Long-form video understanding has always been a challenging problem due to the significant redundancy in both temporal and spatial contents. This challenge is further exacerbated by the limited context length of Multimodal Large Language Models (MLLMs). To address this issue, many previous works have attempted to extract key video information, where the "key" is typically semantic-aware and heavily dependent on the CLIP model as prior. In this paper, we propose Flow4Agent, a novel framework that pioneeringly incorporates motion priors from optical flow to facilitate LLM-based long video understanding. Flow4Agent mitigates the redundancy in long videos at both temporal and spatial levels through two core modules: Temporal Granularity Optimization (TGO) adaptively refines frame-level hierarchies, which first leverages coarse flow priors to group similar visual contents and then applies semantic priors to filter out highly irrelevant scene information. Motion Token Pruning (MTP) further refines the intra-frame visual representations, pruning high-redundancy video tokens using fine-grained optical flow information. Extensive experiments demonstrate that our Flow4Agent outperforms existing methods across a wide range of video MLLM benchmarks, especially for hour-level video understanding tasks, achieving 64.7% on Video-MME, 71.4% on MLVU and 60.4% on LongVideoBench.
论文检索
输入标题、作者或关键词，从 2,006 篇学术成果中精准定位
Understanding user actions and their possible mistakes is essential for successful operation of task assistants. In this paper, we develop a unified framework for joint temporal action segmentation and error recognition (recognizing when and which type of error happens) in procedural task videos. We propose a Generalized Task Graph (GTG) whose nodes encode correct steps and background (task-irrelevant actions). We then develop a GTG-Video Alignment algorithm (GTG2Vid) to jointly segment videos into actions and detect frames containing errors. Given that it is infeasible to gather many videos and their annotations for different types of errors, we study a framework that only requires normal (error-free) videos during training. More specifically, we leverage large language models (LLMs) to obtain error descriptions and subsequently use video-language models (VLMs) to generate visually-aligned textual features, which we use for error recognition. We then propose an Error Recognition Module (ERM) to recognize the error frames predicted by GTG2Vid using the generated error features. By extensive experiments on two egocentric datasets of EgoPER and CaptainCook4D, we show that our framework outperforms other baselines on action segmentation, error detection and recognition.
Multimodal Large Language Models (MLLMs) have demonstrated significant success in visual understanding tasks. However, challenges persist in adapting these models for video comprehension due to the large volume of data and temporal complexity. Existing Video-LLMs using uniform frame sampling often struggle to capture the query-related crucial spatiotemporal clues of videos effectively. In this paper, we introduce Q-Frame, a novel approach for adaptive frame selection and multi-resolution scaling tailored to the video's content and the specific query. Q-Frame employs a training-free, plug-and-play strategy generated by a text-image matching network like CLIP, utilizing the Gumbel-Max trick for efficient frame selection. Q-Frame allows Video-LLMs to process more frames without exceeding computational limits, thereby preserving critical temporal and spatial information. We demonstrate Q-Frame's effectiveness through extensive experiments on benchmark datasets, including MLVU, LongVideoBench, and Video-MME, illustrating its superiority over existing methods and its applicability across various video understanding tasks.
Procedural videos are critical for learning new tasks. Temporal action segmentation (TAS), which classifies the action in every video frame, has become essential for understanding procedural videos. Existing TAS models, however, learn a fixed-set of tasks at training and unable to adapt to novel tasks at test time. Thus, we introduce the new problem of Multi-Modal Few-shot Temporal Action Segmentation (MMF-TAS) to learn open-set models that can generalize to novel procedural tasks with minimal visual/textual examples. We propose the first MMF-TAS framework, by designing a Prototype Graph Network (PGNet). In PGNet, a Prototype Building Block summarizes action information from support videos of the novel tasks via an Action Relation Graph, and encodes this information into action prototypes via a Dynamic Graph Transformer. Next, a Matching Block compares action prototypes with query videos to infer framewise action labels. To exploit the advantages of both visual and textual modalities, we compute separate action prototypes for each modality and combine the two modalities through prediction fusion to avoid overfitting on one modality. By extensive experiments on procedural datasets, we show our method successfully adapts to novel tasks during inference and significantly outperforms baselines. Our code is available at https://github.com/ZijiaLewisLu/ICCV2025-MMF-TAS.
