Composed Video Retrieval (CoVR) retrieves a target video given a query video and a modification text describing the intended change. Existing CoVR benchmarks emphasize appearance shifts or coarse event changes and therefore do not test the ability to capture subtle, fast-paced temporal differences. We introduce TF-CoVR, the first large-scale benchmark dedicated to temporally fine-grained CoVR. TF-CoVR focuses on gymnastics and diving and provides 180K triplets drawn from FineGym and FineDiving. Previous CoVR benchmarks focusing on temporal aspect, link each query to a single target segment taken from the same video, limiting practical usefulness. In TF-CoVR, we instead construct each <query, modification> pair by prompting an LLM with the label differences between clips drawn from different videos; every pair is thus associated with multiple valid target videos (3.9 on average), reflecting real-world tasks such as sports-highlight generation. To model these temporal dynamics we propose TF-CoVR-Base, a concise two-stage training framework: (i) pre-train a video encoder on fine-grained action classification to obtain temporally discriminative embeddings; (ii) align the composed query with candidate videos using contrastive learning. We conduct the first comprehensive study of image, video, and general multimodal embedding (GME) models on temporally fine-grained composed retrieval in both zero-shot and fine-tuning regimes. On TF-CoVR, TF-CoVR-Base improves zero-shot mAP@50 from 5.92 (LanguageBind) to 7.51, and after fine-tuning raises the state-of-the-art from 19.83 to 25.82.
Traditional approaches in unsupervised or self supervised learning for skeleton-based action classification have concentrated predominantly on the dynamic aspects of skeletal sequences. Yet, the intricate interaction between the moving and static elements of the skeleton presents a rarely tapped discriminative potential for action classification. This paper introduces a novel measurement, referred to as spatial-temporal joint density (STJD), to quantify such interaction. Tracking the evolution of this density throughout an action can effectively identify a subset of discriminative moving and/or static joints termed "prime joints" to steer self-supervised learning. A new contrastive learning strategy named STJD-CL is proposed to align the representation of a skeleton sequence with that of its prime joints while simultaneously contrasting the representations of prime and nonprime joints. In addition, a method called STJD-MP is developed by integrating it with a reconstruction-based framework for more effective learning. Experimental evaluations on the NTU RGB+D 60, NTU RGB+D 120, and PKUMMD datasets in various downstream tasks demonstrate that the proposed STJD-CL and STJD-MP improved performance, particularly by 3.5 and 3.6 percentage points over the state-of-the-art contrastive methods on the NTU RGB+D 120 dataset using X-sub and X-set evaluations, respectively.
Popular video training methods mainly operate on a fixed number of tokens sampled from a predetermined spatiotemporal grid, resulting in sub-optimal accuracy-computation trade-offs due to inherent video redundancy. They also lack adaptability to varying computational budgets for downstream tasks, hindering applications of the most competitive model in real-world scenes. We thus propose a new test setting, Token Optimization, for maximized input information across budgets, which optimizes the size-limited set of input tokens through token selection from more suitably sampled videos. To this end, we propose a novel augmentation tool termed Flux. By making the sampling grid flexible and leveraging token selection, it is easily adopted in most popular video training frameworks, boosting model robustness with nearly no additional cost. We integrate Flux in large-scale video pre-training, and the resulting FluxViT establishes new state-of-the-art results across extensive tasks at standard costs. Notably, with 1/4 tokens only, it can still match the performance of previous state-of-the-art models with Token Optimization, yielding nearly 90\% savings. All models and data are available at https://github.com/OpenGVLab/FluxViT.
Temporal action localization (TAL) involves dual tasks to classify and localize actions within untrimmed videos. However, the two tasks often have conflicting requirements for features. Existing methods typically employ separate heads for classification and localization tasks but share the same input feature, leading to suboptimal performance. To address this issue, we propose a novel TAL method with Cross Layer Task Decoupling and Refinement (CLTDR). Based on the feature pyramid of video, CLTDR strategy integrates semantically strong features from higher pyramid layers and detailed boundary-aware boundary features from lower pyramid layers to effectively disentangle the action classification and localization tasks. Moreover, the multiple features from cross layers are also employed to refine and align the disentangled classification and regression results. At last, a lightweight Gated Multi-Granularity (GMG) module is proposed to comprehensively extract and aggregate video features at instant, local, and global temporal granularities. Benefiting from the CLTDR and GMG modules, our method achieves state-of-the-art performance on five challenging benchmarks: THUMOS14, MultiTHUMOS, EPIC-KITCHENS-100, ActivityNet-1.3, and HACS. Our code and pre-trained models are publicly available at: https://github.com/LiQiang0307/CLTDR-GMG.
