Understanding real-world videos with complex semantics and long temporal dependencies remains a fundamental challenge in computer vision. Recent progress in multimodal large language models (MLLMs) has demonstrated strong capabilities in vision-language tasks, while reinforcement learning tuning (RLT) has further improved their reasoning abilities. In this work, we explore RLT as a post-training strategy to enhance the video-specific reasoning capabilities of MLLMs. Built upon the Group Relative Policy Optimization (GRPO) framework, we propose a dual-reward formulation that supervises both semantic and temporal reasoning through discrete and continuous reward signals. To facilitate effective preference-based optimization, we introduce a variance-aware data selection strategy based on repeated inference to identify samples that provide informative learning signals. We evaluate our approach across eight representative video understanding tasks, including VideoQA, Temporal Video Grounding, and Grounded VideoQA. Our method consistently outperforms supervised fine-tuning and existing RLT baselines, achieving superior performance with significantly less training data. These results underscore the importance of reward design and data selection in advancing reasoning-centric video understanding with MLLMs. Notably, The initial code release (two months ago) has now been expanded with updates, including optimized reward mechanisms and additional datasets. The latest version is available at https://github.com/appletea233/Temporal-R1 .
Long Video Temporal Grounding (LVTG) aims at identifying specific moments within lengthy videos based on user-provided text queries for effective content retrieval. The approach taken by existing methods of dividing video into clips and processing each clip via a full-scale expert encoder is challenging to scale due to prohibitive computational costs of processing a large number of clips in long videos. To address this issue, we introduce DeCafNet, an approach employing ``delegate-and-conquer'' strategy to achieve computation efficiency without sacrificing grounding performance. DeCafNet introduces a sidekick encoder that performs dense feature extraction over all video clips in a resource-efficient manner, while generating a saliency map to identify the most relevant clips for full processing by the expert encoder. To effectively leverage features from sidekick and expert encoders that exist at different temporal resolutions, we introduce DeCaf-Grounder, which unifies and refines them via query-aware temporal aggregation and multi-scale temporal refinement for accurate grounding. Experiments on two LTVG benchmark datasets demonstrate that DeCafNet reduces computation by up to 47\% while still outperforming existing methods, establishing a new state-of-the-art for LTVG in terms of both efficiency and performance. Our code is available at https://github.com/ZijiaLewisLu/CVPR2025-DeCafNet.
Egocentric video grounding is a crucial task for embodied intelligence applications, distinct from exocentric video moment localization. Existing methods primarily focus on the distributional differences between egocentric and exocentric videos but often neglect key characteristics of egocentric videos and the fine-grained information emphasized by question-type queries. To address these limitations, we propose OSGNet, an Object-Shot enhanced Grounding Network for egocentric video. Specifically, we extract object information from videos to enrich video representation, particularly for objects highlighted in the textual query but not directly captured in the video features. Additionally, we analyze the frequent shot movements inherent to egocentric videos, leveraging these features to extract the wearer's attention information, which enhances the model's ability to perform modality alignment. Experiments conducted on three datasets demonstrate that OSGNet achieves state-of-the-art performance, validating the effectiveness of our approach. Our code can be found at https://github.com/Yisen-Feng/OSGNet.
We introduce TimeZero, a reasoning-guided LVLM designed for the temporal video grounding (TVG) task. This task requires precisely localizing relevant video segments within long videos based on a given language query. TimeZero tackles this challenge by extending the inference process, enabling the model to reason about video-language relationships solely through reinforcement learning. To evaluate the effectiveness of TimeZero, we conduct experiments on two benchmarks, where TimeZero achieves state-of-the-art performance on Charades-STA. Code is available at https://github.com/www-Ye/TimeZero.
