Large language models (LLMs) exhibit uneven performance across languages. In language-specific applications, practitioners often rely on target-language corpora or cross-lingual transfer to achieve better performance. However, traditional linguistic typology, commonly used as a transfer language selection strategy in previous studies, may not align with LLM’s perception of language similarity. This work proposes LLM-based language similarity as a novel perspective for selecting effective fine-tuning languages. We construct a framework to quantify the similarity within each language pair through both the lenses of language-specific performance patterns and cross-lingual transferability, ultimately deriving three similarity score matrices. Moreover, we observe a counter-intuitive phenomenon: super-additive transfer effect, where fine-tuning on a certain language yields higher performance than fine-tuning directly on the target language. Additionally, due to the absence of an existing dataset meeting our experimental requirements, we construct and release M4CQ-Pro dataset, which features domain-diverse distribution of 135 tasks and content consistency across 31 languages (including over 20 medium- and low-resource languages), with 61518 manually reviewed high-quality questions per language. We evaluate our approach on representative multilingual LLMs and results show that all three LLM-based similarity measures effectively guide fine-tuning language selection, outperforming traditional linguistic similarity, with the integrated measure achieving the best results. Our approach provides not only a novel perspective on language similarity, but also practical baselines for selecting fine-tuning languages.

Visual Language Models (VLMs) have become a robust foundation for document question answering.Processing long documents remains challenging due to limited context windows and computational budgets.Existing page-level retrieval methods offer a practical solution, typically encoding pages and queries into vectors and ranking them via cosine similarity.However, such embedding-based methods (i) lack query-page interaction before similarity scoring and (ii) usually require a large-scale dataset to align visual and textual embeddings.In this paper, we observe that the cross-modal attention maps of well-trained VLMs are able to highlight semantically relevant regions.Building on this insight, we present CAPS (Crossmodal Attention as Page Selector), a retrieval framework that utilizes attention mechanisms inside VLMs for page selection.Specifically, CAPS first enhances attention-based retrieval capability with a small amount of contrastive data, then identifies the most effective attention head through expert head selection, and finally employs an adaptive filtering mechanism to obtain an appropriate number of relevant page candidates.Extensive experiments on four long-document benchmarks demonstrate that CAPS outperforms state-of-the-art embedding-based methods in both retrieval precision and downstream DocQA accuracy.Notably, CAPS achieves these gains using less than 10% of the training data required by competing baselines, highlighting the data efficiency of attention-based page retrieval.

The rapidly developing Large Vision Language Models (LVLMs) still face the hallucination phenomena where the generated responses do not align with the given contexts, significantly restricting the usages of LVLMs. Most previous work detects and mitigates hallucination at the coarse-grained level or requires expensive annotation (e.g., labeling by human experts or proprietary models). To address these issues, we propose detecting and mitigating hallucinations in LVLMs via fine-grained AI feedback. The basic idea is that we generate a small-size sentence-level hallucination annotation dataset by proprietary models, whereby we train a detection model which can perform sentence-level hallucination detection. Then, we propose a detect-then-rewrite pipeline to automatically construct preference dataset for hallucination mitigation training. Furthermore, we propose differentiating the severity of hallucinations, and introducing a Hallucination Severity-Aware Direct Preference Optimization (HSA-DPO) which prioritizes the mitigation of critical hallucination in LVLMs by incorporating the severity of hallucinations into preference learning. Extensive experiments on hallucination detection and mitigation benchmarks demonstrate that our method sets a new state-of-the-art in hallucination detection on MHaluBench, surpassing GPT-4V and Gemini, and reduces the hallucination rate by 36.1% on AMBER and 76.3% on Object HalBench compared to the base model.

Recent advances in vision-language models have shown remarkable potential, yet creating scalable systems that can effectively understand and generate across modalities remains challenging. This talk will present our contributions to advancing scalable vision-language systems, focusing on three key themes: (1) efficient vision-language understanding, including our work on temporal perceiving video-language pre-training and knowledge-enhanced zero-shot retrieval; (2) scalable generation frameworks, encompassing our innovations in zero-shot captioning and co-speech gesture generation; and (3) practical applications and deployments of these technologies. We will discuss how these advances have enabled both better performance and improved efficiency in real-world scenarios, and explore future directions for scalable multimodal systems.

Recently, large reasoning models (LRMs) have demonstrated state-of-the-art performance across a wide range of benchmarks.However, a common challenge for these models is the "overthinking" problem, which leads to excessive reasoning steps and significant computational overhead.Furthermore, the issues with long Chain-of-Thought (CoT) are especially pronounced in smaller models (≤ 3B parameters).Aside from producing excessively verbose "reflection words", they often exhibit repetition and get trapped in unproductive generation loops.Existing solutions typically involve either using flexible reasoning chains as training data or leveraging the model's latent space to bypass intermediate reasoning steps, but none of these methods have considered directly optimizing reasoning trajectories during the sampling phase of training.In our work, we introduce the Multi-Turn Intervention Sampling Framework (MuTIS).Our framework leverages multi-turn interventions to produce concise reasoning chains.It fine-tunes reasoning models through reinforcement learning, demonstrably breaking the accuracy-efficiency trade-off.It also demonstrates strong scalability, exhibiting excellent performance on 7B models.

Despite recent advances in feed-forward 3D Gaussian Splatting, generalizable 3D reconstruction remains challenging, particularly in multi-view correspondence modeling. Existing approaches face a fundamental trade-off: explicit methods achieve geometric precision but struggle with ambiguous regions, while implicit methods provide robustness but suffer from slow convergence. We present H3R, a hybrid framework that addresses this limitation by integrating volumetric latent fusion with attention-based feature aggregation. Our framework consists of two complementary components: an efficient latent volume that enforces geometric consistency through epipolar constraints, and a camera-aware Transformer that leverages Plücker coordinates for adaptive correspondence refinement. By integrating both paradigms, our approach enhances generalization while converging 2$\times$ faster than existing methods. Furthermore, we show that spatial-aligned foundation models (e.g., SD-VAE) substantially outperform semantic-aligned models (e.g., DINOv2), resolving the mismatch between semantic representations and spatial reconstruction requirements. Our method supports variable-number and high-resolution input views while demonstrating robust cross-dataset generalization. Extensive experiments show that our method achieves state-of-the-art performance across multiple benchmarks, with significant PSNR improvements of 0.59 dB, 1.06 dB, and 0.22 dB on the RealEstate10K, ACID, and DTU datasets, respectively. Code is available at https://github.com/JiaHeng-DLUT/H3R.

Multi-modal Large language models (MLLMs) show remarkable ability in video understanding. Nevertheless, understanding long videos remains challenging as the models can only process a finite number of frames in a single inference, potentially omitting crucial visual information. To address the challenge, we propose generating multiple predictions through visual context sampling, followed by a scoring mechanism to select the final prediction. Specifically, we devise a bin-wise sampling strategy that enables MLLMs to generate diverse answers based on various combinations of keyframes, thereby enriching the visual context. To determine the final prediction from the sampled answers, we employ a self-reward by linearly combining three scores: (1) a frequency score indicating the prevalence of each option, (2) a marginal confidence score reflecting the inter-intra sample certainty of MLLM predictions, and (3) a reasoning score for different question types, including clue-guided answering for global questions and temporal self-refocusing for local questions. The frequency score ensures robustness through majority correctness, the confidence-aligned score reflects prediction certainty, and the typed-reasoning score addresses cases with sparse key visual information using tailored strategies. Experiments show that this approach covers the correct answer for a high percentage of long video questions, on seven datasets show that our method improves the performance of three MLLMs.

While multimodal large language models (MLLMs) have demonstrated extraordinary vision-language understanding capabilities, their abilities to solve instance-level visual-language problems beyond a single image warrant further exploration. To assess these unproven abilities of MLLMs, this paper proposes a new visual grounding task called multi-context visual grounding, which aims to localize instances of interest across multiple images based on open-ended text prompts. In order to facilitate this research, we construct a new dataset MC-Bench that features 2K high-quality and manually annotated samples. Each sample consists of an instance-level labeled image pair and a corresponding text prompt that indicates the target instances in the images. These text prompts are highly open-ended and follow three distinct styles, covering 20 practical skills. We benchmark over 20 state-of-the-art MLLMs and foundation models with potential multi-context visual grounding capabilities, along with our developed simple yet effective agentic baseline and a finetuned baseline by multi-context instruction tuning. Our evaluation reveals a non-trivial performance gap between existing MLLMs and humans, along with some insightful observations that suggest potential future directions. We hope that MC-Bench and our empirical findings encourage the research community to further advance the untapped potentials of MLLMs in instance-level tasks, particularly in multi-image contexts. Project page: https://xuyunqiu.github.io/MC-Bench.

Visual instructions for long-horizon tasks are crucial as they intuitively clarify complex concepts and enhance retention across extended steps. Directly generating a series of images using text-to-image models without considering the context of previous steps results in inconsistent images, increasing cognitive load. Additionally, the generated images often miss objects or the attributes such as color, shape, and state of the objects are inaccurate. To address these challenges, we propose LIGER, the first training-free framework for Long-horizon Instruction GEneration with logic and attribute self-Reflection. LIGER first generates a draft image for each step with the historical prompt and visual memory of previous steps. This step-by-step generation approach maintains consistency between images in long-horizon tasks. Moreover, LIGER utilizes various image editing tools to rectify errors including wrong attributes, logic errors, object redundancy, and identity inconsistency in the draft images. Through this self-reflection mechanism, LIGER improves the logic and object attribute correctness of the images. To verify whether the generated images assist human understanding, we manually curated a new benchmark consisting of various long-horizon tasks. Human-annotated ground truth expressions reflect the human-defined criteria for how an image should appear to be illustrative. Experiments demonstrate the visual instructions generated by LIGER are more comprehensive compared with baseline methods. The code and dataset will be available once accepted.

Recent advancements in diffusion models have significantly improved video generation and editing capabilities. However, multi-grained video editing, which encompasses class-level, instance-level, and part-level modifications, remains a formidable challenge. The major difficulties in multi-grained editing include semantic misalignment of text-to-region control and feature coupling within the diffusion model. To address these difficulties, we present VideoGrain, a zero-shot approach that modulates space-time (cross- and self-) attention mechanisms to achieve fine-grained control over video content. We enhance text-to-region control by amplifying each local prompt's attention to its corresponding spatial-disentangled region while minimizing interactions with irrelevant areas in cross-attention. Additionally, we improve feature separation by increasing intra-region awareness and reducing inter-region interference in self-attention. Extensive experiments demonstrate our method achieves state-of-the-art performance in real-world scenarios. Our code, data, and demos are available on the [project page](https://knightyxp.github.io/VideoGrain_project_page/).

Recent advances in operator learning have produced two distinct approaches for solving partial differential equations (PDEs): attention-based methods offering point-level adaptability but lacking spectral constraints, and spectral-based methods providing domain-level continuity priors but limited in local flexibility. This dichotomy has hindered the development of PDE solvers with both strong flexibility and generalization capability. This work introduces Holistic Physics Mixer (HPM), a novel framework that bridges this gap by integrating spectral and physical information in a unified space. HPM unifies both approaches as special cases while enabling more powerful spectral-physical interactions beyond either method alone. This enables HPM to inherit both the strong generalization of spectral methods and the flexibility of attention mechanisms while avoiding their respective limitations. Through extensive experiments across diverse PDE problems, we demonstrate that HPM consistently outperforms state-of-the-art methods in both accuracy and computational efficiency, while maintaining strong generalization capabilities with limited training data and excellent zero-shot performance on unseen resolutions.

Inverse design aims to design the input variables of a physical system to optimize a specified objective function, typically formulated as a search or optimization problem. However, in 3D domains, the design space grows exponentially, rendering exhaustive grid-based searches infeasible. Recent advances in deep learning have accelerated inverse design by providing powerful generative priors and differentiable surrogate models. Nevertheless, current methods tend to approximate the 3D design space using 2D projections or fine-tune existing 3D shapes. These approaches sacrifice volumetric detail and constrain design exploration, preventing true 3D design from scratch. In this paper, we propose a 3D Inverse Design (3DID) framework that directly navigates the 3D design space by coupling a continuous latent representation with a physics-aware optimization strategy. We first learn a unified physics–geometry embedding that compactly captures shape and physical field data in a continuous latent space. Then, we introduce a two-stage strategy to perform physics-aware optimization. In the first stage, a gradient-guided diffusion sampler explores the global latent manifold. In the second stage, an objectivedriven, topology-preserving refinement further sculpts each candidate toward the target objective. This enables 3DID to generate high-fidelity 3D geometries, outperforming existing methods in both solution quality and design versatility.