Modeling human-scene interactions (HSI) is essential for understanding and simulating everyday human behaviors. Recent approaches utilizing generative modeling have made progress in this domain; however, they are limited in controllability and flexibility for real-world applications. To address these challenges, we propose reformulating the HSI modeling problem as Scene-aware Motion In-betweening---a more tractable and practical task. We introduce SceneMI, a framework that supports several practical applications, including keyframe-guided character animation in 3D scenes and enhancing the motion quality of imperfect HSI data. SceneMI employs dual scene descriptors to comprehensively encode global and local scene context. Furthermore, our framework leverages the inherent denoising nature of diffusion models to generalize on noisy keyframes. Experimental results demonstrate SceneMI's effectiveness in scene-aware keyframe in-betweening and generalization to the real-world GIMO dataset, where motions and scenes are acquired by noisy IMU sensors and smartphones. We further showcase SceneMI's applicability in HSI reconstruction from monocular videos.
DynImg: Key Frames with Visual Prompts are Good Representation for Multi-Modal Video Understanding
PDF ↗In recent years, the introduction of Multi-modal Large Language Models (MLLMs) into video understanding tasks has become increasingly prevalent. However, how to effectively integrate temporal information remains a critical research focus. Traditional approaches treat spatial and temporal information separately. Due to issues like motion blur, it is challenging to accurately represent the spatial information of rapidly moving objects. This can lead to temporally important regions being underemphasized during spatial feature extraction, which in turn hinders accurate spatio-temporal interaction and video understanding. To address this limitation, we propose an innovative video representation method called Dynamic-Image (DynImg). Specifically, we introduce a set of non-key frames as temporal prompts to highlight the spatial areas containing fast-moving objects. During the process of visual feature extraction, these prompts guide the model to pay additional attention to the fine-grained spatial features corresponding to these regions. Moreover, to maintain the correct sequence for DynImg, we employ a corresponding 4D video Rotary Position Embedding. This retains both the temporal and spatial adjacency of DynImg, helping MLLM understand the spatio-temporal order within this combined format. Experimental evaluations reveal that DynImg surpasses the state-of-the-art methods by approximately 2% across multiple video understanding benchmarks, proving the effectiveness of our temporal prompts in enhancing video comprehension.
Dynamic Scene Graph Generation (DSGG) aims to comprehensively understand videos by abstracting them into visual triplets <subject, predicate, object>. Most existing methods focus on capturing temporal dependencies, but overlook crucial visual relationship dependencies between entities and predicates, as well as among predicate subclasses. These dependencies are essential for a deeper contextual understanding of scenarios. Additionally, current approaches do not support end-to-end training and instead rely on a two-stage pipeline, which incurs higher computational costs. To address these issues, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN leverages CLIP's semantic priors to model fine-grained triplet cues to generate scene graph. In addition, we design a Predicate Association Parsing (PAP) module that employs a conditional weight mapping mechanism to structure entity and predicate representations. We further introduce a Hierarchical Attention (HA) mechanism to integrate spatio-temporal context with entity and predicate representations, enabling effective associative reasoning. Extensive experiments on the Action Genome dataset demonstrate significant performance improvements over existing methods. The source code is available in \hyperref[URL] https://github.com/wlw951226/ARN .
Spatiotemporal video grounding aims to localize target entities in videos based on textual queries. While existing research has made significant progress in exocentric videos, the egocentric setting remains relatively underexplored, despite its growing importance in applications such as augmented reality and robotics. In this work, we conduct a systematic analysis of the discrepancies between egocentric and exocentric videos, revealing key challenges such as shorter object durations, sparser trajectories, smaller object sizes, and larger positional shifts. To address these challenges, we introduce EgoMask, the first pixel-level benchmark for fine-grained spatiotemporal grounding in egocentric videos. It is constructed by our proposed automatic annotation pipeline, which annotates referring expressions and object masks across short-, medium-, and long-term videos. Additionally, we create EgoMask-Train, a large-scale training dataset to facilitate model development. Experiments demonstrate that the state-of-the-art spatiotemporal grounding models perform poorly on our benchmark EgoMask, but fine-tuning on EgoMask-Train yields significant improvements, while preserving performance on exocentric datasets. Our work thus provides essential resources and insights for advancing egocentric video understanding.