In this paper, we propose spatio-temporal omni-object video grounding, dubbed OmniSTVG, a new STVG task that aims at localizing spatially and temporally all targets mentioned in the textual query from videos. Compared to classic STVG locating only a single target, OmniSTVG enables localization of not only an arbitrary number of text-referred targets but also their interacting counterparts in the query from the video, making it more flexible and practical in real scenarios for comprehensive understanding. In order to facilitate exploration of OmniSTVG, we introduce BOSTVG, a large-scale benchmark dedicated to OmniSTVG. Specifically, our BOSTVG consists of 10,018 videos with 10.2M frames and covers a wide selection of 287 classes from diverse scenarios. Each sequence in BOSTVG, paired with a free-form textual query, encompasses a varying number of targets ranging from 1 to 10. To ensure high quality, each video is manually annotated with meticulous inspection and refinement. To our best knowledge, BOSTVG is to date the first and the largest benchmark for OmniSTVG. To encourage future research, we introduce a simple yet effective approach, named OmniTube, which, drawing inspiration from Transformer-based STVG methods, is specially designed for OmniSTVG and demonstrates promising results. By releasing BOSTVG, we hope to go beyond classic STVG by locating every object appearing in the query for more comprehensive understanding, opening up a new direction for STVG. Our benchmark, model, and results will be released at https://github.com/JellyYao3000/OmniSTVG.
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.
Transformer has attracted increasing interest in STVG, owing to its end-to-end pipeline and promising result. Existing Transformer-based STVG approaches often leverage a set of object queries, which are initialized simply using zeros and then gradually learn target position information via iterative interactions with multimodal features, for spatial and temporal localization. Despite simplicity, these zero object queries, due to lacking target-specific cues, are hard to learn discriminative target information from interactions with multimodal features in complicated scenarios (\e.g., with distractors or occlusion), resulting in degradation. Addressing this, we introduce a novel Target-Aware Transformer for STVG (TA-STVG), which seeks to adaptively generate object queries via exploring target-specific cues from the given video-text pair, for improving STVG. The key lies in two simple yet effective modules, comprising text-guided temporal sampling (TTS) and attribute-aware spatial activation (ASA), working in a cascade. The former focuses on selecting target-relevant temporal cues from a video utilizing holistic text information, while the latter aims at further exploiting the fine-grained visual attribute information of the object from previous target-aware temporal cues, which is applied for object query initialization. Compared to existing methods leveraging zero-initialized queries, object queries in our TA-STVG, directly generated from a given video-text pair, naturally carry target-specific cues, making them adaptive and better interact with multimodal features for learning more discriminative information to improve STVG. In our experiments on three benchmarks, TA-STVG achieves state-of-the-art performance and significantly outperforms the baseline, validating its efficacy.
Recent advancements in multimodal large language models (MLLMs) have shown promising results, yet existing approaches struggle to effectively handle both temporal and spatial localization simultaneously. This challenge stems from two key issues: first, incorporating spatial-temporal localization introduces a vast number of coordinate combinations, complicating the alignment of linguistic and visual coordinate representations; second, encoding fine-grained temporal and spatial information during video feature compression is inherently difficult. To address these issues, we propose LLaVA-ST, a MLLM for fine-grained spatial-temporal multimodal understanding. In LLaVA-ST, we propose Language-Aligned Positional Embedding, which embeds the textual coordinate special token into the visual space, simplifying the alignment of fine-grained spatial-temporal correspondences. Additionally, we design the Spatial-Temporal Packer, which decouples the feature compression of temporal and spatial resolutions into two distinct point-to-region attention processing streams. Furthermore, we propose ST-Align dataset with 4.3M training samples for fine-grained spatial-temporal multimodal understanding. With ST-align, we present a progressive training pipeline that aligns the visual and textual feature through sequential coarse-to-fine stages.Additionally, we introduce an ST-Align benchmark to evaluate spatial-temporal interleaved fine-grained understanding tasks, which include Spatial-Temporal Video Grounding (STVG) , Event Localization and Captioning (ELC) and Spatial Video Grounding (SVG). LLaVA-ST achieves outstanding performance on 11 benchmarks requiring fine-grained temporal, spatial, or spatial-temporal interleaving multimodal understanding. Our code, data and benchmark will be released at Our code, data and benchmark will be released at https://github.com/appletea233/LLaVA-ST .
