Normalizing Flows (NFs) are a class of generative models distinguished by a mathematically invertible architecture, where the forward pass transforms data into a latent space for density estimation, and the reverse pass generates new samples from this space. This characteristic creates an intrinsic synergy between representation learning and data generation. However, the generative quality of standard NFs is limited by poor semantic representations from log-likelihood optimization. To remedy this, we propose a novel alignment strategy that creatively leverages the invertibility of NFs: instead of regularizing the forward pass, we align the intermediate features of the generative (reverse) pass with representations from a powerful vision foundation model, demonstrating superior effectiveness over naive alignment. We also introduce a novel training-free, test-time optimization algorithm for classification, which provides a more intrinsic evaluation of the NF's embedded semantic knowledge. Comprehensive experiments demonstrate that our approach accelerates the training of NFs by over 3.3x, while simultaneously delivering significant improvements in both generative quality and classification accuracy. New state-of-the-art results for NFs are established on ImageNet 64 x 64 and 256 x 256.

Recent advances in video understanding have been driven by MLLMs. But these MLLMs are good at analyzing short videos, while suffering from difficulties in understanding videos with a longer context. To address this difficulty, several agent paradigms have recently been proposed, using MLLMs as agents for retrieving extra contextual knowledge in a long video. However, most existing agents ignore the key fact that a long video is composed with multiple shots, i.e., to answer the user question from a long video, it is critical to deeply understand its relevant shots like human. Without such insight, these agents often mistakenly find redundant even noisy temporal context, restricting their capacity for long video understanding. To fill this gap, we propose VideoChat-A1, a novel long video agent paradigm. Different from the previous works, our VideoChat-A1 can deeply think with long videos, via a distinct chain-of-shot reasoning paradigm. More specifically, it can progressively select the relevant shots of user question, and look into these shots in a coarse-to-fine partition. By multi-modal reasoning along the shot chain, VideoChat-A1 can effectively mimic step-by-step human thinking process, allowing the interactive discovery of preferable temporal context for thoughtful understanding in long videos. Extensive experiments show that, VideoChat-A1 achieves the state-of-the-art performance on the mainstream long video QA benchmarks, e.g., it achieves 77.0 on VideoMME(w/ subs) and 70.1 on EgoSchema, outperforming its strong baselines (e.g., InternVL2.5-8B and InternVideo2.5-8B), by up to 10.1% and 6.2%. Compared to leading closed-source GPT-4o and Gemini 1.5 Pro, VideoChat-A1 offers competitive accuracy, but only with 7% input frames and 12% inference time on average.

Video temporal grounding (VTG) aims to locate precise segments in videos based on language queries, which is a fundamental challenge in video understanding. While recent Multimodal Large Language Models (MLLMs) have shown promise in tackling VTG through reinforcement learning (RL), they overlook the challenges arising from both the quality and difficulty of training samples. (1) Partially annotated samples. Many manually annotated samples contain relevant segments beyond the annotated interval, introducing ambiguous supervision. (2) Hard-to-ground samples. Samples with poor zero-shot performance produce consistently low and indistinguishable rewards during RL training, exhibiting no clear preference among multiple outputs and thus hindering learning efficiency. To address these challenges, we propose VideoTG-R1, a novel curriculum RL framework with reflected boundary annotations, enabling data-efficient training. Specifically, we propose a Boundary Reflection Agent that utilizes MLLMs to predict query-relevant timestamps outside the annotated intervals, allowing us to identify and filter out partially annotated samples, thereby reducing ambiguity. Furthermore, we introduce a Difficulty Estimation Agent to assess the training difficulty of each sample and design a curriculum RL strategy that dynamically masks the videos of hard-to-ground samples according to the training steps, easing the training difficulty and providing clearer preference. Experiments on the VTG and grounded VideoQA tasks demonstrate the effectiveness of our method. Remarkably, with only 10% of the training samples and 21% of the computational budget, VideoTG-R1 outperforms full-data counterparts under both group relative policy optimization (GRPO) and supervised fine-tuning (SFT). The code is available at https://github.com/ldong1111/VideoTG-R1.

Video generation has witnessed significant advancements, yet evaluating these models remains a challenge. A comprehensive evaluation benchmark for video generation is indispensable for two reasons: 1) Existing metrics do not fully align with human perceptions; 2) An ideal evaluation system should provide insights to inform future developments of video generation. To this end, we present VBench++, a comprehensive benchmark suite that dissects "video generation quality" into specific, hierarchical, and disentangled dimensions, each with tailored prompts and evaluation methods. VBench++ has several appealing properties: 1) Comprehensive Dimensions: VBench++ comprises 16 dimensions in text-to-video generation (e.g., subject identity inconsistency, motion smoothness, temporal flickering, and spatial relationship, etc). The evaluation metrics with fine-grained levels reveal individual models' strengths and weaknesses. 2) Human Alignment: We also provide a dataset of human preference annotations to validate our benchmarks' alignment with human perception, for each evaluation dimension respectively. 3) Valuable Insights: We look into current models' ability across various evaluation dimensions, and various content types. We also investigate the gaps between video and image generation models. 4) Versatile Benchmarking: VBench++ is designed to evaluate a wide range of video generation tasks, including text-to-video and image-to-video. We introduce a high-quality Image Suite with an adaptive aspect ratio to enable fair evaluations across different image-to-video generation settings. Beyond assessing technical quality, VBench++ evaluates the trustworthiness of video generative models, providing a more holistic view of model performance. 5) Full Open-Sourcing: We fully open-source VBench++, including all prompts, the Image Suite, evaluation methods, generated videos, and human preference annotations.

Multi-Object Tracking (MOT) has been a long-standing challenge in video understanding. A natural and intuitive approach is to split this task into two parts: object detection and association. Most mainstream methods employ meticulously crafted heuristic techniques to maintain trajectory information and compute cost matrices for object matching. Although these methods can achieve notable tracking performance, they often require a series of elaborate handcrafted modifications while facing complicated scenarios. We believe that manually assumed priors limit the method’s adaptability and flexibility in learning optimal tracking capabilities from domain-specific data. Therefore, we introduce a new perspective that treats Multiple Object Tracking as an in-context ID Prediction task, transforming the aforementioned object association into an end-to-end trainable task. Based on this, we propose a simple yet effective method termed MOTIP. Given a set of trajectories carried with ID information, MOTIP directly decodes the ID labels for current detections to accomplish the association process. Without using tailored or sophisticated architectures, our method achieves state-of-the-art results across multiple benchmarks by solely leveraging object-level features as tracking cues. The simplicity and impressive results of MOTIP leave substantial room for future advancements, thereby making it a promising baseline for subsequent research. Our code and checkpoints are released at https://github.com/MCG-NJU/MOTIP.

Multimodal Large Language Models (MLLMs) have significantly progressed in offline video understanding. However, applying these models to real-world scenarios, such as autonomous driving and human-computer interaction, presents unique challenges due to the need for real-time processing of continuous online video streams. To this end, this paper presents systematic efforts from three perspectives: evaluation benchmark, model architecture, and training strategy. First, we introduce OVBench, a comprehensive question-answering benchmark designed to evaluate models’ ability to perceive, memorize, and reason within online video contexts. It features 6 core task types across three temporal contexts—past, current, and future—forming 16 subtasks from diverse datasets. Second, we propose a new Pyramid Memory Bank (PMB) that effectively retains key spatiotemporal information in video streams. Third, we proposed an offline-to-online learning paradigm, designing an interleaved dialogue format for online video data and constructing an instruction-tuning dataset tailored for online video training. This framework led to the development of VideoChat-Online, a robust and efficient model for online video understanding. Despite the lower computational cost and higher efficiency, VideoChat-Online outperforms existing state-of-the-art offline and online models across popular offline video benchmarks and OVBench, demonstrating the effectiveness of our model architecture and training strategy.

The intuitive nature of drag-based interaction has led to its growing adoption for controlling object trajectories in image-to-video synthesis. Still, existing methods that perform dragging in the 2D space usually face ambiguity when handling out-of-plane movements. In this work, we augment the interaction with a new dimension, i.e., the depth dimension, such that users are allowed to assign a relative depth for each point on the trajectory. That way, our new interaction paradigm not only inherits the convenience from 2D dragging, but facilitates trajectory control in the 3D space, broadening the scope of creativity. We propose a pioneering method for 3D trajectory control in image-to-video synthesis by abstracting object masks into a few cluster points. These points, accompanied by the depth information and the instance information, are finally fed into a video diffusion model as the control signal. Extensive experiments validate the effectiveness of our approach, dubbed LeviTor, in precisely manipulating the object movements when producing photo-realistic videos from static images. Our code is available at: https://github.com/ant-research/LeviTor.

The Audio Description (AD) task aims to generate descriptions of visual elements for visually impaired individuals to help them access long-form video content, like movies. With video feature, text, character bank and context information as inputs, the generated ADs are able to correspond to the characters by name and provide reasonable, contextual descriptions to help audience understand the storyline of movie. To achieve this goal, we propose to leverage pre-trained foundation models through a simple and unified framework to generate ADs with interleaved multimodal sequence as input, termed as Uni-AD. To enhance the alignment of features across various modalities with finer granularity, we introduce a simple and lightweight module that maps video features into the textual feature space. Moreover, we also propose a character-refinement module to provide more precise information by identifying the main characters who play more significant roles in the video context. With these unique designs, we further incorporate contextual information and a contrastive loss into our architecture to generate smoother and more contextually appropriate ADs. Experiments on multiple AD datasets show that Uni-AD performs well on AD generation, which demonstrates the effectiveness of our approach. Our code is available at: https://github.com/ant-research/UniAD.

Current multimodal large language models (MLLMs) struggle with fine-grained or precise understanding of visuals although they give comprehensive perception and reasoning in a spectrum of vision applications. Recent studies either develop tool-using or unify specific visual tasks into the autoregressive framework, often at the expense of overall multimodal performance. To address this issue and enhance MLLMs with visual tasks in a scalable fashion, we propose Task Preference Optimization (TPO), a novel method that utilizes differentiable task preferences derived from typical fine-grained visual tasks. TPO introduces learnable task tokens that establish connections between multiple task-specific heads and the MLLM. By leveraging rich visual labels during training, TPO significantly enhances the MLLM’s multimodal capabilities and task-specific performance. Through multi-task co-training within TPO, we observe synergistic benefits that elevate individual task performance beyond what is achievable through single-task training methodologies. Our instantiation of this approach with VideoChat and LLaVA demonstrates an overall 14.6% improvement in multimodal performance compared to baseline models. Additionally, MLLM-TPO demonstrates robust zero-shot capabilities across various tasks, performing comparably to state-of-the-art supervised models.

Learning from multiple domains is a primary factor that influences the generalization of a single unified robot system. In this paper, we aim to learn the trajectory prediction model by using broad out-of-domain data to improve its performance and generalization ability. Trajectory model is designed to predict any-point trajectories in the current frame given an instruction and can provide detailed control guidance for robotic policy learning. To handle the diverse out-of-domain data distribution, we propose a sparsely-gated MoE (Top-1 gating strategy) architecture for trajectory model, coined as Tra-MoE. The sparse activation design enables good balance between parameter cooperation and specialization, effectively benefiting from large-scale out-of-domain data while maintaining constant FLOPs per token. In addition, we further introduce an adaptive policy conditioning technique by learning 2D mask representations for predicted trajectories, which is explicitly aligned with image observations to guide action prediction more flexibly. We perform extensive experiments on both simulation and real-world scenarios to verify the effectiveness of Tra-MoE and adaptive policy conditioning technique. We also conduct a comprehensive empirical study to train Tra-MoE, demonstrating that our Tra-MoE consistently exhibits superior performance compared to the dense baseline model, even when the latter is scaled to match Tra-MoE’s parameter count.

Efficient lightweight neural networks are with increasing attention due to their faster reasoning speed and easier deployment on mobile devices. However, existing video pre-trained models still focus on the common ViT architecture with high latency, and few works attempt to build efficient architecture on mobile devices. This paper bridges this gap by introducing temporal structural reparameterization into an efficient image-text model and training it on a large-scale high-quality video-text dataset, resulting in an efficient video-text model that can run on mobile devices with strong zero-shot classification and retrieval capabilities, termed as MobileViCLIP. In particular, in terms of inference speed on mobile devices, our MobileViCLIP-Small is 55.4x times faster than InternVideo2-L14 and 6.7x faster than InternVideo2-S14. In terms of zero-shot retrieval performance, our MobileViCLIP-Small obtains similar performance as InternVideo2-L14 and obtains 6.9\% better than InternVideo2-S14 on MSR-VTT. The code is available at https://github.com/MCG-NJU/MobileViCLIP.

We present VRBench, the first long narrative video benchmark crafted for evaluating large models' multi-step reasoning capabilities, addressing limitations in existing evaluations that overlook temporal reasoning and procedural validity. It comprises 960 long videos (with an average duration of 1.6 hours), along with 8,243 human-labeled multi-step question-answering pairs and 25,106 reasoning steps with timestamps. These videos are curated via a multi-stage filtering process including expert inter-rater reviewing to prioritize plot coherence. We develop a human-AI collaborative framework that generates coherent reasoning chains, each requiring multiple temporally grounded steps, spanning seven types (e.g., event attribution, implicit inference). VRBench designs a multi-phase evaluation pipeline that assesses models at both the outcome and process levels. Apart from the MCQs for the final results, we propose a progress-level LLM-guided scoring metric to evaluate the quality of the reasoning chain from multiple dimensions comprehensively. Through extensive evaluations of 12 LLMs and 19 VLMs on VRBench, we undertake a thorough analysis and provide valuable insights that advance the field of multi-step reasoning.

Despite the remarkable performance of multimodal large language models (MLLMs) across diverse tasks, the substantial training and inference costs impede their advancement. In this paper, we propose p-MoD, an efficient MLLM architecture that significantly reduces training and inference costs while maintaining model performance. The majority of computation in MLLMs stems from the overwhelming volume of vision tokens processed by the transformer-based LLM. Accordingly, we leverage the Mixture-of-Depths (MoD) mechanism, where each LLM layer selects essential vision tokens to process while skipping redundant ones. However, integrating MoD into MLLMs is non-trivial. To address the challenges of training and inference stability as well as limited training data, we adapt the MoD module with two novel designs: tanh-gated weight normalization (TanhNorm) and symmetric token reweighting (STRing). Moreover, we observe that vision tokens exhibit higher redundancy in deeper layers and thus design a progressive ratio decay (PRD) strategy, which gradually reduces the token retention ratio layer by layer, employing a shifted cosine schedule. This crucial design fully unleashes the potential of MoD, significantly boosting the efficiency and performance of our models. Extensive experiments on two baseline models across 15 benchmarks show that our model matches or even surpasses the performance of corresponding baselines, while requiring only 55.6% TFLOPs and 53.7% KV cache storage during inference, and 77.7% GPU hours during training.

The existing video understanding benchmarks for multimodal large language models (MLLMs) mainly focus on short videos. The few benchmarks for long video understanding often rely on multiple-choice questions (MCQs). Due to the limitations of MCQ evaluations and the advanced reasoning abilities of MLLMs, models can often answer correctly by combining short video insights with elimination, without truly understanding the content. To bridge this gap, we introduce CG-Bench, a benchmark for clue-grounded question answering in long videos. CG-Bench emphasizes the model's ability to retrieve relevant clues, enhancing evaluation credibility. It includes 1,219 manually curated videos organized into 14 primary, 171 secondary, and 638 tertiary categories, making it the largest benchmark for long video analysis. The dataset features 12,129 QA pairs in three question types: perception, reasoning, and hallucination. To address the limitations of MCQ-based evaluation, we develop two novel clue-based methods: clue-grounded white box and black box evaluations, assessing whether models generate answers based on accurate video understanding. We evaluated multiple closed-source and open-source MLLMs on CG-Bench. The results show that current models struggle significantly with long videos compared to short ones, and there is a notable gap between open-source and commercial models. We hope CG-Bench will drive the development of more reliable and capable MLLMs for long video comprehension.

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.

Creating high-quality data for training robust language-instructed agents is a long-lasting challenge in embodied AI. In this paper, we introduce a Self-Refining Data Flywheel (SRDF) that generates high-quality and large-scale navigational instruction-trajectory pairs by iteratively refining the data pool through the collaboration between two models, the instruction generator and the navigator, without any human-in-the-loop annotation. Specifically, SRDF starts with using a base generator to create an initial data pool for training a base navigator, followed by applying the trained navigator to filter the data pool. This leads to higher-fidelity data to train a better generator, which can, in turn, produce higher-quality data for training the next-round navigator. Such a flywheel establishes a data self-refining process, yielding a continuously improved and highly effective dataset for large-scale language-guided navigation learning. Our experiments demonstrate that after several flywheel rounds, the navigator elevates the performance boundary from 70\% to 78\% SPL on the classic R2R test set, surpassing human performance (76\%) for the first time. Meanwhile, this process results in a superior generator, evidenced by a SPICE increase from 23.5 to 26.2, better than all previous VLN instruction generation methods. Finally, we demonstrate the scalability of our method through increasing environment and instruction diversity, and the generalization ability of our pre-trained navigator across various downstream navigation tasks, surpassing state-of-the-art methods by a large margin in all cases.

