Effective video tokenization is critical for scaling transformer models for long videos. Current approaches tokenize videos using space-time patches, leading to excessive tokens and computational inefficiencies. The best token reduction strategies degrade performance and barely reduce the number of tokens when the camera moves. We introduce grounded video tokenization, a paradigm that organizes tokens based on panoptic sub-object trajectories rather than fixed patches. Our method aligns with fundamental perceptual principles, ensuring that tokenization reflects scene complexity rather than video duration. We propose TrajViT, a video encoder that extracts object trajectories and converts them into semantically meaningful tokens, significantly reducing redundancy while maintaining temporal coherence. Trained with contrastive learning, TrajViT significantly outperforms space-time ViT (ViT3D) across multiple video understanding benchmarks, e.g., TrajViT outperforms ViT3D by a large margin of 6% top-5 recall in average at video-text retrieval task with 10x token deduction. We also show TrajViT as a stronger model than ViT3D for being the video encoder for modern VideoLLM, obtaining an average of 5.2% performance improvement across 6 VideoQA benchmarks while having 4x faster training time and 20x less inference FLOPs. TrajViT is the first efficient encoder to consistently outperform ViT3D across diverse video analysis tasks, making it a robust and scalable solution.
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
LLaFEA: Frame-Event Complementary Fusion for Fine-Grained Spatiotemporal Understanding in LMMs
PDF ↗Large multimodal models (LMMs) excel in scene understanding but struggle with fine-grained spatiotemporal reasoning due to weak alignment between linguistic and visual representations. Existing methods map textual positions and durations into the visual space encoded from frame-based videos, but suffer from temporal sparsity that limits language-vision temporal coordination. To address this issue, we introduce LLaFEA (Large Language and Frame-Event Assistant) to leverage event cameras for temporally dense perception and frame-event fusion. Our approach employs a cross-attention mechanism to integrate complementary spatial and temporal features, followed by self-attention matching for global spatio-temporal associations. We further embed textual position and duration tokens into the fused visual space to enhance fine-grained alignment. This unified framework ensures robust spatio-temporal coordinate alignment, enabling LMMs to interpret scenes at any position and any time. In addition, we construct a dataset of real-world frames-events with coordinate instructions and conduct extensive experiments to validate the effectiveness of the proposed method.
Sign language translation (SLT) is a challenging task that involves translating sign language images into spoken language. For SLT models to perform this task successfully, they must bridge the modality gap and identify subtle variations in sign language components to understand their meanings accurately. To address these challenges, we propose a novel gloss-free SLT framework called Multimodal Sign Language Translation (MMSLT), which leverages the representational capabilities of off-the-shelf multimodal large language models (MLLMs). Specifically, we use MLLMs to generate detailed textual descriptions of sign language components. Then, through our proposed multimodal-language pre-training module, we integrate these description features with sign video features to align them within the spoken sentence space. Our approach achieves state-of-the-art performance on benchmark datasets PHOENIX14T and CSL-Daily, highlighting the potential of MLLMs to be utilized effectively in SLT. Code is available at https://github.com/helpmeIamnewbie/MMSLT.
Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2x speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3x speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The code is available at https://github.com/HYUNJS/STTM.
Recent progress in multimodal large language models has markedly enhanced the understanding of short videos (typically under one minute), and several evaluation datasets have emerged accordingly. However, these advancements fall short of meeting the demands of real-world applications such as embodied intelligence for long-term decision-making, in-depth movie reviews and discussions, and live sports commentary, all of which require comprehension of long videos spanning several hours. To address this gap, we introduce LVBench, a benchmark specifically designed for long video understanding. Our dataset comprises publicly sourced videos and encompasses a diverse set of tasks aimed at long video comprehension and information extraction. LVBench is designed to challenge multimodal models to demonstrate long-term memory and extended comprehension capabilities. Our extensive evaluations reveal that current multimodal models still underperform on these demanding long video understanding tasks. Through LVBench, we aim to spur the development of more advanced models capable of tackling the complexities of long video comprehension.
