{"id": "2606.20559", "title": "UNIEGO: Proxies as Mediators for Unified Egocentric Video Representation Learning", "abstract": "Egocentric video understanding is inherently limited by the narrow perspective of wearable cameras: a single viewpoint, a single modality, a single model cannot capture the full richness of human action. We argue that a truly expressive egocentric representation must subsume complementary knowledge across viewpoints, modalities, and foundation model representations, yet remain deployable from egocentric video alone. To this end, we introduce a hierarchical multi-teacher distillation framework that produces UNIEGO, a unified egocentric encoder trained with nine teachers spanning ego-exo viewpoints, RGB, depth, and skeleton modalities, and four foundation models. Rather than distilling directly from heterogeneous teachers whose incompatible architectures and feature geometries induce conflicting gradients, our framework interposes a layer of representation-specific Proxy models that translate diverse teacher knowledge into a homogeneous egocentric space. A second distillation stage, Selective Proxy Distillation (SPD), then adaptively selects, for each training sample, the subset of proxies that are both correct and confident, distilling exclusively from reliable supervision and suppressing erroneous signals. SPD is further stabilized by initializing UNIEGO as a learned convex combination of proxy parameters, placing the unified model in a well-conditioned region of the loss landscape before distillation begins. UNIEGO achieves state-of-the-art performance across three egocentric video understanding tasks - action recognition, video retrieval, and action segmentation on three challenging ego-exo benchmarks, outperforming naive multi-teacher distillation baselines and demonstrating that structured, proxy-mediated knowledge transfer yields richer and more discriminative egocentric representations.", "date": "2026-06-18", "categories": []}
{"id": "2601.17053", "title": "Synthetic Data Guided Feature Selection for Robust Activity Recognition in Older Adults", "abstract": "Physical activity during hip fracture rehabilitation is essential for mitigating long-term functional decline in geriatric patients. However, it is rarely quantified in clinical practice. Existing continuous monitoring systems with commercially available wearable activity trackers are typically developed in middle-aged adults and therefore perform unreliably in older adults with slower and more variable gait patterns. This study aimed to develop a robust human activity recognition (HAR) system to improve continuous physical activity recognition in the context of hip fracture rehabilitation. 24 healthy older adults aged over 80 years were included to perform activities of daily living (walking, standing, sitting, lying down, and postural transfers) under simulated free-living conditions for 75 minutes while wearing two accelerometers positioned on the lower back and anterior upper thigh. Model robustness was evaluated using leave-one-subject-out cross-validation. The synthetic data demonstrated potential to improve generalization across participants. The resulting feature intervention model (FIM), aided by synthetic data guidance, achieved reliable activity recognition with mean F1-scores of 0.896 for walking, 0.927 for standing, 0.997 for sitting, 0.937 for lying down, and 0.816 for postural transfers. Compared with a control condition model without synthetic data, the FIM significantly improved the postural transfer detection, i.e., an activity class of high clinical relevance that is often overlooked in existing HAR literature. In conclusion, these preliminary results demonstrate the feasibility of robust activity recognition in older adults. Further validation in hip fracture patient populations is required to assess the clinical utility of the proposed monitoring system.", "date": "2026-01-21", "categories": []}
{"id": "2606.14096", "title": "A New Multi-Domain Benchmark for Micro-Action Recognition and Detection", "abstract": "Micro-actions are short-duration, low-amplitude subtle body movements at the whole-body level that can reveal latent intentions, involuntary reactions, and fine-grained affective changes. Our previous MA-52 benchmark has provided an important foundation for micro-action recognition, but it remains limited in scale, scene diversity, task coverage, and evaluation protocols. To advance micro-action analysis toward more realistic and comprehensive settings, we introduce MMA-82, a large-scale multi-domain extension of MA-52. MMA-82 expands the label space from 52 to 82 fine-grained micro-action categories and covers four distinct domains, including laboratory interviews, street interviews, psychiatric patient interviews, and