In this paper, we propose a novel zero-shot compositional video understanding method inspired by how young children efficiently learn new concepts and flexibly expand their existing knowledge framework. While recent large-scale visual language models (VLMs) have achieved remarkable advancements and demonstrated impressive performance improvements across various tasks, they require massive amounts of data and computational resources. However, despite their strong benchmark performance, they often fail to solve simple zero-shot composition tasks. Moreover, VLMs designed for video data demand even greater computational resources. We introduce a new video representation learning method inspired by human compositional learning to address these challenges. Specifically, we demonstrate that achieving zero-shot compositional learning requires effective representation learning that disentangles given data into meaningful semantic units. We propose a novel method that learns such disentangled representations based on an information-theoretic measure. By optimizing coding rate reduction, we successfully learn spatio-temporally disentangled features from videos, one of the most challenging data. Our approach significantly enhances compositional generalizability, demonstrating its effectiveness in zero-shot learning scenarios.
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
We introduce Hierarchical Streaming Video Understanding, a task that combines online temporal action localization with free-form description generation. Given the scarcity of datasets with hierarchical and fine-grained temporal annotations, we demonstrate that LLMs can effectively group atomic actions into higher-level events, enriching existing datasets.We then propose OpenHOUSE (Open-ended Hierarchical Online Understanding System for Events), which extends streaming action perception beyond action classification. OpenHOUSE features a specialized streaming module that accurately detects boundaries between closely adjacent actions, nearly doubling the performance of direct extensions of existing methods.We envision the future of streaming action perception in the integration of powerful generative models, with OpenHOUSE representing a key step in that direction.
Video Large Language Models (VideoLLMs) have recently demonstrated remarkable progress in general video understanding. However, existing models primarily focus on high-level comprehension and are limited to text-only responses, restricting the flexibility for object-centric, multi-round interactions. In this paper, we make three contributions:(i) we address these limitations by introducing a VideoLLM, termed as **RGA3**, capable of performing both object referring and grounding for video reasoning tasks in a multi-round conversational manner, i.e., allowing users to iteratively interact with videos using both textual and visual queries; (ii) we propose **STOM** (Spatial-Temporal Overlay Module), a novel approach that allows arbitrary visual prompts to be processed at any timestamp within a video;(iii) we present **VideoInfer**, a manually curated object-centric video instruction dataset featuring question-answering pairs that require reasoning. We conduct comprehensive experiments on VideoInfer and other existing benchmarks across video question answering and referring video object segmentation. The results on 12 benchmarks spanning 6 tasks show that RGA3 consistently outperforms baseline models in both video question answering and segmentation, underscoring its robustness in multimodal, object-centric video and image understanding. The code, dataset, and web demo will be publicly released.
Autonomous driving (AD) systems are becoming increasingly capable of handling complex tasks, mainly due to recent advances in deep learning and AI. As interactions between autonomous systems and humans increase, the interpretability of decision-making processes in driving systems becomes increasingly crucial for ensuring safe driving operations. Successful human-machine interaction requires understanding the underlying representations of the environment and the driving task, which remains a significant challenge in deep learning-based systems. To address this, we introduce the task of interpretability in maneuver prediction before they occur for driver safety, i.e., driver intent prediction (DIP), which plays a critical role in AD systems. To foster research in interpretable DIP, we curate the eXplainable Driving Action Anticipation Dataset (DAAD-X), a new multimodal, ego-centric video dataset to provide hierarchical, high-level textual explanations as causal reasoning for the driver's decisions. These explanations are derived from both the driver's eye-gaze and the ego-vehicle's perspective. Next, we propose Video Concept Bottleneck Model (VCBM), a framework that generates spatio-temporally coherent explanations inherently, without relying on post-hoc techniques. Finally, through extensive evaluations of the proposed VCBM on the DAAD-X dataset, we demonstrate that transformer-based models exhibit greater interpretability than conventional CNN-based models. Additionally, we introduce a multilabel t-SNE visualization technique to illustrate the disentanglement and causal correlation among multiple explanations.