We introduce MOSCATO: a new benchmark for predicting the evolving states of multiple objects through long procedural videos with multiple actions. While prior work in object state prediction has typically focused on a single object undergoing one or a few state changes, real-world tasks require tracking many objects whose states evolve over multiple actions. Given the high cost of gathering framewise object-state labels for many videos, we develop a weakly-supervised multiple object state prediction framework, which only uses action labels during training. Specifically, we propose a novel Pseudo-Label Acquisition (PLA) pipeline that integrates large language models, vision-language models, and action segment annotations to generate fine-grained, per-frame object-state pseudo-labels for training a Multiple Object State Prediction (MOSP) network. We further devise a State-Action Interaction (SAI) module that explicitly models the correlations between actions and object states, thereby improving MOSP. To facilitate comprehensive evaluation, we create the MOSCATO benchmark by augmenting three egocentric video datasets with framewise object-state annotations. Experiments show that our multi-stage pseudo-labeling approach and SAI module significantly boost performance over zero-shot VLM baselines and naive extensions of existing methods, underscoring the importance of holistic action-state modeling for fine-grained procedural video understanding.
High-quality image-text data is critical in enhancing Vision-Language Models (VLMs), but traditional image-based pretraining approaches face limitations. These methods are resource-intensive, relying on curated, high-quality interleaved data that is costly and challenging to collect at scale. Additionally, while such datasets improve static image-text understanding, they fail to develop the temporal and motion comprehension needed for video understanding. To address these gaps, we propose incorporating video pretraining into VLMs to improve the model's ability to capture temporal dynamics and general visual perception, which requires reconciling spatial redundancy with strict temporal causality. Therefore, we propose Causal Hierarchical Aggregation to separate computation-heavy spatial encoding from lightweight temporal propagation and construct hierarchical receptive fields at varying granularities. As we scale video context to more than 100 billion tokens, our method excels in high throughput and state-of-the-art performances on both Image and Video understanding, as shown in Figure 1, providing a scalable solution to enhance multimodal learning in dynamic contexts.
HumanSAM: Classifying Human-centric Forgery Videos in Human Spatial, Appearance, and Motion Anomaly
PDF ↗Numerous synthesized videos from generative models, especially human-centric ones that simulate realistic human actions, pose significant threats to human information security and authenticity. While progress has been made in binary forgery video detection, the lack of fine-grained understanding of forgery types raises concerns regarding both reliability and interpretability, which are critical for real-world applications. To address this limitation, we propose HumanSAM, a new framework that builds upon the fundamental challenges of video generation models. Specifically, HumanSAM aims to classify human-centric forgeries into three distinct types of artifacts commonly observed in generated content: spatial, appearance, and motion anomaly.To better capture the features of geometry, semantics and spatiotemporal consistency, we propose to generate the human forgery representation by fusing two branches of video understanding and spatial depth. We also adopt a rank-based confidence enhancement strategy during the training process to learn more robust representation by introducing three prior scores. For training and evaluation, we construct the first public benchmark, the Human-centric Forgery Video (HFV) dataset, with all types of forgeries carefully annotated semi-automatically. In our experiments, HumanSAM yields promising results in comparison with state-of-the-art methods, both in binary and multi-class forgery classification.
Benefiting from the advances in large language models and cross-modal alignment, existing multimodal large language models have achieved prominent performance in image and short video understanding. However, the understanding of long videos is still challenging, as their long-context nature results in significant computational and memory overhead. Most existing work treats long videos in the same way as short videos, which is inefficient for real-world applications and hard to generalize to even longer videos. To address these issues, we propose Flash-VStream, an efficient video language model capable of processing extremely long videos and responding to user queries in real time. Particularly, we design a Flash Memory module, containing a low-capacity context memory to aggregate long-context temporal information and model the distribution of information density, and a high-capacity augmentation memory to retrieve detailed spatial information based on this distribution. Compared to existing models, Flash-VStream achieves significant reductions in inference latency. Extensive experiments on long video benchmarks and comprehensive video benchmarks, i.e., EgoSchema, MLVU, LVBench, MVBench and Video-MME, demonstrate the state-of-the-art performance and outstanding efficiency of our method. Code is available at https://github.com/IVGSZ/Flash-VStream.