We introduce Tarsier2, a state-of-the-art large vision-language model (LVLM) designed for generating detailed and accurate video descriptions, while also exhibiting superior general video understanding capabilities. Tarsier2 achieves significant advancements through three key upgrades: (1) Scaling pre-training data from 11M to 40M video-text pairs, enriching both volume and diversity; (2) Performing fine-grained temporal alignment during supervised fine-tuning; (3) Using model-based sampling to automatically construct preference data and applying DPO training for optimization. Extensive experiments show that Tarsier2-7B consistently outperforms leading proprietary models, including GPT-4o and Gemini 1.5 Pro, in detailed video description tasks. On the DREAM-1K benchmark, Tarsier2-7B improves F1 by 2.8% over GPT-4o and 5.8% over Gemini-1.5-Pro. In human side-by-side evaluations, Tarsier2-7B shows a +8.6% performance advantage over GPT-4o and +24.9% over Gemini-1.5-Pro. Tarsier2-7B also sets new state-of-the-art results across 15 public benchmarks, spanning tasks such as video question-answering, video grounding, hallucination test, and embodied question-answering, demonstrating its versatility as a robust generalist vision-language model.
Despite the advancements of Video Large Language Models (VideoLLMs) in various tasks, they struggle with fine-grained temporal understanding, such as Dense Video Captioning (DVC). DVC is a complicated task of describing all events within a video while also temporally localizing them, which integrates multiple fine-grained tasks, including video segmentation, video captioning, and temporal video grounding. Previous VideoLLMs attempt to solve DVC in a single step, failing to utilize their reasoning capability. Moreover, previous training objectives for VideoLLMs do not fully reflect the evaluation metrics, therefore not providing supervision directly aligned to target tasks. To address such a problem, we propose a novel framework named VidChain comprised of Chain-of-Tasks (CoTasks) and Metric-based Direct Preference Optimization (M-DPO). CoTasks decompose a complex task into a sequence of sub-tasks, allowing VideoLLMs to leverage their reasoning capabilities more effectively. M-DPO aligns a VideoLLM with evaluation metrics, providing fine-grained supervision to each task that is well-aligned with metrics. Applied to two different VideoLLMs, VidChain consistently improves their fine-grained video understanding, thereby outperforming previous VideoLLMs on two different DVC benchmarks and also on the temporal video grounding task. Code is available at \url{https://github.com/mlvlab/VidChain}.
Compositional generalization is the capability of a model to understand novel compositions composed of seen concepts. There are multiple levels of novel compositions including phrase-phrase level, phrase-word level, and word-word level. Existing methods achieve promising compositional generalization, but the consistency of compositional generalization across multiple levels of novel compositions remains unexplored. The consistency refers to that a model should generalize to a phrase-phrase level novel composition, and phrase-word/word-word level novel compositions that can be derived from it simultaneously. In this paper, we propose a meta-learning based framework, for achieving consistent compositional generalization across multiple levels. The basic idea is to progressively learn compositions from simple to complex for consistency. Specifically, we divide the original training set into multiple validation sets based on compositional complexity, and introduce multiple meta-weight-nets to generate sample weights for samples in different validation sets. To fit the validation sets in order of increasing compositional complexity, we optimize the parameters of each meta-weight-net independently and sequentially in a multilevel optimization manner. We build a GQA-CCG dataset to quantitatively evaluate the consistency. Experimental results on visual question answering and temporal video grounding, demonstrate the effectiveness of the proposed framework. We release GQA-CCG at https://github.com/NeverMoreLCH/CCG.
Recent researches on video large language models (VideoLLM) predominantly focus on model architectures and training datasets, leaving the interaction format between the user and the model under-explored. In existing works, users often interact with VideoLLMs by using the entire video and a query as input, after which the model generates a response. This interaction format constrains the application of VideoLLMs in scenarios such as live-streaming comprehension where videos do not end and responses are required in a real-time manner, and also results in unsatisfactory performance on time-sensitive tasks that requires localizing video segments. In this paper, we focus on a video-text duet interaction format. This interaction format is characterized by the continuous playback of the video, and both the user and the model can insert their text messages at any position during the video playback. When a text message ends, the video continues to play, akin to the alternative of two performers in a duet. We construct MMDuetIT, a video-text training dataset designed to adapt VideoLLMs to video-text duet interaction format. We also introduce the Multi-Answer Grounded Video Question Answering (MAGQA) task to benchmark the real-time response ability of VideoLLMs. Trained on MMDuetIT, MMDuet demonstrates that adopting the video-text duet interaction format enables the model to achieve significant improvements in various time-sensitive tasks (76% CIDEr on YouCook2 dense video captioning, 90\% mAP on QVHighlights highlight detection and 25% [email omitted] on Charades-STA temporal video grounding) with minimal training efforts, and also enable VideoLLMs to reply in a real-time manner as the video plays. Code, data and demo are available at: https://github.com/yellow-binary-tree/MMDuet.