emotion-rich television videos, resulting in 77,856 annotated instances from 454 subjects. Built upon MMA-82, we establish two core tasks: Micro-Action Recognition and Multi-label Micro-Action Detection. For recognition, we further define in-domain and cross-domain protocols, including few-shot and zero-shot settings, to evaluate model robustness, transferability, and generalization. Extensive experiments show that current methods still struggle with realistic micro-action understanding, especially under domain shift, long-tailed category distributions, and complex temporal localization. Beyond benchmarking, we investigate the relationship between micro-actions and emotion, showing that micro-actions are strongly associated with emotional states and provide complementary cues to facial micro-expressions for improved emotion recognition. These results demonstrate that MMA-82 serves as a comprehensive and challenging benchmark for realistic micro-action analysis and a valuable resource for human-centered AI. MMA-82 is available at https://lpynow.github.io/MMA-82-AIM/.", "date": "2026-06-12", "categories": []}
{"id": "2606.17627", "title": "Divide, Deliberate, Decide: A Multi-Agent Framework for Fine-Grained Egocentric Action Recognition", "abstract": "Fine-grained action recognition in egocentric video is challenging for Vision-Language Models (VLMs): actions often differ only in small visual cues, and a single model tends to be biased toward a subset of these cues. We propose Divide, Deliberate, Decide, a fully-local, zero-shot multi-agent framework in which (i) a VLM orchestrator chunks the video and proposes a top-k candidate label list per segment, (ii) an ensemble of heterogeneous VLM specialists, drawn from different open model families, engages in a structured deliberation that includes a peer-consultation round of questions, and (iii) agent rankings are aggregated with a Borda count and the orchestrator re-ranks its own prediction in light of the specialists' evidence. The entire pipeline runs locally with no fine-tuning. Experiments show that our method positively improves zero-shot action recognition performance over the baseline, highlighting the influence of a heterogeneous deliberation step, showing that the gain stems from decorrelated model priors rather than from additional compute.", "date": "2026-06-16", "categories": []}
{"id": "2603.16970", "title": "MAND: Modality-Aware Novelty Detection for Open-World Egocentric Activity Recognition", "abstract": "Multimodal egocentric activity recognition integrates visual and inertial cues for robust first-person behavior understanding. However, deploying such systems in open-world environments requires detecting novel activities while continuously learning from non-stationary data streams. Existing methods rely on the main fused logits for novelty scoring, without fully exploiting the complementary evidence available from individual modalities. Because these logits are often dominated by RGB, cues from other modalities, particularly IMU, remain underutilized, and this imbalance worsens as catastrophic forgetting accumulates. To address this, we propose MAND, a modality-aware framework for multimodal egocentric open-world continual learning. At inference, Modality-aware Adaptive Scoring (MoAS) adaptively adjusts modality contributions using sample-wise reliability and refines novelty scoring with deviation and disagreement penalties. During training, Modality-aware Representation Stabilization Training (MoRST) preserves the discriminative capacity of each modality across tasks through modality-specific heads and modality-wise logit distillation. Experiments on a public multimodal egocentric benchmark show that MAND consistently improves novel activity detection and known-class accuracy while substantially reducing FPR95, indicating more reliable open-world recognition. The source code is available at \\href{https://github.com/HyeJeongIm/MAND}{github.com/HyeJeongIm/MAND}.", "date": "2026-03-17", "categories": []}
{"id": "2606.14730", "title": "Hierarchical GRU with Input-Conditioned Slot Queries for Ball Action Anticipation", "abstract": "We present a hierarchical model for ball action anticipation in football broadcast video. Given a 30-second observation window, the system predicts actions occurring in the subsequent 5-second window across 10 classes. A shared local Transformer encodes clip-level features within each 5-second sub-window; a GRU then aggregates temporal context across all sub-windows; finally, a Transformer decoder with K input-conditioned event slots decodes the anticipation target via three decoupled heads (objectness, class, temporal offset). We introduce frequency-reweighted Hungarian matching that systematically favours rare action classes, and Gaussian soft targets for temporal bin supervision. On the SoccerNet Ball Action Anticipation benchmark, our method achieves 17.91% mAP on the test server.", "date": "2026-06-02", "categories": []}