Wireless sensing has recently found widespread applications in diverse environments, including homes, offices, and public spaces. By analyzing patterns in channel state information (CSI), it is possible to infer human actions for tasks such as person identification, gesture recognition, and fall detection. However, CSI is highly sensitive to environmental changes, where even minor alterations can significantly distort the CSI patterns. This sensitivity often leads to performance degradation or outright failure when applying wireless sensing models trained in one environment to another. To address this challenge, Domain Alignment (DAL) has been widely adopted for cross-domain classification tasks, as it focuses on aligning the global distributions of the source and target domains in feature space. Despite its popularity, DAL often neglects inter-category relationships, which can lead to misalignment between categories across domains, even when global alignment is achieved. To overcome these limitations, we propose K-Nearest Neighbors Maximum Mean Discrepancy (KNN-MMD), a novel few-shot method for cross-domain wireless sensing. Our approach begins by constructing a help set using KNN from the target domain, enabling local alignment between the source and target domains within each category using MMD. Additionally, we address a key instability issue commonly observed in cross-domain methods, where model performance fluctuates sharply between epochs. Further, most existing methods struggle to determine an optimal stopping point during training due to the absence of labeled data from the target domain. Our method resolves this by excluding the support set from the target domain during training and employing it as a validation set to determine the stopping criterion.
As a globally celebrated sport, soccer has attracted widespread interest from fans all over the world. This paper aims to develop a comprehensive multi-modal framework for soccer video understanding. Specifically, we make the following contributions in this paper: (i) we introduce SoccerReplay-1988, the largest multi-modal soccer dataset to date, featuring videos and detailed annotations from 1,988 complete matches, with an automated annotation pipeline; (ii) we present the first visual-language foundation model in the soccer domain, MatchVision, which leverages spatiotemporal information across soccer videos and excels in various downstream tasks; (iii) we conduct extensive experiments and ablation studies on event classification, commentary generation, and multi-view foul recognition. MatchVision demonstrates state-of-the-art performance on all of them, substantially outperforming existing models, which highlights the superiority of our proposed data and model. We believe that this work will offer a standard paradigm for sports understanding research.
The lack of occlusion data in commonly used action recognition video datasets limits model robustness and impedes sustained performance improvements. We construct OccludeNet, a large-scale occluded video dataset that includes both real-world and synthetic occlusion scene videos under various natural environments. OccludeNet features dynamic tracking occlusion, static scene occlusion, and multi-view interactive occlusion, addressing existing gaps in data. Our analysis reveals that occlusion impacts action classes differently, with actions involving low scene relevance and partial body visibility experiencing greater accuracy degradation. To overcome the limitations of current occlusion-focused approaches, we propose a structural causal model for occluded scenes and introduce the Causal Action Recognition (CAR) framework, which employs backdoor adjustment and counterfactual reasoning. This framework enhances key actor information, improving model robustness to occlusion. We anticipate that the challenges posed by OccludeNet will stimulate further exploration of causal relations in occlusion scenarios and encourage a reevaluation of class correlations, ultimately promoting sustainable performance improvements. The code and full dataset will be released soon.
Robots need to perceive persons in their surroundings for safety and to interact with them. In this paper, we present a person segmentation and action classification approach that operates on 3D scans of hemisphere field of view LiDAR sensors. We recorded a data set with an Ouster OSDome-64 sensor consisting of scenes where persons perform three different actions and annotated it. We propose a method based on a MaskDINO model to detect and segment persons and to recognize their actions from combined spherical projected multi-channel representations of the LiDAR data with an additional positional encoding. Our approach demonstrates good performance for the person segmentation task and further performs well for the estimation of the person action states walking, waving, and sitting. An ablation study provides insights about the individual channel contributions for the person segmentation task. The trained models, code and dataset are made publicly available.