Long video understanding poses unique challenges due to its temporal complexity and low information density. Recent works address this task by sampling numerous frames or incorporating auxiliary tools using LLMs, both of which result in high computational costs. In this work, we introduce a curiosity-driven video agent with self-exploration capability, dubbed as "VCA". Built upon VLMs, VCA autonomously navigates video segments and efficiently builds a comprehensive understanding of complex video sequences.Instead of directly sampling frames, VCA employs a tree-search structure to explore video segments and collect frames. Rather than relying on external feedback or reward, VCA leverages VLM's self-generated intrinsic reward to guide its exploration, enabling it to capture the most crucial information for reasoning. Experimental results on multiple long video benchmarks demonstrate our approach's superior effectiveness and efficiency.
Recently, text-driven diffusion models have significantly promoted the development of video editing. However, there still remain two practical challenges: (1) existing text-to-video editing methods struggle to understand negative text prompt, resulting in ineffective suppression of undesirable content in edited video; (2) these methods are difficult to maintain the temporal consistency of the edited video, leading to inter-frame flickering. To address the above challenges, we propose SUV, a novel semantic modulation method based on text embeddings to suppress undesired content in the edited video. Specifically, on the one hand, we discover that the end embeddings (EE) contain substantial coupled positive and negative embeddings, which is the primary reason for the appearance of undesirable content in the edited video. Based on this discovery, we advocate decoupling the negative embeddings from the EE by employing singular value decomposition and propose an exponential suppression operator to decrease the singular values of negative embeddings, thereby restraining the effect of negative embeddings on the edited video content. Subsequently, two constraints are designed to further suppress negative content while keep positive content unchanged via pushing negative embeddings apart and pulling positive embeddings closer. On the other hand, to boost the temporal consistency of edited video, we devise a fuzzy feature selection strategy to fuse similar features in different frame for avoiding inter-frame flickering. Benefiting from the above elaborate designs, our method not only effectively suppresses undesired content of video, but also maintains inter-frame consistency. Extensive experiments demonstrate that our SUV significantly improves edit accuracy and temporal consistency of edited videos compared to existing methods.
Vision language models (VLMs) have shown remarkable capabilities in integrating linguistic and visual reasoning but remain fundamentally limited in understanding dynamic spatiotemporal interactions. Humans effortlessly track and reason about object movements, rotations, and perspective shifts--abilities essential for robust dynamic real-world understanding yet notably lacking in current VLMs. In this paper, we introduce VLM4D, the first benchmark specifically designed to evaluate the spatiotemporal reasoning capabilities of VLMs. Our benchmark comprises diverse real-world and synthetic videos accompanied by carefully curated question-answer pairs emphasizing translational and rotational motions, perspective awareness, and motion continuity. Through comprehensive evaluations of state-of-the-art open and closed-source VLMs, we identify significant performance gaps compared to human baselines, highlighting fundamental deficiencies in existing models. Extensive analysis reveals that VLMs struggle particularly with integrating multiple visual cues and maintaining temporal coherence. We further explore promising directions, such as leveraging 4D feature field reconstruction and targeted spatiotemporal supervised fine-tuning, demonstrating their effectiveness in enhancing spatiotemporal comprehension. Our work aims to encourage deeper exploration into improving VLMs' spatial and temporal grounding, paving the way towards more capable and reliable visual intelligence for dynamic environments.