{"id": "2606.15275", "title": "MamBOA: State-Space Architecture for Video Recognition", "abstract": "Fine-grained action recognition demands temporal reasoning that general-purpose architectures address through different cost-accuracy tradeoffs: 3D dense operators couple computation to the input volume, while difference-based methods approximate motion through rigid, hand-crafted subtraction of uncontextualized features - each reflecting a deliberate design choice with corresponding limitations in expressiveness or flexibility. We present MamBOA, a backbone-agnostic temporal framework built upon a novel interleaved scan structure that recasts the selective state-space recurrence (S6) as a native motion synthesizer. By interleaving consecutive feature representations extracted from a pretrained backbone into a single alternating sequence, the proposed scan structurally drives the recurrence to encode both temporal observations of each position within a shared hidden state, separated by only a single decay step - rendering the inter-frame transition an intrinsic component of the state dynamics rather than an externally computed quantity. A cascade of dedicated alignment and decoding operations then distills this joint encoding into an explicit motion representation, which a dual-path pooling mechanism adaptively aggregates by balancing attention-driven selection with uniform temporal coverage. The framework interfaces seamlessly with CNN, Transformer, and Mamba backbone families, adding only ~2.1 GFLOPs per feature pair. On Diving48, MamBOA achieves 85.02% Top-1 accuracy with an image-pretrained backbone and 86.24% with a video-pretrained backbone processing the entire video in a single forward pass - demonstrating that structurally induced state-space dynamics constitute a principled and general foundation for motion modeling.", "date": "2026-06-13", "categories": []}
{"id": "2505.04907", "title": "VaCDA: Variational Contrastive Alignment-based Scalable Human Activity Recognition", "abstract": "Technological advancements have led to the rise of wearable devices with sensors that continuously monitor user activities, generating vast amounts of unlabeled data. This data is challenging to interpret, and manual annotation is labor-intensive and error-prone. Additionally, data distribution is often heterogeneous due to device placement, type, and user behavior variations. As a result, traditional transfer learning methods perform suboptimally, making it difficult to recognize daily activities. To address these challenges, we use a variational autoencoder (VAE) to learn a shared, low-dimensional latent space from available sensor data. This space generalizes data across diverse sensors, mitigating heterogeneity and aiding robust adaptation to the target domain. We integrate contrastive learning to enhance feature representation by aligning instances of the same class across domains while separating different classes. We propose Variational Contrastive Domain Adaptation (VaCDA), a multi-source domain adaptation framework combining VAEs and contrastive learning to improve feature representation and reduce heterogeneity between source and target domains. We evaluate VaCDA on multiple publicly available datasets across three heterogeneity scenarios: cross-person, cross-position, and cross-device. VaCDA outperforms the baselines in cross-position and cross-device scenarios.", "date": "2025-05-08", "categories": []}
{"id": "2604.15173", "title": "Boundary-Centric Clip-Budgeted Active Learning for Temporal Action Segmentation", "abstract": "Temporal action segmentation (TAS) in untrimmed videos requires dense temporal supervision. However, most of the annotation cost is spent identifying action transitions where segmentation errors concentrate and small temporal shifts can disproportionately degrade segment-level metrics. We introduce B-ACT, a clip-budgeted active learning framework that explicitly allocates supervision to these error-prone boundary regions. B-ACT operates in a hierarchical two-stage loop: (i) it ranks and queries unlabeled videos using predictive uncertainty, and (ii) within each selected video, it detects candidate transitions from the current model predictions and selects the top-$K$ boundaries via a novel boundary score. The boundary score fuses neighborhood uncertainty, class ambiguity, and temporal prediction dynamics to reveal the underlying importance of each frame. Importantly, our annotation protocol requests labels only at the boundary frames while still training on boundary-centered clips to exploit temporal context through the model's receptive field. Extensive experiments on GTEA, 50Salads, and Breakfast demonstrate that boundary-centric supervision delivers strong label efficiency and consistently surpasses representative TAS active learning baselines and prior state of the art under sparse budgets. Gains are largest on datasets where performance is highly sensitive to boundary placement, as measured by edit and overlap-based F1 metrics.", "date": "2026-04-16", "categories": []}