Whether snipping with scissors or opening a box, humans can quickly understand the 3D configurations of familiar objects. For novel objects, we can resort to long-form inspection to build intuition. The more we observe the object, the better we get at predicting its 3D state immediately. Existing systems, however, are limited to either optimizing underlying representations from multi-view observations or training a feed-forward predictor from supervised datasets. We introduce Predict-Optimize-Distill (POD), a self-improving framework that interleaves prediction and optimization in a mutually reinforcing cycle to achieve better 4D object understanding with increasing observation time. Given a multi-view object scan and a long-form monocular video of human-object interaction, POD iteratively trains a neural network to predict local part poses from RGB frames, uses this predictor to initialize a global optimization which refines output poses through inverse rendering, then finally distills the results of optimization back into the model by generating synthetic self-labeled training data from novel viewpoints. Each iteration improves both the predictive model and the optimized motion trajectory, creating a virtuous cycle that bootstraps its own training data to learn about the pose configurations of an object. We also introduce a quasi-multiview mining strategy for reducing depth ambiguity by leveraging long video. We evaluate POD on 14 real-world and 5 synthetic objects with various joint types, including revolute and prismatic joints as well as multi-body configurations where parts detach or reattach independently. POD demonstrates significant improvement over a pure optimization baseline which gets stuck in local minima, particularly for longer videos. We also find that POD's performance improves with both video length and successive iterations of the self-improving cycle, highlighting its ability to scale performance with additional observations and compute.
Despite advances in general video understanding, Video Large Language Models (Video-LLMs) face challenges in precise temporal localization due to discrete time representations and limited temporally aware datasets. Existing methods for temporal expression either conflate time with text-based numerical values, add a series of dedicated temporal tokens, or regress time using specialized temporal grounding heads. To address these issues, we introduce DisTime, a lightweight framework designed to enhance temporal comprehension in Video-LLMs. DisTime employs a learnable token to create a continuous temporal embedding space and incorporates a Distribution-based Time Decoder that generates temporal probability distributions, effectively mitigating boundary ambiguities and maintaining temporal continuity. Additionally, the Distribution-based Time Encoder re-encodes timestamps to provide time markers for Video-LLMs. To overcome temporal granularity limitations in existing datasets, we propose an automated annotation paradigm that combines the captioning capabilities of Video-LLMs with the localization expertise of dedicated temporal models. This leads to the creation of InternVid-TG, a substantial dataset with 1.25M temporally grounded events across 179k videos, surpassing ActivityNet-Caption by 55 times. Extensive experiments demonstrate that DisTime achieves state-of-the-art performance across benchmarks in three time-sensitive tasks while maintaining competitive performance in Video QA tasks. DisTime is released at https://github.com/josephzpng/DisTime.
Vision language models have received increasing attention for their ability to integrate visual and textual understanding, with some capable of processing native-resolution images and long videos. While the capacity to process large visual data unlocks numerous downstream applications, it often introduces significant latency challenges, as the visual tokens dominate the resource consumption. In this work, we introduce SparseVILA, a novel method of query-aware token retrieval to dynamically accelerate the underlying LLM by pruning tokens in the prefill stage while attending to a sparse subset of visual tokens during the decoding phase. By decoupling the context and generation compression, we can migrate the majority of sparsity into the generation stage, enabling query-aware support for multi-turn conversation while achieving a 1.4x speedup on image benchmarks. This approach leads to +5.9% average accuracy improvements on image-centric benchmarks over previous works. Finally, SparseVILA enables efficient long-context/long-generation tasks by achieving a 3.6x and 1.7x speedup in prefill and decoding, respectively.
RobAVA: A Large-scale Dataset and Baseline Towards Video based Robotic Arm Action Understanding
PDF ↗Understanding the behaviors of robotic arms is essential for various robotic applications such as logistics management, precision agriculture, and automated manufacturing. However, the lack of large-scale and diverse datasets significantly hinders progress in video-based robotic arm action understanding, highlighting the need for collecting a new large-scale dataset. In particular, our RobAVA contains 40k video sequences with video-level fine-grained annotations, covering basic actions such as picking, pushing, and placing, as well as their combinations in different orders and interactions with various objects. Distinguished to existing action recognition benchmarks, RobAVA includes instances of both normal and anomalous executions for each action category. Our further analysis reveals that the primary challenge in robotic arm action recognition lies in the fact that a complete action consists of a sequence of fundamental, atomic behaviors, requiring models to learn the inter-relationships among them. To this end, we propose a novel baseline approach, AGPT-Net, which re-defines the problem of understanding robotic arm actions as a task of aligning video sequences with atomic attributes.To enhance AGPT-Net's ability to distinguish normal and anomalous action instances, we introduce a joint semantic space constraint between category and attribute semantics, thereby amplifying the separation between normal and anomalous attribute representations for each action. We conduct extensive experiments to demonstrate AGPT-Net's superiority over other mainstream recognition models.