The rise of Large Vision-Language Models (LVLMs) has significantly advanced video understanding. However, efficiently processing long videos remains a challenge due to the "Sampling Dilemma": low-density sampling risks missing critical information, while high-density sampling introduces redundancy. To address this issue, we introduce LSDBench, the first benchmark designed to evaluate LVLMs on long-video tasks by constructing high Necessary Sampling Density (NSD) questions--where NSD represents the minimum sampling density required to accurately answer a given question. LSDBench focuses on dense, short-duration actions to rigorously assess the sampling strategies employed by LVLMs. To tackle the challenges posed by high-NSD questions, we propose a novel Reasoning-Driven Hierarchical Sampling (RHS) framework, which combines global localization of question-relevant cues with local dense sampling for precise inference. Additionally, we develop a lightweight Semantic-Guided Frame Selector to prioritize informative frames, enabling RHS to achieve comparable or superior performance with significantly fewer sampled frames. Together, our LSDBench and RHS framework address the unique challenges of high-NSD long-video tasks, setting a new standard for evaluating and improving LVLMs in this domain.
Composed video retrieval is a challenging task that strives to retrieve a target video based on a query video and a textual description detailing specific modifications. Standard retrieval frameworks typically struggle to handle the complexity of fine-grained compositional queries and variations in temporal understanding limiting their retrieval ability in the fine-grained setting. To address this issue, we introduce a novel dataset that captures both fine-grained and composed actions across diverse video segments, enabling more detailed compositional changes in retrieved video content.The proposed dataset, named Dense-WebVid-CoVR, consists of 1.6 million samples with dense modification text that is around seven times more than its existing counterpart. We further develop a new model that integrates visual and textual information through Cross-Attention (CA) fusion using grounded text encoder, enabling precise alignment between dense query modifications and target videos. The proposed model achieves state-of-the-art results surpassing existing methods on all metrics. Notably, it achieves 71.3% Recall@1 in visual+text setting and outperforms the state-of-the-art by 3.4%, highlighting its efficacy in terms of leveraging detailed video descriptions and dense modification texts. Our proposed dataset, code, and model will be publicly released.
Boosted by Multi-modal Large Language Models (MLLMs), text-guided universal segmentation models for the image and video domains have made rapid progress recently. However, these methods are often developed separately for specific domains, overlooking the similarities in task settings and solutions across these two areas. In this paper, we define the union of referring segmentation and reasoning segmentation at both the image and video levels as Instructed Visual Segmentation (IVS). Correspondingly, we propose InstructSeg, an end-to-end segmentation pipeline equipped with MLLMs for IVS. Specifically, we employ an object-aware video perceiver to extract temporal and object information from reference frames, facilitating comprehensive video understanding. Additionally, we introduce vision-guided multi-granularity text fusion to better integrate global and detailed text information with fine-grained visual guidance. By leveraging multi-task and end-to-end training, InstructSeg demonstrates superior performance across diverse image and video segmentation tasks, surpassing both segmentation specialists and MLLM-based methods with a single model.
Video understanding has made huge strides in recent years, relying largely on the power of transformers. As this architecture is notoriously expensive and video data is highly redundant, research into improving efficiency has become particularly relevant. Some creative solutions include token selection and merging. While most methods succeed in reducing the cost of the model and maintaining accuracy, an interesting pattern arises: most methods do not outperform the baseline of randomly discarding tokens. In this paper we take a closer look at this phenomenon and observe 5 principles of the nature of visual tokens. For example, we observe that the value of tokens follows a clear Pareto-distribution where most tokens have remarkably low value, and just a few carry most of the perceptual information. We build on these and further insights to propose a lightweight video model, LITE, that can select a small number of tokens effectively, outperforming state-of-the-art and existing baselines across datasets (Kinetics-400 and Something-Something-V2) in the challenging trade-off of computation (GFLOPs) vs accuracy. Experiments also show that LITE generalizes across datasets and even other tasks without the need for retraining. The code is released at https://github.com/maggieHao/Efficient-LITE.
Beyond Brain Decoding: Visual-Semantic Reconstructions to Mental Creation Extension Based on fMRI
PDF ↗Decoding visual information from fMRI signals is an important pathway to understand how the brain represents the world, and is a cutting-edge field of artificial general intelligence. Decoding fMRI should not be limited to reconstructing visual stimuli, but also further transforming them into descriptions, creating actions, and even generating unseen content. We purposefully propose a novel and efficient brain multimodal architecture, NeuroCreat, which combines the powerful visual and textual abilities of LLM to capture fine-grained semantic information from fMRI, transformed it into embodied implementation of different neural representations. Specifically, we innovatively designed a brain expert adaption (BEA) module, effectively capturing commonalities and individual differences among subjects through the collaborative learning of shared/routed experts. Inspired by human visual working memory, we extracted "creation" information from higher visual cortex for idea generation. We further constructed a prompt variant alignment module, seamlessly integrates fMRI-visual-semantic-creation into LLM to achieve flexible incorporation of different semantics in the decoding of neural representations. Experiments on different fMRI datasets show that NeuroCreat achieves SOTA performance on multiple brain decoding tasks. More importantly, we have innovatively achieved few-shot brain video creation, which opens up a new direction for demonstrating the brain's imaginative ability.