Multimodal Large Language Models (MLLMs) have demonstrated impressive performance in short video understanding. However, understanding long-form videos still remains challenging for MLLMs. This paper proposes TimeSuite, a collection of new designs to adapt the existing short-form video MLLMs for long video understanding, including a simple yet efficient framework to process long video sequence, a high-quality video dataset for grounded tuning of MLLMs, and a carefully-designed instruction tuning task to explicitly incorporate the grounding supervision in the traditional QA format. Specifically, based on VideoChat, we propose our long-video MLLM, coined as VideoChat-T, by implementing a token shuffling to compress long video tokens and introducing Temporal Adaptive Position Encoding (TAPE) to enhance the temporal awareness of visual representation. Meanwhile, we introduce the TimePro, a comprehensive grounding-centric instruction tuning dataset composed of 9 tasks and 349k high-quality grounded annotations. Notably, we design a new instruction tuning task type, called Temporal Grounded Caption, to peform detailed video descriptions with the corresponding time stamps prediction. This explicit temporal location prediction will guide MLLM to correctly attend on the visual content when generating description, and thus reduce the hallucination risk caused by the LLMs. Experimental results demonstrate that our TimeSuite provides a successful solution to enhance the long video understanding capability of short-form MLLM, achieving improvement of 5.6% and 6.8% on the benchmarks of Egoschema and VideoMME, respectively. In addition, VideoChat-T exhibits robust zero-shot temporal grounding capabilities, significantly outperforming the existing state-of-the-art MLLMs. After fine-tuning, it performs on par with the traditional supervised expert models.

In this paper, we introduce SPA, a novel representation learning framework that emphasizes the importance of 3D spatial awareness in embodied AI. Our approach leverages differentiable neural rendering on multi-view images to endow a vanilla Vision Transformer (ViT) with intrinsic spatial understanding. We present the most comprehensive evaluation of embodied representation learning to date, covering 268 tasks across 8 simulators with diverse policies in both single-task and language-conditioned multi-task scenarios. The results are compelling: SPA consistently outperforms more than 10 state-of-the-art representation methods, including those specifically designed for embodied AI, vision-centric tasks, and multi-modal applications, while using less training data. Furthermore, we conduct a series of real-world experiments to confirm its effectiveness in practical scenarios. These results highlight the critical role of 3D spatial awareness for embodied representation learning. Our strongest model takes more than 6000 GPU hours to train and we are committed to open-sourcing all code and model weights to foster future research in embodied representation learning.

Recent Vision Mamba (Vim) models exhibit nearly linear complexity in sequence length, making them highly attractive for processing visual data. However, the training methodologies and their potential are still not sufficiently explored. In this paper, we investigate strategies for Vim and propose Stochastic Layer-Wise Shuffle (SLWS), a novel regularization method that can effectively improve the Vim training. Without architectural modifications, this approach enables the non-hierarchical Vim to get leading performance on ImageNet-1K compared with the similar type counterparts. Our method operates through four simple steps per layer: probability allocation to assign layer-dependent shuffle rates, operation sampling via Bernoulli trials, sequence shuffling of input tokens, and order restoration of outputs. SLWS distinguishes itself through three principles: \textit{(1) Plug-and-play:} No architectural modifications are needed, and it is deactivated during inference. \textit{(2) Simple but effective:} The four-step process introduces only random permutations and negligible overhead. \textit{(3) Intuitive design:} Shuffling probabilities grow linearly with layer depth, aligning with the hierarchical semantic abstraction in vision models. Our work underscores the importance of tailored training strategies for Vim models and provides a helpful way to explore their scalability. Code and models are available at https://github.com/huangzizheng01/ShuffleMamba

Diffusion models have demonstrated remarkable generation quality but at the cost of numerous function evaluations. Recently, advanced ODE-based solvers have been developed to mitigate the substantial computational demands of reverse-diffusion solving under limited sampling steps. However, these solvers, heavily inspired by Adams-like multistep methods, rely solely on t-related Lagrange interpolation. We show that t-related Lagrange interpolation is suboptimal for diffusion model and reveal a compact search space comprised of time steps and solver coefficients. Building on our analysis, we propose a novel differentiable solver search algorithm to identify more optimal solver. Equipped with the searched solver, rectified-flow models, e.g., SiT-XL/2 and FlowDCN-XL/2, achieve FID scores of 2.40 and 2.35, respectively, on ImageNet-$256\times256$ with only 10 steps. Meanwhile, DDPM model, DiT-XL/2, reaches a FID score of 2.33 with only 10 steps. Notably, our searched solver outperforms traditional solvers by a significant margin. Moreover, our searched solver demonstrates generality across various model architectures, resolutions, and model sizes.

We introduce Eagle2.5, a frontier vision-language model (VLM) for long-context multimodal learning. Our work addresses the challenges in long video comprehension and high-resolution image understanding, introducing a generalist framework for both tasks. The proposed training framework incorporates Automatic Degrade Sampling and Image Area Preservation, two techniques that preserve contextual integrity and visual details. The framework also includes numerous efficiency optimizations in the pipeline for long-context data training. Finally, we propose Eagle-Video-110K, a novel dataset that integrates both story-level and clip-level annotations, facilitating long-video understanding. Eagle2.5 demonstrates substantial improvements on long-context multimodal benchmarks, providing a robust solution to the limitations of existing VLMs. Notably, our best model Eagle2.5-8B achieves 72.4\% on Video-MME with 512 input frames, matching the results of top-tier commercial model such as GPT-4o and large-scale open-source models like Qwen2.5-VL-72B and InternVL2.5-78B.

We propose the Multi-modal Untrimmed Video Retrieval task, along with a new benchmark (MUVR) to advance video retrieval for long-video platforms. MUVR aims to retrieve untrimmed videos containing relevant segments using multi-modal queries. It has the following features: 1) Practical retrieval paradigm: MUVR supports video-centric multi-modal queries, expressing fine-grained retrieval needs through long text descriptions, video tag prompts, and mask prompts. It adopts a one-to-many retrieval paradigm and focuses on untrimmed videos, tailored for long-video platform applications. 2) Multi-level visual correspondence: To cover common video categories (e.g., news, travel, dance) and precisely define retrieval matching criteria, we construct multi-level visual correspondence based on core video content (e.g., news events, travel locations, dance moves) which users are interested in and want to retrieve. It covers six levels: copy, event, scene, instance, action, and others. 3) Comprehensive evaluation criteria: We develop 3 versions of MUVR (i.e., Base, Filter, QA). MUVR-Base/Filter evaluates retrieval models, while MUVR-QA assesses MLLMs in a question-answering format. We also propose a Reranking Score to evaluate the reranking ability of MLLMs. MUVR consists of 53K untrimmed videos from the video platform Bilibili, with 1,050 multi-modal queries and 84K matches. Extensive evaluations of 3 state-of-the-art video retrieval models, 6 image-based VLMs, and 10 MLLMs are conducted. MUVR reveals the limitations of retrieval methods in processing untrimmed videos and multi-modal queries, as well as MLLMs in multi-video understanding and reranking. Our code and benchmark is available at https://github.com/debby-0527/MUVR.

We present LongVPO, a novel two‑stage Direct Preference Optimization framework that enables short‑context vision‑language models to robustly understand ultra‑long videos without any long‑video annotations. In Stage 1, we synthesize preference triples by anchoring questions to individual short clips, interleaving them with distractors, and applying visual‑similarity and question‑specificity filtering to mitigate positional bias and ensure unambiguous supervision. We also approximate the reference model’s scoring over long contexts by evaluating only the anchor clip, reducing computational overhead. In Stage 2, we employ a recursive captioning pipeline on long videos to generate scene-level metadata, and then use a large language model to craft multi-segment reasoning queries and dispreferred responses, aligning the model's preferences through multi-segment reasoning tasks. With only 16K synthetic examples and no costly human labels, \model{} outperforms the state‑of‑the‑art open‑source models on multiple long‑video benchmarks, while maintaining strong short‑video performance (e.g., on MVBench), offering a scalable paradigm for efficient long‑form video understanding.

Inducing reasoning in multimodal large language models (MLLMs) is critical for achieving human-level perception and understanding. Existing methods mainly leverage LLM reasoning to analyze parsed visuals, often limited by static perception stages. This paper introduces Visual Test-Time Scaling (VTTS), a novel approach to enhance MLLMs' reasoning via iterative perception during inference. VTTS mimics humans' hierarchical attention by progressively refining focus on high-confidence spatio-temporal regions, guided by updated textual predictions. Specifically, VTTS employs an Iterative Perception (ITP) mechanism, incorporating reinforcement learning with spatio-temporal supervision to optimize reasoning. To support this paradigm, we also present VTTS-80K, a dataset tailored for iterative perception. These designs allows a MLLM to enhance its performance by increasing its perceptual compute. Extensive experiments validate VTTS's effectiveness and generalization across diverse tasks and benchmarks. Our newly introduced Videochat-R1.5 model has achieved remarkable improvements, with an average increase of over 5\%, compared to robust baselines such as Qwen2.5VL-3B and -7B, across more than 15 benchmarks that encompass video conversation, video reasoning, and spatio-temporal perception.

Multimodal Large Language Models (MLLMs) have recently achieved remarkable progress in video understanding. However, their effectiveness in real-time streaming scenarios remains limited due to storage constraints of historical visual features and insufficient real-time spatiotemporal reasoning. To address these challenges, we propose StreamForest, a novel architecture specifically designed for streaming video understanding. Central to StreamForest is the Persistent Event Memory Forest, a memory mechanism that adaptively organizes video frames into multiple event-level tree structures. This process is guided by penalty functions based on temporal distance, content similarity, and merge frequency, enabling efficient long-term memory retention under limited computational resources. To enhance real-time perception, we introduce a Fine-grained Spatiotemporal Window, which captures detailed short-term visual cues to improve current scene perception. Additionally, we present OnlineIT, an instruction-tuning dataset tailored for streaming video tasks. OnlineIT significantly boosts MLLM performance in both real-time perception and future prediction. To evaluate generalization in practical applications, we introduce ODV-Bench, a new benchmark focused on real-time streaming video understanding in autonomous driving scenarios. Experimental results demonstrate that StreamForest achieves the state-of-the-art performance, with accuracies of 77.3% on StreamingBench, 60.5% on OVBench, and 55.6% on OVO-Bench. In particular, even under extreme visual token compression (limited to 1024 tokens), the model retains 96.8% of its average accuracy in eight benchmarks relative to the default setting. These results underscore the robustness, efficiency, and generalizability of StreamForest for streaming video understanding.

Image-to-video generation has made remarkable progress with the advancements in diffusion models, yet generating videos with realistic motion remains highly challenging. This difficulty arises from the complexity of accurately modeling motion, which involves capturing physical constraints, object interactions, and domain-specific dynamics that are not easily generalized across diverse scenarios. To address this, we propose MotionRAG, a retrieval-augmented framework that enhances motion realism by adapting motion priors from relevant reference videos through Context-Aware Motion Adaptation (CAMA). The key technical innovations include: (i) a retrieval-based pipeline extracting high-level motion features using video encoder and specialized resamplers to distill semantic motion representations; (ii) an in-context learning approach for motion adaptation implemented through a causal transformer architecture; (iii) an attention-based motion injection adapter that seamlessly integrates transferred motion features into pretrained video diffusion models. Extensive experiments demonstrate that our method achieves significant improvements across multiple domains and various base models, all with negligible computational overhead during inference. Furthermore, our modular design enables zero-shot generalization to new domains by simply updating the retrieval database without retraining any components. This research enhances the core capability of video generation systems by enabling the effective retrieval and transfer of motion priors, facilitating the synthesis of realistic motion dynamics.

Abstract Audio-Visual Question Answering (AVQA) requires the model to answer questions with complex dynamic audio-visual information. Prior works on this task mainly consider only using single question-answer pairs during training, overlooking the rich semantic associations between questions. In this work, we propose a novel Collective Question-Guided Network (CoQo), which accepts multiple question-answer pairs as input and leverages the reasoning over these questions to assist the model training process. The core module is the proposed Question Guided Transformer (QGT), which uses collective question reasoning to perform question-guided feature extraction. Since multiple question-answer pairs are not always available, especially during inference, our QGT uses a set of learnable tokens to learn the collective information from multiple questions during training. At inference time, these learnable tokens bring additional reasoning information even when only one question is used as input. We employ QGT in both spatial and temporal dimensions to extract question-related features effectively and efficiently. To better capture detailed audio-visual associations, we train the model in a finer level by distinguishing feature pairs of different questions within the same video. Extensive experiments demonstrate that our method can achieve state-of-the-art performance on three AVQA datasets while reducing training time significantly. We also observe strong performances of our method on three VQA benchmarks. Detailed ablation studies further confirm the effectiveness of our proposed collective question reasoning scheme, both quantitatively and qualitatively.

The relation modeling between actors and scene context advances video action detection where the correlation of multiple actors makes their action recognition challenging. Existing studies model each actor and scene relation to improve action recognition. However, the scene variations and background interference limit their effectiveness. In this paper, we propose to select actor-related scene context, rather than directly laveraging raw video scenario, to improve relation modeling. We develop a Cycle Actor-Context Relation network (CycleACR) where there is a symmetric graph that models the actor and context relations in a bidirectional form. Specifically, our CycleACR is constituted of two modules: 1) Actor-to-Context Reorganization (A2C-R), which adaptively collects actor features for context feature reorganizations, and 2) Context-to-Actor Enhancement (C2A-E), which dynamically utilizes the reorganized context features for actor feature enhancement. Stacking multiple CycleACR modules is able to effectively capture the high-order relation and efficiently exchange useful information between actors and context. To fully exploit time-dependent and holistic context information, we further design a parallel local and global temporal context modeling branch. The outputs of the two branches are integrated as the final context-enhanced actor feature representations. Finally, we propose a context-aware memory bank for long-term relation modeling. The proposed bank can effectively store actor-related scene context from other clips without additional memory overhead. Compared to existing designs that focus on C2A-E, our CycleACR introduces the core design of A2C-R for more effective relation modeling. This cycle modeling enablesour CycleACR to achieve state-of-the-art performance on two popular action detection datasets: AVA (40.6 mAP) and UCF101-24 (84.7 mAP). We also provide ablation studies and visualizations to show how our cycle actor-context relation modeling improves video action detection.

Query-based object detectors and segmenters have made great progress in their respective tasks by employing an iterative refinement decoder. These query-based methods directly represent object instances with a set of learnable queries. These query vectors are progressively refined to stable, meaningful representations through a sequence of decoder layers, and then used to directly predict object locations (mask or box) and categories with customized heads. In this paper, we present a novel query-based object decoder design with infinite refinement (DEQ-Decoder) through a deep equilibrium model (DEQ). Our DEQ-Decoder models the query vector refinement as the fixed point solving of an implicit (DEQ) layer. To be more specific to query refinement, we use a two-step unrolled equilibrium equation to explicitly capture the query vector refinement. Accordingly, we are able to incorporate refinement awareness into the DEQ-Decoder training with the inexact gradient back-propagation (RAG). In addition, to stabilize the training of our DEQ-Decoder and improve its generalization ability, we devise a deep supervision scheme on the optimization path of DEQ-Decoder with refinement-aware perturbation (RAP). To demonstrate the effectiveness of DEQ-Decoder, we apply it to object detection and instance segmentation. For object detection, we propose DEQDet based on our DEQ-Decode. DEQDet converges faster, consumes less memory, and achieves better results than the baseline counterpart (AdaMixer). In particular, our DEQDet with ResNet50 backbone and 300 queries achieves the 49.6 mAP and 33.9 AP$_{s}$s on the MS COCO benchmark under $2\times$2× training scheme (24 epochs). For instance segmentation, Our DEQSeg achieves much better box mAP metrics and slightly better mask metrics for different mask decoding branches.

In this paper, we study the problem of procedure planning in instructional videos, which aims to make a plan (i.e. a sequence of actions) given the current visual observation and the desired goal. Previous works cast this as a sequence modeling problem and leverage either intermediate visual observations or language instructions as supervision to make autoregressive planning, resulting in complex learning schemes and expensive annotation costs. To avoid intermediate supervision annotation and error accumulation caused by planning autoregressively, we propose a diffusion-based framework, coined as PDPP (Projected Diffusion model for Procedure Planning), to directly model the whole action sequence distribution with task label as supervision instead. Our core idea is to treat procedure planning as a distribution fitting problem under the given observations, thus transform the planning problem to a sampling process from this distribution during inference. The diffusion-based modeling approach also effectively addresses the uncertainty issue in procedure planning. Based on PDPP, we further apply joint training to our framework to generate plans with varying horizon lengths using a single model and reduce the number of training parameters required. We instantiate our PDPP with three popular diffusion models and investigate a serious of condition-introducing methods in our framework, including condition embeddings, Mixture-of-Experts (MoEs), two-stage prediction and Classifier-Free Guidance strategy. Finally, we apply our PDPP to the Visual Planners for human Assistance (VPA) problem which requires the goal specified in natural language rather than visual observation. We conduct experiments on challenging datasets of different scales and our PDPP model achieves the state-of-the-art performance on multiple metrics, even compared with those strongly-supervised counterparts. These results further demonstrates the effectiveness and generalization ability of our model. Code and trained models are available at https://github.com/MCG-NJU/PDPP.

Current prevailing Video Object Segmentation (VOS) methods follow the pipeline of extraction-then-matching, which first extracts features on current and reference frames independently, and then performs dense matching between them. This decoupled pipeline limits information propagation between frames to high-level features, and fails to capture fine-grained details for matching. Furthermore, the pixel-wise matching lacks holistic target understanding, making it prone to disturbance by similar distractors. To address these issues, we propose a unified VOS framework, coined JointFormer, for jointly modeling feature extraction, correspondence matching, and a compressed memory. The Joint Modeling Block leverages attention operations to simultaneously extract and propagate the target information from the reference frame to the current frame and a compressed memory token.This joint modeling scheme enables extensive multi-layer propagation beyond high-level feature space and facilitates robust instance-distinctive feature learning. In addition, to incorporate the long-term and holistic target information, we introduce a compressed memory token with a customized online updating mechanism, which aggregates target features and performs temporal information propagation in a frame-wise manner, enhancing the global modeling consistency. Our JointFormer achieves a new state-of-the-art performance on the DAVIS 2017 val/test-dev (89.7% and 87.6%) benchmarks and the YouTube-VOS 2018/2019 val (87.0% and 87.0%) benchmarks. To demonstrate the generalizability of JointFormer, it is further evaluated on four new benchmarks with various challenges, including MOSE for complex scenes, VISOR for egocentric videos, VOST for complex transformations, and LVOS for long-term videos. Without specific design to address these unusual difficulties, our model achieves the best performance across all benchmarks when compared with several current best models, illustrating its excellent generalization and robustness. Further extensive ablations and visualizations indicate our JointFormer enables more comprehensive and effective feature learning and matching.