{"id": "2606.13022", "title": "Quality-Preserving Imperceptible Adversarial Attack on Skeleton-based Human Action Recognition", "abstract": "Adversarial attacks on skeletal human action recognition have received significant attention. However, existing methods typically introduce noise-like perturbations that degrade motion quality post-attack, and thereby are inherently perceptible with recent advancements in S-HAR systems. We discover that this degradation stems from the gap between empirical and true risks during the optimization process of previous adversarial attacks. To address this issue, we propose an attack where adversarial motions are obtained without compromising their motion quality. To minimize the risk gap and preserve motion quality, we propose a distribution-based adversarial attack method without introducing noise-like perturbations. To faithfully evaluate the motion quality, we propose a new metric that aligns with human perception on real-world naturalness. Experiments have been conducted on the state-of-the-art S-HAR methods across two datasets, demonstrating the superiority of our method in both the attack success rate and the post-attack motion quality through qualitative and quantitative analyses. The success of our quality-preserving attack application and distribution-based method raises serious concerns about the robustness of action recognizers, highlighting the need for further enhancements in this domain.", "date": "2026-06-11", "categories": []}
{"id": "2606.13030", "title": "A Multi-Modal Framework with Cross-Subject Pseudo-Labeling and Semantic Alignment for Micro-Gesture Recognition", "abstract": "Micro-gestures (MGs) are spontaneous and subtle body movements that frequently convey hidden human emotions. Recognizing MGs in untrimmed videos remains highly challenging due to their extremely low signal-to-noise ratio, severe long-tailed class distribution, and the inherent domain shift encountered in cross-subject evaluation scenarios. In this paper, we propose a comprehensive multi-modal framework for Track 1 of the 4th MiGA-IJCAI Challenge. To capture fine-grained representations, we design a saliency-guided multi-modal extraction pipeline integrating 68-keypoint skeleton joint coordinates, 3D heatmap volumes, and high-resolution RGB visual features. We introduce a gentle square-root smoothed weighting mechanism paired with an Orthogonal Semantic Embedding Loss to protect tail classes without compromising overall recognition capabilities. More importantly, to bridge the cross-subject generalization gap, we propose a Cross-Modal Pseudo-Labeling (CMPL) strategy for unsupervised domain adaptation, which significantly boosts single-modal robustness. A temperature-scaled soft-voting mechanism is finally utilized to alleviate overconfidence during late fusion. Extensive experiments demonstrate that our framework achieves a competitive F1-score of 68.13\\%, securing the 4th place.", "date": "2026-06-11", "categories": []}
{"id": "2606.13194", "title": "WHAR Arena: Benchmarking the State of the Art in Efficient Wearable Human Activity Recognition", "abstract": "Deep learning has become the dominant paradigm in Wearable Human Activity Recognition (WHAR), yet progress is obscured by a comparability crisis. Results are often reported using inconsistent datasets, custom data processing, and varying evaluation protocols, making state-of-the-art claims fragile. We address this with a large-scale, open-source benchmark that integrates 30 diverse datasets under standardized processing, unified model interfaces, and a shared cross-subject evaluation protocol. Evaluating 17 representative architectures across 4760 training runs, we jointly measure predictive performance alongside on-device latency, peak memory, and model size on an Android reference device. Our results reveal that the WHAR state of the art is distributed rather than dominated by a single architecture. While CNN-HAR achieves the highest mean macro-F1, top-performing models cluster tightly, indicating contemporary architectures have converged near a predictive performance ceiling. When accounting for deployment efficiency, compact neural models, such as TinierHAR, and classical Random Forests define the practically relevant Pareto frontier, whereas larger recurrent and hybrid models incur high hardware costs without corresponding performance gains. Consequently, while predictive performance has plateaued, substantial potential for future progress remains in optimizing deployment efficiency and improving adaptation to domain shifts. We release our full framework to support transparent reuse and extension.", "date": "2026-06-11", "categories": []}