Recent approaches to 3D reconstruction in dynamic scenes primarily rely on the integration of separate geometry estimation and matching modules, where the latter plays a critical role in distinguishing dynamic regions and mitigating the interference caused by moving objects. Furthermore, the matching module explicitly models object motion, enabling the tracking of specific targets and advancing motion understanding in complex scenarios. Recently, the proposed representation of pointmap in DUSt3R suggests a potential solution to unify both geometry estimation and matching in 3D space, effectively reducing computational overhead by eliminating the need for redundant auxiliary modules. However, it still struggles with ambiguous correspondences in dynamic regions, which limits reconstruction performance in such scenarios. In this work, we present POMATO, a unified framework for dynamic 3D reconstruction by marrying POintmap MAtching with Temporal mOtion. Specifically, our method first learns an explicit matching relationship by mapping RGB pixels across different views to 3D pointmaps within a unified coordinate system. Furthermore, we introduce a temporal motion module for dynamic motions that ensures scale consistency across different frames and enhances performance in 3D reconstruction tasks requiring both precise geometry and reliable matching, most notably 3D point tracking. We show the effectiveness of our proposed POMATO by demonstrating the remarkable performance across multiple downstream tasks, including video depth estimation, 3D point tracking, and pose estimation. Code and models are publicly available at https://github.com/wyddmw/POMATO.
Understanding continuous video streams plays a fundamental role in real-time applications, including embodied AI and autonomous driving. Unlike offline video processing, streaming video understanding requires the ability to process video streams frame by frame, preserve historical information, and make low-latency decisions. To address these challenges, our main contributions are three-fold. (i) we develop a novel streaming video backbone, termed as StreamFormer, by incorporating causal temporal attention into a pre-trained vision transformer. This enables efficient streaming video processing while maintaining image representation capability. (ii) to train StreamFormer, we propose to unify diverse spatial-temporal video understanding tasks within a multitask visual-language alignment framework. Hence, StreamFormer learns global semantics, temporal dynamics, and fine-grained spatial relationships simultaneously. (iii) we conduct extensive experiments for online action detection, online video instance segmentation, and video question answering. StreamFormer achieves competitive performance while maintaining efficiency, demonstrating its potential for real-time applications.
Training robot policies within a learned world model is trending due to the inefficiency of real-world interactions. The established image-based world models and policies have shown prior success, but lack robust geometric information that requires consistent spatial and physical understanding of the three-dimensional world, even pre-trained on internet-scale video sources. To this end, we propose a novel branch of world model named Gaussian World Model (GWM) for robotic manipulation, which reconstructs the future state by inferring the propagation of Gaussian primitives under the effect of robot actions. At its core is a latent Diffusion Transformer (DiT) combined with a 3D variational autoencoder, enabling fine-grained scene-level future state reconstruction with Gaussian Splatting. GWM can not only enhance the visual representation for imitation learning agent by self-supervised future prediction training, but can serve as a neural simulator that supports model-based reinforcement learning. Both simulated and real-world experiments depict that GWM can precisely predict future scenes conditioned on diverse robot actions, and can be further utilized to train policies that outperform the state-of-the-art by impressive margins, showcasing the initial data scaling potential of 3D world model.
In video Multimodal Large Language Models (video MLLMs), the visual encapsulation process plays a pivotal role in converting video contents into representative tokens for LLM input. While linear projectors are widely employed for encapsulation, they introduce semantic indistinctness and temporal incoherence when applied to videos. Conversely, the structure of resamplers shows promise in tackling these challenges, but an effective solution remains unexplored. Drawing inspiration from resampler structures, we introduce DisCo, a novel visual encapsulation method designed to yield semantically distinct and temporally coherent visual tokens for video MLLMs. DisCo integrates two key components: (1) A Visual Concept Discriminator (VCD) module, assigning unique semantics for visual tokens by associating them in pair with discriminative concepts in the video. (2) A Temporal Focus Calibrator (TFC) module, ensuring consistent temporal focus of visual tokens to video elements across every video frame. Through extensive experiments on multiple video MLLM frameworks, we demonstrate that DisCo remarkably outperforms previous state-of-the-art methods across a variety of video understanding benchmarks, while also achieving higher token efficiency thanks to the reduction of semantic indistinctness.