Videos carry rich visual information including object description, action, interaction, etc., but the existing multimodal large language models (MLLMs) fell short in referential understanding scenarios such as video-based referring. In this paper, we present Artemis, an MLLM that pushes video-based referential understanding to a finer level. Given a video, Artemis receives a natural-language question with a bounding box in any video frame and describes the referred target in the entire video. The key to achieving this goal lies in extracting compact, target-specific video features, where we set a solid baseline by tracking and selecting spatiotemporal features from the video. We train Artemis on the newly established VideoRef45K dataset with 45K video-QA pairs and design a computationally efficient, three-stage training procedure. Results are promising both quantitatively and qualitatively. Additionally, we show that \model can be integrated with video grounding and text summarization tools to understand more complex scenarios. Code and data are available at https://github.com/qiujihao19/Artemis.
In this paper, we investigate the feasibility of leveraging large language models (LLMs) for integrating general knowledge and incorporating pseudo-events as priors for temporal content distribution in video moment retrieval (VMR) models. The motivation behind this study arises from the limitations of using LLMs as decoders for generating discrete textual descriptions, which hinders their direct application to continuous outputs like salience scores and inter-frame embeddings that capture inter-frame relations. To overcome these limitations, we propose utilizing LLM encoders instead of decoders. Through a feasibility study, we demonstrate that LLM encoders effectively refine inter-concept relations in multimodal embeddings, even without being trained on textual embeddings. We also show that the refinement capability of LLM encoders can be transferred to other embeddings, such as BLIP and T5, as long as these embeddings exhibit similar inter-concept similarity patterns to CLIP embeddings. We present a general framework for integrating LLM encoders into existing VMR architectures, specifically within the fusion module. Through experimental validation, we demonstrate the effectiveness of our proposed methods by achieving state-of-the-art performance in VMR. The source code can be accessed at https://github.com/fletcherjiang/LLMEPET.
We introduce InternVideo2, a new family of video foundation models (ViFM) that achieve the state-of-the-art results in video recognition, video-text tasks, and video-centric dialogue. Our core design is a progressive training approach that unifies the masked video modeling, crossmodal contrastive learning, and next token prediction, scaling up the video encoder size to 6B parameters. At the data level, we prioritize spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. Through extensive experiments, we validate our designs and demonstrate superior performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related dialogue and long video understanding benchmarks, highlighting its ability to reason and comprehend longer contexts. Code and models are available at https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo2/.
Inspired by the activity-silent and persistent activity mechanisms in human visual perception biology, we design a Unified Static and Dynamic Network (UniSDNet), to learn the semantic association between the video and text/audio queries in a cross-modal environment for efficient video grounding. For static modeling, we devise a novel residual structure (ResMLP) to boost the global comprehensive interaction between the video segments and queries, achieving more effective semantic enhancement/supplement. For dynamic modeling, we effectively exploit three characteristics of the persistent activity mechanism in our network design for a better video context comprehension. Specifically, we construct a diffusely connected video clip graph on the basis of 2D sparse temporal masking to reflect the "short-term effect" relationship. We innovatively consider the temporal distance and relevance as the joint "auxiliary evidence clues" and design a multi-kernel Temporal Gaussian Filter to expand the context clue into high-dimensional space, simulating the "complex visual perception", and then conduct element level filtering convolution operations on neighbour clip nodes in message passing stage for finally generating and ranking the candidate proposals. Our UniSDNet is applicable to both Natural Language Video Grounding (NLVG) and Spoken Language Video Grounding (SLVG) tasks. Our UniSDNet achieves SOTA performance on three widely used datasets for NLVG, as well as three datasets for SLVG, e.g., reporting new records at 38.88% R@1,[email omitted] on ActivityNet Captions and 40.26% R@1,[email omitted] on TACoS. To facilitate this field, we collect two new datasets (Charades-STA Speech and TACoS Speech) for SLVG task. Meanwhile, the inference speed of our UniSDNet is 1.56$\times$ faster than the strong multi-query benchmark. Code is available at: https://github.com/xian-sh/UniSDNet.