The task of weakly supervised temporal sentence grounding (WSTSG) aims to detect temporal intervals corresponding to a language description from untrimmed videos with only video-level video-language correspondence. For an anchor sample, most existing approaches generate negative samples either from other videos or within the same video for contrastive learning. However, some training samples are highly similar to the anchor sample, directly regarding them as negative samples leads to difficulties for optimization and ignores the correlations between these similar samples and the anchor sample. To address this, we propose Positive Sample Mining (PSM), a novel framework that mines positive samples from the training set to provide more discriminative supervision. Specifically, for a given anchor sample, we partition the remaining training set into semantically similar and dissimilar subsets based on the similarity of their text queries. To effectively leverage these correlations, we introduce a PSM-guided contrastive loss to ensure that the anchor proposal is closer to similar samples and further from dissimilar ones. Additionally, we design a PSM-guided rank loss to ensure that similar samples are closer to the anchor proposal than to the negative intra-video proposal, aiming to distinguish the anchor proposal and the negative intra-video proposal. Experiments on the WSTSG and grounded VideoQA tasks demonstrate the effectiveness and superiority of our method.

Diffusion models have made tremendous progress in text-driven image and video generation. Now text-to-image foundation models are widely applied to various down-stream image synthesis tasks, such as controllable image generation and image editing, while downstream video synthesis tasks are less explored for several reasons. First, it requires huge memory and computation overhead to train a video generation foundation model. Even with video foundation models, additional costly training is still required for downstream video synthesis tasks. Second, although some works extend image diffusion models into videos in a training-free manner, temporal consistency cannot be well preserved. Finally, these adaption methods are specifically designed for one task and fail to generalize to different tasks. To mitigate these issues, we propose a training-free general-purpose video synthesis framework, coined as BIVDiff, via bridging specific image diffusion models and general text-to-video foundation diffusion models. Specifically, we first use a specific image diffusion model (e.g., ControlNet and Instruct Pix2Pix) for frame-wise video generation, then perform Mixed Inversion on the generated video, and finally input the inverted latents into the video diffusion models (e.g., VidRD and ZeroScope) for temporal smoothing. This decoupled framework enables flexible image model selection for different purposes with strong task generalization and high efficiency. To validate the effectiveness and general use of BIVDiff, we perform a wide range of video synthesis tasks, including controllable video generation, video editing, video inpainting, and outpainting.

With the rapid development of Multi-modal Large language Models (MLLMs), a number of diagnostic bench-marks have recently emerged to evaluate the comprehension capabilities of these models. However, most bench-marks predominantly assess spatial understanding in the static image tasks, while overlooking temporal understanding in the dynamic video tasks. To alleviate this issue, we introduce a comprehensive Multi-modal Video understanding Benchmark, namely MVBench, which covers 20 chal-lenging video tasks that cannot be effectively solved with a single frame. Specifically, we first introduce a novel static-to-dynamic method to define these temporal-related tasks. By transforming various static tasks into dynamic ones, we enable the systematic generation of video tasks that require a broad spectrum of temporal skills, ranging from perception to cognition. Then, guided by the task definition, we au-tomatically convert public video annotations into multiple-choice QA to evaluate each task. On one hand, such a distinct paradigm allows us to build MVBench efficiently, without much manual intervention. On the other hand, it guarantees evaluation fairness with ground-truth video an-notations, avoiding the biased scoring of LLMs. More-over, we further develop a robust video MLLM baseline, i.e., VideoChat2, by progressive multi-modal training with di-verse instruction-tuning data. The extensive results on our MVBench reveal that, the existing MLLMs are far from sat-isfactory in temporal understanding, while our VideoChat2 largely surpasses these leading models by over 15% on MVBench. All models and data are available at https://github.com/OpenGVLab/Ask-Anything.

Neural rendering methods have significantly advanced photo-realistic 3D scene rendering in various academic and industrial applications. The recent 3D Gaussian Splatting method has achieved the state-of-the-art rendering quality and speed combining the benefits of both primitive-based representations and volumetric representations. However, it often leads to heavily redundant Gaussians that try to fit every training view, neglecting the underlying scene ge-ometry. Consequently, the resulting model becomes less robust to significant view changes, texture-less area and lighting effects. We introduce Scaffold-GS, which uses an-chor points to distribute local 3D Gaussians, and predicts their attributes on-the-fly based on viewing direction and distance within the view frustum. Anchor growing and pruning strategies are developed based on the importance of neural Gaussians to reliably improve the scene cover-age. We show that our method effectively reduces redun-dant Gaussians while delivering high-quality rendering. We also demonstrates an enhanced capability to accommodate scenes with varying levels-of-detail and view-dependent ob-servations, without sacrificing the rendering speed. Project page: https://city-super.github.iolscaffold-gsl.

Being able to map the activities of others into one's own point of view is a fundamental human skill even from a very early age. Taking a step toward understanding this human ability, we introduce EgoExoLearn, a large-scale dataset that emulates the human demonstration following process, in which individuals record egocentric videos as they execute tasks guided by exocentric-view demonstration videos. Focusing on the potential applications in daily assistance and professional support, EgoExoLearn contains egocentric and demonstration video data spanning 120 hours captured in daily life scenarios and specialized laboratories. Along with the videos we record high-quality gaze data and provide detailed multimodal annotations, formulating a playground for modeling the human ability to bridge asynchronous procedural actions from different viewpoints. To this end, we present benchmarks such as crossview association, cross-view action planning, and crossview referenced skill assessment, along with detailed analysis. We expect EgoExoLearn can serve as an important resource for bridging the actions across views, thus paving the way for creating AI agents capable of seamlessly learning by observing humans in the real world. The dataset and benchmark codes are available at https://github.com/OpenGVLab/EgoExoLearn.

Video generation has witnessed significant advance-ments, yet evaluating these models remains a challenge. A comprehensive evaluation benchmark for video generation is indispensable for two reasons: 1) Existing metrics do not fully align with human perceptions; 2) An ideal eval-uation system should provide insights to inform future de-velopments of video generation. To this end, we present VBench, a comprehensive benchmark suite that dissects “video generation quality” into specific, hierarchical, and disentangled dimensions, each with tailored prompts and evaluation methods. VBench has three appealing proper-ties: 1) Comprehensive Dimensions: VBench comprises 16 dimensions in video generation (e.g., subject identity in-consistency, motion smoothness, temporal flickering, and spatial relationship, etc.). The evaluation metrics with fine-grained levels reveal individual models' strengths and weaknesses. 2) Human Alignment: We also provide a dataset of human preference annotations to validate our benchmarks' alignment with human perception, for each evaluation dimension respectively. 3) Valuable Insights: We look into current models' ability across various evaluation dimensions, and various content types. We also investi-gate the gaps between video and image generation models. We will open-source VBench, including all prompts, evaluation methods, generated videos, and human preference an-notations, and also include more video generation models in VBench to drive forward the field of video generation.

Large motion poses a critical challenge in Video Frame Interpolation (VFI) task. Existing methods are often constrained by limited receptive fields, resulting in sub-optimal performance when handling scenarios with large motion. In this paper, we introduce a new pipeline for VFI, which can effectively integrate global-level information to alleviate issues associated with large motion. Specifically, we first estimate a pair of initial intermediate flows using a high-resolution feature map for extracting local details. Then, we incorporate a sparse global matching branch to compen-sate for flow estimation, which consists of identifying flaws in initial flows and generating sparse flow compensation with a global receptive field. Finally, we adaptively merge the initial flow estimation with global flow compensation, yielding a more accurate intermediate flow. To evaluate the effectiveness of our method in handling large motion, we carefully curate a more challenging subset from commonly used benchmarks. Our method demonstrates the state-of-the-art performance on these VFI subsets with large motion.

Video-based visual relation detection tasks, such as video scene graph generation, play important roles in fine-grained video understanding. However, current video visual relation detection datasets have two main limitations that hinder the progress of research in this area. First, they do not explore complex human-human interactions in multi-person scenarios. Second, the relation types of existing datasets have relatively low-level semantics and can be often recognized by appearance or simple prior information, without the need for detailed spatio-temporal context reasoning. Nevertheless, comprehending high-level interactions between humans is crucial for understanding complex multi-person videos, such as sports and surveillance videos. To address this issue, we propose a new video visual relation detection task: video human-human interaction detection, and build a dataset named SportsHHI for it. SportsHHI contains 34 high-level interaction classes from basketball and volleyball sports. 118,075 human bounding boxes and 50,649 interaction instances are annotated on 11,398 keyframes. To benchmark this, we propose a two-stage baseline method and conduct extensive experiments to reveal the key factors for a successful human-human interaction detector. We hope that SportsHHI can stimulate research on human interaction understanding in videos and promote the development of spatio-temporal context modeling techniques in video visual relation detection.

Vision Transformer (ViT) has shown high potential in video recognition, owing to its flexible design, adaptable self-attention mechanisms, and the efficacy of masked pretraining. Yet, it remains unclear how to adapt these pretrained short-term ViTs for temporal action detection (TAD) in untrimmed videos. The existing works treat them as off-the-shelf feature extractors for each short-trimmed snippet without capturing the fine-grained relation among different snippets in a broader temporal context. To mitigate this issue, this paper focuses on designing a new mechanism for adapting these pre-trained ViT models as a unified long-form video transformer to fully unleash its modeling power in capturing inter-snippet relation, while still keeping low computation overhead and memory consumption for efficient TAD. To this end, we design effective crosssnippet propagation modules to gradually exchange short-term video information among different snippets from two levels. For inner-backbone information propagation, we introduce a cross-snippet propagation strategy to enable multi-snippet temporal feature interaction inside the backbone. For post-backbone information propagation, we propose temporal transformer layers for further clip-level modeling. With the plain ViT-B pre-trained with VideoMAE, our end-to-end temporal action detector (ViT-TAD) yields a very competitive performance to previous temporal action detectors, riching up to 69.5 average mAP on THUMOS14, 37.40 average mAP on ActivityNet-1.3 and 17.20 average mAP on FineAction.

Self-supervised foundation models have shown great potential in computer vision thanks to the pre-training paradigm of masked autoencoding. Scale is a primary factor influencing the performance of these foundation models. However, these large foundation models often result in high computational cost. This paper focuses on pre-training relatively small vision transformer models that could be efficiently adapted to downstream tasks. Specifically, taking inspiration from knowledge distillation in model compression, we propose a new asymmetric masked distillation (AMD) framework for pre-training relatively small models with au-toencoding. The core of AMD is to devise an asymmetric masking strategy, where the teacher model is enabled to see more context information with a lower masking ratio, while the student model is still equipped with a high masking ratio. We design customized multi-layer feature alignment between the teacher encoder and student encoder to regularize the pre-training of student MAE. To demonstrate the effectiveness and versatility of AMD, we apply it to both ImageMAE and VideoMAE for pre-training relatively small ViT models. AMD achieved 84.6% classification accuracy on IN1K using the ViT-B model. And AMD achieves 73.3% classification accuracy using the ViT-B model on the Something-in-Something V2 dataset, a 3.7% improvement over the original ViT-B model from VideoMAE. We also transfer AMD pre-trained models to downstream tasks and obtain consistent performance improvement over the original masked autoencoding. The code and models are available at https://github.com/MCG-NJU/AMD.

Temporal Action Detection (TAD) aims to identify the action boundaries and the corresponding category within untrimmed videos. Inspired by the success of DETR in object detection, several methods have adapted the query-based framework to the TAD task. However, these approaches primarily followed DETR to predict actions at the instance level (i.e., identify each action by its center point), leading to sub-optimal boundary localization. To address this issue, we propose a new Dual-level query-based TAD framework, namely DualDETR, to detect actions from both instance-level and boundary-level. Decoding at different levels requires semantics of different granularity, therefore we introduce a two-branch decoding structure. This structure builds distinctive decoding processes for different lev-els, facilitating explicit capture of temporal cues and se-mantics at each level. On top of the two-branch design, we present a joint query initialization strategy to align queries from both levels. Specifically, we leverage encoder propos-als to match queries from each level in a one-to-one man-ner. Then, the matched queries are initialized using position and content prior from the matched action proposal. The aligned dual-level queries can refine the matched proposal with complementary cues during subsequent decoding. We evaluate DualDETR on three challenging multi-label TAD benchmarks. The experimental results demonstrate the su-perior performance of DualDETR to the existing state-of-the-art methods, achieving a substantial improvement under det-mAP and delivering impressive results under seg-mAP.

Human visual recognition is a sparse process, where only a few salient visual cues are attended to rather than every detail being traversed uniformly. However, most current vision networks follow a dense paradigm, processing every single visual unit (such as pixels or patches) in a uniform manner. In this paper, we challenge this dense convention and present a new vision transformer, coined SparseFormer, to explicitly imitate human's sparse visual recognition in an end-to-end manner. SparseFormer learns to represent images using a highly limited number of tokens (e.g., down to $9$) in the latent space with sparse feature sampling procedure instead of processing dense units in the original image space. Therefore, SparseFormer circumvents most of dense operations on the image space and has much lower computational costs. Experiments on the ImageNet-1K classification show that SparseFormer delivers performance on par with canonical or well-established models while offering more favorable accuracy-throughput tradeoff. Moreover, the design of our network can be easily extended to the video classification task with promising performance with lower compute. We hope our work can provide an alternative way for visual modeling and inspire further research on sparse vision architectures. Code and weights are available at https://github.com/showlab/sparseformer.

This paper introduces InternVid, a large-scale video-centric multimodal dataset that enables learning powerful and transferable video-text representations for multimodal understanding and generation. InternVid contains over 7 million videos lasting nearly 760K hours, yielding 234M video clips accompanied by detailed descriptions of total 4.1B words. Our core contribution is to develop a scalable approach to autonomously build a high-quality video-text dataset with large language models (LLM), thereby showcasing its efficacy in learning video-language representation at scale. Specifically, we utilize a multi-scale approach to generate video-related descriptions. Furthermore, we introduce ViCLIP, a video-text representation learning model based on ViT-L. Learned on InternVid via contrastive learning, this model demonstrates leading zero-shot action recognition and competitive video retrieval performance. Beyond basic video understanding tasks like recognition and retrieval, our dataset and model have broad applications. They are particularly beneficial for generating interleaved video-text data for learning a video-centric dialogue system, advancing video-to-text and text-to-video generation research. These proposed resources provide a tool for researchers and practitioners interested in multimodal video understanding and generation.

Arbitrary-resolution image generation still remains a challenging task in AIGC, as it requires handling varying resolutions and aspect ratios while maintaining high visual quality. Existing transformer-based diffusion methods suffer from quadratic computation cost and limited resolution extrapolation capabilities, making them less effective for this task. In this paper, we propose FlowDCN, a purely convolution-based generative model with linear time and memory complexity, that can efficiently generate high-quality images at arbitrary resolutions. Equipped with a new design of learnable group-wise deformable convolution block, our FlowDCN yields higher flexibility and capability to handle different resolutions with a single model. FlowDCN achieves the state-of-the-art 4.30 sFID on $256\times256$ ImageNet Benchmark and comparable resolution extrapolation results, surpassing transformer-based counterparts in terms of convergence speed (only $\frac{1}{5}$ images), visual quality, parameters ($8\%$ reduction) and FLOPs ($20\%$ reduction). We believe FlowDCN offers a promising solution to scalable and flexible image synthesis.

Inter-frame modeling is pivotal in generating intermediate frames for video frame interpolation (VFI). Current approaches predominantly rely on convolution or attention-based models, which often either lack sufficient receptive fields or entail significant computational overheads. Recently, Selective State Space Models (S6) have emerged, tailored specifically for long sequence modeling, offering both linear complexity and data-dependent modeling capabilities. In this paper, we propose VFIMamba, a novel frame interpolation method for efficient and dynamic inter-frame modeling by harnessing the S6 model. Our approach introduces the Mixed-SSM Block (MSB), which initially rearranges tokens from adjacent frames in an interleaved fashion and subsequently applies multi-directional S6 modeling. This design facilitates the efficient transmission of information across frames while upholding linear complexity. Furthermore, we introduce a novel curriculum learning strategy that progressively cultivates proficiency in modeling inter-frame dynamics across varying motion magnitudes, fully unleashing the potential of the S6 model. Experimental findings showcase that our method attains state-of-the-art performance across diverse benchmarks, particularly excelling in high-resolution scenarios. In particular, on the X-TEST dataset, VFIMamba demonstrates a noteworthy improvement of 0.80 dB for 4K frames and 0.96 dB for 2K frames.

Video understanding relies on accurate action detection for temporal analysis. However, existing mainstream methods have limitations in real-world applications due to their offline and closed-set evaluation approaches, as well as their dependence on manual annotations. To address these challenges and enable real-time action understanding in open-world scenarios, we propose OV-OAD, a zero-shot online action detector that leverages vision-language models and learns solely from text supervision. By introducing an object-centered decoder unit into a Transformer-based model, we aggregate frames with similar semantics using video-text correspondence. Extensive experiments on four action detection benchmarks demonstrate that OV-OAD outperforms other advanced zero-shot methods. Specifically, it achieves 37.5\% mean average precision on THUMOS’14 and 73.8\% calibrated average precision on TVSeries. This research establishes a robust baseline for zero-shot transfer in online action detection, enabling scalable solutions for open-world temporal understanding. The code will be available for download at \url{https://github.com/OpenGVLab/OV-OAD}.