FrameFusion: Combining Similarity and Importance for Video Token Reduction on Large Vision Language Models
PDF ↗The increasing demand to process long and high-resolution videos significantly burdens Large Vision-Language Models (LVLMs) due to the enormous number of visual tokens. Existing token reduction methods primarily prune tokens based on importance metrics, such as cumulative attention scores. However, even important tokens may exhibit high redundancy caused by similarity among adjacent video frames and repetitive visual elements. To address this limitation, we propose FrameFusion, a novel token reduction approach integrating similarity-based merging with importance-based pruning. We conduct a thorough study on token similarity characteristics, revealing three key insights: (1) spatially corresponding visual tokens between adjacent frames have higher cosine similarities compared to other token pairs; (2) high token similarities prominently decrease in deeper model layers; and (3) token similarity rankings are highly consistent across different layers. Guided by these observations, FrameFusion computes token similarities exclusively between corresponding visual tokens from adjacent frames, applies token merging at initial successive layers followed by pruning in deeper layers, and adopts a cascaded merging strategy to further enhance efficiency. We evaluate FrameFusion comprehensively across six diverse LVLMs, ranging from 2B to 72B parameters, using five video benchmarks encompassing video retrieval, question-answering, and spatial-temporal understanding tasks. Experiments show that FrameFusion reduces visual tokens by 70%, achieving 1.6-3.6x end-to-end speedups, with an average performance impact of less than 3%. Our code is available at: https://github.com/thu-nics/FrameFusion.
Large language models (LLMs) have taken a great step towards AGI. Meanwhile, an increasing number of domain-specific problems such as math and programming boost these general-purpose models to continuously evolve via learning deeper expertise. Now is thus the time further to extend the diversity of specialized applications for knowledgeable LLMs, though collecting high quality data with unexpected and informative tasks is challenging. In this paper, we propose to use advertisement (ad) videos as a challenging test-bed to probe the ability of LLMs in perceiving beyond the objective physical content of common visual domain. Our motivation is to take full advantage of the clue-rich and information-dense ad videos' traits, e.g., marketing logic, persuasive strategies, and audience engagement. Our contribution is three-fold: (1) To our knowledge, this is the first attempt to use ad videos with well-designed tasks to evaluate LLMs. We contribute AdsQA, a challenging ad Video QA benchmark derived from 1,544 ad videos with 10,962 clips, totaling 22.7 hours, providing 5 challenging tasks. (2) We propose ReAd-R, a Deepseek-R1 styled RL model that reflects on questions, and generates answers via reward-driven optimization. (3) We benchmark 14 top-tier LLMs on AdsQA, and our ReAd-R achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities by a clear margin.
Vision-Language Models (VLMs) have shown remarkable capabilities across diverse visual tasks, including image recognition, video understanding, and Visual Question Answering (VQA) when explicitly trained for these tasks. Despite these advances, we find that present-day VLMs (including the proprietary GPT-4o) lack a fundamental cognitive ability: learning to localize specific objects in a scene by taking into account the context.In this work, we focus on the task of few-shot personalized localization, where a model is given a small set of annotated images (in-context examples) -- each with a category label and bounding box -- and is tasked with localizing the same object type in a query image. Personalized localization can be particularly important in cases of ambiguity of several related objects that can respond to a text or an object that is hard to describe with words.To provoke personalized localization abilities in models, we present a data-centric solution that fine-tunes them using carefully curated data from video object tracking datasets. By leveraging sequences of frames tracking the same object across multiple shots, we simulate instruction-tuning dialogues that promote context awareness. To reinforce this, we introduce a novel regularization technique that replaces object labels with pseudo-names, ensuring the model relies on visual context rather than prior knowledge. Our method significantly enhances few-shot localization performance of recent VLMs ranging from 7B to 72B in size, without sacrificing generalization, as demonstrated on several benchmarks tailored towards evaluating personalized localization abilities. This work is the first to explore and benchmark personalized few-shot localization for VLMs -- exposing critical weaknesses in present-day VLMs, and lays a foundation for future research in context-driven vision-language applications.