Extracting multiscale contextual information and higher-order correlations among skeleton sequences using Graph Convolutional Networks (GCNs) alone is inadequate for effective action classification. Hypergraph convolution addresses the above issues but cannot harness the long-range dependencies. Transformer proves to be effective in capturing these dependencies and making complex contextual features accessible. We propose an Autoregressive Adaptive HyperGraph Transformer (AutoregAd-HGformer) model for in-phase (autoregressive and discrete) and out-phase (adaptive) hypergraph generation. The vector quantized in-phase hypergraph equipped with powerful autoregressive learned priors produces a more robust and informative representation suitable for hyperedge formation. The out-phase hypergraph generator provides a model-agnostic hyperedge learning technique to align the attributes with input skeleton embedding. The hybrid (supervised and unsupervised) learning in AutoregAd-HGformer explores the action-dependent feature along spatial, temporal, and channel dimensions. The extensive experimental results and ablation study indicate the superiority of our model over state-of-the-art hypergraph architectures on NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets.
Human action recognition has been a topic of interest across multiple fields ranging from security to entertainment systems. Tracking the motion and identifying the action being performed on a real time basis is necessary for critical security systems. In entertainment, especially gaming, the need for immediate responses for actions and gestures are paramount for the success of that system. We show that Motion History image has been a well established framework to capture the temporal and activity information in multi dimensional detail enabling various usecases including classification. We utilize MHI to produce sample data to train a classifier and demonstrate its effectiveness for action classification across six different activities in a single multi-action video. We analyze the classifier performance and identify usecases where MHI struggles to generate the appropriate activity image and discuss mechanisms and future work to overcome those limitations.
Skeleton-based Temporal Action Segmentation involves the dense action classification of variable-length skeleton sequences. Current approaches primarily apply graph-based networks to extract framewise, whole-body-level motion representations, and use one-hot encoded labels for model optimization. However, whole-body motion representations do not capture fine-grained part-level motion representations and the one-hot encoded labels neglect the intrinsic semantic relationships within the language-based action definitions. To address these limitations, we propose a novel method named Language-assisted Human Part Motion Representation Learning (LPL), which contains a Disentangled Part Motion Encoder (DPE) to extract dual-level (i.e., part and whole-body) motion representations and a Language-assisted Distribution Alignment (LDA) strategy for optimizing spatial relations within representations. Specifically, after part-aware skeleton encoding via DPE, LDA generates dual-level action descriptions to construct a textual embedding space with the help of a large-scale language model. Then, LDA motivates the alignment of the embedding space between text descriptions and motions. This alignment allows LDA not only to enhance intra-class compactness but also to transfer the language-encoded semantic correlations among actions to skeleton-based motion learning. Moreover, we propose a simple yet efficient Semantic Offset Adapter to smooth the cross-domain misalignment. Our experiments indicate that LPL achieves state-of-the-art performance across various datasets (e.g., +4.4\% Accuracy, +5.6\% F1 on the PKU-MMD dataset). Moreover, LDA is compatible with existing methods and improves their performance (e.g., +4.8\% Accuracy, +4.3\% F1 on the LARa dataset) without additional inference costs.
In recent years, Wi-Fi sensing has garnered significant attention due to its numerous benefits, such as privacy protection, low cost, and penetration ability. Extensive research has been conducted in this field, focusing on areas such as gesture recognition, people identification, and fall detection. However, many data-driven methods encounter challenges related to domain shift, where the model fails to perform well in environments different from the training data. One major factor contributing to this issue is the limited availability of Wi-Fi sensing datasets, which makes models learn excessive irrelevant information and over-fit to the training set. Unfortunately, collecting large-scale Wi-Fi sensing datasets across diverse scenarios is a challenging task. To address this problem, we propose CrossFi, a siamese network-based approach that excels in both in-domain scenario and cross-domain scenario, including few-shot, zero-shot scenarios, and even works in few-shot new-class scenario where testing set contains new categories. The core component of CrossFi is a sample-similarity calculation network called CSi-Net, which improves the structure of the siamese network by using an attention mechanism to capture similarity information, instead of simply calculating the distance or cosine similarity. Based on it, we develop an extra Weight-Net that can generate a template for each class, so that our CrossFi can work in different scenarios. Experimental results demonstrate that our CrossFi achieves state-of-the-art performance across various scenarios. In gesture recognition task, our CrossFi achieves an accuracy of 98.17% in in-domain scenario, 91.72% in one-shot cross-domain scenario, 64.81% in zero-shot cross-domain scenario, and 84.75% in one-shot new-class scenario. The code for our model is publicly available at https://github.com/RS2002/CrossFi.