Can a model distinguish between the sound of a spoon hitting a hardwood floor versus a carpeted one? Everyday object interactions produce sounds unique to the objects involved. We introduce the sounding object detection task to evaluate a model's ability to link these sounds to the objects directly involved. Inspired by human perception, our multimodal object-aware framework learns from in-the-wild egocentric videos. To encourage an object-centric approach, we first develop an automatic pipeline to compute segmentation masks of the objects involved to guide the model's focus during training towards the most informative regions of the interaction. A slot attention visual encoder is used to further enforce an object prior. We demonstrate state of the art performance on our new task along with existing multimodal action understanding tasks.
Egocentric videos provide valuable insights into human interactions with the physical world, which has sparked growing interest in the computer vision and robotics communities. A critical challenge in fully understanding the geometry and dynamics of egocentric videos is dense scene reconstruction. However, the lack of high-quality labeled datasets in this field has hindered the effectiveness of current supervised learning methods. In this work, we aim to address this issue by exploring an self-supervised dynamic scene reconstruction approach. We introduce **EgoMono4D**, a novel model that unifies the estimation of multiple variables necessary for *Ego*centric *Mono*cular *4D* reconstruction, including camera intrinsic, camera poses, and video depth, all within a fast feed-forward framework. Starting from pretrained single-frame depth and intrinsic estimation model, we extend it with camera poses estimation and align multi-frame results on large-scale unlabeled egocentric videos. We evaluate EgoMono4D in both in-domain and zero-shot generalization settings, achieving superior performance in dense pointclouds sequence reconstruction compared to all baselines. EgoMono4D represents the first attempt to apply self-supervised learning for pointclouds sequence reconstruction to the label-scarce egocentric field, enabling fast, dense, and generalizable reconstruction. The interactable visualization, code and trained models are released https://egomono4d.github.io/.
Advertisement videos serve as a rich and valuable source of purpose-driven information, encompassing high-quality visual, textual, and contextual cues designed to engage viewers. They are often more complex than general videos of similar duration due to their structured narratives and rapid scene transitions, posing significant challenges to multi-modal large language models (MLLMs). In this work, we introduce VideoAds, the first dataset tailored for benchmarking the performance of MLLMs on advertisement videos. VideoAds comprises well-curated advertisement videos with complex temporal structures, accompanied by manually annotated diverse questions across three core tasks: visual finding, video summary, and visual reasoning. We propose a quantitative measure to compare VideoAds against existing benchmarks in terms of video complexity. Through extensive experiments, we find that Qwen2.5-VL-72B, an open-source MLLM, achieves 73.35% accuracy on VideoAds, outperforms GPT-4o (66.82%) and Gemini-1.5 Pro (69.66%); the two proprietary models especially fall behind the open-source model in video summarization and reasoning, but perform the best in visual finding. Gemini-2.5 Pro leads with an accuracy of 80.04%. Notably, human experts easily achieve a remarkable accuracy of 94.27%. These results underscore the necessity of advancing MLLMs' temporal modeling capabilities and highlight VideoAds as a potentially pivotal benchmark for future research in understanding video that requires high FPS sampling. The dataset and evaluation code will be publicly available at https://videoadsbenchmark.netlify.app.
Recent advances in Large Multi-modal Models (LMMs) are primarily focused on offline video understanding. Instead, streaming video understanding poses great challenges to recent models due to its time-sensitive, omni-modal and interactive characteristics. In this work, we aim to extend the streaming video understanding from a new perspective and propose a novel task named Visual Instruction Feedback in which models should be aware of visual contents and learn to extract instructions from them. For example, when users wave their hands to agents, agents should recognize the gesture and start conversations with welcome information. Thus, following instructions in visual modality greatly enhances user-agent interactions. To facilitate research, we define seven key subtasks highly relevant to visual modality and collect the ViSpeak-Instruct dataset for training and the ViSpeak-Bench for evaluation. Further, we propose the ViSpeak model, which is a SOTA streaming video understanding LMM with GPT-4o-level performance on various streaming video understanding benchmarks. After finetuning on our ViSpeak-Instruct dataset, ViSpeak is equipped with basic visual instruction feedback ability, serving as a solid baseline for future research.