Video-text Large Language Models (video-text LLMs) have shown remarkable performance in answering questions and holding conversations on simple videos. However, they perform almost the same as random on grounding text queries in long and complicated videos, having little ability to understand and reason about temporal information, which is the most fundamental difference between videos and images. In this paper, we propose HawkEye, one of the first video-text LLMs that can perform temporal video grounding in a fully text-to-text manner. To collect training data that is applicable for temporal video grounding, we construct InternVid-G, a large-scale video-text corpus with segment-level captions and negative spans, with which we introduce two new time-aware training objectives to video-text LLMs. We also propose a coarse-grained method of representing segments in videos, which is more robust and easier for LLMs to learn and follow than other alternatives. Extensive experiments show that HawkEye is better at temporal video grounding and comparable on other video-text tasks with existing video-text LLMs, which verifies its superior video-text multi-modal understanding abilities.
Spatio-temporal video grounding (or STVG) task aims at locating a spatio-temporal tube for a specific instance given a text query. Despite advancements, current methods easily suffer the distractors or heavy object appearance variations in videos due to insufficient object information from the text, leading to degradation. Addressing this, we propose a novel framework, context-guided STVG (CG-STVG), which mines discriminative instance context for object in videos and applies it as a supplementary guidance for target localization. The key of CG-STVG lies in two specially designed modules, including instance context generation (ICG), which focuses on discovering visual context information (in both appearance and motion) of the instance, and instance context refinement (ICR), which aims to improve the instance context from ICG by eliminating irrelevant or even harmful information from the context. During grounding, ICG, together with ICR, are deployed at each decoding stage of a Transformer architecture for instance context learning. Particularly, instance context learned from one decoding stage is fed to the next stage, and leveraged as a guidance containing rich and discriminative object feature to enhance the target-awareness in decoding feature, which conversely benefits generating better new instance context for improving localization finally. Compared to existing methods, CG-STVG enjoys object information in text query and guidance from mined instance visual context for more accurate target localization. In our experiments on three benchmarks, including HCSTVG-v1/-v2 and VidSTG, CG-STVG sets new state-of-the-arts in m_tIoU and m_vIoU on all of them, showing its efficacy. The code will be released at https://github.com/HengLan/CGSTVG.
In the weakly supervised temporal video grounding study, previous methods use predetermined single Gaussian proposals which lack the ability to express diverse events described by the sentence query. To enhance the expression ability of a proposal, we propose a Gaussian mixture proposal (GMP) that can depict arbitrary shapes by learning importance, centroid, and range of every Gaussian in the mixture. In learning GMP, each Gaussian is not trained in a feature space but is implemented over a temporal location. Thus the conventional feature-based learning for Gaussian mixture model is not valid for our case. In our special setting, to learn moderately coupled Gaussian mixture capturing diverse events, we newly propose a pull-push learning scheme using pulling and pushing losses, each of which plays an opposite role to the other. The effects of components in our scheme are verified in-depth with extensive ablation studies and the overall scheme achieves state-of-the-art performance. Our code is available at https://github.com/sunoh-kim/pps.
Video moment retrieval is a challenging task requiring fine-grained interactions between video and text modalities. Recent work in image-text pretraining has demonstrated that most existing pretrained models suffer from information asymmetry due to the difference in length between visual and textual sequences. We question whether the same problem also exists in the video-text domain with an auxiliary need to preserve both spatial and temporal information. Thus, we evaluate a recently proposed solution involving the addition of an asymmetric co-attention network for video grounding tasks. Additionally, we incorporate momentum contrastive loss for robust, discriminative representation learning in both modalities. We note that the integration of these supplementary modules yields better performance compared to state-of-the-art models on the TACoS dataset and comparable results on ActivityNet Captions, all while utilizing significantly fewer parameters with respect to baseline.