{"id": "2606.13332", "title": "OR-Action: Multi-Role Video Understanding with Fine-Grained Actions", "abstract": "Fine-grained understanding of operating room (OR) activity could enable workflow-aware assistance, yet remains difficult due to clutter, occlusions, and limited sensing. The prevailing approach to model this environment is scene graphs as an interpretable representation of OR interactions. Converting their frame-wise relational predictions into temporally extended, fine-grained actions however, is challenging without explicit temporal modeling. To enable a principled temporal evaluation of current OR understanding methods, we introduce the first action-centric benchmark built on a publicly available ego-exocentric OR dataset by defining a fine-grained, multi-role action taxonomy and generating dense action segments via distillation from ground-truth scene graph state changes. Experiments on this benchmark show that current scene graph prediction methods struggle to model temporal structure, even when adding explicit modeling through Graph Neural Networks. We therefore introduce a vision-only temporal model that outperforms graph-based methods significantly when using all available egocentric video as input. Building on this model we also introduce a novel multi- to single-view feature alignment strategy that improves single-view performance on multi-role action recognition, mitigating the need for extensive egocentric video capture. Benchmark and code will be released upon acceptance.", "date": "2026-06-11", "categories": []}
{"id": "2605.19031", "title": "KAN-MLP-Mixer: A comprehensive investigation of the usage of Kolmogorov-Arnold Networks (KANs) for improving IMU-based Human Activity Recognition", "abstract": "Kolmogorov-Arnold Networks (KANs) have demonstrated an exceptional ability to learn complex functions on clean, low-dimensional data but struggle to maintain performance on noisy and imperfect real-world datasets. In contrast, conventional multi-layer perceptrons (MLPs) are far more tolerant to noise and computationally efficient. Replacing all MLP components with KANs in HAR models often degrades accuracy and computation efficiency, highlighting an open challenge: how to combine KANs' precision with MLPs' noise robustness and efficiency. To address this, we systematically explore various placements of KAN modules within deep HAR networks and propose a hybrid architecture that strategically synergizes the strengths of both paradigms, which uses a KAN-based input embedding layer, retains MLP layers for intermediate feature mixing, and introduces a specialized LarctanKAN module for final activity classification. Across eight public HAR datasets, the hybrid KAN-MLP model achieves an average macro F1 score relative improvement of 5.33\\% compared pure-MLP model, significantly outperforming standalone KAN and MLP baselines. Furthermore, integrating this hybrid strategy into other state-of-the-art HAR architectures consistently boosts their performance. Our findings demonstrate that a carefully orchestrated combination of KAN, MLP, or other conventional neural components yields more robust and accurate HAR models for real-world wearable sensing environments.", "date": "2026-05-18", "categories": []}
{"id": "2606.11450", "title": "Exploring Adaptive Masked Reconstruction for Self-Supervised Skeleton-Based Action Recognition", "abstract": "Recently, masked skeleton reconstruction models have emerged as strong action representation learners, driving significant progress in self-supervised skeleton-based action recognition. However, existing state-of-the-art methods must predict an exceedingly large number of spatiotemporal patches, significantly prolonging training time. Besides, by treating all spatiotemporal regions equally during reconstruction, these models are distracted from learning the critical motion patterns that underlie action semantics. To address these challenges, we propose Adaptive Masked Reconstruction (AMR), a faster and stronger pre-training framework. We first decouple the decoder from the encoder, enabling flexible prediction of larger spatiotemporal patches and dramatically reducing reconstruction complexity. Given that larger patches contain more complex information, which is challenging to predict and consequently degrades performance, we accordingly introduce an adaptive guidance module. This module identifies regions of high motion informativeness, guiding the model to focus on the most discriminative parts of each patch and alleviating reconstruction difficulty. Experiments on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD datasets demonstrate that AMR not only accelerates pre-training substantially but also improves downstream recognition accuracy, surpassing current state-of-the-art approaches.", "date": "2026-06-09", "categories": []}