Multimodal large language models (MLLMs) are broadly empowering various fields. Despite their advancements, the explainability of MLLMs remains less explored, hindering deeper understanding, model credibility, and effective visualization. Unlike conventional vision models (e.g., CNNs, ViTs, CLIP) that produce a single output, MLLMs generate sequences of tokens progressively, where each generated token depends on the previous context. Therefore, earlier context tokens can introduce redundant activations that interfere with the explanation of later tokens beyond their original information. Existing studies often overlook this issue, but our observations reveal that these redundant correlations can significantly hurt the reliability of explanations. To address this, we propose an estimated causal inference method to mitigate the interference of context to achieve high-quality MLLM explanation, with a novel rank Gaussian filter to further reduce activation noises. We term this method Token Activation Map (TAM) to highlight the consideration of interactions between tokens. TAM also indicates that it excels at explaining multiple tokens of MLLM, which is different from the Class Activation Map (CAM) for a single prediction. Our TAM method significantly outperforms existing SoTA methods, showcasing high-quality visualization results that can be utilized for various scenarios, such as object localization, failure case analysis, video visualization, MLLMs visual comparison, and model understanding (e.g., color, shape, action, location, visual reasoning, multi-turn conversation, etc). The code is available at github.com/xmed-lab/TAM.
Video understanding requires the extraction of rich spatio-temporal representations, achieved by transformer models through self-attention. Unfortunately, self-attention poses a computational burden. In NLP, Mamba has surfaced as an efficient alternative for transformers. However, Mamba's successes do not trivially extend to vision tasks, including those in video analysis. In this paper, we theoretically analyze the differences between self-attention and Mamba. We identify two limitations in Mamba's token processing: historical decay and element contradiction. We propose VideoMambaPro (VMP) that addresses these limitations by adding masked backward computation and elemental residual connections to a VideoMamba backbone. VideoMambaPro models surpass VideoMamba by 1.6-3.0% and 1.1-1.9% top-1 on Kinetics-400 and Something-Something V2, respectively. Even without extensive pre-training, our models present an attractive and efficient alternative to current transformer models. Moreover, our two solutions are orthogonal to recent advances in Vision Mamba models, and are likely to provide further improvements in future models.
Our objective is the automatic generation of Audio Descriptions (ADs) for edited video material, such as movies and TV series. To achieve this, we propose a two-stage framework that leverages "shots" as the fundamental units of video understanding. This includes extending temporal context to neighboring shots and incorporating film grammar devices, such as shot scales and thread structures, to guide AD generation. Our method is compatible with both open-source and proprietary Visual-Language Models (VLMs), integrating expert knowledge from add-on modules without requiring additional training of the VLMs. We achieve state-of-the-art performance among all prior training-free approaches and even surpass fine-tuned methods on several benchmarks. To evaluate the quality of predicted ADs, we introduce a new evaluation measure -- an action score -- specifically targeted to assessing this important aspect of AD. Additionally, we propose a novel evaluation protocol that treats automatic frameworks as AD generation assistants and asks them to generate multiple candidate ADs for selection.
Do we still need to represent objects explicitly in multimodal large language models (MLLMs)? To one extreme, pre-trained encoders convert images into visual tokens, with which objects and spatiotemporal relationships may be implicitly modeled. To the other extreme, image captions by themselves provide strong empirical performances for understanding tasks, despite missing fine-grained spatiotemporal information. To answer this question, we introduce ObjectMLLM, a framework capable of leveraging arbitrary computer vision algorithm to extract and integrate structured visual representation. Through extensive evaluations on six video question answering benchmarks, we confirm that explicit integration of object-centric representation remains necessary. Surprisingly, we observe that the simple approach of quantizing the continuous, structured object information and representing them as plain text performs the best, offering a data-efficient approach to integrate other visual perception modules into MLLM design. Our code and models are released at https://github.com/brown-palm/ObjectMLLM.