Weakly supervised temporal action localization (WTAL) aims to detect action instances in untrimmed videos using only video-level annotations. Since many existing works optimize WTAL models based on action classification labels, they encounter the task discrepancy problem (i.e., localization-by-classification). To tackle this issue, recent studies have attempted to utilize action category names as auxiliary semantic knowledge through vision-language pre-training (VLP). However, there are still areas where existing research falls short. Previous approaches primarily focused on leveraging textual information from language models but overlooked the alignment of dynamic human action and VLP knowledge in a joint space. Furthermore, the deterministic representation employed in previous studies struggles to capture fine-grained human motions. To address these problems, we propose a novel framework that aligns human action knowledge and VLP knowledge in a probabilistic embedding space. Moreover, we propose intra- and inter-distribution contrastive learning to enhance the probabilistic embedding space based on statistical similarities. Extensive experiments and ablation studies reveal that our method significantly outperforms all previous state-of-the-art methods. Code is available at https://github.com/sejong-rcv/PVLR.
Existing multimodal-based human action recognition approaches are either computationally expensive, which limits their applicability in real-time scenarios, or fail to exploit the spatial temporal information of multiple data modalities. In this work, we present an efficient pose-driven attention-guided multimodal network (EPAM-Net) for action recognition in videos. Specifically, we adapted X3D networks for both RGB and pose streams to capture spatio-temporal features from RGB videos and their skeleton sequences. Then skeleton features are utilized to help the visual network stream focusing on key frames and their salient spatial regions using a spatial temporal attention block. Finally, the scores of the two streams of the proposed network are fused for final classification. The experimental results show that our method achieves competitive performance on NTU-D 60 and NTU RGB-D 120 benchmark datasets. Moreover, our model provides a 6.2--9.9x reduction in FLOPs (floating-point operation, in number of multiply-adds) and a 9--9.6x reduction in the number of network parameters. The code will be available at https://github.com/ahmed-nady/Multimodal-Action-Recognition.
We present EgoExo-Fitness, a new full-body action understanding dataset, featuring fitness sequence videos recorded from synchronized egocentric and fixed exocentric (third-person) cameras. Compared with existing full-body action understanding datasets, EgoExo-Fitness not only contains videos from first-person perspectives, but also provides rich annotations. Specifically, two-level temporal boundaries are provided to localize single action videos along with sub-steps of each action. More importantly, EgoExo-Fitness introduces innovative annotations for interpretable action judgement--including technical keypoint verification, natural language comments on action execution, and action quality scores. Combining all of these, EgoExo-Fitness provides new resources to study egocentric and exocentric full-body action understanding across dimensions of "what", "when", and "how well". To facilitate research on egocentric and exocentric full-body action understanding, we construct benchmarks on a suite of tasks (i.e., action classification, action localization, cross-view sequence verification, cross-view skill determination, and a newly proposed task of guidance-based execution verification), together with detailed analysis. Code and data will be available at https://github.com/iSEE-Laboratory/EgoExo-Fitness/tree/main.
We introduce InternVideo2, a new family of video foundation models (ViFM) that achieve the state-of-the-art results in video recognition, video-text tasks, and video-centric dialogue. Our core design is a progressive training approach that unifies the masked video modeling, crossmodal contrastive learning, and next token prediction, scaling up the video encoder size to 6B parameters. At the data level, we prioritize spatiotemporal consistency by semantically segmenting videos and generating video-audio-speech captions. This improves the alignment between video and text. Through extensive experiments, we validate our designs and demonstrate superior performance on over 60 video and audio tasks. Notably, our model outperforms others on various video-related dialogue and long video understanding benchmarks, highlighting its ability to reason and comprehend longer contexts. Code and models are available at https://github.com/OpenGVLab/InternVideo/tree/main/InternVideo2/.