Recent advances in Large Language Models (LLMs) have shown promise in function-level code generation, yet repository-level software engineering tasks remain challenging. Current solutions predominantly rely on proprietary LLM agents, which introduce unpredictability and limit accessibility, raising concerns about data privacy and model customization. This paper investigates whether open-source LLMs can effectively address repository-level tasks without requiring agent-based approaches. We demonstrate this is possible by enabling LLMs to comprehend functions and files within codebases through their semantic information and structural dependencies. To this end, we introduce Code Graph Models (CGMs), which integrate repository code graph structures into the LLM's attention mechanism and map node attributes to the LLM's input space using a specialized adapter. When combined with an agentless graph RAG framework, our approach achieves a 43.00% resolution rate on the SWE-bench Lite benchmark using the open-source Qwen2.5-72B model. This performance ranks first among open weight models, second among methods with open-source systems, and eighth overall, surpassing the previous best open-source model-based method by 12.33%.

The limited availability of high-quality training data poses a major obstacle in data-driven PDE solving, where expensive data collection and resolution constraints severely impact the ability of neural operator networks to learn and generalize the underlying physical system. To address this challenge, we propose DeltaPhi, a novel learning framework that transforms the PDE solving task from learning direct input-output mappings to learning the residuals between similar physical states, a fundamentally different approach to neural operator learning. This reformulation provides implicit data augmentation by exploiting the inherent stability of physical systems where closer initial states lead to closer evolution trajectories. DeltaPhi is architecture-agnostic and can be seamlessly integrated with existing neural operators to enhance their performance. Extensive experiments demonstrate consistent and significant improvements across diverse physical systems including regular and irregular domains, different neural architectures, multiple training data amount, and cross-resolution scenarios, confirming its effectiveness as a general enhancement for neural operators in data-limited PDE solving.

Long-form video understanding poses a significant challenge for video large language models (VideoLLMs) due to prohibitively high computational and memory demands. In this paper, We propose $\textbf{FlexSelect}$, a flexible and efficient token selection strategy for processing long videos. FlexSelect identifies and retains the most semantically relevant content by leveraging cross-modal attention patterns from a reference transformer layer. It comprises two key components: (1) $\textbf{a training-free token ranking pipeline}$ that leverages faithful cross-modal attention weights to estimate each video token’s importance, and (2) $\textbf{a rank-supervised lightweight selector}$ that is trained to replicate these rankings and filter redundant tokens. This generic approach can be seamlessly integrated into various VideoLLM architectures, such as LLaVA-Video, InternVL and Qwen-VL, serving as a plug-and-play module to extend their temporal context length. Empirically, FlexSelect delivers strong gains across multiple long-video benchmarks – including VideoMME, MLVU, LongVB, and LVBench. Morever, it achieves significant speed-ups ($\textit{e.g.,}$ up to 9 $\times$ on a LLaVA-Video-7B model), highlighting FlexSelect’s promise for efficient long-form video understanding. Project page: https://flexselect.github.io

Video-language pre-training models have recently achieved remarkable results on various multi-modal downstream tasks. However, most of these models rely on contrastive learning or masking modeling to align global features across modalities, neglecting the local associations between video frames and text tokens. This limits the model’s ability to perform fine-grained matching and generalization, especially for tasks that selecting segments in long videos based on query texts. To address this issue, we propose a novel stitching and matching pre-text task for video-language pre-training that encourages fine-grained interactions between modalities. Our task involves stitching video frames or sentences into longer sequences and predicting the positions of cross-model queries in the stitched sequences. The individual frame and sentence representations are thus aligned via the stitching and matching strategy, encouraging the fine-grained interactions between videos and texts. in the stitched sequences for the cross-modal query. We conduct extensive experiments on various benchmarks covering text-to-video retrieval, video question answering, video captioning, and moment retrieval. Our results demonstrate that the proposed method significantly improves the generalization capacity of the video-text pre-training models.

Text-video retrieval is a critical multi-modal task to find the most relevant video for a text query. Although pretrained models like CLIP have demonstrated impressive potential in this area, the rising cost of fully finetuning these models due to increasing model size continues to pose a problem. To address this challenge, prompt tuning has emerged as an alternative. However, existing works still face two problems when adapting pretrained image-text models to downstream video-text tasks: (1) The visual encoder could only encode frame-level features and failed to extract global-level general video information. (2) Equipping the visual and text encoder with separated prompts failed to mitigate the visual-text modality gap. To this end, we propose DGL, a cross-modal Dynamic prompt tuning method with Global-Local video attention. In contrast to previous prompt tuning methods, we employ the shared latent space to generate local-level text and frame prompts that encourage inter-modal interaction. Furthermore, we propose modeling video in a global-local attention mechanism to capture global video information from the perspective of prompt tuning. Extensive experiments reveal that when only 0.67% parameters are tuned, our cross-modal prompt tuning strategy DGL outperforms or is comparable to fully finetuning methods on MSR-VTT, VATEX, LSMDC, and ActivityNet datasets. Code will be available at https://github.com/knightyxp/DGL.

We concentrate on a novel human-centric image synthesis task, that is, given only one reference facial photograph, it is expected to generate specific individual images with diverse head positions, poses, facial expressions, and illuminations in different contexts. To accomplish this goal, we argue that our generative model should be capable of the following favorable characteristics: (1) a strong visual and semantic understanding of our world and human society for basic object and human image generation. (2) generalizable identity preservation ability. (3) flexible and fine-grained head control. Recently, large pre-trained text-to-image diffusion models have shown remarkable results, serving as a powerful generative foundation. As a basis, we aim to unleash the above two capabilities of the pre-trained model. In this work, we present a new framework named CapHuman. We embrace the “encode then learn to align” paradigm, which enables generalizable identity preservation for new individuals without cumbersome tuning at inference. CapHuman encodes identity features and then learns to align them into the latent space. Moreover, we introduce the 3D facial prior to equip our model with control over the human head in a flexible and 3D-consistent manner. Extensive qualitative and quantitative analyses demonstrate our CapHuman can produce well-identity-preserved, photo-realistic, and high-fidelity portraits with content-rich representations and various head renditions, superior to established baselines. Code and checkpoint will be released at https://github.com/VamosC/CapHuman.

We study the zero-shot Composed Image Retrieval (ZS- CIR) task, which is to retrieve the target image given a reference image and a description without training on the triplet datasets. Previous works generate pseudo-word tokens by projecting the reference image features to the text embedding space. However, they focus on the global visual representation, ignoring the representation of detailed attributes, e.g., color, object number and layout. To address this challenge, we propose a Knowledge-Enhanced Dual-stream zero-shot composed image retrieval framework (KEDs). KEDs implicitly models the attributes of the reference images by incorporating a database. The database enriches the pseudo-word tokens by providing relevant images and captions, emphasizing shared attribute information in various aspects. In this way, KEDs recognizes the reference image from diverse perspectives. Moreover, KEDs adopts an extra stream that aligns pseudo-word tokens with textual concepts, leveraging pseudo-triplets mined from image-text pairs. The pseudo-word tokens generated in this stream are explicitly aligned with fine-grained semantics in the text embedding space. Extensive experiments on widely used benchmarks, i.e. ImageNet-R, COCO object, Fashion-IQ and CIRR, show that KEDs outperforms previous zero-shot composed image retrieval methods. Code is available at https://github.com/suoych/KEDs.

One fascinating aspect of pre-trained vision-language models (VLMs) learning under language supervision is their impressive zero-shot generalization capability. However, this ability is hindered by distribution shifts between the training and testing data. Previous test time adaptation (TTA) methods for VLMs in zero-shot classification rely on minimizing the entropy of model outputs, tending to be stuck in incorrect model predictions. In this work, we propose TTA with feedback to rectify the model output and prevent the model from becoming blindly confident. Specifically, a CLIP model is adopted as the reward model during TTA and provides feedback for the VLM. Given a single test sample, the VLM is forced to maximize the CLIP reward between the input and sampled results from the VLM output distribution. The proposed \textit{reinforcement learning with CLIP feedback~(RLCF)} framework is highly flexible and universal. Beyond the classification task, with task-specific sampling strategies and a proper reward baseline choice, RLCF can be easily extended to not only discrimination tasks like retrieval but also generalization tasks like image captioning, improving the zero-shot generalization capacity of VLMs. According to the characteristics of these VL tasks, we build different fully TTA pipelines with RLCF to improve the zero-shot generalization ability of various VLMs. Extensive experiments along with promising empirical results demonstrate the effectiveness of RLCF. The code is available at https://github.com/mzhaoshuai/RLCF.

Video captioning is a challenging task and typically requires paired video-text data for training. However, manually annotating coherent textual descriptions for videos is laborious and time-consuming. To address this challenge, we propose a novel approach that enhances video captioning using only synthetic text data. Leveraging the exceptional text generation capabilities of large language models (LLMs), we produce high-quality and diverse video captions tailored to the target domain. Our approach employs a two-stage prompting strategy: first prompt GPT-4 with few-shot target-domain captions to create a set of high-quality captions, and then continue prompting with the generated captions to acquire large-scale synthetic data. To effectively utilize these captions, we introduce Mixture of Scale and Shift experts (MoS2), an efficient adaptation method for pre-trained captioning models. MoS2 employs lightweight routing networks to estimate probability distributions over a collection of scale and shift experts, dynamically allocating tokens to the appropriate experts. This dynamic adjustment mechanism enhances the model's ability to handle data variations and mitigates the distribution shift between synthetic and real captions. Moreover, our method reduces the number of learnable parameters, facilitating more efficient adaptation. Our method achieves superior performance with only synthetic text data, narrowing the gap between zero-shot and fine-tuned models and reducing the dependency on paired data from the target domain.

Maintaining temporal stability is crucial in multi-agent trajectory prediction. Insufficient regularization to uphold this temporal stability often results in fluctuations in kinematic states, leading to inconsistent predictions and the amplification of errors. In this study, we introduce a framework called Multi-Agent Trajectory prediction via neural interaction Energy (MATE). This framework assesses the interactive motion of agents by employing neural interaction energy, which captures the dynamics of interactions and illustrates their influence on the future trajectories of agents. To bolster temporal stability, we introduce two constraints: inter-agent interaction constraint and intra-agent motion constraint. These constraints work together to ensure temporal stability at both the system and agent levels, effectively mitigating prediction fluctuations inherent in multi-agent systems. Comparative evaluations against previous methods on four diverse datasets, including simulated and real-world scenarios, highlight the superior prediction accuracy and generalization capabilities of our model.

Text-driven 3D avatar customization has attracted increasing attention in recent years, where precisely editing specific local parts of avatars with only text prompts is particularly challenging. Previous editing methods usually use segmentation or cross-attention masks as constraints for local editing. Although these masks tightly cover existing objects/parts, they may limit editing methods to create drastic geometry deformations beyond the covered contents. From a different perspective, this paper presents a GPT-guided local avatar editing framework, namely GG-Editor. Specifically, GG-Editor progressively mines more reasonable candidate editing regions via harnessing multimodal large language models which already organically assimilate common-sense human knowledge. In order to improve the editing quality of the local areas, GG-Editor explicitly decouples the geometry/appearance optimization, and adopts a global-local synergy editing strategy with GPT-generated local prompts. Moreover, to preserve concepts residing in source avatars, GG-Editor proposes an orthogonal denoising score that orthogonally decomposes editing directions and introduce an explicit term for preservation. Comprehensive experiments demonstrate that GG-Editor with only textual prompts achieves realistic and high-fidelity local editing results, significantly surpassing prior works. Project page: https://xuyunqiu.github.io/GG-Editor/.

Video diffusion models have made substantial progress in various video generation applications. However, training models for long video generation tasks require significant computational and data resources, posing a challenge to developing long video diffusion models. This paper investigates a straightforward and training-free approach to extend an existing short video diffusion model (e.g. pre-trained on 16-frame videos) for consistent long video generation (e.g. 128 frames). Our preliminary observation has found that directly applying the short video diffusion model to generate long videos can lead to severe video quality degradation. Further investigation reveals that this degradation is primarily due to the distortion of high-frequency components in long videos, characterized by a decrease in spatial high-frequency components and an increase in temporal high-frequency components. Motivated by this, we propose a novel solution named FreeLong to balance the frequency distribution of long video features during the denoising process. FreeLong blends the low-frequency components of global video features, which encapsulate the entire video sequence, with the high-frequency components of local video features that focus on shorter subsequences of frames. This approach maintains global consistency while incorporating diverse and high-quality spatiotemporal details from local videos, enhancing both the consistency and fidelity of long video generation. We evaluated FreeLong on multiple base video diffusion models and observed significant improvements. Additionally, our method supports coherent multi-prompt generation, ensuring both visual coherence and seamless transitions between scenes. Our project page is at: https://yulu.net.cn/freelong.