X2I: Seamless Integration of Multimodal Understanding into Diffusion Transformer via Attention Distillation
PDF ↗Text-to-image (T2I) models are well known for their ability to produce highly realistic images, while multimodal large language models (MLLMs) are renowned for their proficiency in understanding and integrating multiple modalities. However, currently there is no straightforward and efficient framework to transfer the multimodal comprehension abilities of MLLMs to T2I models to enable them to understand multimodal inputs. In this paper, we propose the X2I framework, which endows Diffusion Transformer (DiT) models with the capability to comprehend various modalities, including multilingual text, screenshot documents, images, videos, and audio. X2I is trained using merely 100K English corpus with 160 GPU hours. Building on the DiT teacher model, we adopt an innovative distillation method to extract the inference capabilities of the teacher model and design a lightweight AlignNet structure to serve as an intermediate bridge. Compared to the teacher model, X2I shows a decrease in performance degradation of less than 1% while gaining various multimodal understanding abilities. Furthermore, it is applicable for LoRA training in the context of image-text to image generation, filling a void in the industry in this area. We further design a simple LightControl to enhance the fidelity of instructional image editing. Finally, extensive experiments demonstrate the effectiveness, efficiency, multifunctionality, and transferability of our X2I.
3D spatial reasoning is the ability to analyze and interpret the positions, orientations, and spatial relationships of objects within the 3D space. This allows models to develop a comprehensive understanding of the 3D scene, enabling their applicability to a broader range of applications, such as autonomous navigation, robotics, and AR/VR. Despite the remarkable improvements achieved by large multi-modal models (LMMs) in a wide range of image and video understanding tasks, their abilities to perform 3D spatial reasoning are less studied. In this work we present the first comprehensive 3D spatial reasoning benchmark, 3DSRBench, with 3,000 annotated image question answering triplets from 12 question types. We balance the data distribution by collecting complimentary images that lead to opposite answers given the same question. We also adopt a novel FlipEval for robust evaluation of 3D spatial reasoning capabilities. Moreover, to study the robustness of 3D spatial reasoning w.r.t. camera 3D viewpoints, our 3DSRBench involves two subsets with 3D spatial reasoning questions on images from the same scene with common and uncommon viewpoints. We benchmark a wide range of open-sourced and proprietary LMMs, revealing their limitations in different types of 3D awareness, i.e., height, orientation, location, and multi-object reasoning. Our 3DSRBench also allows us to study the design choices of developing LMMs with strong 3D reasoning capabilities, such as the vision encoders, connectors, and training recipes.
Structured Policy Optimization: Enhance Large Vision-Language Model via Self-referenced Dialogue
PDF ↗Preference optimization algorithms typically enhance LLM response quality by leveraging human feedback on multiple answers given a fixed instruction. However, these methods often lack capturing the dynamic nature of conversational exchanges. For large vision-language models (LVLMs), direct preference optimization (DPO) can over-emphasize linguistic nuances while overlooking visual context. To address this challenge, we introduce structured policy optimization (SPO) -- a novel preference optimization method that simultaneously aligns preference instructions, responses, and dialogue interactions to improve multi-modal understanding and reasoning capabilities. The efficacy of SPO is attributed to one key design:treating the questioning and answering as a sequential action and binding them through a trajectory reward. This reward formulation better aligns with real-world dialogue studies and eliminates the need for fixed instructions. We evaluate our models on interleaved benchmarks, including image, multi-image, and video-based understanding and reasoning tasks. Experimental results show that the proposed SPO fine-tuning LVLM with multi-modal preference data can align with human preference more efficiently than DPO.
The challenge of long video understanding lies in its high computational complexity and prohibitive memory cost, since the memory and computation required by transformer-based LLMs scale quadratically with input sequence length. We propose AuroraLong to address this challenge by replacing the LLM component in MLLMs with a linear RNN language model that handles input sequence of arbitrary length with constant-size hidden states. To further increase throughput and efficiency, we combine visual token merge with linear RNN models by reordering the visual tokens by their sizes in ascending order. Despite having only 2B parameters and being trained exclusively on public data, AuroraLong achieves performance comparable to Transformer-based models of similar size trained on private datasets across multiple video benchmarks. This demonstrates the potential of efficient, linear RNNs to democratize long video understanding by lowering its computational entry barrier. To our best knowledge, we are the first to use a linear RNN based LLM backbone in a LLaVA-like model for open-ended video understanding.