive experiments on two real-world image datasets well demonstrate the superiority of the proposed scheme over other state-of-the-art relevance feedback based CBIR solutions.

An image is worth of thousand words. Automatic Web image annotation is a practical and effective way for both Web image retrieval and image understanding. However, current annotation techniques are very difficult to get natural language interpretation for images such as ldquopandas eat bamboordquo. In this paper, we proposed an approach to interpret image semantics through semi-supervised mining annotated words. The idea in this approach mainly consists of three parts: at first, the visibility of annotated words of target image is calculated by semi-supervised learning approach from the landmark words in WordNet; then the annotated words are used as queries to retrieve matched Web pages; at last, the meaningful sentences in the matched Web pages are ranked as the interpretation of target image by semi-supervised learning approach. Experiments conducted on real-world Web images demonstrate the effectiveness of the proposed approach.