Efficient long-video understanding~(LVU) remains a challenging task in computer vision. Current long-context vision-language models~(LVLMs) suffer from information loss due to compression and brute-force downsampling. While retrieval-augmented generation (RAG) methods mitigate this issue, their applicability is limited due to explicit query dependency. To overcome this challenge, we introduce a novel memory-enhanced RAG-based approach called MemVid, which is inspired by the cognitive memory of human beings. Our approach operates in four basic steps: 1) memorizing holistic video information, 2) reasoning about the task's information needs based on memory, 3) retrieving critical moments based on the information needs, and 4) focusing on the retrieved moments to produce the final answer. To enhance the system's memory-grounded reasoning capabilities while achieving optimal end-to-end performance, we propose a curriculum learning strategy. This approach begins with supervised learning on well-annotated reasoning results, then progressively explores and reinforces more plausible reasoning outcomes through reinforcement learning. We perform extensive evaluations on popular LVU benchmarks, including MLVU, VideoMME and LVBench. In our experiments, MemVid demonstrates superior efficiency and effectiveness compared to both LVLMs and RAG methods.
Despite recent advances in Video Large Language Models (VideoLLMs), effectively understanding long-form videos remains a significant challenge. Perceiving lengthy videos containing thousands of frames poses substantial computational burden. To mitigate this issue, this paper introduces Generative Frame Sampler (GenS), a plug-and-play module integrated with VideoLLMs to facilitate efficient lengthy video perception. Built upon a lightweight VideoLLM, GenS leverages its inherent vision-language capabilities to identify question-relevant frames. To facilitate effective retrieval, we construct GenS-Video-150K, a large-scale video instruction dataset with dense frame relevance annotations. Extensive experiments demonstrate that GenS consistently boosts the performance of various VideoLLMs, including open-source models (Qwen2-VL-7B, Aria-25B, VILA-40B, LLaVA-Video-7B/72B) and proprietary assistants (GPT-4o, Gemini). When equipped with GenS, open-source VideoLLMs achieve impressive state-of-the-art results on long-form video benchmarks: LLaVA-Video-72B reaches 66.8 (+4.3) on LongVideoBench and 77.0 (+2.7) on MLVU, while Aria obtains 39.2 on HourVideo surpassing the Gemini-1.5-pro by 1.9 points. We will release all datasets and models at https://generative-sampler.github.io.
Recent advances in long video understanding typically mitigate visual redundancy through visual token pruning based on attention distribution. However, while existing methods employ post-hoc low-response token pruning in decoder layers, they overlook the input-level semantic correlation between visual tokens and instructions (query). In this paper, we propose QuoTA, an ante-hoc training-free modular that extends existing large video-language models (LVLMs) for visual token assignment based on query-oriented frame-level importance assessment. The query-oriented token selection is crucial as it aligns visual processing with task-specific requirements, optimizing token budget utilization while preserving semantically relevant content. Specifically, (i) QuoTA strategically allocates frame-level importance scores based on query relevance, enabling one-time visual token assignment before cross-modal interactions in decoder layers, (ii) we decouple the query through Chain-of-Thoughts reasoning to facilitate more precise LVLM-based frame importance scoring, and (iii) QuoTA offers a plug-and-play functionality that extends to existing LVLMs. Extensive experimental results demonstrate that implementing QuoTA with LLaVA-Video-7B yields an average performance improvement of 3.2% across six benchmarks (including Video-MME and MLVU) while operating within an identical visual token budget as the baseline. Codes are open-sourced at https://github.com/MAC-AutoML/QuoTA.