Efficient lightweight neural networks have received increasing attention due to their faster reasoning speed and easier deployment on mobile devices. However, existing video models still focus on the larger ViT architecture, and few works attempt to build efficient architecture. Since many efficient contrastive language-image pre-training (CLIP) models have shown strong zero-shot classification and retrieval capability, we attempt to fill the gap in video-text understanding models and propose a fast and efficient video-text model MobileViCLIP with strong zero-shot reasoning capability that can be deployed on mobile devices. In particular, our MobileViCLIP-Small obtains similar zero-shot retrieval performance as InternVideo2-L14 on text-to-video dataset MSR-VTT while being 46.7x faster when deployed on the mobile device. Furthermore, MobileViCLIP-Small can generalize to zero-shot action recognition task and obtains 1.0% better Top-1 accuracy than InternVideo2-S14 while being 5.6x faster on the mobile device. The code is available at https://github.com/MCG-NJU/MobileViCLIP.
Towards Video Thinking Test: A Holistic Benchmark for Advanced Video Reasoning and Understanding
PDF ↗Human intelligence requires both correctness and robustness, with the former being foundational for the latter. In video understanding, correctness ensures the accurate interpretation of visual content, and robustness maintains consistent performance in challenging conditions. Despite advances in video large language models (video LLMs), existing benchmarks inadequately reflect the gap between these models and human intelligence in maintaining correctness and robustness in video interpretation. We introduce the Video Turing Test (Video-TT), a benchmark designed to assess if video LLMs can interpret real-world videos as effectively as humans.Video-TT differentiates between errors due to inadequate frame sampling and 1) genuine gaps in understanding complex visual narratives, and 2) evaluates robustness against natural adversarial questions. Video-TT comprises 1,000 YouTube Shorts videos, each with one open-ended question and four adversarial questions that probe visual and narrative complexity. Our evaluation shows a significant gap between video LLMs and human performance, underscoring the need for benchmarks like Video-TT to advance video understanding.
Accurate driving behavior recognition and reasoning are critical for autonomous driving video understanding. However, existing methods often tend to dig out the shallow causal, fail to address spurious correlations across modalities, and ignore the ego-vehicle level causality modeling. To overcome these limitations, we propose a novel Multimodal Causal Analysis Model (MCAM) that constructs latent causal structures between visual and language modalities. Firstly, we design a multi-level feature extractor to capture long-range dependencies. Secondly, we design a causal analysis module that dynamically models driving scenarios using a directed acyclic graph (DAG) of driving states. Thirdly, we utilize a vision-language transformer to align critical visual features with their corresponding linguistic expressions. Extensive experiments on the BDD-X, and CoVLA datasets demonstrate that MCAM achieves SOTA performance in visual-language causal relationship learning. Furthermore, the model exhibits superior capability in capturing causal characteristics within video sequences, showcasing its effectiveness for autonomous driving applications. The code is available at https://github.com/SixCorePeach/MCAM
Referring audio-visual segmentation (RAVS) has recently seen significant advancements, yet challenges remain in integrating multimodal information and deeply understanding and reasoning about audiovisual content. To extend the boundaries of RAVS and facilitate future research in this field, we propose Omnimodal Referring Audio-Visual Segmentation (OmniAVS), a new dataset containing 2,104 videos and 61,095 multimodal referring expressions. OmniAVS stands out with three key innovations: (1) 8 types of multimodal expressions that flexibly combine text, speech, sound, and visual cues; (2) an emphasis on understanding audio content beyond just detecting their presence; and (3) the inclusion of complex reasoning and world knowledge in expressions. Furthermore, we introduce Omnimodal Instructed Segmentation Assistant (OISA), to address the challenges of multimodal reasoning and fine-grained understanding of audiovisual content in OmniAVS. OISA uses MLLM to comprehend complex cues and perform reasoning-based segmentation. Extensive experiments show that OISA outperforms existing methods on OmniAVS and achieves competitive results on other related tasks.