{"id": "2606.11645", "title": "Motion Reinforces Appearance: RGB-Skeleton Gated Residual Fusion for Micro-Gesture Online Recognition", "abstract": "Micro-gesture analysis attracts increasing attention for inferring spontaneous emotion from subtle body movements. Micro-gesture online recognition, which localizes and classifies each gesture instance in untrimmed videos, is a core task in the 4th EI-MiGA-IJCAI Challenge. Compared with typical temporal action detection, MGR emphasizes the localization and classification of actions, requiring the model to output the start time, end time, and category of each micro-gesture. Moreover, since micro-gestures are highly spontaneous, relying solely on a single modality makes it difficult to capture the complete and accurate multi-modal cues. In this work, we propose DyFADet+, which extends DyFADet into a dual-stream RGB-skeleton framework. In our model, both modalities are projected into shared multi-scale temporal embeddings and fused through a gated residual module, which adaptively injects skeleton motion into the RGB representation rather than using naive concatenation. Finally, these fused features are decoded by a Dynamic TAD head for online classification and boundary regression. On the SMG dataset, our method achieves an F1 score of 40.88, ranking 2nd in the Micro-gesture Online Recognition track.", "date": "2026-06-10", "categories": []}
{"id": "2606.12033", "title": "SpikeTAD: Spiking Neural Networks for End-to-End Temporal Action Detection", "abstract": "Video understanding is a crucial part of computer vision, with numerous application scenarios. With the increasing popularity of mobile devices, an increasing number of efforts are trying to deploy video understanding models on them. However, existing video understanding models are difficult to deploy due to their large size and prohibitive power consumption. Spiking Neural Networks (SNNs) have shown bioplausibility and low power advantages over Artificial Neural Networks (ANNs), especially on neuromorphic chips which are regarded as essential components of future mobile devices. However, excessively long conversion time-steps and severe performance degradation problems limit their application. To solve the problems above, we explore the application of SNNs on temporal action detection (TAD), which is an important task in video understanding, and propose the first SNN-based end-to-end TAD architecture coined as SpikeTAD. While maintaining extremely low power consumption, SpikeTAD achieves an average mAP of 67.2% in THUMOS14 and 37.42% in ActivityNet-1.3, demonstrating the feasibility of a low-power TAD model. Our code is available at https://github.com/MCG-NJU/SpikeTAD.", "date": "2026-06-10", "categories": []}
{"id": "2606.12374", "title": "Semantically-Aware Diver Activity Recognition Framework for Effective Underwater Multi-Human-Robot Collaboration", "abstract": "Effective multi-human-robot collaboration is essential for expanding human-led operations in the challenging and high-risk underwater environment. For autonomous underwater vehicles (AUVs) to become true teammates, they must be able to comprehend their surroundings and recognize a diver's activities to offer assistance and ensure safety. Towards this goal, we introduce DAR-Net, a novel transformer-based framework that analyzes complex underwater scenes to classify diver activities. Our contribution lies in a semantically guided learning formulation that couples transformer-based temporal reasoning with pixel-level scene supervision. This multi-loss training strategy explicitly aligns global activity recognition with local human-robot interaction semantics, which is particularly critical in low-visibility underwater conditions. To address the significant challenge of data scarcity in this domain, we present the first-ever Underwater Diver Activity (UDA) dataset, a foundational resource containing over 2,600 annotated images with pixel-level masks. Through rigorous experimental evaluations in a controlled environment, we demonstrate that DAR-Net achieves promising accuracy in recognizing six distinct diver activities, outperforming state-of-the-art models. While this dataset provides a crucial baseline, our work serves as a pioneering step, laying the groundwork for future research and facilitating the development of more intelligent, collaborative underwater robotic systems.", "date": "2026-06-10", "categories": []}