Pre-trained vision-language models (VLMs) have shown impressive results in various visual classification tasks. However, we often fail to fully unleash their potential when adapting them for new concept understanding due to limited information on new classes. To address this limitation, we introduce a novel adaptation framework, AWT (Augment, Weight, then Transport). AWT comprises three key components: augmenting inputs with diverse visual perspectives and enriched class descriptions through image transformations and language models; dynamically weighting inputs based on the prediction entropy; and employing optimal transport to mine semantic correlations in the vision-language space. AWT can be seamlessly integrated into various VLMs, enhancing their zero-shot capabilities without additional training and facilitating few-shot learning through an integrated multimodal adapter module. We verify AWT in multiple challenging scenarios, including zero-shot and few-shot image classification, zero-shot video action recognition, and out-of-distribution generalization. AWT consistently outperforms the state-of-the-art methods in each setting. In addition, our extensive studies further demonstrate AWT's effectiveness and adaptability across different VLMs, architectures, and scales.

Visual object tracking often employs a multi-stage pipeline of feature extraction, target information integration, and bounding box estimation. To simplify this pipeline and unify the process of feature extraction and target information integration, in this paper, we present a compact tracking framework, termed as MixFormer, built upon transformers. Our core design is to utilize the flexibility of attention operations, and we propose a Mixed Attention Module (MAM) for simultaneous feature extraction and target information integration. This synchronous modeling scheme allows us to extract target-specific discriminative features and perform extensive communication between target and search area. Based on MAM, we build our MixFormer trackers simply by stacking multiple MAMs and placing a localization head on top. Specifically, we instantiate two types of MixFormer trackers, a hierarchical tracker MixCvT, and a non-hierarchical simple tracker MixViT. For these two trackers, we investigate a series of pre-training methods and uncover the different behaviors between supervised pre-training and self-supervised pre-training in our MixFormer trackers. We also extend the masked autoencoder pre-training to our MixFormer trackers and design the new competitive TrackMAE pre-training technique. Finally, to handle multiple target templates during online tracking, we devise an asymmetric attention scheme in MAM to reduce computational cost, and propose an effective score prediction module to select high-quality templates. Our MixFormer trackers set a new state-of-the-art performance on seven tracking benchmarks, including LaSOT, TrackingNet, VOT2020, GOT-10 k, OTB100, TOTB and UAV123. In particular, our MixViT-L achieves AUC scores of 73.3% on LaSOT, 86.1% on TrackingNet and 82.8% on TOTB.

In this paper, we study the problem of jointly estimating the optical flow and scene flow from synchronized 2D and 3D data. Previous methods either employ a complex pipeline that splits the joint task into independent stages, or fuse 2D and 3D information in an "early-fusion" or "late-fusion" manner. Such one-size-fits-all approaches suffer from a dilemma of failing to fully utilize the characteristic of each modality or to maximize the inter-modality complementarity. To address the problem, we propose a novel end-to-end framework, which consists of 2D and 3D branches with multiple bidirectional fusion connections between them in specific layers. Different from previous work, we apply a point-based 3D branch to extract the LiDAR features, as it preserves the geometric structure of point clouds. To fuse dense image features and sparse point features, we propose a learnable operator named bidirectional camera-LiDAR fusion module (Bi-CLFM). We instantiate two types of the bidirectional fusion pipeline, one based on the pyramidal coarse-to-fine architecture (dubbed CamLiPWC), and the other one based on the recurrent all-pairs field transforms (dubbed CamLiRAFT). On FlyingThings3D, both CamLiPWC and CamLiRAFT surpass all existing methods and achieve up to a 47.9% reduction in 3D end-point-error from the best published result. Our best-performing model, CamLiRAFT, achieves an error of 4.26% on the KITTI Scene Flow benchmark, ranking 1st among all submissions with much fewer parameters. Besides, our methods have strong generalization performance and the ability to handle non-rigid motion.

Multimodal transformer exhibits high capacity and flexibility to align image and text for visual grounding. However, the existing encoder-only grounding framework (e.g., TransVG) suffers from heavy computation due to the self-attention operation with quadratic time complexity. To address this issue, we present a new multimodal transformer architecture, coined as Dynamic Mutilmodal detection transformer (DETR) (Dynamic MDETR), by decoupling the whole grounding process into encoding and decoding phases. The key observation is that there exists high spatial redundancy in images. Thus, we devise a new dynamic multimodal transformer decoder by exploiting this sparsity prior to speed up the visual grounding process. Specifically, our dynamic decoder is composed of a 2D adaptive sampling module and a text guided decoding module. The sampling module aims to select these informative patches by predicting the offsets with respect to a reference point, while the decoding module works for extracting the grounded object information by performing cross attention between image features and text features. These two modules are stacked alternatively to gradually bridge the modality gap and iteratively refine the reference point of grounded object, eventually realizing the objective of visual grounding. Extensive experiments on five benchmarks demonstrate that our proposed Dynamic MDETR achieves competitive trade-offs between computation and accuracy. Notably, using only 9% feature points in the decoder, we can reduce ∼ 44% GFLOPs of the multimodal transformer, but still get higher accuracy than the encoder-only counterpart. With the same number of encoder layers as TransVG, our Dynamic MDETR (ResNet-50) outperforms TransVG (ResNet-101) but only brings marginal extra computational cost relative to TransVG. In addition, to verify its generalization ability and scale up our Dynamic MDETR, we build the first one-stage CLIP empowered visual grounding framework, and achieve the state-of-the-art performance on these benchmarks.

Traditional video action detectors typically adopt the two-stage pipeline, where a person detector is first employed to generate actor boxes and then 3D RoIAlign is used to extract actor-specific features for action recognition. This detection paradigm requires multi-stage training and inference, and the feature sampling is only constrained inside the box, failing to effectively leverage richer context information outside. Recently, several query-based action detectors have been proposed to predict action instances in an end-to-end manner. However, they still lack adaptability in feature sampling and decoding, thus suffering from the issues of inferior performance or slower convergence. In this paper, we propose two core designs for a more flexible one-stage sparse action detector. First, we present a query-based adaptive feature sampling module, which endows the detector with the flexibility of mining a group of discriminative features from the entire spatio-temporal domain. Second, we devise a decoupled feature mixing module, which dynamically attends to and mixes video features along the spatial and temporal dimensions respectively for better feature decoding. Based on these designs, we instantiate two detection pipelines, that is, STMixer-K for keyframe action detection and STMixer-T for action tubelet detection. Without bells and whistles, our STMixer detectors obtain the state-of-the-art results on five challenging spatio-temporal action detection benchmarks for keyframe action detection or action tube detection.

Action tube detection is a challenging task as it requires not only to locate action instances in each frame, but also link them in time. Existing action tube detection methods often employ multi-stage pipelines with complex designs and time-consuming linking procedure. In this paper, we present a simple end-to-end action tube detection method, termed as Sparse Tube Detector (STDet). Unlike those dense action detectors, our core idea is to use a set of learnable tube queries and directly decode them into action tubes (i.e., a set of tracked boxes with action label) from video content. This sparse detection paradigm shares several advantages. First, the large number of hand-crafted anchor candidates in dense action detectors is greatly reduced to a small number of learnable tubes, which results in a more efficient detection framework. Second, our learnable tube queries directly attend the whole video content, which endows our method with the capacity of capturing long-range information for action detection. Finally, our action detector is an end-to-end tube detection without requiring the linking procedure, which directly and explicitly predicts the action boundary instead of depending on the linking strategy. Extensive experiments shows that our STDet outperforms the previous state-of-the-art methods on two challenging untrimmed video action detection datasets of UCF101-24 and MultiSports. We hope our method will be an simple end-to-end tube detection baseline and can inspire new ideas in this direction.

Discovering temporal boundary is critical for untrimmed video tasks, such as temporal sentence grounding and action detection. Due to the labor-intensive boundary annotations, the recent studies focus on the weakly-supervised setting, with only sentences or action tags in the training videos. However, how to align temporal boundaries and textual descriptions is problematic in most weakly-supervised approaches. To alleviate this difficulty, we propose a novel Dual Masked Modeling (DM2) framework, which can effectively enhance clip-text alignment to boost temporal boundary discovery, by cross-modal masked modeling in the dual fashion. Specifically, we introduce two coupled reconstruction branches, i.e., Clip-Aware Masked Text Modeling (C-MTM), and Text-Aware Masked Clip Modeling (T-MCM), after generating a temporal proposal of the underlying clip. In C-MTM, we recover the masked text with visual assistance of the clip proposal. In T-MCM, we recover the masked clip proposal with lingual assistance of the text. Via such complementary reconstruction supervision, our DM2 can cooperatively exploit robust matching between the video clip and the referred text, allowing to unify grounding and localization in a concise manner. Finally, we perform extensive experiments on the popular temporal benchmarks, i.e., Charades-STA, ActivityNet Captions, ActivityNet-v1.3 and THUMOS-14. Our DM2 achieves state-of-the-art for both weakly-supervised temporal grounding and localization. Codes and models will be released afterward.

Current RGB-D scene recognition approaches often train two standalone backbones for RGB and depth modalities with the same Places or ImageNet pre-training. However, the pre-trained depth network is still biased by RGB-based models which may result in a suboptimal solution. In this paper, we present a single-model self-supervised hybrid pre-training framework for RGB and depth modalities, termed as CoMAE. Our CoMAE presents a curriculum learning strategy to unify the two popular self-supervised representation learning algorithms: contrastive learning and masked image modeling. Specifically, we first build a patch-level alignment task to pre-train a single encoder shared by two modalities via cross-modal contrastive learning. Then, the pre-trained contrastive encoder is passed to a multi-modal masked autoencoder to capture the finer context features from a generative perspective. In addition, our single-model design without requirement of fusion module is very flexible and robust to generalize to unimodal scenario in both training and testing phases. Extensive experiments on SUN RGB-D and NYUDv2 datasets demonstrate the effectiveness of our CoMAE for RGB and depth representation learning. In addition, our experiment results reveal that CoMAE is a data-efficient representation learner. Although we only use the small-scale and unlabeled training set for pre-training, our CoMAE pre-trained models are still competitive to the state-of-the-art methods with extra large-scale and supervised RGB dataset pre-training. Code will be released at https://github.com/MCG-NJU/CoMAE.

Extending the success of 2D Large Kernel to 3D perception is challenging due to: 1. the cubically-increasing overhead in processing 3D data; 2. the optimization difficulties from data scarcity and sparsity. Previous work has taken the first step to scale up the kernel size from 3 × 3 × 3 to 7 × 7 × 7 by introducing block-shared weights. However, to reduce the feature variations within a block, it only employs modest block size and fails to achieve larger kernels like the 21 × 21 × 21. To address this issue, we propose a new method, called LinK, to achieve a wider-range perception receptive field in a convolution-like manner with two core designs. The first is to replace the static kernel matrix with a linear kernel generator, which adaptively provides weights only for non-empty voxels. The second is to reuse the pre-computed aggregation results in the overlapped blocks to reduce computation complexity. The proposed method successfully enables each voxel to perceive context within a range of 21 × 21 × 21. Extensive experiments on two basic perception tasks, 3D object detection and 3D semantic segmentation, demonstrate the effectiveness of our method. Notably, we rank 1st on the public leaderboard of the 3D detection benchmark of nuScenes (LiDAR track), by simply incorporating a LinK-based backbone into the basic detector, CenterPoint. We also boost the strong segmentation baseline's mIoU with 2.7% in the SemanticKITTI test set. Code is available at https://github.com/MCG-NJU/LinK.

Scale is the primary factor for building a powerful foundation model that could well generalize to a variety of downstream tasks. However, it is still challenging to train video foundation models with billions of parameters. This paper shows that video masked autoencoder (VideoMAE) is a scalable and general self-supervised pre-trainer for building video foundation models. We scale the VideoMAE in both model and data with a core design. Specifically, we present a dual masking strategy for efficient pre-training, with an encoder operating on a subset of video tokens and a decoder processing another subset of video tokens. Although VideoMAE is very efficient due to high masking ratio in encoder, masking decoder can still further reduce the overall computational cost. This enables the efficient pre-training of billion-level models in video. We also use a progressive training paradigm that involves an initial pre-training on a diverse multi-sourced unlabeled dataset, followed by a post-pre-training on a mixed labeled dataset. Finally, we successfully train a video ViT model with a billion parameters, which achieves a new state-of-the-art performance on the datasets of Kinetics (90.0% on K400 and 89.9% on K600) and Something-Something (68.7% on V1 and 77.0% on V2). In addition, we extensively verify the pre-trained video ViT models on a variety of downstream tasks, demonstrating its effectiveness as a general video representation learner.

In this paper, we study the problem of procedure planning in instructional videos, which aims to make goal-directed plans given the current visual observations in unstructured real-life videos. Previous works cast this problem as a sequence planning problem and leverage either heavy intermediate visual observations or natural language instructions as supervision, resulting in complex learning schemes and expensive annotation costs. In contrast, we treat this problem as a distribution fitting problem. In this sense, we model the whole intermediate action sequence distribution with a diffusion model (PDPP), and thus transform the planning problem to a sampling process from this distribution. In addition, we remove the expensive intermediate supervision, and simply use task labels from instructional videos as supervision instead. Our model is a U-Net based diffusion model, which directly samples action sequences from the learned distribution with the given start and end observations. Furthermore, we apply an efficient projection method to provide accurate conditional guides for our model during the learning and sampling process. Experiments on three datasets with different scales show that our PDPP model can achieve the state-of-the-art performance on multiple metrics, even without the task supervision. Code and trained models are available at https://github.com/MCG-NJU/PDPP.

Traditional video action detectors typically adopt the two-stage pipeline, where a person detector is first employed to generate actor boxes and then 3D RoIAlign is used to extract actor-specific features for classification. This detection paradigm requires multi-stage training and inference, and cannot capture context information outside the bounding box. Recently, a few query-based action detectors are proposed to predict action instances in an end-to-end manner. However, they still lack adaptability in feature sampling and decoding, thus suffering from the issues of inferior performance or slower convergence. In this paper, we propose a new one-stage sparse action detector, termed STMixer. STMixer is based on two core designs. First, we present a query-based adaptive feature sampling module, which endows our STMixer with the flexibility of mining a set of discriminative features from the entire spatiotemporal domain. Second, we devise a dual-branch feature mixing module, which allows our STMixer to dynamically attend to and mix video features along the spatial and the temporal dimension respectively for better feature decoding. Coupling these two designs with a video backbone yields an efficient end-to-end action detector. Without bells and whistles, our STMixer obtains the state-of-the-art results on the datasets of AVA, UCF101-24, and JHMDB.

Effectively extracting inter-frame motion and appearance information is important for video frame interpolation (VFI). Previous works either extract both types of information in a mixed way or devise separate modules for each type of information, which lead to representation ambiguity and low efficiency. In this paper, we propose a new module to explicitly extract motion and appearance information via a unified operation. Specifically, we rethink the information process in inter-frame attention and reuse its attention map for both appearance feature enhancement and motion information extraction. Furthermore, for efficient VFI, our proposed module could be seamlessly integrated into a hybrid CNN and Transformer architecture. This hybrid pipeline can alleviate the computational complexity of inter-frame attention as well as preserve detailed low-level structure information. Experimental results demonstrate that, for both fixed- and arbitrary-timestep interpolation, our method achieves state-of-the-art performance on various datasets. Meanwhile, our approach enjoys a lighter computation overhead over models with close performance. The source code and models are available at https://github.com/MCG-NJU/EMA-VFI.

Streaming video clips with large-scale video tokens impede vision transformers (ViTs) for efficient recognition, especially in video action detection where sufficient spatiotemporal representations are required for precise actor identification. In this work, we propose an end-to-end framework for efficient video action detection (EVAD) based on vanilla ViTs. Our EVAD consists of two specialized designs for video action detection. First, we propose a spatiotemporal token dropout from a keyframe-centric perspective. In a video clip, we maintain all tokens from its keyframe, preserve tokens relevant to actor motions from other frames, and drop out the remaining tokens in this clip. Second, we refine scene context by leveraging remaining tokens for better recognizing actor identities. The region of interest (RoI) in our action detector is expanded into temporal domain. The captured spatiotemporal actor identity representations are refined via scene context in a decoder with the attention mechanism. These two designs make our EVAD efficient while maintaining accuracy, which is validated on three benchmark datasets (i.e., AVA, UCF101-24, JHMDB). Compared to the vanilla ViT backbone, our EVAD reduces the overall GFLOPs by 43% and improves real-time inference speed by 40% with no performance degradation. Moreover, even at similar computational costs, our EVAD can improve the performance by 1.1 mAP with higher resolution inputs. Code is available at https://github.com/MCG-NJU/EVAD.

Multi-object tracking (MOT) in sports scenes plays a critical role in gathering players statistics, supporting further applications, such as automatic tactical analysis. Yet existing MOT benchmarks cast little attention on this domain. In this work, we present a new large-scale multi-object tracking dataset in multiple sports scenes, coined as SportsMOT, where all players on the court are supposed to be tracked. It consists of 240 video sequences, over 150K frames (almost 15x MOT17) and over 1.6M bounding boxes (3x MOT17) collected from 3 sports categories, including basketball, volleyball and football. Our dataset is characterized with two key properties: 1) fast and variable-speed motion and 2) similar yet distinguishable appearance. We expect SportsMOT to encourage the MOT trackers to promote in both motion-based association and appearance-based association. We benchmark several state-of-the-art trackers and reveal the key challenge of SportsMOT lies in object association. To alleviate the issue, we further propose a new multi-object tracking framework, termed as MixSort, introducing a MixFormer-like structure as an auxiliary association model to prevailing tracking-by-detection trackers. By integrating the customized appearance-based association with the original motion-based association, MixSort achieves state-of-the-art performance on SportsMOT and MOT17. Based on MixSort, we give an in-depth analysis and provide some profound insights into SportsMOT.