Video grounding, the process of identifying a specific moment in an untrimmed video based on a natural language query, has become a popular topic in video understanding. However, fully supervised learning approaches for video grounding that require large amounts of annotated data can be expensive and time-consuming. Recently, zero-shot video grounding (ZS-VG) methods that leverage pre-trained object detectors and language models to generate pseudo-supervision for training video grounding models have been developed. However, these approaches have limitations in recognizing diverse categories and capturing specific dynamics and interactions in the video context. To tackle these challenges, we introduce a novel two-stage ZS-VG framework called Lookup-and-Verification (LoVe), which treats the pseudo-query generation procedure as a video-to-concept retrieval problem. Our approach allows for the extraction of diverse concepts from an open-concept pool and employs a verification process to ensure the relevance of the retrieved concepts to the objects or events of interest in the video proposals. Comprehensive experimental results on the Charades-STA, ActivityNet-Captions, and DiDeMo datasets demonstrate the effectiveness of the LoVe framework.

Large-scale pre-trained vision-language models (e.g., CLIP) have shown incredible generalization performance in downstream tasks such as video-text retrieval (VTR). Traditional approaches have leveraged CLIP's robust multi-modal alignment ability for VTR by directly fine-tuning vision and text encoders with clean video-text data. Yet, these techniques rely on carefully annotated video-text pairs, a process that is costly and labor-intensive. In this context, we introduce a new approach, Pseudo-Supervised Selective Contrastive Learning (PS-SCL). PS-SCL minimizes the dependency on manually-labeled text annotations by generating pseudo-supervisions from unlabeled video data for training. We first exploit CLIP's visual recognition capabilities to generate pseudo-texts automatically. These pseudo-texts contain diverse visual concepts from the video and serve as weak textual guidance. Moreover, we introduce Selective Contrastive Learning (SeLeCT), which prioritizes and selects highly correlated video-text pairs from pseudo-supervised video-text pairs. By doing so, SeLeCT enables more effective multi-modal learning under weak pairing supervision. Experimental results demonstrate that our method outperforms CLIP zero-shot performance by a large margin on multiple video-text retrieval benchmarks, e.g., 8.2% R@1 for video-to-text on MSRVTT, 12.2% R@1 for video-to-text on DiDeMo, and 10.9% R@1 for video-to-text on ActivityNet, respectively.

Pre-trained vision-language models (VLMs) are the de-facto foundation models for various downstream tasks. However, scene text recognition methods still prefer backbones pre-trained on a single modality, namely, the visual modality, despite the potential of VLMs to serve as powerful scene text readers. For example, CLIP can robustly identify regular (horizontal) and irregular (rotated, curved, blurred, or occluded) text in images. With such merits, we transform CLIP into a scene text reader and introduce CLIP4STR, a simple yet effective STR method built upon image and text encoders of CLIP. It has two encoder-decoder branches: a visual branch and a cross-modal branch. The visual branch provides an initial prediction based on the visual feature, and the cross-modal branch refines this prediction by addressing the discrepancy between the visual feature and text semantics. To fully leverage the capabilities of both branches, we design a dual predict-and-refine decoding scheme for inference. We scale CLIP4STR in terms of the model size, pre-training data, and training data, achieving state-of-the-art performance on 13 STR benchmarks. Additionally, a comprehensive empirical study is provided to enhance the understanding of the adaptation of CLIP to STR. We believe our method establishes a simple yet strong baseline for future STR research with VLMs.

Video captioning is a more challenging task compared to image captioning, primarily due to differences in content density. Video data contains redundant visual content, making it difficult for captioners to generalize diverse content and avoid being misled by irrelevant elements. Moreover, redundant content is not well-trimmed to match the corresponding visual semantics in the ground truth, further increasing the difficulty of video captioning. Current research in video captioning predominantly focuses on captioner design, neglecting the impact of content density on captioner performance. Considering the differences between videos and images, there exists an another line to improve video captioning by leveraging concise and easily-learned image samples to further diversify video samples. This modification to content density compels the captioner to learn more effectively against redundancy and ambiguity. In this article, we propose a novel approach calledImage-Compounded learning for videoCaptioners (IcoCap) to facilitate better learning of complex video semantics. IcoCap comprises two components: the Image-Video Compounding Strategy (ICS) and Visual-Semantic Guided Captioning (VGC). ICS compounds easily-learned image semantics into video semantics, further diversifying video content and prompting the network to generalize contents in a more diverse sample. Besides, learning with the sample compounded with image contents, the captioner is compelled to better extract valuable video cues in the presence of straightforward image semantics. This helps the captioner further focus on relevant information while filtering out extraneous content. Then, VGC guides the network in flexibly learning ground truth captions based on the compounded samples, helping to mitigate the mismatch between the ground truth and ambiguous semantics in video samples. Our experimental results demonstrate the effectiveness of IcoCap in improving the learning of video captioners. Applied to the widely-used MSVD, MSR-VTT, and VATEX datasets, our approach achieves competitive or superior results compared to state-of-the-art methods, illustrating its capacity to handle the redundant and ambiguous video data

In this article, we present skeleton-based isolated sign language recognition (IsoSLR) with part mixing - SKIM. An IsoSLR model that solely takes the skeleton representation of the human body as input. Previous skeleton-based works either perform worse when compared to RGB-based counterparts or require fusion with other modalities to obtain competitive results. With SKIM, a single skeleton-based model without complex pre-training can obtain similar or even higher accuracy than current state-of-the-art methods. This margin can be further increased by simple late fusion within the same modality. To achieve this, we first develop a novel data augmentation technique called part mixing. It swaps the corresponding keypoints within one region (e.g. hand) between two randomly selected samples and combines their labels linearly as the new label. As regions like hand and face are key articulators for sign language, direct swapping of such parts creates a believable pseudo sign that promotes the model to recognize the true pairs. Secondly, following current advances in skeleton-based action recognition, we devise a channel-wise graph neural network with multi-scale awareness and per-keypoint temporal re-weighting. With this design, the backbone is capable of leveraging both manual and non-manual features. The combination of hand mixing and the channel-wise multi-scale GCN backbone allows us to achieve state-of-the-art accuracy on both WLASL and NMFs-CSL benchmarks.

Illustrating a multi-sentence story with visual content is a significant challenge in multimedia research. While previous works have focused on sequential story-to-visual representations at the image level or representing a single sentence with a video clip, illustrating a long multi-sentence story with coherent videos remains an under-explored area. In this paper, we propose the task of video-based story illustration that focuses on the goal of visually illustrating a story with retrieved video clips. To support this task, we first create a large-scale dataset of coherent video stories in each sample, consisting of 85K narrative stories with 60 pairs of consistent clips and texts. We then propose the Story Context-Enhanced Model, which leverages local and global contextual information within the story, inspired by sequence modeling in language understanding. Through comprehensive quantitative experiments, we demonstrate the effectiveness of our baseline model. In addition, qualitative results and detailed user studies reveal that our method can retrieve coherent video sequences from stories. The dataset and code will be made publicly athttps://nfy-dot.github.io/CVSV-dataset/.

Though significant progress has been achieved on fine-grained visual classification (FGVC), severe overfitting still hinders model generalization. A recent study shows that hard samples in the training set can be easily fit, but most existing FGVC methods fail to classify some hard examples in the test set. The reason is that the model overfits those hard examples in the training set, but does not learn to generalize to unseen examples in the test set. In this article, we propose a moderate hard example modulation (MHEM) strategy to properly modulate the hard examples. MHEM encourages the model to not overfit hard examples and offers better generalization and discrimination. First, we introduce three conditions and formulate a general form of a modulated loss function. Second, we instantiate the loss function and provide a strong baseline for FGVC, where the performance of a naive backbone can be boosted and be comparable with recent methods. Moreover, we demonstrate that our baseline can be readily incorporated into the existing methods and empower these methods to be more discriminative. Equipped with our strong baseline, we achieve consistent improvements on three typical FGVC datasets, i.e., CUB-200-2011, Stanford Cars, and FGVC-Aircraft. We hope the idea of moderate hard example modulation will inspire future research work toward more effective fine-grained visual recognition.

Few-shot image classification aims at exploring transferable features from base classes to recognize images of the unseen novel classes with only a few labeled images. Existing methods usually compare the support features and query features, which are implemented by either matching the global feature vectors or matching the local feature maps at the same position. However, few labeled images fail to capture all the diverse context and intraclass variations, leading to mismatch issues for existing methods. On one hand, due to the misaligned position and cluttered background, existing methods suffer from the object mismatch issue. On the other hand, due to the scale inconsistency between images, existing methods suffer from the scale mismatch issue. In this article, we propose the bilaterally normalized scale-consistent Sinkhorn distance (BSSD) to solve these issues. First, instead of same-position matching, we use the Sinkhorn distance to find an optimal matching between images, mitigating the object mismatch caused by misaligned position. Meanwhile, we propose the intraimage and interimage attentions as the bilateral normalization on the Sinkhorn distance to suppress the object mismatch caused by background clutter. Second, local feature maps are enhanced with the multiscale pooling strategy, making the Sinkhorn distance possible to find a consistent matching scale between images. Experimental results show the effectiveness of the proposed approach, and we achieve the state-of-the-art on three few-shot benchmarks.

This work explores visual recognition models on real-world datasets exhibiting a long-tailed distribution. Most of previous works are based on a holistic perspective that the overall gradient for training model is directly obtained by considering all classes jointly. However, due to the extreme data imbalance in long-tailed datasets, joint consideration of different classes tends to induce the gradient distortion problem; i.e., the overall gradient tends to suffer from shifted direction toward data-rich classes and enlarged variances caused by data-poor classes. The gradient distortion problem impairs the training of our models. To avoid such drawbacks, we propose to disentangle the overall gradient and aim to consider the gradient on data-rich classes and that on data-poor classes separately. We tackle the long-tailed visual recognition problem via a dual-phase-based method. In the first phase, only data-rich classes are concerned to update model parameters, where only separated gradient on data-rich classes is used. In the second phase, the rest data-poor classes are involved to learn a complete classifier for all classes. More importantly, to ensure the smooth transition from phase I to phase II, we propose an exemplar bank and a memory-retentive loss. In general, the exemplar bank reserves a few representative examples from data-rich classes. It is used to maintain the information of data-rich classes when transiting. The memory-retentive loss constrains the change of model parameters from phase I to phase II based on the exemplar bank and data-poor classes. The extensive experimental results on four commonly used long-tailed benchmarks, including CIFAR100-LT, Places-LT, ImageNet-LT, and iNaturalist 2018, highlight the excellent performance of our proposed method.

We focus on the challenging problem of learning an unbiased classifier from a large number of potentially relevant but noisily labeled web images given only a few clean labeled images. This problem is particularly practical because it reduces the expensive annotation costs by utilizing freely accessible web images with noisy labels. Typically, prototypes are representative images or features used to classify or identify other images. However, in the few clean and many noisy scenarios, the class prototype can be severely biased due to the presence of irrelevant noisy images. The resulting prototypes are less compact and discriminative, as previous methods do not take into account the diverse range of images in the noisy web image collections. On the other hand, the relation modeling between noisy and clean images is not learned for the class prototype generation in an end-to-end manner, which results in a suboptimal class prototype. In this article, we introduce a similarity maximization loss named SimNoiPro. Our SimNoiPro first generates noise-tolerant hybrid prototypes composed of clean and noise-tolerant prototypes and then pulls them closer to each other. Our approach considers the diversity of noisy images by explicit division and overcomes the optimization discrepancy issue. This enables better relation modeling between clean and noisy images and helps extract judicious information from the noisy image set. The evaluation results on two extended few-shot classification benchmarks confirm that our SimNoiPro outperforms prior methods in measuring image relations and cleaning noisy data.