From image to video understanding, the capabilities of Multi-modal LLMs (MLLMs) are increasingly powerful. However, most existing video understanding benchmarks are relatively short, which makes them inadequate for effectively evaluating the long-sequence modeling capabilities of MLLMs. This highlights the urgent need for a comprehensive and integrated long video understanding benchmark to assess the ability of MLLMs thoroughly. To this end, we propose ALLVB (ALL-in-One Long Video Understanding Benchmark). ALLVB's main contributions include: 1) It integrates 9 major video understanding tasks. These tasks are converted into video QA formats, allowing a single benchmark to evaluate 9 different video understanding capabilities of MLLMs, highlighting the versatility, comprehensiveness, and challenging nature of ALLVB. 2) A fully automated annotation pipeline using GPT-4o is designed, requiring only human quality control, which facilitates the maintenance and expansion of the benchmark. 3) It contains 1,376 videos across 16 categories, averaging nearly 2 hours each, with a total of 252k QAs. To the best of our knowledge, it is the largest long video understanding benchmark in terms of the number of videos, average duration, and number of QAs. We have tested various mainstream MLLMs on ALLVB, and the results indicate that even the most advanced commercial models have significant room for improvement. This reflects the benchmark's challenging nature and demonstrates the substantial potential for development in long video understanding.
Recent advances in human preference alignment have significantly enhanced multimodal generation and understanding. A key approach is training reward models to guide preference optimization. However, existing models are often task-specific, limiting their adaptability across diverse visual applications. We also argue that jointly learning to assess multiple tasks may foster a synergistic effect, where improved image understanding enhances image generation assessment, and refined image evaluation benefits video assessment through better frame analysis. To this end, this paper proposes UnifiedReward, the first unified reward model for multimodal understanding and generation assessment, enabling both pairwise ranking and pointwise scoring, which can be employed for vision model preference alignment. Specifically, (1) we first develop UnifiedReward on our constructed large-scale human preference dataset, including both image and video generation/understanding tasks. (2) Then, it is utilized to automatically construct high-quality preference pair data based on the vision models, fine-gradually filtering their outputs through pair ranking and point sifting. (3) Finally, these data are used for their preference alignment through Direct Preference Optimization (DPO). Experimental results demonstrate that joint learning to assess diverse visual tasks can lead to substantial mutual benefits and we apply our pipeline to both image and video understanding/generation tasks, significantly improving the performance in each domain.
In the rapidly evolving domain of video understanding, Video Question Answering (VideoQA) remains a focal point. However, existing datasets exhibit gaps in temporal and spatial granularity, which consequently limits the capabilities of existing VideoQA methods. This paper introduces the Multi-Object Multi-Actor Question Answering (MOMA-QA) dataset, which is designed to address these shortcomings by emphasizing temporal localization, spatial relationship reasoning, and entity-centric queries. With ground truth scene graphs and temporal interval annotations, MOMA-QA is ideal for developing models for fine-grained video understanding. Furthermore, we present a novel video-language model, SGVLM, which incorporates a scene graph predictor, an efficient frame retriever, and a pre-trained large language model for temporal localization and fine-grained relationship understanding. Evaluations on MOMA-QA and other public datasets demonstrate the superior performance of our model, setting new benchmarks for VideoQA.
Temporal localization in untrimmed videos, which aims to identify specific timestamps, is crucial for video understanding but remains challenging. This task encompasses several subtasks, including temporal action localization, temporal video grounding, moment retrieval, and generic event boundary detection. Existing methods in each subfield are typically designed for specific tasks and lack generalizability across domains. In this paper, we propose TimeLoc, a unified end-to-end framework for timestamp localization that can handle multiple tasks. First, our approach employs a simple yet effective one-stage localization model that supports text queries as input and multiple actions as output. Second, we jointly train the video encoder and localization model in an end-to-end manner. To efficiently process long videos, we introduce temporal chunking, enabling the handling of videos with over 30k frames. Third, we find that fine-tuning pre-trained text encoders with a multi-stage training strategy further enhances text-conditioned localization. TimeLoc achieves state-of-the-art results across multiple benchmarks: +1.3% and +1.9% mAP over previous best methods on THUMOS14 and EPIC-Kitchens-100, +1.1% on Kinetics-GEBD, +2.94% mAP on QVHighlights, and significant improvements in temporal video grounding (+11.5% on TACoS and +6.7% on Charades-STA under [email omitted]). Our code and checkpoints will be released at https://github.com/sming256/TimeLoc.