As a video task, Multiple Object Tracking (MOT) is expected to capture temporal information of targets effectively. Unfortunately, most existing methods only explicitly exploit the object features between adjacent frames, while lacking the capacity to model long-term temporal information. In this paper, we propose MeMOTR, a long-term memory-augmented Transformer for multi-object tracking. Our method is able to make the same object's track embedding more stable and distinguishable by leveraging longterm memory injection with a customized memory-attention layer. This significantly improves the target association ability of our model. Experimental results on DanceTrack show that MeMOTR impressively surpasses the state-of-the-art method by 7.9% and 13.0% on HOTA and AssA metrics, respectively. Furthermore, our model also outperforms other Transformer-based methods on association performance on MOT17 and generalizes well on BDD100K. Code is available at https://github.com/MCG-NJU/MeMOTR.

Masked autoencoding has shown excellent performance on self-supervised video representation learning. Temporal redundancy has led to a high masking ratio and customized masking strategy in VideoMAE. In this paper, we aim to further improve the performance of video masked autoencoding by introducing a motion guided masking strategy. Our key insight is that motion is a general and unique prior in video, which should be taken into account during masked pre-training. Our motion guided masking explicitly incorporates motion information to build temporal consistent masking volume. Based on this masking volume, we can track the unmasked tokens in time and sample a set of temporal consistent cubes from videos. These temporal aligned unmasked tokens will further relieve the information leakage issue in time and encourage the MGMAE to learn more useful structure information. We implement our MGMAE with an online efficient optical flow estimator and backward masking map warping strategy. We perform experiments on the datasets of Something-Something V2 and Kinetics-400, demonstrating the superior performance of our MGMAE to the original VideoMAE. In addition, we provide the visualization analysis to illustrate that our MGMAE can sample temporal consistent cubes in a motion-adaptive manner for more effective video pre-training.

Video Foundation Models (VFMs) have received limited exploration due to high computational costs and data scarcity. Previous VFMs rely on Image Foundation Models (IFMs), which face challenges in transferring to the video domain. Although VideoMAE has trained a robust ViT from limited data, its low-level reconstruction poses convergence difficulties and conflicts with high-level cross-modal alignment. This paper proposes a training-efficient method for temporal-sensitive VFMs that integrates the benefits of existing methods. To increase data efficiency, we mask out most of the low-semantics video tokens, but selectively align the unmasked tokens with IFM, which serves as the UnMasked Teacher (UMT). By providing semantic guidance, our method enables faster convergence and multi-modal friendliness. With a progressive pre-training framework, our model can handle various tasks including scene-related, temporal-related, and complex video-language understanding. Using only public sources for pre-training in 6 days on 32 A100 GPUs, our scratch-built ViT-L/16 achieves state-of-the-art performances on various video tasks.

The prolific performances of Vision Transformers (ViTs) in image tasks have prompted research into adapting the image ViTs for video tasks. However, the substantial gap between image and video impedes the spatiotemporal learning of these image-pretrained models. Though video-specialized models like UniFormer can transfer to the video domain more seamlessly, their unique architectures require prolonged image pretraining, limiting the scalability. Given the emergence of powerful open-source image ViTs, we propose unlocking their potential for video understanding with efficient UniFormer designs. We call the resulting model UniFormerV2, since it inherits the concise style of the Uni-Former block, while redesigning local and global relation aggregators that seamlessly integrate advantages from both ViTs and UniFormer. Our UniFormerV2 achieves state-of-the-art performances on 8 popular video benchmarks, including scene-related Kinetics-400/600/700, heterogeneous Moments in Time, temporal-related Something-Something V1/V2, and untrimmed ActivityNet and HACS. It is note-worthy that to the best of our knowledge, UniFormerV2 is the first to elicit 90% top-1 accuracy on Kinetics-400.

Camera-based 3D object detection in BEV (Bird’s Eye View) space has drawn great attention over the past few years. Dense detectors typically follow a two-stage pipeline by first constructing a dense BEV feature and then performing object detection in BEV space, which suffers from complex view transformations and high computation cost. On the other side, sparse detectors follow a query-based paradigm without explicit dense BEV feature construction, but achieve worse performance than the dense counterparts. In this paper, we find that the key to mitigate this performance gap is the adaptability of the detector in both BEV and image space. To achieve this goal, we propose SparseBEV, a fully sparse 3D object detector that outperforms the dense counterparts. SparseBEV contains three key designs, which are (1) scale-adaptive self attention to aggregate features with adaptive receptive field in BEV space, (2) adaptive spatio-temporal sampling to generate sampling locations under the guidance of queries, and (3) adaptive mixing to decode the sampled features with dynamic weights from the queries. On the test split of nuScenes, SparseBEV achieves the state-of-the-art performance of 67.5 NDS. On the val split, SparseBEV achieves 55.8 NDS while maintaining a real-time inference speed of 23.5 FPS. Code is available at https://github.com/MCG-NJU/SparseBEV.

Previous object detectors make predictions based on dense grid points or numerous preset anchors. Most of these detectors are trained with one-to-many label assignment strategies. On the contrary, recent query-based object detectors are based a sparse set of learnable queries refined by a series of decoder layers. The one-to-one label assignment is independently applied on each layer for deep supervision during training. Despite the great success of query-based object detection, however, this vanilla one-to-one label assignment strategy requires the detectors to have strong fine-grained discrimination and modeling capacity. In this paper, we propose a new query-based object detector with cross-stage interaction, coined as StageInter-actor. During the forward pass, we come up with an efficient way to improve this modeling ability by reusing dynamic operators with lightweight adapters. As for the label assignment, a cross-stage label assigner is designed to improve the one-to-one label assignment. With this assigner, the training target class labels are gathered across stages and then reallocated to proper predictions at each decoder layer. On MS COCO benchmark, our model improves the baseline counterpart by 2.2 AP, and achieves a 44.8 AP with ResNet-50 as backbone, 100 queries and 12 training epochs. With longer training time and 300 queries, StageIn-teractor achieves 51.3 AP and 52.7 AP with ResNeXt-101-DCN and Swin-S, respectively. The code and models are made available at https://github.com/MCG-NJU/StageInteractor.

Most existing forecasting systems are memory-based methods, which attempt to mimic human forecasting ability by employing various memory mechanisms and have progressed in temporal modeling for memory dependency. Nevertheless, an obvious weakness of this paradigm is that it can only model limited historical dependence and can not transcend the past. In this paper, we rethink the temporal dependence of event evolution and propose a novel memory-anticipation-based paradigm to model an entire temporal structure, including the past, present, and future. Based on this idea, we present Memory-and-Anticipation Transformer (MAT), a memory-anticipation-based approach, to address the online action detection and anticipation tasks. In addition, owing to the inherent superiority of MAT, it can process online action detection and anticipation tasks in a unified manner. The proposed MAT model is tested on four challenging benchmarks TVSeries, THUMOS’14, HDD, and EPIC-Kitchens-100, for online action detection and anticipation tasks, and it significantly outperforms all existing methods. Code is available at https://github.com/Echo0125/Memory-and-Anticipation-Transformer.

Query-based object detectors directly decode image features into object instances with a set of learnable queries. These query vectors are progressively refined to stable meaningful representations through a sequence of decoder layers, and then used to directly predict object locations and categories with simple FFN heads. In this paper, we present a new query-based object detector (DEQDet) by designing a deep equilibrium decoder. Our DEQ decoder models the query vector refinement as the fixed point solving of an implicit layer and is equivalent to applying infinite steps of refinement. To be more specific to object decoding, we use a two-step unrolled equilibrium equation to explicitly capture the query vector refinement. Accordingly, we are able to incorporate refinement awareness into the DEQ training with the inexact gradient back-propagation (RAG). In addition, to stabilize the training of our DEQDet and improve its generalization ability, we devise the deep supervision scheme on the optimization path of DEQ with refinement-aware perturbation (RAP). Our experiments demonstrate DEQDet converges faster, consumes less memory, and achieves better results than the baseline counterpart (AdaMixer). In particular, our DEQDet with ResNet50 backbone and 300 queries achieves the 49.5 mAP and 33.0 APSon the MS COCO benchmark under 2× training scheme (24 epochs).

Temporal action detection (TAD) aims to localize the start and end frames of actions in untrimmed videos, which is a challenging task due to the similarity of adjacent frames and the ambiguity of action boundaries. Previous methods often generate coarse proposals first and then perform proposal-based refinement, which is coupled with prior action detectors and leads to proposal-oriented offsets. However, this paradigm increases the training difficulty of the TAD model and is heavily influenced by the quantity and quality of the proposals. To address the above issues, we decouple the refinement process from conventional TAD methods and propose a learnable, proposal-free refinement method for fine boundary localization, named RefineTAD. We first propose a multi-level refinement module to generate multi-scale boundary offsets, score offsets and boundary-aware probability at each time point based on the feature pyramid. Then, we propose an offset focusing strategy to progressively refine the predicted results of TAD models in a coarse-to-fine manner with our multi-scale offsets. We perform extensive experiments on three challenging datasets and demonstrate that our RefineTAD significantly improves the state-of-the-art TAD methods with minimal computational overhead.

Open set action recognition (OSAR) is a challenging task that requires a classifier to identify actions that do not belong to any of the classes in its training set. Existing methods employ the Evidential Neural Network (ENN) as an open-set classifier, which is trained in a supervised manner on feature representations from known classes to quantify the predictive uncertainty of human actions. In this paper, we propose a novel framework for OSAR that enriches the discriminative representation from a backbone with a reconstructive one to further improve performance. Our approach involves augmenting the input features with their reconstruction obtained from a reconstruction-based model in unsupervised training on known classes. We then use the correspondence between the two features to learn the open-set classifier, forcing it to associate low correspondence both when the feature is from unknown classes as well as when the input feature and its reconstruction variant are inconsistent with each other. Our experimental results on standard OSAR benchmarks demonstrate that our end-to-end trained model significantly outperforms state-of-the-art methods. Our proposed approach shows the effectiveness of combining discriminative and reconstructive representations for OSAR.

Transformer-based trackers have achieved strong accuracy on the standard benchmarks. However, their efficiency remains an obstacle to practical deployment on both GPU and CPU platforms. In this paper, to overcome this issue, we propose a fully transformer tracking framework, coined as \emph{MixFormerV2}, without any dense convolutional operation and complex score prediction module. Our key design is to introduce four special prediction tokens and concatenate them with the tokens from target template and search areas. Then, we apply the unified transformer backbone on these mixed token sequence. These prediction tokens are able to capture the complex correlation between target template and search area via mixed attentions. Based on them, we can easily predict the tracking box and estimate its confidence score through simple MLP heads. To further improve the efficiency of MixFormerV2, we present a new distillation-based model reduction paradigm, including dense-to-sparse distillation and deep-to-shallow distillation. The former one aims to transfer knowledge from the dense-head based MixViT to our fully transformer tracker, while the latter one is used to prune some layers of the backbone. We instantiate two types of MixForemrV2, where the MixFormerV2-B achieves an AUC of 70.6\% on LaSOT and AUC of 56.7\% on TNL2k with a high GPU speed of 165 FPS, and the MixFormerV2-S surpasses FEAR-L by 2.7\% AUC on LaSOT with a real-time CPU speed.

While recent advancements in vision-language models have had a transformative impact on multi-modal comprehension, the extent to which these models possess the ability to comprehend generated images remains uncertain. Synthetic images, in comparison to real data, encompass a higher level of diversity in terms of both content and style, thereby presenting significant challenges for the models to fully grasp. In light of this challenge, we introduce a comprehensive dataset, referred to as JourneyDB, that caters to the domain of generative images within the context of multi-modal visual understanding. Our meticulously curated dataset comprises 4 million distinct and high-quality generated images, each paired with the corresponding text prompts that were employed in their creation. Furthermore, we additionally introduce an external subset with results of another 22 text-to-image generative models, which makes JourneyDB a comprehensive benchmark for evaluating the comprehension of generated images. On our dataset, we have devised four benchmarks to assess the performance of generated image comprehension in relation to both content and style interpretation. These benchmarks encompass prompt inversion, style retrieval, image captioning, and visual question answering. Lastly, we evaluate the performance of state-of-the-art multi-modal models when applied to the JourneyDB dataset, providing a comprehensive analysis of their strengths and limitations in comprehending generated content. We anticipate that the proposed dataset and benchmarks will facilitate further research in the field of generative content understanding. The dataset is publicly available at https://journeydb.github.io.

Generic Boundary Detection (GBD) aims at locating the general boundaries that divide videos into semantically coherent and taxonomy-free units, and could serve as an important pre-processing step for long-form video understanding. Previous works often separately handle these different types of generic boundaries with specific designs of deep networks from simple CNN to LSTM. Instead, in this paper, we presentTemporal Perceiver, a general architecture with Transformer, offering a unified solution to the detection of arbitrary generic boundaries, ranging from shot-level, event-level, to scene-level GBDs. The core design is to introduce a small set of latent feature queries as anchors to compress the redundant video input into a fixed dimension via cross-attention blocks. Thanks to this fixed number of latent units, it greatly reduces the quadratic complexity of attention operation to a linear form of input frames. Specifically, to explicitly leverage the temporal structure of videos, we construct two types of latent feature queries: boundary queries and context queries, which handle the semantic incoherence and coherence accordingly. Moreover, to guide the learning of latent feature queries, we propose an alignment loss on the cross-attention maps to explicitly encourage the boundary queries to attend on the top boundary candidates. Finally, we present a sparse detection head on the compressed representation, and directly output the final boundary detection results without any post-processing module. We test our Temporal Perceiver on a variety of GBD benchmarks. Our method obtains the state-of-the-art results on all benchmarks with RGB single-stream features: SoccerNet-v2 (81.9 percent average-mAP), Kinetics-GEBD (86.0 percent average-f1), TAPOS (73.2 percent average-f1), MovieScenes (51.9 percent AP and 53.1 percent$M_{iou}$) and MovieNet (53.3 percent AP and 53.2 percent$M_{iou}$), demonstrating the generalization ability of our Temporal Perceiver. To further pursue a general GBD model, we combined various tasks to train a class-agnostic Temporal perceiver and evaluate its performance across all benchmarks. Results show that the class-agnostic Perceiver achieves comparable detection accuracy and even better generalization ability compared to dataset-specific Temporal Perceiver.

Estimating human pose and shape from monocular images is a long-standing problem in computer vision. Since the release of statistical body models, 3D human mesh recovery has been drawing broader attention. With the same goal of obtaining well-aligned and physically plausible mesh results, two paradigms have been developed to overcome challenges in the 2D-to-3D lifting process: i) an optimization-based paradigm, where different data terms and regularization terms are exploited as optimization objectives; and ii) a regression-based paradigm, where deep learning techniques are embraced to solve the problem in an end-to-end fashion. Meanwhile, continuous efforts are devoted to improving the quality of 3D mesh labels for a wide range of datasets. Though remarkable progress has been achieved in the past decade, the task is still challenging due to flexible body motions, diverse appearances, complex environments, and insufficient in-the-wild annotations. To the best of our knowledge, this is the first survey that focuses on the task of monocular 3D human mesh recovery. We start with the introduction of body models and then elaborate recovery frameworks and training objectives by providing in-depth analyses of their strengths and weaknesses. We also summarize datasets, evaluation metrics, and benchmark results. Open issues and future directions are discussed in the end, hoping to motivate researchers and facilitate their research in this area.

Aggregating neighbor features is essential for point cloud neural network. In the existing work, each point in the cloud may inevitably be selected as the neighbors of multiple aggregation centers, as all centers will gather neighbor features from the whole point cloud independently. Thus, each point has to participate in the calculation repeatedly, generating redundant duplicates in the memory, leading to intensive computation costs and memory consumption. Meanwhile, to pursue higher accuracy, previous methods often rely on a complex local aggregator to extract fine geometric representation, further slowing down the processing pipeline. To address these issues, we propose a new local aggregator of linear complexity for point cloud analysis, coined as APP. Specifically, we introduce an auxiliary container as an anchor to exchange features between the source point and the aggregating center. Each source point pushes its feature to only one auxiliary container, and each center point pulls features from only one auxiliary container. This avoids the re-computation issue of each source point. To facilitate the learning of the local structure of point cloud, we use an online normal estimation module to provide explainable geometric information to enhance our APP modeling capability. Our built network is more efficient than all the previous baselines with a clear margin while still consuming a lower memory. Experiments on classification and semantic segmentation demonstrate that APP-Net reaches comparable accuracies to other networks. In the classification task, it can process more than 10,000 samples per second with less than 10GB of memory on a single GPU. We will release the code at https://github.com/MCG-NJU/ APP-Net.

Temporal grounding aims to localize a video moment which is semantically aligned with a given natural language query. Existing methods typically apply a detection or regression pipeline on the fused representation with the research focus on designing complicated prediction heads or fusion strategies. Instead, from a perspective on temporal grounding as a metric-learning problem, we present a Mutual Matching Network (MMN), to directly model the similarity between language queries and video moments in a joint embedding space. This new metric-learning framework enables fully exploiting negative samples from two new aspects: constructing negative cross-modal pairs in a mutual matching scheme and mining negative pairs across different videos. These new negative samples could enhance the joint representation learning of two modalities via cross-modal mutual matching to maximize their mutual information. Experiments show that our MMN achieves highly competitive performance compared with the state-of-the-art methods on four video grounding benchmarks. Based on MMN, we present a winner solution for the HC-STVG challenge of the 3rd PIC workshop. This suggests that metric learning is still a promising method for temporal grounding via capturing the essential cross-modal correlation in a joint embedding space. Code is available at https://github.com/MCG-NJU/MMN.