In this paper, we tackle the problem of sign language translation (SLT) without gloss annotations.Although intermediate representation like gloss has been proven effective, gloss annotations are hard to acquire, especially in large quantities.This limits the domain coverage of translation datasets, thus handicapping real-world applications.To mitigate this problem, we design the Gloss-Free End-to-end sign language translation framework (GloFE).Our method improves the performance of SLT in the gloss-free setting by exploiting the shared underlying semantics of signs and the corresponding spoken translation.Common concepts are extracted from the text and used as a weak form of intermediate representation.The global embedding of these concepts is used as a query for cross-attention to find the corresponding information within the learned visual features.In a contrastive manner, we encourage the similarity of query results between samples containing such concepts and decrease those that do not.We obtained state-of-the-art results on large-scale datasets, including OpenASL and How2Sign. 1

This paper presents a whitening-based contrastive learning method for sentence embedding learning (WhitenedCSE), which combines contrastive learning with a novel shuffled group whitening.Generally, contrastive learning pulls distortions of a single sample (i.e., positive samples) close and push negative samples far away, correspondingly facilitating the alignment and uniformity in the feature space.A popular alternative to the "pushing" operation is whitening the feature space, which scatters all the samples for uniformity.Since the whitening and the contrastive learning have large redundancy w.r.t. the uniformity, they are usually used separately and do not easily work together.For the first time, this paper integrates whitening into the contrastive learning scheme and facilitates two benefits.1) Better uniformity.We find that these two approaches are not totally redundant but actually have some complementarity due to different uniformity mechanism.2) Better alignment.We randomly divide the feature into multiple groups along the channel axis and perform whitening independently within each group.By shuffling the group division, we derive multiple distortions of a single sample and thus increase the positive sample diversity.Consequently, using multiple positive samples with enhanced diversity further improves contrastive learning due to better alignment.Extensive experiments on seven semantic textual similarity tasks show our method achieves consistent improvement over the contrastive learning baseline and sets new states of the art, e.g., 78.78% (+2.53% based on BERT base ) Spearman correlation on STS tasks. 1

Deep neural networks on regular 1D lists (e.g., natural languages) and irregular 3D sets (e.g., point clouds) have made tremendous achievements. The key to natural language processing is to model words and their regular order dependency in texts. For point cloud understanding, the challenge is to understand the geometry via irregular point coordinates, in which point-feeding orders do not matter. However, there are a few kinds of data that exhibit both regular 1 D list and irregular 3D set structures, such as proteins and non-coding RNAs. In this paper, we refer to them as 3D point lists and propose a Transformer-style PointListNet to model them. First, PointListNet employs non-parametric distance-based attention because we find sometimes it is the distance, instead of the feature or type, that mainly determines how much two points, e.g., amino acids, are correlated in the micro world. Second, different from the vanilla Transformer that directly performs a simple linear transformation on inputs to generate values and does not explicitly model relative relations, our PointListNet integrates the 1D order and 3D Euclidean displacements into values. We conduct experiments on protein fold classification and enzyme reaction classification. Experimental results show the effectiveness of the proposed PointListNet.

To build Video Question Answering (VideoQA) systems capable of assisting humans in daily activities, seeking answers from long-form videos with diverse and complex events is a must. Existing multi-modal VQA models achieve promising performance on images or short video clips, especially with the recent success of large-scale multi-modal pre-training. However, when extending these methods to long-form videos, new challenges arise. On the one hand, using a dense video sampling strategy is computationally prohibitive. On the other hand, methods relying on sparse sampling struggle in scenarios where multi-event and multi-granularity visual reasoning are required. In this work, we introduce a new model named$\mathcal{M}ulti{-}$· modal Iterative$\mathcal{S}$.patial-temporal Transformer$(\mathcal{MIST})$) to better adapt pre-trained models for long-form VideoQA. Specifically,$\mathcal{MIST}$decomposes traditional dense spatial-temporal self-attention into cascaded segment and region selection modules that adaptively select frames and image regions that are closely relevant to the question itself. Visual concepts at different granularities are then processed efficiently through an attention module. In addition,$\mathcal{MIST}$iteratively conducts selection and attention over multiple layers to support reasoning over multiple events. The experimental results on four VideoQA datasets, including AGQA, NExT-QA, STAR, and Env-QA, show that$\mathcal{MIST}$achieves state-of-the-art performance and is superior at efficiency. The code is available at github.com/showlab/mist.

Large-scale pre-training has brought unimodal fields such as computer vision and natural language processing to a new era. Following this trend, the size of multimodal learning models constantly increases, leading to an urgent need to reduce the massive computational cost of finetuning these models for downstream tasks. In this paper, we propose an efficient and flexible multimodal fusion method, namely PMF, tailored for fusing unimodally pretrained transformers. Specifically, we first present a modular multimodal fusion framework that exhibits high flexibility and facilitates mutual interactions among different modalities. In addition, we disentangle vanilla prompts into three types in order to learn different optimizing objectives for multimodal learning. It is also worth noting that we propose to add prompt vectors only on the deep layers of the unimodal transformers, thus significantly reducing the training memory usage. Experiment results show that our proposed method achieves comparable performance to several other multimodal finetuning methods with less than 3% trainable parameters and up to 66% saving of training memory usage.

Inferring affordance for 3D articulated objects is a challenging and practical problem. It is a primary problem for applying robots to real-world scenarios. The exploration can be summarized as figuring out where to act and how to act. Correspondingly, the task mainly requires producing actionability scores, action proposals, and success likelihood scores according to the given 3D object information and robotic information. Current works usually directly process multi-modal inputs with early fusion and apply critic networks to produce scores, which leads to insufficient multi-modal learning ability and inefficiently iterative training in multiple stages. This paper proposes a novel Multimodality-Aware Autoencoder-based affordance Learning (MAAL) for the 3D object affordance problem. It is an efficient pipeline, trained in one go, and only requires a few positive samples in training data. More importantly, MAAL contains a MultiModal Energized Encoder (MME) for better multi-modal learning. It comprehensively models all multi-modal inputs from 3D objects and robotic actions. Jointly considering information from multiple modalities, the encoder further learns interactions between robots and objects. MME empowers the better multi-modal learning ability for understanding object affordance. Experimental results and visualizations, based on a large-scale dataset PartNet-Mobility, show the effectiveness of MAAL in learning multi-modal data and solving the 3D articulated object affordance problem.

Temporal grounding is the task of locating a specific segment from an untrimmed video according to a query sentence. This task has achieved significant momentum in the computer vision community as it enables activity grounding beyond pre-defined activity classes by utilizing the semantic diversity of natural language descriptions. The semantic diversity is rooted in the principle of compositionality in linguistics, where novel semantics can be systematically described by combining known words in novel ways (compositional generalization). However, existing temporal grounding datasets are not carefully designed to evaluate the compositional generalizability. To systematically benchmark the compositional generalizability of temporal grounding models, we introduce a new Compositional Temporal Grounding task and construct two new dataset splits, i.e., Charades-CG and ActivityNet-CG. We empirically find that they fail to generalize to queries with novel combinations of seen words. We argue that the inherent compositional structure (i.e., composition constituents and their relationships) inside the videos and language is the crucial factor to achieve compositional generalization. Based on this insight, we propose a variational cross-graph reasoning framework that explicitly decomposes video and language into hierarchical semantic graphs, respectively, and learns fine-grained semantic correspondence between the two graphs. Meanwhile, we introduce a novel adaptive structured semantics learning approach to derive the structure-informed and domain-generalizable graph representations, which facilitate the fine-grained semantic correspondence reasoning between the two graphs. To further evaluate the understanding of the compositional structure, we also introduce a more challenging setting, where one of the components in the novel composition is unseen. This requires more sophisticated understanding of the compositional structure to infer the potential semantics of the unseen word based on the other learned composition constituents appearing in both the video and language context, and their relationships. Extensive experiments validate the superior compositional generalizability of our approach, demonstrating its ability to handle queries with novel combinations of seen words as well as novel words in the testing composition.

In this paper, we propose to tackle egocentric action recognition by suppressing background distractors and enhancing action-relevant interactions. The existing approaches usually utilize two independent branches to recognize egocentric actions, i.e., a verb branch and a noun branch. However, the mechanism to suppress distracting objects and exploit local human-object correlations is missing. To this end, we introduce two extra sources of information, i.e., the candidate objects spatial location and their discriminative features, to enable concentration on the occurring interactions. We design a Symbiotic Attention with Object-centric feature Alignment framework (SAOA) to provide meticulous reasoning between the actor and the environment. First, we introduce an object-centric feature alignment method to inject the local object features to the verb branch and noun branch. Second, we propose a symbiotic attention mechanism to encourage the mutual interaction between the two branches and select the most action-relevant candidates for classification. The framework benefits from the communication among the verb branch, the noun branch, and the local object information. Experiments based on different backbones and modalities demonstrate the effectiveness of our method. Notably, our framework achieves the state-of-the-art on the largest egocentric video dataset.

Compared with 2D image salient object detection (SOD), SOD in omnidirectional images (or 360° images) usually suffers from geometric distortion. Although existing omnidirectional image SOD (ODI-SOD) methods have improved the detection accuracy obviously, their application may be cumbersome in real scenes due to their high computational cost. To avoid distortion and reduce the computational cost simultaneously in ODI-SOD, we propose a novel lightweight distortion-aware network, named LDNet, in this letter. First, to extract features with less distortion from ODIs, we integrate the distortion-aware convolution and depth-wise separable convolution (DSConv) into distortion-aware DSConv (DDSConv) and replace the regular convolutions in the last two blocks of the ResNet-18 with DDSConvs to obtain our lightweight backbone network (LD-ResNet-18). To enhance spatial information in each channel of the extracted features at each level comprehensively, then, we propose a lightweight distortion-aware channel-wise enhancement (DCE) module (only 0.05M parameters) including DDSConvs with various dilation rates, channel shuffle operation and attention mechanism, and employ a high-to-low dense connection structure to modulate the enhanced multi-level features. Besides, we design a distortion-aware self-correlation (DSC) module (only 0.02M parameters) for mining the contextual dependency of the features via a coarse-fine strategy, and the correlated features are refined by DCE modules and integrated by another dense connection structure. The final saliency map is predicted from the densely integrated features. Compared with 12 state-of-the-art methods on two public datasets, our lightweight LDNet achieves competitive or even better performance with only 2.9M parameters and 3.4G FLOPs, which balances the efficiency and performance.

Domain adaptation methods reduce domain shift typically by learning domain-invariant features. Most existing methods are built on distribution matching, e.g., adversarial domain adaptation, which tends to corrupt feature discriminability. In this paper, we propose Discriminative Radial Domain Adaptation (DRDA) which bridges source and target domains via a shared radial structure. It's motivated by the observation that as the model is trained to be progressively discriminative, features of different categories expand outwards in different directions, forming a radial structure. We show that transferring such an inherently discriminative structure would enable to enhance feature transferability and discriminability simultaneously. Specifically, we represent each domain with a global anchor and each category a local anchor to form a radial structure and reduce domain shift via structure matching. It consists of two parts, namely isometric transformation to align the structure globally and local refinement to match each category. To enhance the discriminability of the structure, we further encourage samples to cluster close to the corresponding local anchors based on optimal-transport assignment. Extensively experimenting on multiple benchmarks, our method is shown to consistently outperforms state-of-the-art approaches on varied tasks, including the typical unsupervised domain adaptation, multi-source domain adaptation, domain-agnostic learning, and domain generalization.

Weakly supervised person search involves training a model with only bounding box annotations, without human-annotated identities. Clustering algorithms are commonly used to assign pseudo-labels to facilitate this task. However, inaccurate pseudo-labels and imbalanced identity distributions can result in severe label and sample noise. In this work, we propose a novel Collaborative Contrastive Refining (CCR) weakly-supervised framework for person search that jointly refines pseudo-labels and the sample-learning process with different contrastive strategies. Specifically, we adopt a hybrid contrastive strategy that leverages both visual and context clues to refine pseudo-labels, and leverage the sample-mining and noise-contrastive strategy to reduce the negative impact of imbalanced distributions by distinguishing positive samples and noise samples. Our method brings two main advantages: 1) it facilitates better clustering results for refining pseudo-labels by exploring the hybrid similarity; 2) it is better at distinguishing query samples and noise samples for refining the sample-learning process. Extensive experiments demonstrate the superiority of our approach over the state-of-the-art weakly supervised methods by a large margin (more than 3% mAP on CUHK-SYSU). Moreover, by leveraging more diverse unlabeled data, our method achieves comparable or even better performance than the state-of-the-art supervised methods.

In real-world scenarios, collected and annotated data often exhibit the characteristics of multiple classes and long-tailed distribution. Additionally, label noise is inevitable in large-scale annotations and hinders the applications of learning-based models. Although many deep learning based methods have been proposed for handling long-tailed multi-label recognition or label noise respectively, learning with noisy labels in long-tailed multi-label visual data has not been well-studied because of the complexity of long-tailed distribution entangled with multi-label correlation. To tackle such a critical yet thorny problem, this paper focuses on reducing noise based on some inherent properties of multi-label classification and long-tailed learning under noisy cases. In detail, we propose a Stitch-Up augmentation to synthesize a cleaner sample, which directly reduces multi-label noise by stitching up multiple noisy training samples. Equipped with Stitch-Up, a Heterogeneous Co-Learning framework is further designed to leverage the inconsistency between long-tailed and balanced distributions, yielding cleaner labels for more robust representation learning with noisy long-tailed data. To validate our method, we build two challenging benchmarks, named VOC-MLT-Noise and COCO-MLT-Noise, respectively. Extensive experiments are conducted to demonstrate the effectiveness of our proposed method. Compared to a variety of baselines, our method achieves superior results.