Humans detect real-world object anomalies by perceiving, interacting, and reasoning based on object-conditioned physical knowledge. The long-term goal of Industrial Anomaly Detection (IAD) is to enable machines to autonomously replicate this skill. However, current IAD algorithms are largely developed and tested on static, semantically simple datasets, which diverge from real-world scenarios where physical understanding and reasoning are essential. To bridge this gap, we introduce the Physics Anomaly Detection (Phys-AD) dataset, the first large-scale, real-world, physics-grounded video dataset for industrial anomaly detection. Collected using a real robot arm and motor, Phys-AD provides a diverse set of dynamic, semantically rich scenarios. The dataset includes more than 6400 videos across 22 real-world object categories, interacting with robot arms and motors, and exhibits 47 types of anomalies. Anomaly detection in Phys-AD requires visual reasoning, combining both physical knowledge and video content to determine object abnormality. We benchmark state-of-the-art anomaly detection methods under three settings: unsupervised AD, weakly-supervised AD, and video-understanding AD, highlighting their limitations in handling physics-grounded anomalies. Additionally, we introduce the Physics Anomaly Explanation (PAEval) metric, designed to assess the ability of visual-language foundation models to not only detect anomalies but also provide accurate explanations for their underlying physical causes. Our project is available at https://guyao2023.github.io/Phys-AD/.
In egocentric video understanding, the motion of hands and objects as well as their interactions play a significant role by nature. However, existing egocentric video representation learning methods mainly focus on aligning video representation with high-level narrations, overlooking the intricate dynamics between hands and objects. In this work, we aim to integrate the modeling of fine-grained hand-object dynamics into the video representation learning process. Since no suitable data is available, we introduce HOD, a novel pipeline employing a hand-object detector and a large language model to generate high-quality narrations with detailed descriptions of hand-object dynamics. To learn these fine-grained dynamics, we propose EgoVideo, a model with a new lightweight motion adapter to capture fine-grained hand-object motion information. Through our co-training strategy, EgoVideo effectively and efficiently leverages the fine-grained hand-object dynamics in the HOD data. Extensive experiments demonstrate that our method achieves state-of-the-art performance across multiple egocentric downstream tasks, including improvements of 6.3% in EK-100 multi-instance retrieval, 5.7% in EK-100 classification, and 16.3% in EGTEA classification in zero-shot settings. Furthermore, our model exhibits robust generalization capabilities in hand-object interaction and robot manipulation tasks. Code and data are available at https://github.com/OpenRobotLab/EgoHOD/.
In recent years, online lecture videos have become an increasingly popular resource for acquiring new knowledge. Systems capable of effectively understanding/indexing lecture videos are thus highly desirable, enabling downstream tasks like question answering to help users efficiently locate specific information within videos. This work proposes PreMind, a novel multi-agent multimodal framework that leverages various large models for advanced understanding/indexing of presentation-style videos. PreMind first segments videos into slide-presentation segments using a Vision-Language Model (VLM) to enhance modern shot-detection techniques. Each segment is then analyzed to generate multimodal indexes through three key steps: (1) extracting slide visual content, (2) transcribing speech narratives, and (3) consolidating these visual and speech contents into an integrated understanding. Three innovative mechanisms are also proposed to improve performance: leveraging prior lecture knowledge to refine visual understanding, detecting/correcting speech transcription errors using a VLM, and utilizing a critic agent for dynamic iterative self-reflection in vision analysis. Compared to traditional video indexing methods, PreMind captures rich, reliable multimodal information, allowing users to search for details like abbreviations shown only on slides. Systematic evaluations on the public LPM dataset and an internal enterprise dataset are conducted to validate PreMind's effectiveness, supported by detailed analyses.
Video object segmentation (VOS) is a critical task in the development of video perception and understanding. The Segment-Anything Model 2 (SAM 2), released by Meta AI, is the current state-of-the-art architecture for end-to-end VOS. SAM 2 performs very well on both clean video data and augmented data, and completely intelligent video perception requires an understanding of how this architecture is capable of achieving such quality results. To better understand how each step within the SAM 2 architecture permits high-quality video segmentation, a variety of complex video transformations are passed through the architecture, and the impact at each stage of the process is measured. It is observed that each progressive stage enables the filtering of complex transformation noise and the emphasis of the object of interest. Contributions include the creation of complex transformation video datasets, an analysis of how each stage of the SAM 2 architecture interprets these transformations, and visualizations of segmented objects through each stage. By better understanding how each model structure impacts overall video understanding, VOS development can work to improve real-world applicability and performance tracking, localizing, and segmenting objects despite complex cluttered scenes and obscurations.