Temporal action detection aims to locate the boundaries of action in the video. The current method based on boundary matching enumerates and calculates all possible boundary matchings to generate proposals. However, these methods neglect the long-range context aggregation in boundary prediction. At the same time, due to the similar semantics of adjacent matchings, local semantic aggregation of densely-generated matchings cannot improve semantic richness and discrimination. In this paper, we propose the end-to-end proposal generation method named Dual Context Aggregation Network (DCAN) to aggregate context on two levels, namely, boundary level and proposal level, for generating high-quality action proposals, thereby improving the performance of temporal action detection. Specifically, we design the Multi-Path Temporal Context Aggregation (MTCA) to achieve smooth context aggregation on boundary level and precise evaluation of boundaries. For matching evaluation, Coarse-to-fine Matching (CFM) is designed to aggregate context on the proposal level and refine the matching map from coarse to fine. We conduct extensive experiments on ActivityNet v1.3 and THUMOS-14. DCAN obtains an average mAP of 35.39% on ActivityNet v1.3 and reaches mAP 54.14% at [email protected] on THUMOS-14, which demonstrates DCAN can generate high-quality proposals and achieve state-of-the-art performance. We release the code at https://github.com/cg1177/DCAN.

In this paper, we present a new cross-architecture contrastive learning (CACL) framework for self-supervised video representation learning. CACL consists of a 3D CNN and a video transformer which are used in parallel to generate diverse positive pairs for contrastive learning. This allows the model to learn strong representations from such diverse yet meaningful pairs. Furthermore, we introduce a temporal self-supervised learning module able to predict an Edit distance explicitly between two video sequences in the temporal order. This enables the model to learn a rich temporal representation that compensates strongly to the video-level representation learned by the CACL. We evaluate our method on the tasks of video retrieval and action recognition on UCF101 and HMDB51 datasets, where our method achieves excellent performance, surpassing the state-of-the-art methods such as VideoMoCo [23] and MoCo+BE [34] by a large margin.

Tracking often uses a multistage pipeline of feature extraction, target information integration, and bounding box estimation. To simplify this pipeline and unify the process of feature extraction and target information integration, we present a compact tracking framework, termed as MixFormer, built upon transformers. Our core design is to utilize the flexibility of attention operations, and propose a Mixed Attention Module (MAM) for simultaneous feature extraction and target information integration. This synchronous modeling scheme allows to extract target-specific discriminative features and perform extensive communication between target and search area. Based on MAM, we build our MixFormer tracking framework simply by stacking multiple MAMs with progressive patch embedding and placing a localization head on top. In addition, to handle multiple target templates during online tracking, we devise an asymmetric attention scheme in MAM to reduce computational cost, and propose an effective score prediction module to select high-quality templates. Our MixFormer sets a new state-of-the-art performance on five tracking benchmarks, including LaSOT, TrackingNet, VOT2020, GOT-10k, and UAV123. In particular, our MixFormer-L achieves NP score of 79.9% on LaSOT, 88.9% on TrackingNet and EAO of 0.555 on VOT2020. We also perform in-depth ablation studies to demonstrate the effectiveness of simultaneous feature extraction and information integration. Code and trained models are publicly available at https://github.com/MCG-NJU/MixFormer.

Traditional object detectors employ the dense paradigm of scanning over locations and scales in an image. The recent query-based object detectors break this convention by decoding image features with a set of learnable queries. However, this paradigm still suffers from slow convergence, limited performance, and design complexity of extra networks between backbone and decoder. In this paper, we find that the key to these issues is the adaptability of decoders for casting queries to varying objects. Accordingly, we propose a fast-converging query-based detector, named AdaMixer, by improving the adaptability of query-based decoding processes in two aspects. First, each query adaptively samples features over space and scales based on estimated offsets, which allows AdaMixer to efficiently attend to the coherent regions of objects. Then, we dynamically decode these sampled features with an adaptive MLP-Mixer under the guidance of each query. Thanks to these two critical designs, AdaMixer enjoys architectural simplicity without requiring dense attentional encoders or explicit pyramid networks. On the challenging MS COCO benchmark, AdaMixer with ResNet-50 as the backbone, with 12 training epochs, reaches up to 45.0 AP on the validation set along with 27.9 APs in detecting small objects. With the longer training scheme, AdaMixer with ResNeXt-101-DCN and Swin-S reaches 49.5 and 51.3 AP. Our work sheds light on a simple, accurate, and fast converging architecture for query-based object detectors. The code is made available at https://github.com/MCG-NJU/AdaMixer.

Videos incorporate rich semantics as well as redundant information. Seeking a compact yet effective video representation, e.g., sample informative frames from the entire video, is critical to efficient video recognition. There have been works that formulate frame sampling as a sequential decision task by selecting frames one by one according to their importance. In this paper, we present a more efficient framework named OCSampler, which explores such a representation with one short clip. OCSampler designs a new paradigm of learning instance-specific video condensation policies to select frames only in a single step. Rather than picking up frames sequentially like previous methods, we simply process a whole sequence at once. Accordingly, these policies are derived from a light-weighted skim network together with a simple yet effective policy network. Moreover, we extend the proposed method with a frame number budget, enabling the framework to produce correct predictions in high confidence with as few frames as possible. Experiments on various benchmarks demonstrate the effectiveness of OCSampler over previous methods in terms of accuracy and efficiency. Specifically, it achieves 76.9% mAP and 21.7 GFLOPs on ActivityNet with an impressive throughput: 123.9 Video/s on a single TITAN Xp GPU.

Generic event boundary detection (GEBD) is an important yet challenging task in video understanding, which aims at detecting the moments where humans naturally perceive event boundaries. The main challenge of this task is perceiving various temporal variations of diverse event boundaries. To this end, this paper presents an effective and end-to-end learnable framework (DDM-Net). To tackle the diversity and complicated semantics of event boundaries, we make three notable improvements. First, we construct a feature bank to store multi-level features of space and time, prepared for difference calculation at multiple scales. Second, to alleviate inadequate temporal modeling of pre-vious methods, we present dense difference maps (DDM) to comprehensively characterize the motion pattern. Finally, we exploit progressive attention on multi-level DDM to jointly aggregate appearance and motion clues. As a result, DDM-Net respectively achieves a significant boost of 14% and 8% on Kinetics-GEBD and TAPOS benchmark, and outperforms the top-1 winner solution of LOVEU Challenge@CVPR 2021 without bells and whistles. The state-of-the-art result demonstrates the effectiveness of richer motion representation and more sophisticated aggregation, in handling the diversity of GEBD. The code is made available at https://github.com/MCG-NJU/DDM.

Scene graph generation (SGG) is to detect object pairs with their relations in an image. Existing SGG approaches often use multi-stage pipelines to decompose this task into object detection, relation graph construction, and dense or dense-to-sparse relation prediction. Instead, from a perspective on SGG as a direct set prediction, this paper presents a simple, sparse, and unified framework, termed as Structured Sparse R-CNN. The key to our method is a set of learnable triplet queries and a structured triplet detector which could be jointly optimized from the training set in an end-to-end manner. Specifically, the triplet queries encode the general prior for object pairs with their relations, and provide an initial guess of scene graphs for subsequent refinement. The triplet detector presents a cascaded architecture to progressively refine the detected scene graphs with the customized dynamic heads. In addition, to relieve the training difficulty of our method, we propose a relaxed and enhanced training strategy based on knowledge distillation from a Siamese Sparse R-CNN. We perform experiments on several datasets: Visual Genome and Open Images V4/V6, and the results demonstrate that our method achieves the state-of-the-art performance. In addition, we also perform in-depth ablation studies to provide insights on our structured modeling in triplet detector design and training strategies. The code and models are made available at https://github.com/MCG-NJU/Structured-Sparse-RCNN.

Detectors trained with massive labeled data often exhibit dramatic performance degradation in some particular scenarios with data distribution gap. To alleviate this problem of domain shift, conventional wisdom typically concentrates solely on reducing the discrepancy between the source and target domains via attached domain classifiers, yet ignoring the difficulty of such transferable features in coping with both classification and localization subtasks in object detection. To address this issue, in this paper, we propose Task-specific Inconsistency Alignment (TIA), by developing a new alignment mechanism in separate task spaces, improving the performance of the detector on both subtasks. Specifically, we add a set of auxiliary predictors for both classification and localization branches, and exploit their behavioral inconsistencies as finer-grained domain-specific measures. Then, we devise task-specific losses to align such cross-domain disagreement of both subtasks. By optimizing them individually, we are able to well approximate the category- and boundary-wise discrepancies in each task space, and therefore narrow them in a decoupled manner. TIA demonstrates superior results on various scenarios to the previous state-of-the-art methods. It is also observed that both the classification and localization capabilities of the detector are sufficiently strengthened, further demonstrating the effectiveness of our TIA method. Code and trained models are publicly available at https://github.com/MCG-NJU/TIA.

Action spotting and classification consist in detecting the exact moments at which events occur in long videos. The current mainstream spotting practices generally use a two-stage pipeline that performs feature collection and integration, then salient action detection and postprocessing. Following that, we present SpotFormer, a simple yet effective framework, capable of precise action spotting. Specifically, we employ several most advanced backbone networks as auxiliary feature extractors, and reduce feature dimensionality in a straightforward and efficient way. The frame-wise features are fed into a transformer-based spotting network devised to leverage spatiotemporal information. We obtain 0.609 tight mAP score via model ensemble and achieve the state-of-the-art performance on the SoccerNet-v2 dataset.

Traditional temporal action detection (TAD) usually handles untrimmed videos with small number of action instances from a single label (e.g., ActivityNet, THUMOS). However, this setting might be unrealistic as different classes of actions often co-occur in practice. In this paper, we focus on the task of multi-label temporal action detection that aims to localize all action instances from a multi-label untrimmed video. Multi-label TAD is more challenging as it requires for fine-grained class discrimination within a single video and precise localization of the co-occurring instances. To mitigate this issue, we extend the sparse query-based detection paradigm from the traditional TAD and propose the multi-label TAD framework of PointTAD. Specifically, our PointTAD introduces a small set of learnable query points to represent the important frames of each action instance. This point-based representation provides a flexible mechanism to localize the discriminative frames at boundaries and as well the important frames inside the action. Moreover, we perform the action decoding process with the Multi-level Interactive Module to capture both point-level and instance-level action semantics. Finally, our PointTAD employs an end-to-end trainable framework simply based on RGB input for easy deployment. We evaluate our proposed method on two popular benchmarks and introduce the new metric of detection-mAP for multi-label TAD. Our model outperforms all previous methods by a large margin under the detection-mAP metric, and also achieves promising results under the segmentation-mAP metric.

Pre-training video transformers on extra large-scale datasets is generally required to achieve premier performance on relatively small datasets. In this paper, we show that video masked autoencoders (VideoMAE) are data-efficient learners for self-supervised video pre-training (SSVP). We are inspired by the recent ImageMAE and propose customized video tube masking with an extremely high ratio. This simple design makes video reconstruction a more challenging and meaningful self-supervision task, thus encouraging extracting more effective video representations during the pre-training process. We obtain three important findings with VideoMAE: (1) An extremely high proportion of masking ratio (i.e., 90% to 95%) still yields favorable performance for VideoMAE. The temporally redundant video content enables higher masking ratio than that of images. (2) VideoMAE achieves impressive results on very small datasets (i.e., around 3k-4k videos) without using any extra data. This is partially ascribed to the challenging task of video reconstruction to enforce high-level structure learning. (3) VideoMAE shows that data quality is more important than data quantity for SSVP. Domain shift between pre-training and target datasets is an important factor. Notably, our VideoMAE with the vanilla ViT backbone can achieve 87.4% on Kinects-400, 75.4% on Something-Something V2, 91.3% on UCF101, and 62.6% on HMDB51, without using any extra data. Code is available at https://github.com/MCG-NJU/VideoMAE.

Temporal action localization (TAL) is an important and challenging problem in video understanding. However, most existing TAL benchmarks are built upon the coarse granularity of action classes, which exhibits two major limitations in this task. First, coarse-level actions can make the localization models overfit in high-level context information, and ignore the atomic action details in the video. Second, the coarse action classes often lead to the ambiguous annotations of temporal boundaries, which are inappropriate for temporal action localization. To tackle these problems, we develop a novel large-scale and fine-grained video dataset, coined as FineAction, for temporal action localization. In total, FineAction contains 103K temporal instances of 106 action categories, annotated in 17K untrimmed videos. Compared to the existing TAL datasets, our FineAction takes distinct characteristics of fine action classes with rich diversity, dense annotations of multiple instances, and co-occurring actions of different classes, which introduces new opportunities and challenges for temporal action localization. To benchmark FineAction, we systematically investigate the performance of several popular temporal localization methods on it, and deeply analyze the influence of fine-grained instances in temporal action localization. As a minor contribution, we present a simple baseline approach for handling the fine-grained action detection, which achieves an mAP of 13.17% on our FineAction. We believe that FineAction can advance research of temporal action localization and beyond. The dataset is available at https://deeperaction.github.io/datasets/fineaction.

Temporal modeling still remains challenging for action recognition in videos. To mitigate this issue, this paper presents a new video architecture, termed as Temporal Difference Network (TDN), with a focus on capturing multi-scale temporal information for efficient action recognition. The core of our TDN is to devise an efficient temporal module (TDM) by explicitly leveraging a temporal difference operator, and systematically assess its effect on short-term and long-term motion modeling. To fully capture temporal information over the entire video, our TDN is established with a two-level difference modeling paradigm. Specifically, for local motion modeling, temporal difference over consecutive frames is used to supply 2D CNNs with finer motion pattern, while for global motion modeling, temporal difference across segments is incorporated to capture long-range structure for motion feature excitation. TDN provides a simple and principled temporal modeling framework and could be instantiated with the existing CNNs at a small extra computational cost. Our TDN presents a new state of the art on the Something-Something V1 & V2 datasets and is on par with the best performance on the Kinetics-400 dataset. In addition, we conduct in-depth ablation studies and plot the visualization results of our TDN, hopefully providing insightful analysis on temporal difference modeling. We release the code at https://github.com/MCG-NJU/TDN.

Previous point cloud semantic segmentation networks use the same process to aggregate features from neighbors of the same category and different categories. However, the joint area between two objects usually only occupies a small percentage in the whole scene. Thus the networks are well- trained for aggregating features from the same category point while not fully trained on aggregating points of different categories. To address this issue, this paper proposes to utilize different aggregation strategies between the same category and different categories. Specifically, it presents a customized module, termed as Category Guided Aggregation (CGA), where it first identifies whether the neighbors belong to the same category with the center point or not, and then handles the two types of neighbors with two carefully-designed modules. Our CGA presents a general network module and could be leveraged in any existing semantic segmentation network. Experiments on three different backbones demonstrate the effectiveness of our method.

The classification and regression head are both indispensable components to build up a dense object detector, which are usually supervised by the same training samples and thus expected to have consistency with each other for detecting objects accurately in the detection pipeline. In this paper, we break the convention of the same training samples for these two heads in dense detectors and explore a novel supervisory paradigm, termed as Mutual Supervision (MuSu), to respectively and mutually assign training samples for the classification and regression head to ensure this consistency. MuSu defines training samples for the regression head mainly based on classification predicting scores and in turn, defines samples for the classification head based on localization scores from the regression head. Experimental results show that the convergence of detectors trained by this mutual supervision is guaranteed and the effectiveness of the proposed method is verified on the challenging MS COCO benchmark. We also find that tiling more anchors at the same location benefits detectors and leads to further improvements under this training scheme. We hope this work can inspire further researches on the interaction of the classification and regression task in detection and the supervision paradigm for detectors, especially separately for these two heads.

Deep learning has achieved remarkable progress for visual recognition on large-scale balanced datasets but still performs poorly on real-world long-tailed data. Previous methods often adopt class re-balanced training strategies to effectively alleviate the imbalance issue, but might be a risk of over-fitting tail classes. The recent decoupling method overcomes over-fitting issues by using a multi-stage training scheme, yet, it is still incapable of capturing tail class information in the feature learning stage. In this paper, we show that soft label can serve as a powerful solution to incorporate label correlation into a multi-stage training scheme for long-tailed recognition. The intrinsic relation between classes embodied by soft labels turns out to be helpful for long-tailed recognition by transferring knowledge from head to tail classes.Specifically, we propose a conceptually simple yet particularly effective multi-stage training scheme, termed as Self Supervised to Distillation (SSD). This scheme is composed of two parts. First, we introduce a self-distillation framework for long-tailed recognition, which can mine the label relation automatically. Second, we present a new distillation label generation module guided by self-supervision. The distilled labels integrate information from both label and data domains that can model long-tailed distribution effectively. We conduct extensive experiments and our method achieves the state-of-the-art results on three long-tailed recognition benchmarks: ImageNet-LT, CIFAR100-LT and iNaturalist 2018. Our SSD outperforms the strong LWS baseline by from 2.7% to 4.5% on various datasets.

Spatio-temporal action detection is an important and challenging problem in video understanding. The existing action detection benchmarks are limited in aspects of small numbers of instances in a trimmed video or low-level atomic actions. This paper aims to present a new multi-person dataset of spatio-temporal localized sports actions, coined as MultiSports. We first analyze the important ingredients of constructing a realistic and challenging dataset for spatio-temporal action detection by proposing three criteria: (1) multi-person scenes and motion dependent identification, (2) with well-defined boundaries, (3) relatively fine-grained classes of high complexity. Based on these guide-lines, we build the dataset of MultiSports v1.0 by selecting 4 sports classes, collecting 3200 video clips, and annotating 37701 action instances with 902k bounding boxes. Our datasets are characterized with important properties of high diversity, dense annotation, and high quality. Our Multi-Sports, with its realistic setting and detailed annotations, exposes the intrinsic challenges of spatio-temporal action detection. To benchmark this, we adapt several baseline methods to our dataset and give an indepth analysis on the action detection results in our dataset. We hope our MultiSports can serve as a standard benchmark for spatio-temporal action detection in the future. Our dataset website is at https://deeperaction.github.io/multisports/.