Thanks to the inherent spatial or sequential structures underlying the data like images and texts, deep architectures such as convolutional neural networks (CNNs) and the Transformer have been recognized as the preeminent approaches in image processing and language modeling. In the real world, there are a large number of tabular data without any explicit structures, which breaks the inductive bias of most neural networks like CNNs. Although multi-layer perceptrons (MLPs) obtain empirical success on tabular data, they cannot well explain the underlying relationship between multiple variables. Compared with other fields, research on deep models toward tabular data has received relatively less scrutiny. To bridge this gap, we propose Dual-Route Structure-Adaptive Graph Networks (DRSA-Net) to model the nonlinearity in tabular feature vectors without any prior. DRSA-Net adaptively learns a sparse graph structure between variables and then characterizes interactions between them from the view of dual-route message passing. We demonstrate that DRSA-Net could easily degenerate into the typical MLPs and factorization machines (FMs). Extensive experiments on recommendations, images (no spatial information after preprocessing), and some benchmark machine learning datasets show that DRSA-Net achieves comparable or superior performance with many classic algorithms and recently proposed deep models.

Temporal sentence grounding in videos is a crucial task in vision-language learning. Its goal is retrieving a video segment from an untrimmed video that semantically corresponds to a natural language query. A video usually contains multiple semantic events, which are rarely isolated. They tend to be temporally ordered and semantically correlated (e.g., some event is often the precursor of another event). To precisely localize a semantic moment from a video, it is critical to effectively extract and aggregate multi-granularity contextual information, including the fine-grained local context around the moment-related video segment (in short snippet-level) and coarse-grained semantic correlation (in segment-level). Additionally, a second main insight in this work is that the above context aggregation should be favorably guided by the queries, rather than fully query-agnostic. Putting above ideas together, we here present a new network that does language-guided multi-granularity context aggregation. It is comprised of two major modules. The core of the first module is a novel language-guided temporal adaptive convolution (LTAC) devised to extract fine-grained information over video snippets around the ground-truth video segment. It decomposes a convolution into two channel-oriented / temporal-oriented ones. In particular, the convolutional channels are supposed to be more susceptible to queries, thus we learn to generate a dynamic channel-oriented kernel with respect to the querying sentence. As a second module, we propose a language-guided global relation block (LGRB) that extracts video-level context. It augments the contextual feature by using a multi-scale temporal attention that tackles the scale variation of ground-truth video segments, and a multi-modal semantic attention that relies on syntactic of the query. For the validation purpose, we have conducted comprehensive experiments on two popularly-adopted video benchmarks (i.e., ActivityNet Captions and Charades-STA). All experimental results and ablation studies have clearly corroborated the effectiveness of our model designs, outstripping prior state-of-the-art methods in terms of major performance metrics for the task.

Text-video retrieval is one of the basic tasks for multimodal research and has been widely harnessed in many real-world systems. Most existing approaches directly compare the global representation between videos and text descriptions and utilize the global contrastive loss to train the model. These designs overlook the local alignment and the word-level supervision signal. In this paper, we propose a new framework, called Align and Tell, for text-video retrieval. Compared to the previous work, our framework contains additional modules,i.e., two transformer decoders for local alignment and one captioning head to enhance the representation learning. First, we introduce a set of learnable queries to interact with both textual representations and video representations and project them to a fixed number of local features. After that, local contrastive learning is performed to complement the global comparison. Moreover, we design a video captioning head to provide additional supervision signals during training. This word-level supervision can enhance the visual presentation and alleviate the cross-modal gap. The captioning head can be removed during inference and does not introduce extra computational costs. Extensive empirical results demonstrate that our Align and Tell model can achieve state-of-the-art performance on four text-video retrieval datasets, including MSR-VTT, MSVD, LSMDC, and ActivityNet-Captions.

Filter pruning is effective to reduce the computational costs of neural networks. Existing methods show that updating the previous pruned filter would enable large model capacity and achieve better performance. However, during the iterative pruning process, even if the network weights are updated to new values, the pruning criterion remains the same. In addition, when evaluating the filter importance, only the magnitude information of the filters is considered. However, in neural networks, filters do not work individually, but they would affect other filters. As a result, the magnitude information of each filter, which merely reflects the information of an individual filter itself, is not enough to judge the filter importance. To solve the above problems, we propose meta-attribute-based filter pruning (MFP). First, to expand the existing magnitude information-based pruning criteria, we introduce a new set of criteria to consider the geometric distance of filters. Additionally, to explicitly assess the current state of the network, we adaptively select the most suitable criteria for pruning via a meta-attribute, a property of the neural network at the current state. Experiments on two image classification benchmarks validate our method. For ResNet-50 on ILSVRC-2012, we could reduce more than 50% FLOPs with only 0.44% top-5 accuracy loss.

It is crucial to sample a small portion of relevant frames for efficient video classification. The existing methods mainly develop hand-designed sampling strategies or learn sequential selection policies. However, there are two challenges to be solved. First, hand-designed sampling strategies are intrinsically non-adaptive to different video backbones. Second, sequential frame selection policies ignore temporal relations among all video frames. The sequential selection process also hinders the application of these video samplers in speed-critical systems. In this article, we propose a differentiable parallel video sampling network (PSN) to tackle the aforementioned challenges, First, we optimize the video sampler with a differentiable surrogate loss, allowing to dynamically learn the sampler with the cooperation from the video classification model. Our sampler considers the feedback from all frames jointly, eliminating the learning difficulties of sequential decision making. The learning process is fully gradient-based, making the sampler be learned efficiently. Our video sampler can assess a set of frames swiftly and determine the importance of each frame in parallel. Second, we propose to model the inter-relation among contextual frames, which encourages the sampler to select frames based on a comprehensive inspection of the entire video. We observe that a simple context relation mining instantiation would significantly improve the classification performance. The experimental results on three standard video recognition benchmarks demonstrate the efficacy and efficiency of our framework.

Temporal grounding in videos aims to localize one target video segment that semantically corresponds to a given query sentence. Thanks to the semantic diversity of natural language descriptions, temporal grounding allows activity grounding beyond pre-defined classes and has received increasing attention in recent years. The semantic diversity is rooted in the principle of compositionality in linguistics, where novel semantics can be systematically described by combining known words in novel ways (compositional generalization). However, current temporal grounding datasets do not specifically test for the compositional generalizability. To systematically measure the compositional generalizability of temporal grounding models, we introduce a new Compositional Temporal Grounding task and construct two new dataset splits, i.e., Charades-CG and ActivityNet-CG. Evaluating the state-of-the-art methods on our new dataset splits, we empirically find that they fail to generalize to queries with novel combinations of seen words. To tackle this challenge, we propose a variational cross-graph reasoning framework that explicitly decomposes video and language into multiple structured hierarchies and learns fine-grained semantic correspondence among them. Experiments illustrate the superior compositional generalizability of our approach. The repository of this work is at ht tps: / / gi thub. com/YYJMJC/ Composi tional- Temporal-Grounding.

Profiting from the advance of deep convolutional networks, current state-of-the-art video action recognition models have achieved remarkable progress. Nevertheless, most of existing models suffer from low interpretability of the predicted actions. Inspired by the observation that temporally-configured human-object interactions often serve as a key indicator of many actions, this work crafts an action reasoning framework that performs Markov Logic Network (MLN) based probabilistic logical inference. Crucially, we propose to encode an action by first-order logical rules that correspond to the temporal changes of visual relationships in videos. The main contributions of this work are two-fold: 1) Different from existing black-box models, the proposed model simultaneously implements the localization of temporal boundaries and the recognition of action categories by grounding the logical rules of MLN in videos. The weight associated with each such rule further provides an estimate of confidence. These collectively make our model more explainable and robust. 2) Instead of using hand-crafted logical rules in conventional MLN, we develop a data-driven instantiation of the MLN. In specific, a hybrid learning scheme is proposed. It combines MLN's weight learning and reinforcement learning, using the former's results as a self-critic for guiding the latter's training. Additionally, by treating actions as logical predicates, the proposed framework can also be integrated with deep models for further performance boost. Comprehensive experiments on two complex video action datasets (Charades & CAD-120) clearly demonstrate the effectiveness and explainability of our proposed method.

Talking gesture generation is a practical yet challenging task that aims to synthesize gestures in line with speech. Gestures with meaningful signs can better convey useful information and arouse sympathy in the audience. Current works focus on aligning gestures with the speech rhythms, which are difficult to mine the semantics and model semantic gestures explicitly. This paper proposes a novel semantic Energized Generation (SEEG) method for semantic-aware gesture generation. Our method contains two parts: DEcoupled Mining module (DEM) and Semantic Energizing Module (SEM). DEM decouples the semantic-irrelevant information from inputs and separately mines information for the beat and semantic gestures. SEM conducts semantic learning and produces semantic gestures. Apart from representational similarity, SEM requires the predictions to express the same semantics as the ground truth. Besides, a semantic prompter is designed in SEM to leverage the semantic-aware supervision to predictions. This promotes the networks to learn and generate semantic gestures. Experimental results reported in three metrics on different benchmarks prove that SEEG efficiently mines semantic cues and generates semantic gestures. SEEG outperforms other methods in all semantic-aware evaluations on different datasets. Qualitative evaluations also indicate the superiority of SEEG in semantic expressiveness. Code is available via https://github.com/akira-l/SEEG.

Understanding network generalization and feature discrimination is an open research problem in visual recognition. Many studies have been conducted to assess the quality of feature representations. One of the simple strategies is to utilize a linear probing classifier to quantitatively evaluate the class accuracy under the obtained features. The typical linear probe is only applied as a proxy at the inference time, but its efficacy in measuring features' suitability for linear classification is largely neglected in training. In this paper, we propose an episodic linear probing (ELP) classifier to reflect the generalization of visual rep-resentations in an online manner. ELP is trained with detached features from the network and re-initialized episodically. It demonstrates the discriminability of the visual representations in training. Then, an ELP-suitable Regularization term (ELP-SR) is introduced to reflect the distances of probability distributions between the ELP classifier and the main classifier. ELP-SR leverages are-scaling factor to regularize each sample in training, which modulates the loss function adaptively and encourages the features to be discriminative and generalized. We observe significant improvements in three real-world visual recognition tasks: fine-grained visual classification, long-tailed visual recognition, and generic object recognition. The performance gains show the effectiveness of our method in im-proving network generalization and feature discrimination.

As an important area in computer vision, object tracking has formed two separate communities that respectively study Single Object Tracking (SOT) and Multiple Object Tracking (MOT). However, current methods in one tracking scenario are not easily adapted to the other due to the divergent training datasets and tracking objects of both tasks. Although UniTrack [45] demonstrates that a shared appearance model with multiple heads can be used to tackle individual tracking tasks, it fails to exploit the large-scale tracking datasets for training and performs poorly on the single object tracking. In this work, we present the Unified Transformer Tracker (UTT) to address tracking problems in different scenarios with one paradigm. A track transformer is developed in our UTT to track the target in both SOT and MOT where the correlation between the target feature and the tracking frame feature is exploited to localize the target. We demonstrate that both SOT and MOT tasks can be solved within this framework, and the model can be simultaneously end-to-end trained by alternatively optimizing the SOT and MOT objectives on the datasets of individual tasks. Extensive experiments are conducted on several benchmarks with a unified model trained on both SOT and MOT datasets.

Understanding human emotions is a crucial ability for intelligent robots to provide better human-robot interactions. The existing works are limited to trimmed video-level emotion classification, failing to locate the temporal window corresponding to the emotion. In this paper, we introduce a new task, named Temporal Emotion Localization in videos (TEL), which aims to detect human emotions and localize their corresponding temporal boundaries in untrimmed videos with aligned subtitles. TEL presents three unique challenges compared to temporal action localization: 1) The emotions have extremely varied temporal dynamics; 2) The emotion cues are embedded in both appearances and complex plots; 3) The fine-grained temporal annotations are complicated and labor-intensive. To address the first two challenges, we propose a novel dilated context integrated network with a coarse-fine two-stream architecture. The coarse stream captures varied temporal dynamics by modeling multi-granularity temporal contexts. The fine stream achieves complex plots understanding by reasoning the dependency between the multi-granularity temporal contexts from the coarse stream and adaptively integrates them into fine-grained video segment features. To address the third challenge, we introduce a cross-modal consensus learning paradigm, which leverages the inherent semantic consensus between the aligned video and subtitle to achieve weakly-supervised learning. We contribute a new testing set with 3,000 manually-annotated temporal boundaries so that future research on the TEL problem can be quantitatively evaluated. Extensive experiments show the effectiveness of our approach on temporal emotion localization. The repository of this work is at https://github.com/YYJMJC/TemporalEmotion-Localization-in-Videos.