V2PE: Improving Multimodal Long-Context Capability of Vision-Language Models with Variable Visual Position Encoding
PDF ↗Vision-Language Models (VLMs) have shown promising capabilities in handling various multimodal tasks, yet they struggle in long-context scenarios, particularly tasks involving videos, high-resolution images, or lengthy image-text documents. In our work, we first conduct an empirical analysis of VLMs' long-context capabilities using our augmented long-context multimodal datasets. Our findings reveal that directly applying the positional encoding mechanism used for textual tokens to visual tokens is suboptimal, and VLM performance degrades sharply when the position encoding exceeds the model's context window. To address this, we propose Variable Visual Position Encoding (V2PE), a novel positional encoding approach that employs variable and smaller increments for visual tokens, enabling more efficient management of long multimodal sequences. Our experiments demonstrate the effectiveness of V2PE in enhancing VLMs' ability to effectively understand and reason over long multimodal contexts. We further integrate V2PE with our augmented long-context multimodal datasets to fine-tune the open-source VLMs. The fine-tuned model achieves strong performance on both standard and long-context multimodal tasks. Notably, when the sequence length of the training dataset is increased to 256K tokens, the model is capable of processing multimodal sequences up to 1M tokens, highlighting its potential for real-world long-context applications. We shall release the code, model weights, and datasets to facilitate further research.
Embodied VideoAgent: Persistent Memory from Egocentric Videos and Embodied Sensors Enables Dynamic Scene Understanding
PDF ↗This paper investigates the problem of understanding dynamic 3D scenes from egocentric observations, a key challenge in robotics and embodied AI. Unlike prior studies that explored this as long-form video understanding and utilized egocentric video only, we instead propose an LLM-based agent, Embodied VideoAgent, which constructs scene memory from both egocentric video and embodied sensory inputs (e.g. depth and pose sensing). We further introduce a VLM-based approach to automatically update the memory when actions or activities over objects are perceived. Embodied VideoAgent attains significant advantages over counterparts in challenging reasoning and planning tasks in 3D scenes, achieving gains of 6.5% on Ego4D-VQ3D, 2.6% on OpenEQA, and 15.3% on EnvQA. We have also demonstrated its potential in various embodied AI tasks including generating embodied interactions and perception for robot manipulation. The code and demo will be made public.
Recognizing subtle similarities and differences among sets of similar activities is central to many real-world applications, including skill acquisition, sports performance evaluation, and anomaly detection. Humans excel at such fine-grained analysis, which requires comprehensive video understanding and cross-video reasoning about action attributes, poses, positions, and emotional states. Yet existing video-based large language models typically address only single-video recognition, leaving their capacity for multi-video reasoning largely unexplored. We introduce VideoSetDiff, a curated dataset designed to test detail-oriented recognition across diverse activities, from subtle action attributes to viewpoint transitions. Our evaluation of current video-based LLMs on VideoSetDiff reveals critical shortcomings, particularly in fine-grained detail recognition and multi-video reasoning. To mitigate these issues, we propose an automatically generated dataset for instruction tuning alongside a novel multi-video recognition framework. While instruction tuning and specialized multi-video reasoning improve performance, all tested models remain far from satisfactory. These findings underscore the need for more robust video-based LLMs capable of handling complex multi-video tasks, enabling diverse real-world applications.
Co-speech gestures play a vital role in non-verbal communication. In this paper, we introduce a new framework for co-speech gesture understanding in the wild. Specifically, we propose three new tasks and benchmarks to evaluate a model's capability to comprehend gesture-speech-text associations: (i) gesture based retrieval, (ii) gestured word spotting, and (iii) active speaker detection using gestures. We present a new approach that learns a tri-modal video-gesture-speech-text representation to solve these tasks. By leveraging a combination of global phrase contrastive loss and local gesture-word coupling loss, we demonstrate that a strong gesture representation can be learned in a weakly supervised manner from videos in the wild. Our learned representations outperform previous methods, including large vision-language models (VLMs). Further analysis reveals that speech and text modalities capture distinct gesture related signals, underscoring the advantages of learning a shared tri-modal embedding space. The dataset, model, and code are available at: https://www.robots.ox.ac.uk/ vgg/research/jegal.