Video data is with complex temporal dynamics due to various factors such as camera motion, speed variation, and different activities. To effectively capture this diverse motion pattern, this paper presents a new temporal adaptive module (TAM) to generate video-specific temporal kernels based on its own feature map. TAM proposes a unique two-level adaptive modeling scheme by decoupling the dynamic kernel into a location sensitive importance map and a location invariant aggregation weight. The importance map is learned in a local temporal window to capture short-term information, while the aggregation weight is generated from a global view with a focus on long-term structure. TAM is a modular block and could be integrated into 2D CNNs to yield a powerful video architecture (TANet) with a very small extra computational cost. The extensive experiments on Kinetics-400 and Something-Something datasets demonstrate that our TAM outperforms other temporal modeling methods consistently, and achieves the state-of-the-art performance under the similar complexity. The code is available at https://github.com/liu-zhy/temporal-adaptive-module.

Temporal action proposal generation is an important and challenging task in video understanding, which aims at detecting all temporal segments containing action in-stances of interest. The existing proposal generation approaches are generally based on pre-defined anchor windows or heuristic bottom-up boundary matching strategies. This paper presents a simple and efficient framework (RTD-Net) for direct action proposal generation, by re-purposing a Transformer-alike architecture. To tackle the essential visual difference between time and space, we make three important improvements over the original transformer detection framework (DETR). First, to deal with slowness prior in videos, we replace the original Transformer en-coder with a boundary attentive module to better capture long-range temporal information. Second, due to the ambiguous temporal boundary and relatively sparse annotations, we present a relaxed matching scheme to relieve the strict criteria of single assignment to each groundtruth. Finally, we devise a three-branch head to further improve the proposal confidence estimation by explicitly predicting its completeness. Extensive experiments on THUMOS14 and ActivityNet-1.3 benchmarks demonstrate the effectiveness of RTD-Net, on both tasks of temporal action proposal generation and temporal action detection. Moreover, due to its simplicity in design, our framework is more efficient than previous proposal generation methods, without non-maximum suppression post-processing. The code and models are made available at https://github.com/MCG-NJU/RTD-Action.

This paper deals with a challenging task of video scene graph generation (VidSGG), which could serve as a structured video representation for high-level understanding tasks. We present a new detect-to-track paradigm for this task by decoupling the context modeling for relation prediction from the complicated low-level entity tracking. Specifically, we design an efficient method for frame-level VidSGG, termed as Target Adaptive Context Aggregation Network (TRACE), with a focus on capturing spatio-temporal context information for relation recognition. Our TRACE framework streamlines the VidSGG pipeline with a modular design, and presents two unique blocks of Hierarchical Relation Tree (HRTree) construction and Target-adaptive Context Aggregation. More specific, our HRTree first provides an adpative structure for organizing possible relation candidates efficiently, and guides context aggregation module to effectively capture spatio-temporal structure information. Then, we obtain a contextualized feature representation for each relation candidate and build a classification head to recognize its relation category. Finally, we provide a simple temporal association strategy to track TRACE detected results to yield the video-level VidSGG. We perform experiments on two VidSGG benchmarks: ImageNet-VidVRD and Action Genome, and the results demonstrate that our TRACE achieves the state-of-the-art performance. The code and models are made available at https://github.com/MCG-NJU/TRACE.

Regression-based methods have recently shown promising results in reconstructing human meshes from monocular images. By directly mapping raw pixels to model parameters, these methods can produce parametric models in a feed-forward manner via neural networks. However, minor deviation in parameters may lead to noticeable mis-alignment between the estimated meshes and image evidences. To address this issue, we propose a Pyramidal Mesh Alignment Feedback (PyMAF) loop to leverage a feature pyramid and rectify the predicted parameters explicitly based on the mesh-image alignment status in our deep regressor. In PyMAF, given the currently predicted parameters, mesh-aligned evidences will be extracted from finer-resolution features accordingly and fed back for parameter rectification. To reduce noise and enhance the reliability of these evidences, an auxiliary pixel-wise supervision is imposed on the feature encoder, which provides mesh-image correspondence guidance for our network to preserve the most related information in spatial features. The efficacy of our approach is validated on several benchmarks, including Human3.6M, 3DPW, LSP, and COCO, where experimental results show that our approach consistently improves the mesh-image alignment of the reconstruction. The project page with code and video results can be found at https://hongwenzhang.github.io/pymaf.

Frame sampling is a fundamental problem in video action recognition due to the essential redundancy in time and limited computation resources. The existing sampling strategy often employs a fixed frame selection and lacks the flexibility to deal with complex variations in videos. In this paper, we present a simple, sparse, and explainable frame sampler, termed as Motion-Guided Sampler (MGSampler). Our basic motivation is that motion is an important and universal signal that can drive us to adaptively select frames from videos. Accordingly, we propose two important properties in our MGSampler design: motion sensitive and motion uniform. First, we present two different motion representations to enable us to efficiently distinguish the motion-salient frames from the background. Then, we devise a motion-uniform sampling strategy based on the cumulative motion distribution to ensure the sampled frames evenly cover all the important segments with high motion salience. Our MGSampler yields a new principled and holistic sampling scheme, that could be incorporated into any existing video architecture. Experiments on five benchmarks demonstrate the effectiveness of our MGSampler over the previous fixed sampling strategies, and its generalization power across different backbones, video models, and datasets. The code is available at https://github.com/MCG-NJU/MGSampler.

This paper presents the method that underlies our submission to the Pre-training for Video Understanding Challenge Track II. We follow the basic pipeline of temporal segment networks [20] and further improve its performance in several aspects. Specifically, we use the latest transformer-based architectures, e.g., Swin Transformer, DeiT, CLIP-ViT, to enhance the representation power. We analyze different pre-training proxy tasks on the official pre-training datasets and other open-source video datasets. With these techniques, we derive an ensemble of deep models to attain a high classification accuracy (Top-1 accuracy 62.28%) on the testing set and secures first place in Track II of this challenge.

Three-dimensional face dense alignment and reconstruction in the wild is a challenging problem as partial facial information is commonly missing in occluded and large pose face images. Large head pose variations also increase the solution space and make the modeling more difficult. Our key idea is to model occlusion and pose to decompose this challenging task into several relatively more manageable subtasks. To this end, we propose an end-to-end framework, termed as Self-aligned Dual face Regression Network (SADRNet), which predicts a pose-dependent face, a pose-independent face. They are combined by an occlusion-aware self-alignment to generate the final 3D face. Extensive experiments on two popular benchmarks, AFLW2000-3D and Florence, demonstrate that the proposed method achieves significant superior performance over existing state-of-the-art methods.

The task of spatial-temporal action detection has attracted increasing researchers. Existing dominant methods solve this problem by relying on short-term information and dense serial-wise detection on each individual frames or clips. Despite their effectiveness, these methods showed inadequate use of long-term information and are prone to inefficiency. In this paper, we propose for the first time, an efficient framework that generates action tube proposals from video streams with a single forward pass in a sparse-to-dense manner. There are two key characteristics in this framework: (1) Both long-term and short-term sampled information are explicitly utilized in our spatio-temporal network, (2) A new dynamic feature sampling module (DTS) is designed to effectively approximate the tube output while keeping the system tractable. We evaluate the efficacy of our model on the UCF101-24, JHMDB-21 and UCFSports benchmark datasets, achieving promising results that are competitive to state-of-the-art methods. The proposed sparse-to-dense strategy rendered our framework about 7.6 times more efficient than the nearest competitor.

Efficiency is an important issue in designing video architectures for action recognition. 3D CNNs have witnessed remarkable progress in action recognition from videos. However, compared with their 2D counterparts, 3D convolutions often introduce a large amount of parameters and cause high computational cost. To relieve this problem, we propose an efficient temporal module, termed as Temporal Enhancement-and-Interaction (TEI Module), which could be plugged into the existing 2D CNNs (denoted by TEINet). The TEI module presents a different paradigm to learn temporal features by decoupling the modeling of channel correlation and temporal interaction. First, it contains a Motion Enhanced Module (MEM) which is to enhance the motion-related features while suppress irrelevant information (e.g., background). Then, it introduces a Temporal Interaction Module (TIM) which supplements the temporal contextual information in a channel-wise manner. This two-stage modeling scheme is not only able to capture temporal structure flexibly and effectively, but also efficient for model inference. We conduct extensive experiments to verify the effectiveness of TEINet on several benchmarks (e.g., Something-Something V1&V2, Kinetics, UCF101 and HMDB51). Our proposed TEINet can achieve a good recognition accuracy on these datasets but still preserve a high efficiency.

In this work, we propose Knowledge Integration Networks (referred as KINet) for video action recognition. KINet is capable of aggregating meaningful context features which are of great importance to identifying an action, such as human information and scene context. We design a three-branch architecture consisting of a main branch for action recognition, and two auxiliary branches for human parsing and scene recognition which allow the model to encode the knowledge of human and scene for action recognition. We explore two pre-trained models as teacher networks to distill the knowledge of human and scene for training the auxiliary tasks of KINet. Furthermore, we propose a two-level knowledge encoding mechanism which contains a Cross Branch Integration (CBI) module for encoding the auxiliary knowledge into medium-level convolutional features, and an Action Knowledge Graph (AKG) for effectively fusing high-level context information. This results in an end-to-end trainable framework where the three tasks can be trained collaboratively, allowing the model to compute strong context knowledge efficiently. The proposed KINet achieves the state-of-the-art performance on a large-scale action recognition benchmark Kinetics-400, with a top-1 accuracy of 77.8%. We further demonstrate that our KINet has strong capability by transferring the Kinetics-trained model to UCF-101, where it obtains 97.8% top-1 accuracy.

We introduce the first method for automatic image generation from scene-level freehand sketches. Our model allows for controllable image generation by specifying the synthesis goal via freehand sketches. The key contribution is an attribute vector bridged Generative Adversarial Network called EdgeGAN, which supports high visual-quality object-level image content generation without using freehand sketches as training data. We have built a large-scale composite dataset called SketchyCOCO to support and evaluate the solution. We validate our approach on the tasks of both object-level and scene-level image generation on SketchyCOCO. Through quantitative, qualitative results, human evaluation and ablation studies, we demonstrate the method's capacity to generate realistic complex scene-level images from various freehand sketches.

Temporal modeling is key for action recognition in videos. It normally considers both short-range motions and long-range aggregations. In this paper, we propose a Temporal Excitation and Aggregation (TEA) block, including a motion excitation (ME) module and a multiple temporal aggregation (MTA) module, specifically designed to capture both short- and long-range temporal evolution. In particular, for short-range motion modeling, the ME module calculates the feature-level temporal differences from spatiotemporal features. It then utilizes the differences to excite the motion-sensitive channels of the features. The long-range temporal aggregations in previous works are typically achieved by stacking a large number of local temporal convolutions. Each convolution processes a local temporal window at a time. In contrast, the MTA module proposes to deform the local convolution to a group of sub-convolutions, forming a hierarchical residual architecture. Without introducing additional parameters, the features will be processed with a series of sub-convolutions, and each frame could complete multiple temporal aggregations with neighborhoods. The final equivalent receptive field of temporal dimension is accordingly enlarged, which is capable of modeling the long-range temporal relationship over distant frames. The two components of the TEA block are complementary in temporal modeling. Finally, our approach achieves impressive results at low FLOPs on several action recognition benchmarks, such as Kinetics, Something-Something, HMDB51, and UCF101, which confirms its effectiveness and efficiency.

The existing action recognition methods are mainly based on clip-level classifiers such as two-stream CNNs or 3D CNNs, which are trained from the randomly selected clips and applied to densely sampled clips during testing. However, this standard setting might be suboptimal for training classifiers and also requires huge computational overhead when deployed in practice. To address these issues, we propose a new framework for action recognition in videos, called Dynamic Sampling Networks (DSN), by designing a dynamic sampling module to improve the discriminative power of learned clip-level classifiers and as well increase the inference efficiency during testing. Specifically, DSN is composed of a sampling module and a classification module, whose objective is to learn a sampling policy to on-the-fly select which clips to keep and train a clip-level classifier to perform action recognition based on these selected clips, respectively. In particular, given an input video, we train an observation network in an associative reinforcement learning setting to maximize the rewards of the selected clips with a correct prediction. We perform extensive experiments to study different aspects of the DSN framework on four action recognition datasets: UCF101, HMDB51, THUMOS14, and ActivityNet v1.3. The experimental results demonstrate that DSN is able to greatly improve the inference efficiency by only using less than half of the clips, which can still obtain a slightly better or comparable recognition accuracy to the state-of-the-art approaches.

Despite the success of deep learning for static image understanding, it remains unclear what are the most effective network architectures for spatial-temporal modeling in videos. In this paper, in contrast to the existing CNN+RNN or pure 3D convolution based approaches, we explore a novel spatialtemporal network (StNet) architecture for both local and global modeling in videos. Particularly, StNet stacks N successive video frames into a super-image which has 3N channels and applies 2D convolution on super-images to capture local spatial-temporal relationship. To model global spatialtemporal structure, we apply temporal convolution on the local spatial-temporal feature maps. Specifically, a novel temporal Xception block is proposed in StNet, which employs a separate channel-wise and temporal-wise convolution over the feature sequence of a video. Extensive experiments on the Kinetics dataset demonstrate that our framework outperforms several state-of-the-art approaches in action recognition and can strike a satisfying trade-off between recognition accuracy and model complexity. We further demonstrate the generalization performance of the leaned video representations on the UCF101 dataset.

Cross-modal transfer is helpful to enhance modality-specific discriminative power for scene recognition. To this end, this paper presents a unified framework to integrate the tasks of cross-modal translation and modality-specific recognition, termed as Translate-to-Recognize Network TRecgNet. Specifically, both translation and recognition tasks share the same encoder network, which allows to explicitly regularize the training of recognition task with the help of translation, and thus improve its final generalization ability. For translation task, we place a decoder module on top of the encoder network and it is optimized with a new layer-wise semantic loss, while for recognition task, we use a linear classifier based on the feature embedding from encoder and its training is guided by the standard cross-entropy loss. In addition, our TRecgNet allows to exploit large numbers of unlabeled RGB-D data to train the translation task and thus improve the representation power of encoder network. Empirically, we verify that this new semi-supervised setting is able to further enhance the performance of recognition network. We perform experiments on two RGB-D scene recognition benchmarks: NYU Depth v2 and SUN RGB-D, demonstrating that TRecgNet achieves superior performance to the existing state-of-the-art methods, especially for recognition solely based on a single modality.

Modeling relation between actors is important for recognizing group activity in a multi-person scene. This paper aims at learning discriminative relation between actors efficiently using deep models. To this end, we propose to build a flexible and efficient Actor Relation Graph (ARG) to simultaneously capture the appearance and position relation between actors. Thanks to the Graph Convolutional Network, the connections in ARG could be automatically learned from group activity videos in an end-to-end manner, and the inference on ARG could be efficiently performed with standard matrix operations. Furthermore, in practice, we come up with two variants to sparsify ARG for more effective modeling in videos: spatially localized ARG and temporal randomized ARG. We perform extensive experiments on two standard group activity recognition datasets: the Volleyball dataset and the Collective Activity dataset, where state-of-the-art performance is achieved on both datasets. We also visualize the learned actor graphs and relation features, which demonstrate that the proposed ARG is able to capture the discriminative relation information for group activity recognition.

Spatial downsampling layers are favored in convolutional neural networks (CNNs) to downscale feature maps for larger receptive fields and less memory consumption. However, for discriminative tasks, there is a possibility that these layers lose the discriminative details due to improper pooling strategies, which could hinder the learning process and eventually result in suboptimal models. In this paper, we present a unified framework over the existing downsampling layers (e.g., average pooling, max pooling, and strided convolution) from a local importance view. In this framework, we analyze the issues of these widely-used pooling layers and figure out the criteria for designing an effective downsampling layer. According to this analysis, we propose a conceptually simple, general, and effective pooling layer based on local importance modeling, termed as Local Importance-based Pooling (LIP). LIP can automatically enhance discriminative features during the downsampling procedure by learning adaptive importance weights based on inputs. Experiment results show that LIP consistently yields notable gains with different depths and different architectures on ImageNet classification. In the challenging MS COCO dataset, detectors with our LIP-ResNets as backbones obtain a consistent improvement (≥1.4%) over the vanilla ResNets, and especially achieve the current state-of-the-art performance in detecting small objects under the single-scale testing scheme1.

Due to the high cost of manual annotation, learning directly from the web has attracted broad attention. One issue that limits their performance is the problem of visual polysemy. To address this issue, we present an adaptive multi-model framework that resolves polysemy by visual disambiguation. Compared to existing methods, the primary advantage of our approach lies in that our approach can adapt to the dynamic changes in the search results. Our proposed framework consists of two major steps: we first discover and dynamically select the text queries according to the image search results, then we employ the proposed saliency-guided deep multi-instance learning network to remove outliers and learn classification models for visual disambiguation. Extensive experiments demonstrate the superiority of our proposed approach.

We present a general and flexible video-level framework for learning action models in videos. This method, called temporal segment network (TSN), aims to model long-range temporal structure with a new segment-based sampling and aggregation scheme. This unique design enables the TSN framework to efficiently learn action models by using the whole video. The learned models could be easily deployed for action recognition in both trimmed and untrimmed videos with simple average pooling and multi-scale temporal window integration, respectively. We also study a series of good practices for the implementation of the TSN framework given limited training samples. Our approach obtains the state-the-of-art performance on five challenging action recognition benchmarks: HMDB51 (71.0 percent), UCF101 (94.9 percent), THUMOS14 (80.1 percent), ActivityNet v1.2 (89.6 percent), and Kinetics400 (75.7 percent). In addition, using the proposed RGB difference as a simple motion representation, our method can still achieve competitive accuracy on UCF101 (91.0 percent) while running at 340 FPS. Furthermore, based on the proposed TSN framework, we won the video classification track at the ActivityNet challenge 2016 among 24 teams.