Large-scale vision-language pre-training has shown impressive advances in a wide range of downstream tasks. Existing methods mainly model the cross-modal alignment by the similarity of the global representations of images and text, or advanced cross-modal attention upon image and text features. However, they fail to explicitly learn the fine-grained semantic alignment between visual regions and textual phrases, as only global image-text alignment information is available. In this paper, we introduce LOUPE, a fine-grained semantically aLigned visiOn-langUage PrE-training framework, which learns fine-grained semantic alignment from the novel perspective of game-theoretic interactions. To efficiently estimate the game-theoretic interactions, we further propose an uncertainty-aware neural Shapley interaction learning module. Experiments show that LOUPE achieves state-of-the-art performance on a variety of vision-language tasks. Without any object-level human annotations and fine-tuning, LOUPE achieves competitive performance on object detection and visual grounding. More importantly, LOUPE opens a new promising direction of learning fine-grained semantics from large-scale raw image-text pairs.

Recently, large-scale pre-training methods like CLIP have made great progress in multi-modal research such as text-video retrieval. In CLIP, transformers are vital for modeling complex multi-modal relations. However, in the vision transformer of CLIP, the essential visual tokenization process, which produces discrete visual token sequences, generates many homogeneous tokens due to the redundancy nature of consecutive and similar frames in videos. This significantly increases computation costs and hinders the deployment of video retrieval models in web applications. In this paper, to reduce the number of redundant video tokens, we design a multi-segment token clustering algorithm to find the most representative tokens and drop the non-essential ones. As the frame redundancy occurs mostly in consecutive frames, we divide videos into multiple segments and conduct segment-level clustering. Center tokens from each segment are later concatenated into a new sequence, while their original spatial-temporal relations are well maintained. We instantiate two clustering algorithms to efficiently find deterministic medoids and iteratively partition groups in high dimensional space. Through this token clustering and center selection procedure, we successfully reduce computation costs by removing redundant visual tokens. This method further enhances segment-level semantic alignment between video and text representations, enforcing the spatio-temporal interactions of tokens from within-segment frames. Our method, coined as CenterCLIP, surpasses existing state-of-the-art by a large margin on typical text-video benchmarks, while reducing the training memory cost by 35% and accelerating the inference speed by 14% at the best case. The code is available at https://github.com/mzhaoshuai/CenterCLIP https://github.com/mzhaoshuai/CenterCLIP.

Most state-of-the-art methods of object detection suffer from poor generalization ability when the training and test data are from different domains. To address this problem, previous methods mainly explore to align distribution between source and target domains, which may neglect the impact of the domain-specific information existing in the aligned features. Besides, when transferring detection ability across different domains, it is important to extract the instance-level features that are domain-invariant. To this end, we explore to extract instance-invariant features by disentangling the domain-invariant features from the domain-specific features. Particularly, a progressive disentangled mechanism is proposed to decompose domain-invariant and domain-specific features, which consists of a base disentangled layer and a progressive disentangled layer. Then, with the help of Region Proposal Network (RPN), the instance-invariant features are extracted based on the output of the progressive disentangled layer. Finally, to enhance the disentangled ability, we design a detached optimization to train our model in an end-to-end fashion. Experimental results on four domain-shift scenes show our method is separately 2.3, 3.6, 4.0, and 2.0 percent higher than the baseline method. Meanwhile, visualization analysis demonstrates that our model owns well disentangled ability.

In this paper, we propose to leverage freely available unlabeled video data to facilitate few-shot video classification. In this semi-supervised few-shot video classification task, millions of unlabeled data are available for each episode during training. These videos can be extremely imbalanced, while they have profound visual and motion dynamics. To tackle the semi-supervised few-shot video classification problem, we make the following contributions. First, we propose a label independent memory (LIM) to cache label related features, which enables a similarity search over a large set of videos. LIM produces a class prototype for few-shot training. This prototype is an aggregated embedding for each class, which is more robust to noisy video features. Second, we integrate a multi-modality compound memory network to capture both RGB and flow information. We propose to store the RGB and flow representation in two separate memory networks, but they are jointly optimized via a unified loss. In this way, mutual communications between the two modalities are leveraged to achieve better classification performance. Third, we conduct extensive experiments on the few-shot Kinetics-100, Something-Something-100 datasets, which validates the effectiveness of leveraging the accessible unlabeled data for few-shot classification.

GloVe learns word embeddings by leveraging statistical information from word co-occurrence matrices. However, word pairs in the matrices are extracted from a predefined local context window, which might lead to limited word pairs and potentially semantic irrelevant word pairs. In this paper, we proposeSemGloVe, which distillssemantic co-occurrencesfrom BERT into static GloVe word embeddings. Particularly, we propose two models to extract co-occurrence statistics based on either the masked language model or the multi-head attention weights of BERT. Our methods can extract word pairs limited by the local window assumption, and can define the co-occurrence weights by directly considering the semantic distance between word pairs. Experiments on several word similarity datasets and external tasks show that SemGloVe can outperform GloVe.

RGB-D salient object detection (SOD) has attracted increasingly more attention as it shows more robust results in complex scenes compared with RGB SOD. However, state-of-the-art RGB-D SOD approaches heavily rely on a large amount of pixel-wise annotated data for training. Such densely labeled annotations are often labor-intensive and costly. To reduce the annotation burden, we investigate RGB-D SOD from a weakly supervised perspective. More specifically, we use annotator-friendly scribble annotations as supervision signals for model training. Since scribble annotations are much sparser compared to ground-truth masks, some critical object structure information might be neglected. To preserve such structure information, we explicitly exploit the complementary edge information from two modalities (i.e., RGB and depth). Specifically, we leverage the dual-modal edge guidance and introduce a new network architecture with a dual-edge detection module and a modality-aware feature fusion module. In order to use the useful information of unlabeled pixels, we introduce a prediction consistency training scheme by comparing the predictions of two networks optimized by different strategies. Moreover, we develop an active scribble boosting strategy to provide extra supervision signals with negligible annotation cost, leading to significant SOD performance improvement. Extensive experiments on seven benchmarks validate the superiority of our proposed method. Remarkably, the proposed method with scribble annotations achieves competitive performance in comparison to fully supervised state-of-the-art methods.

Neighboring frames are more correlated compared to frames from further temporal distances. In this paper, we aim to explore the temporal correlations among neighboring frames and exploit cross-layer multi-scale features for action recognition. First, we present a Temporal Cross-Layer Correlation (TCLC) framework for temporal correlation learning. The unified framework uncovers both local and global structures from video data, enabling a better exploration of temporal context and assisting cross-layer spatio-temporal feature learning. Second, we propose a novel cross-layer attention and a center-guided attention mechanism to integrate features with contextual knowledge from multiple scales. Our method is a two-stage process for effective cross-layer feature learning. The first stage incorporates the cross-layer attention module to decide the importance weight of the convolutional layers. The second stage leverages the center-guided attention mechanism to aggregate local features from each layer for the generation of a final video representation. We leverage global centers to extract shared semantic knowledge among videos. We evaluate TCLC on three action recognition datasets, i.e., UCF-101, HMDB-51 and Kinetics. Our experimental results demonstrate the superiority of our proposed temporal correlation mining method.

Text-video retrieval is a challenging task that aims to search relevant video contents based on natural language descriptions. The key to this problem is to measure text-video similarities in a joint embedding space. However, most existing methods only consider the global cross-modal similarity and overlook the local details. Some works incorporate the local comparisons through cross-modal local matching and reasoning. These complex operations introduce tremendous computation. In this paper, we design an efficient global-local alignment method. The multi-modal video sequences and text features are adaptively aggregated with a set of shared semantic centers. The local cross-modal similarities are computed between the video feature and text feature within the same center. This design enables the meticulous local comparison and reduces the computational cost of the interaction between each text-video pair. Moreover, a global alignment method is proposed to provide a global cross-modal measurement that is complementary to the local perspective. The global aggregated visual features also provide additional supervision, which is indispensable to the optimization of the learnable semantic centers. We achieve consistent improvements on three standard text-video retrieval benchmarks and outperform the state-of-the-art by a clear margin.

It has been shown that deep neural networks are prone to overfitting on biased training data. Towards addressing this issue, meta-learning employs a meta model for correcting the training bias. Despite the promising performances, super slow training is currently the bottleneck in the meta learning approaches. In this paper, we introduce a novel Faster Meta Update Strategy (FaMUS) to replace the most expensive step in the meta gradient computation with a faster layer-wise approximation. We empirically find that FaMUS yields not only a reasonably accurate but also a low-variance approximation of the meta gradient. We conduct extensive experiments to verify the proposed method on two tasks. We show our method is able to save two-thirds of the training time while still maintaining the comparable or achieving even better generalization performance. In particular, our method achieves the state-of-the-art performance on both synthetic and realistic noisy labels, and obtains promising performance on long-tailed recognition on standard benchmarks. Code are released at https://github.com/youjiangxu/FaMUS.

In this paper, we tackle the problem of discovering new classes in unlabeled visual data given labeled data from disjoint classes. Existing methods typically first pre-train a model with labeled data, and then identify new classes in unlabeled data via unsupervised clustering. However, the labeled data that provide essential knowledge are often underexplored in the second step. The challenge is that the labeled and unlabeled examples are from non-overlapping classes, which makes it difficult to build a learning relationship between them. In this work, we introduce Open-Mix to mix the unlabeled examples from an open set and the labeled examples from known classes, where their non-overlapping labels and pseudo-labels are simultaneously mixed into a joint label distribution. OpenMix dynamically compounds examples in two ways. First, we produce mixed training images by incorporating labeled examples with unlabeled examples. With the benefit of unique prior knowledge in novel class discovery, the generated pseudo-labels will be more credible than the original unlabeled predictions. As a result, OpenMix helps preventing the model from overfitting on unlabeled samples that may be assigned with wrong pseudo-labels. Second, the first way encourages the unlabeled examples with high class-probabilities to have considerable accuracy. We introduce these examples as reliable anchors and further integrate them with un-labeled samples. This enables us to generate more combinations in unlabeled examples and exploit finer object relations among the new classes. Experiments on three classification datasets demonstrate the effectiveness of the proposed OpenMix, which is superior to state-of-the-art methods in novel class discovery.

Video-and-Language Inference is a recently proposed task for joint video-and-language understanding. This new task requires a model to draw inference on whether a natural language statement entails or contradicts a given video clip. In this paper, we study how to address three critical challenges for this task: judging the global correctness of the statement involved multiple semantic meanings, joint reasoning over video and subtitles, and modeling long-range relationships and complex social interactions. First, we propose an adaptive hierarchical graph network that achieves in-depth understanding of the video over complex interactions. Specifically, it performs joint reasoning over video and subtitles in three hierarchies, where the graph structure is adaptively adjusted according to the semantic structures of the statement. Secondly, we introduce semantic coherence learning to explicitly encourage the semantic coherence of the adaptive hierarchical graph network from three hierarchies. The semantic coherence learning can further improve the alignment between vision and linguistics, and the coherence across a sequence of video segments. Experimental results show that our method significantly outperforms the baseline by a large margin.

Most existing few-shot classification methods only consider generalization on one dataset (i.e., single-domain), failing to transfer across various seen and unseen domains. In this paper, we consider the more realistic multi-domain few-shot classification problem to investigate the cross-domain generalization. Two challenges exist in this new setting: (1) how to efficiently generate multi-domain feature representation, and (2) how to explore domain correlations for better cross-domain generalization. We propose a parameter-efficient multi-mode modulator to address both challenges. First, the modulator is designed to maintain multiple modulation parameters (one for each domain) in a single network, thus achieving single-network multi-domain representation. Given a particular domain, domain-aware features can be efficiently generated with the well-devised separative selection module and cooperative query module. Second, we further divide the modulation parameters into the domain-specific set and the domain-cooperative set to explore the intra-domain information and inter-domain correlations, respectively. The intra-domain information describes each domain independently to prevent negative interference. The inter-domain correlations guide information sharing among relevant domains to enrich their own representation. Moreover, unseen domains can utilize the correlations to obtain an adaptive combination of seen domains for extrapolation. We demonstrate that the proposed multi-mode modulator achieves state-of-the-art results on the challenging META-DATASET benchmark, especially for unseen test domains.

Egocentric video recognition is a challenging task that requires to identify both the actor’s motion and the active object that the actor interacts with. Recognizing the active object is particularly hard due to the cluttered background with distracting objects, the frequent field of view changes, severe occlusion, etc. To improve the active object classification, most existing methods use object detectors or human gaze information, which are computationally expensive or require labor-intensive annotations. To avoid these additional costs, we propose an end-to-end Interactive Prototype Learning (IPL) framework to learn better active object representations by leveraging the motion cues from the actor. First, we introduce a set of verb prototypes to disentangle active object features from distracting object features. Each prototype corresponds to a primary motion pattern of an egocentric action, offering a distinctive supervision signal for active object feature learning. Second, we design two interactive operations to enable the extraction of active object features, i.e., noun-to-verb assignment and verb-to-noun selection. These operations are parameter-efficient and can learn judicious location-aware features on top of 3D CNN backbones. We demonstrate that the IPL framework can generalize to different backbones and outperform the state-of-the-art on three large-scale egocentric video datasets, i.e., EPIC-KITCHENS-55, EPIC-KITCHENS-100 and EGTEA.