Despite the development of various deep learning methods for Wi-Fi sensing, package loss often results in noncontinuous estimation of the Channel State Information (CSI), which negatively impacts the performance of the learning models. To overcome this challenge, we propose a deep learning model based on Bidirectional Encoder Representations from Transformers (BERT) for CSI recovery, named CSI-BERT. CSI-BERT can be trained in an self-supervised manner on the target dataset without the need for additional data. Furthermore, unlike traditional interpolation methods that focus on one subcarrier at a time, CSI-BERT captures the sequential relationships across different subcarriers. Experimental results demonstrate that CSI-BERT achieves lower error rates and faster speed compared to traditional interpolation methods, even when facing with high loss rates. Moreover, by harnessing the recovered CSI obtained from CSI-BERT, other deep learning models like Residual Network and Recurrent Neural Network can achieve an average increase in accuracy of approximately 15\% in Wi-Fi sensing tasks. The collected dataset WiGesture and code for our model are publicly available at https://github.com/RS2002/CSI-BERT.
Addressing the dual challenges of local redundancy and global dependencies in video understanding, this work innovatively adapts the Mamba to the video domain. The proposed VideoMamba overcomes the limitations of existing 3D convolution neural networks and video transformers. Its linear-complexity operator enables efficient long-term modeling, which is crucial for high-resolution long video understanding. Extensive evaluations reveal VideoMamba's four core abilities: (1) Scalability in the visual domain without extensive dataset pretraining, thanks to a novel self-distillation technique; (2) Sensitivity for recognizing short-term actions even with fine-grained motion differences; (3) Superiority in long-term video understanding, showcasing significant advancements over traditional feature-based models; and (4) Compatibility with other modalities, demonstrating robustness in multi-modal contexts. Through these distinct advantages, VideoMamba sets a new benchmark for video understanding, offering a scalable and efficient solution for comprehensive video understanding. All the code and models are available at https://github.com/OpenGVLab/VideoMamba.
A comprehensive understanding of videos is inseparable from describing the action with its contextual action-object interactions. However, many current video understanding tasks prioritize general action classification and overlook the actors and relationships that shape the nature of the action, resulting in a superficial understanding of the action. Motivated by this, we introduce Open-vocabulary Video Relation Extraction (OVRE), a novel task that views action understanding through the lens of action-centric relation triplets. OVRE focuses on pairwise relations that take part in the action and describes these relation triplets with natural languages. Moreover, we curate the Moments-OVRE dataset, which comprises 180K videos with action-centric relation triplets, sourced from a multi-label action classification dataset. With Moments-OVRE, we further propose a crossmodal mapping model to generate relation triplets as a sequence. Finally, we benchmark existing cross-modal generation models on the new task of OVRE.
Video transformers have become the de facto standard for human action recognition, yet their exclusive reliance on the RGB modality still limits their adoption in certain domains. One such domain is Activities of Daily Living (ADL), where RGB alone is not sufficient to distinguish between visually similar actions, or actions observed from multiple viewpoints. To facilitate the adoption of video transformers for ADL, we hypothesize that the augmentation of RGB with human pose information, known for its sensitivity to fine-grained motion and multiple viewpoints, is essential. Consequently, we introduce the first Pose Induced Video Transformer: PI-ViT (or $\pi$-ViT), a novel approach that augments the RGB representations learned by video transformers with 2D and 3D pose information. The key elements of $\pi$-ViT are two plug-in modules, 2D Skeleton Induction Module and 3D Skeleton Induction Module, that are responsible for inducing 2D and 3D pose information into the RGB representations. These modules operate by performing pose-aware auxiliary tasks, a design choice that allows $\pi$-ViT to discard the modules during inference. Notably, $\pi$-ViT achieves the state-of-the-art performance on three prominent ADL datasets, encompassing both real-world and large-scale RGB-D datasets, without requiring poses or additional computational overhead at inference.