Recent Multi-modal Large Language Models (MLLMs) have made great progress in video understanding. However, their performance on videos involving human actions is still limited by the lack of high-quality data. To address this, we introduce a two-stage data annotation pipeline. First, we design strategies to accumulate videos featuring clear human actions from the Internet. Second, videos are annotated in a standardized caption format that uses human attributes to distinguish individuals and chronologically details their actions and interactions. Through this pipeline, we curate two datasets, namely HAICTrain and HAICBench. \textbf{HAICTrain} comprises 126K video-caption pairs generated by Gemini-Pro and verified for training purposes. Meanwhile, \textbf{HAICBench} includes 412 manually annotated video-caption pairs and 2,000 QA pairs, for a comprehensive evaluation of human action understanding. Experimental results demonstrate that training with HAICTrain not only significantly enhances human understanding abilities across 4 benchmarks, but can also improve text-to-video generation results. Both the HAICTrain and HAICBench are released at https://huggingface.co/datasets/KuaishouHAIC/HAIC.
Temporal action detection (TAD) is a fundamental video understanding task that aims to identify human actions and localize their temporal boundaries in videos. Although this field has achieved remarkable progress in recent years, further progress and real-world applications are impeded by the absence of a standardized framework. Currently, different methods are compared under different implementation settings, evaluation protocols, etc., making it difficult to assess the real effectiveness of a specific technique. To address this issue, we propose \textbf{OpenTAD}, a unified TAD framework consolidating 16 different TAD methods and 9 standard datasets into a modular codebase. In OpenTAD, minimal effort is required to replace one module with a different design, train a feature-based TAD model in end-to-end mode, or switch between the two. OpenTAD also facilitates straightforward benchmarking across various datasets and enables fair and in-depth comparisons among different methods. With OpenTAD, we comprehensively study how innovations in different network components affect detection performance and identify the most effective design choices through extensive experiments. This study has led to a new state-of-the-art TAD method built upon existing techniques for each component. We have made our code and models available at https://github.com/sming256/OpenTAD.
Recent advances in Multi-Modal Large Language Models (M-LLMs) show promising results in video reasoning. Popular Multi-Modal Large Language Model (M-LLM) frameworks usually apply naive uniform sampling to reduce the number of video frames that are fed into an M-LLM, particularly for long context videos. However, it could lose crucial context in certain periods of a video, so that the downstream M-LLM may not have sufficient visual information to answer a question. To attack this pain point, we propose a light-weight M-LLM -based frame selection method that adaptively select frames that are more relevant to users' queries. In order to train the proposed frame selector, we introduce two supervision signals (i) Spatial signal, where single frame importance score by prompting a M-LLM; (ii) Temporal signal, in which multiple frames selection by prompting Large Language Model (LLM) using the captions of all frame candidates. The selected frames are then digested by a frozen downstream video M-LLM for visual reasoning and question answering. Empirical results show that the proposed M-LLM video frame selector improves the performances various downstream video Large Language Model (video-LLM) across medium (ActivityNet, NExT-QA) and long (EgoSchema, LongVideoBench) context video question answering benchmarks.
Modern Video Large Language Models (VLLMs) often rely on uniform frame sampling for video understanding, but this approach frequently fails to capture critical information due to frame redundancy and variations in video content. We propose MaxInfo, a training-free method based on the maximum volume principle, which selects and retains the most representative frames from the input video. By maximizing the geometric volume formed by selected embeddings, MaxInfo ensures that the chosen frames cover the most informative regions of the embedding space, effectively reducing redundancy while preserving diversity. This method enhances the quality of input representations and improves long video comprehension performance across benchmarks. For instance, MaxInfo achieves a 3.28% improvement on LongVideoBench and a 6.4% improvement on EgoSchema for LLaVA-Video-7B. It also achieves a 3.47% improvement for LLaVA-Video-72B. The approach is simple to implement and works with existing VLLMs without the need for additional training, making it a practical and effective alternative to traditional uniform sampling methods.