Spatiotemporal feature learning in videos is a fundamental problem in computer vision. This paper presents a new architecture, termed as Appearance-and-Relation Network (ARTNet), to learn video representation in an end-to-end manner. ARTNets are constructed by stacking multiple generic building blocks, called as SMART, whose goal is to simultaneously model appearance and relation from RGB input in a separate and explicit manner. Specifically, SMART blocks decouple the spatiotemporal learning module into an appearance branch for spatial modeling and a relation branch for temporal modeling. The appearance branch is implemented based on the linear combination of pixels or filter responses in each frame, while the relation branch is designed based on the multiplicative interactions between pixels or filter responses across multiple frames. We perform experiments on three action recognition benchmarks: Kinetics, UCF101, and HMDB51, demonstrating that SMART blocks obtain an evident improvement over 3D convolutions for spatiotemporal feature learning. Under the same training setting, ARTNets achieve superior performance on these three datasets to the existing state-of-the-art methods.1

Visual question answering (VQA) is of significant interest due to its potential to be a strong test of image understanding systems and to probe the connection between language and vision. Despite much recent progress, general VQA is far from a solved problem. In this paper, we focus on the VQA multiple-choice task, and provide some good practices for designing an effective VQA model that can capture language-vision interactions and perform joint reasoning. We explore mechanisms of incorporating part-ofspeech (POS) tag guided attention, convolutional n-grams, triplet attention interactions between the image, question and candidate answer, and structured learning for triplets based on image-question pairs 1. We evaluate our models on two popular datasets: Visual7W and VQA Real Multiple Choice. Our final model achieves the state-of-the-art performance of 68.2% on Visual7W, and a very competitive performance of 69.6% on the test-standard split of VQA Real Multiple Choice.

The two-stream CNNs prove very successful for video based action recognition. However the classical two-stream CNNs are time costly, mainly due to the bottleneck of calculating optical flows. In this paper, we propose a two-stream based real-time action recognition approach by using motion vector to replace optical flow. Motion vectors are encoded in video stream and can be extracted directly without extra calculation. However directly training CNN with motion vectors degrades accuracy severely due to the noise and the lack of fine details in motion vectors. In order to relieve this problem, we propose four training strategies which leverage the knowledge learned from optical flow CNN to enhance the accuracy of motion vector CNN. Our insight is that motion vector and optical flow share inherent similar structures which allows us to transfer knowledge from one domain to another. To fully utilize the knowledge learned in optical flow domain, we develop deeply transferred motion vector CNN. Experimental results on various datasets show the effectiveness of our training strategies. Our approach is significantly faster than optical flow based approaches and achieves processing speed of 390.7 frames per second, surpassing real-time requirement. We release our model and code to facilitate further research.

Deep ConvNets have been shown to be effective for the task of human pose estimation from single images. However, several challenging issues arise in the video-based case such as self-occlusion, motion blur, and uncommon poses with few or no examples in the training data. Temporal information can provide additional cues about the location of body joints and help to alleviate these issues. In this paper, we propose a deep structured model to estimate a sequence of human poses in unconstrained videos. This model can be efficiently trained in an end-to-end manner and is capable of representing the appearance of body joints and their spatio-temporal relationships simultaneously. Domain knowledge about the human body is explicitly incorporated into the network providing effective priors to regularize the skeletal structure and to enforce temporal consistency. The proposed end-to-end architecture is evaluated on two widely used benchmarks for video-based pose estimation (Penn Action and JHMDB datasets). Our approach outperforms several state-of-the-art methods.

Current action recognition methods heavily rely on trimmed videos for model training. However, it is expensive and time-consuming to acquire a large-scale trimmed video dataset. This paper presents a new weakly supervised architecture, called UntrimmedNet, which is able to directly learn action recognition models from untrimmed videos without the requirement of temporal annotations of action instances. Our UntrimmedNet couples two important components, the classification module and the selection module, to learn the action models and reason about the temporal duration of action instances, respectively. These two components are implemented with feed-forward networks, and UntrimmedNet is therefore an end-to-end trainable architecture. We exploit the learned models for action recognition (WSR) and detection (WSD) on the untrimmed video datasets of THUMOS14 and ActivityNet. Although our UntrimmedNet only employs weak supervision, our method achieves performance superior or comparable to that of those strongly supervised approaches on these two datasets.

Detecting actions in untrimmed videos is an important yet challenging task. In this paper, we present the structured segment network (SSN), a novel framework which models the temporal structure of each action instance via a structured temporal pyramid. On top of the pyramid, we further introduce a decomposed discriminative model comprising two classifiers, respectively for classifying actions and determining completeness. This allows the framework to effectively distinguish positive proposals from background or incomplete ones, thus leading to both accurate recognition and localization. These components are integrated into a unified network that can be efficiently trained in an end-to-end fashion. Additionally, a simple yet effective temporal action proposal scheme, dubbed temporal actionness grouping (TAG) is devised to generate high quality action proposals. On two challenging benchmarks, THUMOS14 and ActivityNet, our method remarkably outperforms previous state-of-the-art methods, demonstrating superior accuracy and strong adaptivity in handling actions with various temporal structures.

Convolutional neural networks (CNNs) have recently achieved remarkable successes in various image classification and understanding tasks. The deep features obtained at the top fully connected layer of the CNN (FC-features) exhibit rich global semantic information and are extremely effective in image classification. On the other hand, the convolutional features in the middle layers of the CNN also contain meaningful local information, but are not fully explored for image representation. In this paper, we propose a novel locally supervised deep hybrid model (LS-DHM) that effectively enhances and explores the convolutional features for scene recognition. First, we notice that the convolutional features capture local objects and fine structures of scene images, which yield important cues for discriminating ambiguous scenes, whereas these features are significantly eliminated in the highly compressed FC representation. Second, we propose a new local convolutional supervision layer to enhance the local structure of the image by directly propagating the label information to the convolutional layers. Third, we propose an efficient Fisher convolutional vector (FCV) that successfully rescues the orderless mid-level semantic information (e.g., objects and textures) of scene image. The FCV encodes the large-sized convolutional maps into a fixed-length mid-level representation, and is demonstrated to be strongly complementary to the high-level FC-features. Finally, both the FCV and FC-features are collaboratively employed in the LS-DHM representation, which achieves outstanding performance in our experiments. It obtains 83.75% and 67.56% accuracies, respectively, on the heavily benchmarked MIT Indoor67 and SUN397 data sets, advancing the state-of-the-art substantially.

Convolutional neural networks (CNNs) have made remarkable progress on scene recognition, partially due to these recent large-scale scene datasets, such as the Places and Places2. Scene categories are often defined by multi-level information, including local objects, global layout, and background environment, thus leading to large intra-class variations. In addition, with the increasing number of scene categories, label ambiguity has become another crucial issue in large-scale classification. This paper focuses on large-scale scene recognition and makes two major contributions to tackle these issues. First, we propose a multi-resolution CNN architecture that captures visual content and structure at multiple levels. The multi-resolution CNNs are composed of coarse resolution CNNs and fine resolution CNNs, which are complementary to each other. Second, we design two knowledge guided disambiguation techniques to deal with the problem of label ambiguity: 1) we exploit the knowledge from the confusion matrix computed on validation data to merge ambiguous classes into a super category and 2) we utilize the knowledge of extra networks to produce a soft label for each image. Then, the super categories or soft labels are employed to guide CNN training on the Places2. We conduct extensive experiments on three large-scale image datasets (ImageNet, Places, and Places2), demonstrating the effectiveness of our approach. Furthermore, our method takes part in two major scene recognition challenges, and achieves the second place at the Places2 challenge in ILSVRC 2015, and the first place at the LSUN challenge in CVPR 2016. Finally, we directly test the learned representations on other scene benchmarks, and obtain the new state-of-the-art results on the MIT Indoor67 (86.7%) and SUN397 (72.0%). We release the code and models at https://github.com/wanglimin/MRCNN-Scene-Recognition.

Traditional feature encoding scheme (e.g., Fisher vector) with local descriptors (e.g., SIFT) and recent convolutional neural networks (CNNs) are two classes of successful methods for image recognition. In this paper, we propose a hybrid representation, which leverages the discriminative capacity of CNNs and the simplicity of descriptor encoding schema for image recognition, with a focus on scene recognition. To this end, we make three main contributions from the following aspects. First, we propose a patch-level and end-to-end architecture to model the appearance of local patches, called PatchNet. PatchNet is essentially a customized network trained in a weakly supervised manner, which uses the image-level supervision to guide the patch-level feature extraction. Second, we present a hybrid visual representation, called VSAD, by utilizing the robust feature representations of PatchNet to describe local patches and exploiting the semantic probabilities of PatchNet to aggregate these local patches into a global representation. Third, based on the proposed VSAD representation, we propose a new state-of-the-art scene recognition approach, which achieves an excellent performance on two standard benchmarks: MIT Indoor67 (86.2%) and SUN397 (73.0%).

Actionness was introduced to quantify the likelihood of containing a generic action instance at a specific location. Accurate and efficient estimation of actionness is important in video analysis and may benefit other relevant tasks such as action recognition and action detection. This paper presents a new deep architecture for actionness estimation, called hybrid fully convolutional network (HFCN), which is composed of appearance FCN (A-FCN) and motion FCN (M-FCN). These two FCNs leverage the strong capacity of deep models to estimate actionness maps from the perspectives of static appearance and dynamic motion, respectively. In addition, the fully convolutional nature of H-FCN allows it to efficiently process videos with arbitrary sizes. Experiments are conducted on the challenging datasets of Stanford40, UCF Sports, and JHMDB to verify the effectiveness of H-FCN on actionness estimation, which demonstrate that our method achieves superior performance to previous ones. Moreover, we apply the estimated actionness maps on action proposal generation and action detection. Our actionness maps advance the current state-of-the-art performance of these tasks substantially.

The deep two-stream architecture [23] exhibited excellent performance on video based action recognition. The most computationally expensive step in this approach comes from the calculation of optical flow which prevents it to be real-time. This paper accelerates this architecture by replacing optical flow with motion vector which can be obtained directly from compressed videos without extra calculation. However, motion vector lacks fine structures, and contains noisy and inaccurate motion patterns, leading to the evident degradation of recognition performance. Our key insight for relieving this problem is that optical flow and motion vector are inherent correlated. Transferring the knowledge learned with optical flow CNN to motion vector CNN can significantly boost the performance of the latter. Specifically, we introduce three strategies for this, initialization transfer, supervision transfer and their combination. Experimental results show that our method achieves comparable recognition performance to the state-of-the-art, while our method can process 390.7 frames per second, which is 27 times faster than the original two-stream method.

Recent studies demonstrate the effectiveness of super vector representation in a number of visual recognition tasks. One popular approach along this line is the Vector of Locally Aggregated Descriptor (VLAD) where the super vector is encoded with a codebook generated by k-means. However, the effectiveness of the codebook is often limited, due to the poor clustering solution, the high dimensionality of visual descriptors and the global PCA for data preprocessing. To circumvent these problems, we propose three approaches for codebook enhancement, (i) partition of data, (ii) partition of feature, and (iii) local PCA. Moreover, all these approaches can be effectively integrated together to further boost the recognition performance. In our experiments, we evaluate our enhancement approaches on two challenging visual tasks, i.e., action recognition (HMDB51) and object recognition (PASCAL VOC2007). The results show that our approaches and the fusion versions significantly outperform the baselines.

Visual features are of vital importance for human action understanding in videos. This paper presents a new video representation, called trajectory-pooled deep-convolutional descriptor (TDD), which shares the merits of both hand-crafted features [31] and deep-learned features [24]. Specifically, we utilize deep architectures to learn discriminative convolutional feature maps, and conduct trajectory-constrained pooling to aggregate these convolutional features into effective descriptors. To enhance the robustness of TDDs, we design two normalization methods to transform convolutional feature maps, namely spatiotemporal normalization and channel normalization. The advantages of our features come from (i) TDDs are automatically learned and contain high discriminative capacity compared with those hand-crafted features; (ii) TDDs take account of the intrinsic characteristics of temporal dimension and introduce the strategies of trajectory-constrained sampling and pooling for aggregating deep-learned features. We conduct experiments on two challenging datasets: HMD-B51 and UCF101. Experimental results show that TDDs outperform previous hand-crafted features [31] and deep-learned features [24]. Our method also achieves superior performance to the state of the art on these datasets.

Images and videos are often characterized by multiple types of local descriptors such as SIFT, HOG and HOF, each of which describes certain aspects of object feature. Recognition systems benefit from fusing multiple types of these descriptors. Two widely applied fusion pipelines are descriptor concatenation and kernel average. The first one is effective when different descriptors are strongly correlated, while the second one is probably better when descriptors are relatively independent. In practice, however, different descriptors are neither fully independent nor fully correlated, and previous fusion methods may not be satisfying. In this paper, we propose a new global representation, Multi-View Super Vector (MVSV), which is composed of relatively independent components derived from a pair of descriptors. Kernel average is then applied on these components to produce recognition result. To obtain MVSV, we develop a generative mixture model of probabilistic canonical correlation analyzers (M-PCCA), and utilize the hidden factors and gradient vectors of M-PCCA to construct MVSV for video representation. Experiments on video based action recognition tasks show that MVSV achieves promising results, and outperforms FV and VLAD with descriptor concatenation or kernel average fusion strategy.

Many mid-level representations have been developed to replace traditional bag-of-words model (VQ+k-means) such as sparse coding, OMP-k with k-SVD, and fisher vector with GMM in image domain. These approaches can be split into a dictionary learning phase and a feature encoding phase which are often closely related. In this paper, we jointly evaluate the effect of these two phases for video-based action recognition. Specially, we compare several dictionary learning methods and feature encoding schemes through extensive experiments on the KTH and HMDB51 datasets. Experimental results indicate that fisher vector performs consistently better than the other encoding methods, and sparse coding is robust to different dictionaries even random weights. In addition, we observe that the advantages of sophisticated mid-level representations do not come from their specific dictionaries but the encoding mechanisms, and we can just use randomly selected exemplars as dictionaries for most of encoding methods. Finally, we achieve the state-of-the-art results on the HMDB51 and UCF101 by combining our configurations with improved dense trajectory features.

Modeling the temporal structure of sub-activities is an important yet challenging problem in complex activity classification. This paper proposes a latent hierarchical model (LHM) to describe the decomposition of complex activity into sub-activities in a hierarchical way. The LHM has a tree-structure, where each node corresponds to a video segment (sub-activity) at certain temporal scale. The starting and ending time points of each sub-activity are represented by two latent variables, which are automatically determined during the inference process. We formulate the training problem of the LHM in a latent kernelized SVM framework and develop an efficient cascade inference method to speed up classification. The advantages of our methods come from: 1) LHM models the complex activity with a deep structure, which is decomposed into sub-activities in a coarse-to-fine manner and 2) the starting and ending time points of each segment are adaptively determined to deal with the temporal displacement and duration variation of sub-activity. We conduct experiments on three datasets: 1) the KTH; 2) the Hollywood2; and 3) the Olympic Sports. The experimental results show the effectiveness of the LHM in complex activity classification. With dense features, our LHM achieves the state-of-the-art performance on the Hollywood2 dataset and the Olympic Sports dataset.

This paper proposes motionlet, a mid-level and spatiotemporal part, for human motion recognition. Motion let can be seen as a tight cluster in motion and appearance space, corresponding to the moving process of different body parts. We postulate three key properties of motion let for action recognition: high motion saliency, multiple scale representation, and representative-discriminative ability. Towards this goal, we develop a data-driven approach to learn motion lets from training videos. First, we extract 3D regions with high motion saliency. Then we cluster these regions and preserve the centers as candidate templates for motion let. Finally, we examine the representative and discriminative power of the candidates, and introduce a greedy method to select effective candidates. With motion lets, we present a mid-level representation for video, called motionlet activation vector. We conduct experiments on three datasets, KTH, HMDB51, and UCF50. The results show that the proposed methods significantly outperform state-of-the-art methods.

This paper proposes motion atom and phrase as a mid-level temporal ``part'' for representing and classifying complex action. Motion atom is defined as an atomic part of action, and captures the motion information of action video in a short temporal scale. Motion phrase is a temporal composite of multiple motion atoms with an AND/OR structure, which further enhances the discriminative ability of motion atoms by incorporating temporal constraints in a longer scale. Specifically, given a set of weakly labeled action videos, we firstly design a discriminative clustering method to automatically discover a set of representative motion atoms. Then, based on these motion atoms, we mine effective motion phrases with high discriminative and representative power. We introduce a bottom-up phrase construction algorithm and a greedy selection method for this mining task. We examine the classification performance of the motion atom and phrase based representation on two complex action datasets: Olympic Sports and UCF50. Experimental results show that our method achieves superior performance over recent published methods on both datasets.

In this paper, we present a novel approach for multiclass object detection by combining local appearances and contextual constraints. We first construct a multiclass Hough forest of local patches, which can well deal with multiclass object deformations and local appearance variations, due to randomization and discrimination of the forest. Then, in the object hypothesis space, a new multiclass context model is proposed to capture relative location constraints, disambiguating appearance inputs in multiclass object detection. Finally, multiclass objects are detected with a greedy search algorithm efficiently. Experimental evaluations on two image data sets show that the combination of local appearances and context achieves state-of-the-art performance in multiclass object detection.