Few-shot object detection (FSOD) aims to strengthen the performance of novel object detection with few labeled samples. To alleviate the constraint of few samples, enhancing the generalization ability of learned features for novel objects plays a key role. Thus, the feature learning process of FSOD should focus more on intrinsical object characteristics, which are invariant under different visual changes and therefore are helpful for feature generalization. Unlike previous attempts of the meta-learning paradigm, in this paper, we explore how to enhance object features with intrinsical characteristics that are universal across different object categories. We propose a new prototype, namely universal prototype, that is learned from all object categories. Besides the advantage of characterizing invariant characteristics, the universal prototypes alleviate the impact of unbalanced object categories. After enhancing object features with the universal prototypes, we impose a consistency loss to maximize the agreement between the enhanced features and the original ones, which is beneficial for learning invariant object characteristics. Thus, we develop a new framework of few-shot object detection with universal prototypes (F SODup) that owns the merit of feature generalization towards novel objects. Experimental results on PASCAL VOC and MS COCO show the effectiveness of F SODup. Particularly, for the 1-shot case of VOC Split2, FSODupoutperforms the baseline by 6.8% in terms of mAP.

To improve the generalization of detectors, for domain adaptive object detection (DAOD), recent advances mainly explore aligning feature-level distributions between the source and single-target domain, which may neglect the impact of domain-specific information existing in the aligned features. Towards DAOD, it is important to extract domain-invariant object representations. To this end, in this paper, we try to disentangle domain-invariant representations from domain-specific representations. And we propose a novel disentangled method based on vector decomposition. Firstly, an extractor is devised to separate domain-invariant representations from the input, which are used for extracting object proposals. Secondly, domain-specific representations are introduced as the differences between the input and domain-invariant representations. Through the difference operation, the gap between the domain-specific and domain-invariant representations is enlarged, which promotes domain-invariant representations to contain more domain-irrelevant information. In the experiment, we separately evaluate our method on the single- and compound-target case. For the single-target case, experimental results of four domain-shift scenes show our method obtains a significant performance gain over baseline methods. Moreover, for the compound-target case (i.e., the target is a compound of two different domains without domain labels), our method outperforms baseline methods by around 4%, which demonstrates the effectiveness of our method.

To boost research into cognition-level visual understanding, i.e., making an accurate inference based on a thorough understanding of visual details, visual commonsense reasoning (VCR) has been proposed. Compared with traditional visual question answering which requires models to select correct answers, VCR requires models to select not only the correct answers, but also the correct rationales. Recent research into human cognition has indicated that brain function or cognition can be considered as a global and dynamic integration of local neuron connectivity, which is helpful in solving specific cognition tasks. Inspired by this idea, we propose a directional connective network to achieve VCR by dynamically reorganizing the visual neuron connectivity that is contextualized using the meaning of questions and answers and leveraging the directional information to enhance the reasoning ability. Specifically, we first develop a GraphVLAD module to capture visual neuron connectivity to fully model visual content correlations. Then, a contextualization process is proposed to fuse sentence representations with visual neuron representations. Finally, based on the output of contextualized connectivity, we propose directional connectivity to infer answers and rationales, which includes a ReasonVLAD module. Experimental results on the VCR dataset and visualization analysis demonstrate the effectiveness of our method.

Accurate predictions of future pedestrian trajectory could prevent a considerable number of traffic injuries and improve pedestrian safety. It involves multiple sources of information and real-time interactions, e.g., vehicle speed and ego-motion, pedestrian intention and historical locations. Existing methods directly apply a simple concatenation operation to combine multiple cues while their dynamics over time are less studied. In this paper, we propose a novel Long Short-Term Memory (LSTM), namely, to incorporate multiple sources of information from pedestrians and vehicles adaptively. Different from LSTM, our considers mutual interactions and explores intrinsic relations among multiple cues. First, we introduce extra memory cells to improve the transferability of LSTMs in modeling future variations. These extra memory cells include a speed cell to explicitly model vehicle speed dynamics, an intention cell to dynamically analyze pedestrian crossing intentions and a correlation cell to exploit correlations among temporal frames. These three individual cells uncover the future movement of vehicles, pedestrians and global scenes. Second, we propose a gated shifting operation to learn the movement of pedestrians. The intention of crossing the road or not would significantly affect pedestrian's spatial locations. To this end, global scene dynamics and pedestrian intention information are leveraged to model the spatial shifts. Third, we integrate the speed variations to the output gate and dynamically reweight the output channels via the scaling of vehicle speed. The movement of the vehicle would alter the scale of the predicted pedestrian bounding box: as the vehicle gets closer to the pedestrian, the bounding box is enlarging. Our rescaling process captures the relative movement and updates the size of pedestrian bounding boxes accordingly. Experiments conducted on three pedestrian trajectory forecasting benchmarks show that our achieves state-of-the-art performance.

Anticipating actions before they are executed is crucial for a wide range of practical applications, including autonomous driving and robotics. In this paper, we study the egocentric action anticipation task, which predicts future action seconds before it is performed for egocentric videos. Previous approaches focus on summarizing the observed content and directly predicting future action based on past observations. We believe it would benefit the action anticipation if we could mine some cues to compensate for the missing information of the unobserved frames. We then propose to decompose the action anticipation into a series of future feature predictions. We imagine how the visual feature changes in the near future and then predicts future action labels based on these imagined representations. Differently, our ImagineRNN is optimized in a contrastive learning way instead of feature regression. We utilize a proxy task to train the ImagineRNN, i.e., selecting the correct future states from distractors. We further improve ImagineRNN by residual anticipation, i.e., changing its target to predicting the feature difference of adjacent frames instead of the frame content. This promotes the network to focus on our target, i.e., the future action, as the difference between adjacent frame features is more important for forecasting the future. Extensive experiments on two large-scale egocentric action datasets validate the effectiveness of our method. Our method significantly outperforms previous methods on both the seen test set and the unseen test set of the EPIC Kitchens Action Anticipation Challenge.

Object detection has gained great improvements with the advances of convolutional neural networks and the availability of large amounts of accurate training data. Though the amount of data is increasing significantly, the quality of data annotations is not guaranteed from the existing crowd-sourcing labeling platforms. In addition to noisy category labels, imprecise bounding box annotations are commonly existed for object detection data. When the quality of training data degenerates, the performance of the typical object detectors is severely impaired. In this paper, we propose a Meta-Refine-Net (MRNet) to train object detectors from noisy category labels and imprecise bounding boxes. First, MRNet learns to adaptively assign lower weights to proposals with incorrect labels so as to suppress large loss values generated by these proposals on the classification branch. Second, MRNet learns to dynamically generate more accurate bounding box annotations to overcome the misleading of imprecisely annotated bounding boxes. Thus, the imprecise bounding boxes could impose positive impacts on the regression branch rather than simply be ignored. Third, we propose to refine the imprecise bounding box annotations by jointly learning from both the category and the localization information. By doing this, the approximation of ground-truth bounding boxes is more accurate while the misleading would be further alleviated. Our MRNet is model-agnostic and is capable of learning from noisy object detection data with only a few clean examples (less than 2%). Extensive experiments on PASCAL VOC 2012 and MS COCO 2017 demonstrate the effectiveness and efficiency of our method.

In the few-shot common-localization task, given few support images without bounding box annotations at each episode, the goal is to localize the common object in the query image of unseen categories. The few-shot common-localization task involves common object reasoning from the given images, predicting the spatial locations of the object with different shapes, sizes, and orientations. In this work, we propose a common-centric localization (CCL) network for few-shot common-localization. The motivation of our common-centric localization network is to learn the common object features by dynamic feature relation reasoning via a graph convolutional network with conditional feature aggregation. First, we propose a local common object region generation pipeline to reduce background noises due to feature misalignment. Each support image predicts more accurate object spatial locations by replacing the query with the images in the support set. Second, we introduce a graph convolutional network with dynamic feature transformation to enforce the common object reasoning. To enhance the discriminability during feature matching and enable a better generalization in unseen scenarios, we leverage a conditional feature encoding function to alter visual features according to the input query adaptively. Third, we introduce a common-centric relation structure to model the correlation between the common features and the query image feature. The generated common features guide the query image feature towards a more common object-related representation. We evaluate our common-centric localization network on four datasets, i.e., CL-VOC-07, CL-VOC-12, CL-COCO, CL-VID. We obtain significant improvements compared to state-of-the-art. Our quantitative results confirm the effectiveness of our network.

Egocentric video recognition is a natural testbed for diverse interaction reasoning. Due to the large action vocabulary in egocentric video datasets, recent studies usually utilize a two-branch structure for action recognition, i.e., one branch for verb classification and the other branch for noun classification. However, correlation study between the verb and the noun branches have been largely ignored. Besides, the two branches fail to exploit local features due to the absence of position-aware attention mechanism. In this paper, we propose a novel Symbiotic Attention framework leveraging Privileged information (SAP) for egocentric video recognition. Finer position-aware object detection features can facilitate the understanding of actor's interaction with the object. We introduce these features in action recognition and regard them as privileged information. Our framework enables mutual communication among the verb branch, the noun branch, and the privileged information. This communication process not only injects local details into global features, but also exploits implicit guidance about the spatio-temporal position of an on-going action. We introduce a novel symbiotic attention (SA) to enable effective communication. It first normalizes the detection guided features on one branch to underline the action-relevant information from the other branch. SA adaptively enhances the interactions among the three sources. To further catalyze this communication, spatial relations are uncovered for the selection of most action-relevant information. It identifies the most valuable and discriminative feature for classification. We validate the effectiveness of our SAP quantitatively and qualitatively. Notably, it achieves the state-of-the-art on two large-scale egocentric video datasets.

Typical video classification methods often divide a video into short clips, do inference on each clip independently, then aggregate the clip-level predictions to generate the video-level results. However, processing visually similar clips independently ignores the temporal structure of the video sequence, and increases the computational cost at inference time. In this paper, we propose a novel framework named FASTER, i.e., Feature Aggregation for Spatio-TEmporal Redundancy. FASTER aims to leverage the redundancy between neighboring clips and reduce the computational cost by learning to aggregate the predictions from models of different complexities. The FASTER framework can integrate high quality representations from expensive models to capture subtle motion information and lightweight representations from cheap models to cover scene changes in the video. A new recurrent network (i.e., FAST-GRU) is designed to aggregate the mixture of different representations. Compared with existing approaches, FASTER can reduce the FLOPs by over 10× while maintaining the state-of-the-art accuracy across popular datasets, such as Kinetics, UCF-101 and HMDB-51.

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.

Establishing dense correspondences across semantically similar images is a challenging task. Due to the large intra-class variation and background clutter, two common issues occur in current approaches. First, many pixels in a source image are assigned to one target pixel, i.e., many to one matching. Second, some object pixels are assigned to the background pixels, i.e., background matching. We solve the first issue by global feature matching, which maximizes the total matching correlations between images to obtain a global optimal matching matrix. The row sum and column sum constraints are enforced on the matching matrix to induce a balanced solution, thus suppressing the many to one matching. We solve the second issue by applying a staircase function on the class activation maps to re-weight the importance of pixels into four levels from foreground to background. The whole procedure is combined into a unified optimal transport algorithm by converting the maximization problem to the optimal transport formulation and incorporating the staircase weights into optimal transport algorithm to act as empirical distributions. The proposed algorithm achieves state-of-the-art performance on four benchmark datasets. Notably, a 26\% relative improvement is achieved on the large-scale SPair-71k dataset.

In this work, we propose a generally applicable transformation unit for visual recognition with deep convolutional neural networks. This transformation explicitly models channel relationships with explainable control variables. These variables determine the neuron behaviors of competition or cooperation, and they are jointly optimized with the convolutional weight towards more accurate recognition. In Squeeze-and-Excitation (SE) Networks, the channel relationships are implicitly learned by fully connected layers, and the SE block is integrated at the block-level. We instead introduce a channel normalization layer to reduce the number of parameters and computational complexity. This lightweight layer incorporates a simple l2 normalization, enabling our transformation unit applicable to operator-level without much increase of additional parameters. Extensive experiments demonstrate the effectiveness of our unit with clear margins on many vision tasks, i.e., image classification on ImageNet, object detection and instance segmentation on COCO, video classification on Kinetics.