Video quality assessment tasks rely heavily on the rich features required for video understanding, such as semantic information, texture, and temporal motion. The existing video foundational model, InternVideo2, has demonstrated strong potential in video understanding tasks due to its large parameter size and large-scale multimodal data pertaining. Building on this, we explored the transferability of InternVideo2 to video quality assessment under compression scenarios. To design a lightweight model suitable for this task, we proposed a distillation method to equip the smaller model with rich compression quality priors. Additionally, we examined the performance of different backbones during the distillation process. The results showed that, compared to other methods, our lightweight model distilled from InternVideo2 achieved excellent performance in compression video quality assessment.
We introduce a gradient-based approach for learning task graphs from procedural activities, improving over hand-crafted methods. Our method directly optimizes edge weights via maximum likelihood, enabling integration into neural architectures. We validate our approach on CaptainCook4D, EgoPER, and EgoProceL, achieving +14.5%, +10.2%, and +13.6% F1-score improvements. Our feature-based approach for predicting task graphs from textual/video embeddings demonstrates emerging video understanding abilities. We also achieved top performance on the procedure understanding benchmark on Ego-Exo4D and significantly improved online mistake detection (+19.8% on Assembly101-O, +6.4% on EPIC-Tent-O). Code: https://github.com/fpv-iplab/Differentiable-Task-Graph-Learning.
Large multimodal models (LMMs) have shown remarkable performance in video understanding tasks and can even process videos longer than one hour. However, despite their ability to handle long inputs, generating outputs with corresponding levels of richness remains a challenge. In this paper, we explore the issue of long outputs in LMMs using video captioning as a proxy task, and we find that open-source LMMs struggle to consistently generate outputs exceeding about 300 words. Through controlled experiments, we find that the scarcity of paired examples with long-captions during training is the primary factor limiting the model's output length. However, manually annotating long-caption examples is time-consuming and expensive. To address this, we propose the LongCaption-Agent, a framework that synthesizes long caption data by aggregating multi-level descriptions. Using LongCaption-Agent, we curated a new long-caption dataset, LongCaption-10K. We also develop LongCaption-Bench, a benchmark designed to comprehensively evaluate the quality of long captions generated by LMMs. By incorporating LongCaption-10K into training, we enable LMMs to generate captions exceeding 1,000 words, while maintaining high output quality. In LongCaption-Bench, our 8B parameter model achieved state-of-the-art performance, even surpassing larger proprietary models. We will release the dataset and code after publication.
Recent advances in visual language models (VLMs) have significantly improved image captioning, but extending these gains to video understanding remains challenging due to the scarcity of fine-grained video captioning datasets. To bridge this gap, we propose a novel zero-shot video captioning approach that combines frame-level scene graphs from a video to obtain intermediate representations for caption generation. Our method first generates frame-level captions using an image VLM, converts them into scene graphs, and consolidates these graphs to produce comprehensive video-level descriptions. To achieve this, we leverage a lightweight graph-to-text model trained solely on text corpora, eliminating the need for video captioning annotations. Experiments on the MSR-VTT and ActivityNet Captions datasets show that our approach outperforms zero-shot video captioning baselines, demonstrating that aggregating frame-level scene graphs yields rich video understanding without requiring large-scale paired data or high inference cost.
Traffic accidents present complex challenges for autonomous driving, often featuring unpredictable scenarios that hinder accurate system interpretation and responses.Nonetheless, prevailing methodologies fall short in elucidating the causes of accidents and proposing preventive measures due to the paucity of training data specific to accident scenarios.In this work, we introduce AVD2 (Accident Video Diffusion for Accident Video Description), a novel framework that enhances accident scene understanding by generating accident videos that aligned with detailed natural language descriptions and reasoning, resulting in the contributed EMM-AU (Enhanced Multi-Modal Accident Video Understanding) dataset. Empirical results reveal that the integration of the EMM-AU dataset establishes state-of-the-art performance across both automated metrics and human evaluations, markedly advancing the domains of accident analysis and prevention. Project resources are available at https://an-answer-tree.github.io