There have been increasing interests in modeling long-tailed data. Unlike artificially collected datasets, long-tailed data are naturally existed in the real-world and thus more realistic. To deal with the class imbalance problem, we introduce an Inflated Episodic Memory (IEM) for long-tailed visual recognition. First, our IEM augments the convolutional neural networks with categorical representative features for rapid learning on tail classes. In traditional few-shot learning, a single prototype is usually leveraged to represent a category. However, long-tailed data has higher intra-class variances. It could be challenging to learn a single prototype for one category. Thus, we introduce IEM to store the most discriminative feature for each category individually. Besides, the memory banks are updated independently, which further decreases the chance of learning skewed classifiers. Second, we introduce a novel region self-attention mechanism for multi-scale spatial feature map encoding. It is beneficial to incorporate more discriminative features to improve generalization on tail classes. We propose to encode local feature maps at multiple scales, and the spatial contextual information should be aggregated at the same time. Equipped with IEM and region self-attention, we achieve state-of-the-art performance on four standard long-tailed image recognition benchmarks. Besides, we validate the effectiveness of IEM on a long-tailed video recognition benchmark, i.e., YouTube-8M.

In this paper, we introduce ActBERT for self-supervised learning of joint video-text representations from unlabeled data. First, we leverage global action information to catalyze the mutual interactions between linguistic texts and local regional objects. It uncovers global and local visual clues from paired video sequences and text descriptions for detailed visual and text relation modeling. Second, we introduce an ENtangled Transformer block (ENT) to encode three sources of information, i.e., global actions, local regional objects, and linguistic descriptions. Global-local correspondences are discovered via judicious clues extraction from contextual information. It enforces the joint video-text representation to be aware of fine-grained objects as well as global human intention. We validate the generalization capability of ActBERT on downstream video-and language tasks, i.e., text-video clip retrieval, video captioning, video question answering, action segmentation, and action step localization. ActBERT significantly outperform the state-of-the-arts, demonstrating its superiority in video-text representation learning.

In Visual Dialog, an agent has to parse temporal context in the dialog history and spatial context in the image to hold a meaningful dialog with humans. For example, to answer “what is the man on her left wearing?” the agent needs to (1) analyze the temporal context in the dialog history to infer who is being referred to as “her,” (2) parse the image to attend “her,” and (3) uncover the spatial context to shift the attention to “her left” and check the apparel of the man. In this article, we use a dialog network to memorize the temporal context and an attention processor to parse the spatial context. Since the question and the image are usually very complex, which makes it difficult for the question to be grounded with a single glimpse, the attention processor attends to the image multiple times to better collect visual information. In the Visual Dialog task, the generative decoder (G) is trained under the word-by-word paradigm, which suffers from the lack of sentence-level training. We propose to reinforce G at the sentence level using the discriminative model (D), which aims to select the right answer from a few candidates, to ameliorate the problem. Experimental results on the VisDial dataset demonstrate the effectiveness of our approach.

Predicting future frames in videos has become a promising direction of research for both computer vision and robot learning communities. The core of this problem involves moving object capture and future motion prediction. While object capture specifies which objects are moving in videos, motion prediction describes their future dynamics. Motivated by this analysis, we propose a Cubic Long Short-Term Memory (CubicLSTM) unit for video prediction. CubicLSTM consists of three branches, i.e., a spatial branch for capturing moving objects, a temporal branch for processing motions, and an output branch for combining the first two branches to generate predicted frames. Stacking multiple CubicLSTM units along the spatial branch and output branch, and then evolving along the temporal branch can form a cubic recurrent neural network (CubicRNN). Experiment shows that CubicRNN produces more accurate video predictions than prior methods on both synthetic and real-world datasets.

There has been an increasing interest in 3D indoor navigation, where a robot in an environment moves to a target according to an instruction. To deploy a robot for navigation in the physical world, lots of training data is required to learn an effective policy. It is quite labour intensive to obtain sufficient real environment data for training robots while synthetic data is much easier to construct by render-ing. Though it is promising to utilize the synthetic environments to facilitate navigation training in the real world, real environment are heterogeneous from synthetic environment in two aspects. First, the visual representation of the two environments have significant variances. Second, the houseplans of these two environments are quite different. There-fore two types of information,i.e. visual representation and policy behavior, need to be adapted in the reinforce mentmodel. The learning procedure of visual representation and that of policy behavior are presumably reciprocal. We pro-pose to jointly adapt visual representation and policy behavior to leverage the mutual impacts of environment and policy. Specifically, our method employs an adversarial feature adaptation model for visual representation transfer anda policy mimic strategy for policy behavior imitation. Experiment shows that our method outperforms the baseline by 19.47% without any additional human annotations.

In image captioning, the typical attention mechanisms are arduous to identify the equivalent visual signals especially when predicting highly abstract words. This phenomenon is known as the semantic gap between vision and language. This problem can be overcome by providing semantic attributes that are homologous to language. Thanks to the inherent recurrent nature and gated operating mechanism, Recurrent Neural Network (RNN) and its variants are the dominating architectures in image captioning. However, when designing elaborate attention mechanisms to integrate visual inputs and semantic attributes, RNN-like variants become unflexible due to their complexities. In this paper, we investigate a Transformer-based sequence modeling framework, built only with attention layers and feedforward layers. To bridge the semantic gap, we introduce EnTangled Attention (ETA) that enables the Transformer to exploit semantic and visual information simultaneously. Furthermore, Gated Bilateral Controller (GBC) is proposed to guide the interactions between the multimodal information. We name our model as ETA-Transformer. Remarkably, ETA-Transformer achieves state-of-the-art performance on the MSCOCO image captioning dataset. The ablation studies validate the improvements of our proposed modules.

Prevailing deep convolutional neural networks (CNNs) for person re-IDentification (reID) are usually built upon ResNet or VGG backbones, which were originally designed for classification. Because reID is different from classification, the architecture should be modified accordingly. We propose to automatically search for a CNN architecture that is specifically suitable for the reID task. There are three aspects to be tackled. First, body structural information plays an important role in reID but it is not encoded in backbones. Second, Neural Architecture Search (NAS) automates the process of architecture design without human effort, but no existing NAS methods incorporate the structure information of input images. Third, reID is essentially a retrieval task but current NAS algorithms are merely designed for classification. To solve these problems, we propose a retrieval-based search algorithm over a specifically designed reID search space, named Auto-ReID. Our Auto-ReID enables the automated approach to find an efficient and effective CNN architecture for reID. Extensive experiments demonstrate that the searched architecture achieves state-of-the-art performance while reducing 50% parameters and 53% FLOPs compared to others.

In this paper, we investigate the audio-visual event localization problem. This task is to localize a visible and audible event in a video. Previous methods first divide a video into short segments, and then fuse visual and acoustic features at the segment level. The duration of these segments is usually short, making the visual and acoustic feature of each segment possibly not well aligned. Direct concatenation of the two features at the segment level can be vulnerable to a minor temporal misalignment of the two signals. We propose a Dual Attention Matching (DAM) module to cover a longer video duration for better high-level event information modeling, while the local temporal information is attained by the global cross-check mechanism. Our premise is that one should watch the whole video to understand the high-level event, while shorter segments should be checked in detail for localization. Specifically, the global feature of one modality queries the local feature in the other modality in a bi-directional way. With temporal co-occurrence encoded between auditory and visual signals, DAM can be readily applied in various audio-visual event localization tasks, e.g., cross-modality localization, supervised event localization. Experiments on the AVE dataset show our method outperforms the state-of-the-art by a large margin.

Visual commonsense reasoning (VCR) has been introduced to boost research of cognition-level visual understanding, i.e., a thorough understanding of correlated details of the scene plus an inference with related commonsense knowledge. Recent studies on neuroscience have suggested that brain function or cognition can be described as a global and dynamic integration of local neuronal connectivity, which is context-sensitive to specific cognition tasks. Inspired by this idea, towards VCR, we propose a connective cognition network (CCN) to dynamically reorganize the visual neuron connectivity that is contextualized by the meaning of questions and answers. Concretely, we first develop visual neuron connectivity to fully model correlations of visual content. Then, a contextualization process is introduced to fuse the sentence representation with that of visual neurons. Finally, based on the output of contextualized connectivity, we propose directional connectivity to infer answers or rationales. Experimental results on the VCR dataset demonstrate the effectiveness of our method. Particularly, in $Q \to AR$ mode, our method is around 4\% higher than the state-of-the-art method.

We aim to significantly reduce the computational cost for classification of temporally untrimmed videos while retaining similar accuracy. Existing video classification methods sample frames with a predefined frequency over entire video. Differently, we propose an end-to-end deep reinforcement approach which enables an agent to classify videos by watching a very small portion of frames like what we do. We make two main contributions. First, information is not equally distributed in video frames along time. An agent needs to watch more carefully when a clip is informative and skip the frames if they are redundant or irrelevant. The proposed approach enables the agent to adapt sampling rate to video content and skip most of the frames without the loss of information. Second, in order to have a confident decision, the number of frames that should be watched by an agent varies greatly from one video to another. We incorporate an adaptive stop network to measure confidence score and generate timely trigger to stop the agent watching videos, which improves efficiency without loss of accuracy. Our approach reduces the computational cost significantly for the large-scale YouTube-8M dataset, while the accuracy remains the same.

Given only a few image-text pairs, humans can learn to detect semantic concepts and describe the content. For machine learning algorithms, they usually require a lot of data to train a deep neural network to solve the problem. However, it is challenging for the existing systems to generalize well to the few-shot multi-modal scenario, because the learner should understand not only images and texts but also their relationships from only a few examples. In this paper, we tackle two multi-modal problems, i.e., image captioning and visual question answering (VQA), in the few-shot setting.

Image captioning is a challenging task where the machine automatically describes an image by sentences or phrases. It often requires a large number of paired image-sentence annotations for training. However, a pre-trained captioning model can hardly be applied to a new domain in which some novel object categories exist, i.e., the objects and their description words are unseen during model training. To correctly caption the novel object, it requires professional human workers to annotate the images by sentences with the novel words. It is labor expensive and thus limits its usage in real-world applications. In this paper, we introduce the zero-shot novel object captioning task where the machine generates descriptions without extra training sentences about the novel object. To tackle the challenging problem, we propose a Decoupled Novel Object Captioner (DNOC) framework that can fully decouple the language sequence model from the object descriptions. DNOC has two components. 1) A Sequence Model with the Placeholder (SM-P) generates a sentence containing placeholders. The placeholder represents an unseen novel object. Thus, the sequence model can be decoupled from the novel object descriptions. 2) A key-value object memory built upon the freely available detection model, contains the visual information and the corresponding word for each object. A query generated from the SM-P is used to retrieve the words from the object memory. The placeholder will further be filled with the correct word, resulting in a caption with novel object descriptions. The experimental results on the held-out MSCOCO dataset demonstrate the ability of DNOC in describing novel concepts.

With the tremendous advances made by Convolutional Neural Networks (ConvNets) on object recognition, we can now easily obtain adequately reliable machine-labeled annotations easily from predictions by off-the-shelf ConvNets. In this work, we present an abstraction memory based framework for few-shot learning, building upon machine-labeled image annotations. Our method takes large-scale machine-annotated dataset (e.g., OpenImages) as an external memory bank. In the external memory bank, the information is stored in the memory slots in the form of key-value, in which image feature is regarded as the key and the label embedding serves as the value. When queried by the few-shot examples, our model selects visually similar data from the external memory bank and writes the useful information obtained from related external data into another memory bank, i.e., abstraction memory. Long Short-Term Memory (LSTM) controllers and attention mechanisms are utilized to guarantee the data written to the abstraction memory correlates with the query example. The abstraction memory concentrates information from the external memory bank to make the few-shot recognition effective. In the experiments, we first confirm that our model can learn to conduct few-shot object recognition on clean human-labeled data from the ImageNet dataset. Then, we demonstrate that with our model, machine-labeled image annotations are very effective and abundant resources for performing object recognition on novel categories. Experimental results show that our proposed model with machine-labeled annotations achieves great results, with only a 1% difference in accuracy between the machine-labeled annotations and the human-labeled annotations.

Despite the recent success of neural networks in image feature learning, a major problem in the video domain is the lack of sufficient labeled data for learning to model temporal information. In this paper, we propose an unsupervised temporal modeling method that learns from untrimmed videos. The speed of motion varies constantly, e.g., a man may run quickly or slowly. We therefore train a Multirate Visual Recurrent Model (MVRM) by encoding frames of a clip with different intervals. This learning process makes the learned model more capable of dealing with motion speed variance. Given a clip sampled from a video, we use its past and future neighboring clips as the temporal context, and reconstruct the two temporal transitions, i.e., present-past transition and present-future transition, reflecting the temporal information in different views. The proposed method exploits the two transitions simultaneously by incorporating a bidirectional reconstruction which consists of a backward reconstruction and a forward reconstruction. We apply the proposed method to two challenging video tasks, i.e., complex event detection and video captioning, in which it achieves state-of-the-art performance. Notably, our method generates the best single feature for event detection with a relative improvement of 10.4% on the MEDTest-13 dataset and achieves the best performance in video captioning across all evaluation metrics on the YouTube2Text dataset.