{"id": "2606.10789", "title": "Closing the Modality Gap in Zero-Shot HAR: Contrastive Training and Separability-Optimized Prototypes on IMU Data", "abstract": "Zero-shot learning (ZSL) for inertial measurement unit (IMU)-based human activity recognition (HAR) faces a central challenge: bridging the gap between sensor embeddings and semantic class representations. We systematically evaluate seven configurations combining three inference methods with two training pipelines on the PAMAP2 dataset, using 14 seen and 4 unseen activity classes with subjects 108 and 109 held out for testing. We find that the modality gap is a training-time phenomenon governed by the encoder objective. A temporal convolutional network (TCN) trained with cross-entropy over label-name Sentence- BERT prototypes yields sensor embeddings with a mean cosine similarity of 0.30 to the corresponding text prototypes, while replacing the label-name prototype targets with discriminative activity descriptions raises this to 0.69. This alignment improvement transfers consistently across all three inference methods. The strongest result combines contrastive training with inverted softmax correction, achieving 73.2% accuracy and 0.583 macro F1 on unseen classes, compared to 58.3% accuracy and 0.34 macro F1 for the label-name baseline. A secondary finding is that richer text descriptions reduce inter-prototype separability in Sentence-BERT space, because shared biomechanical vocabulary causes the language model to compress the prototype cloud. This effect does not negate the benefits of contrastive alignment provided prototype descriptions retain sufficient discriminative vocabulary. We also demonstrate that overall accuracy is a misleading primary metric when test-set class distributions are imbalanced, and recommend macro-averaged F1 as the standard reporting metric for ZSL-HAR benchmarks.", "date": "2026-06-09", "categories": []}
{"id": "2603.04125", "title": "A Baseline Study and Benchmark for Few-Shot Open-Set Action Recognition with Feature Residual Discrimination", "abstract": "Few-Shot Action Recognition (FS-AR) has shown promising results but is often limited by a closed-set assumption that fails in real-world open-set scenarios. While Few-Shot Open-Set (FSOS) recognition is well-established for images, its extension to spatio-temporal video data remains underexplored. To address this, we propose an architectural extension based on a Feature-Residual Discriminator (FR-Disc), adapting previous work on skeletal data to the more complex video domain. Extensive experiments on five datasets demonstrate that while common open-set techniques provide only marginal gains, our FR-Disc significantly enhances unknown rejection capabilities without compromising closed-set accuracy, setting a new state-of-the-art for FSOS-AR. The project website, code, and benchmark are available at: https://hsp-iit.github.io/fsosar/.", "date": "2026-03-04", "categories": []}
{"id": "2606.09261", "title": "Self-supervised Learning Matters: A Simple Ensemble Solution for Micro-Gesture Recognition", "abstract": "In this paper, we present XInsight Lab's solution to the micro-gesture classification track of the 4th MiGA Challenge at IJCAI 2026, in which our solution ranked first and achieved a new state-of-the-art result. We propose a multimodal ensemble framework that integrates a self-supervised RGB-based model with supervised multi-stream models from previous solutions. The self-supervised RGB model is pretrained on 120K unlabeled clips via masked video modeling and then fine-tuned on iMiGUE. This simple yet effective RGB baseline achieves 69.224% top-1 accuracy on the iMiGUE test set, demonstrating the benefit of learning transferable representations from unlabeled in-domain videos. By incorporating this model as a complementary branch, the final ensemble reaches 74.419% top-1 accuracy, surpassing the previous state of the art by 1.206 percentage points. Experimental results on iMiGUE, including ablation studies on the ensemble strategy, validate the effectiveness of self-supervised RGB representation learning for micro-gesture recognition.", "date": "2026-06-08", "categories": []}
{"id": "2606.09679", "title": "SoccerNet 2026 Player-Centric Ball-Action Spotting:Retraining and Post-Processing Extensions to the FOOTPASS Baselines", "abstract": "We describe our system for the SoccerNet 2026 Player-Centric Ball-Action Spotting Challenge, which requires predicting who performs which action and when, across eight classes in broadcast soccer. Building on the three FOOTPASS baselines [1] (TAAD, TAAD+GNN, and TAAD+DST), we contribute four extensions: (1) gradient check pointing to enable full-backbone fine-tuning on a single GPU; (2) fusion of GNN logits into the DST encoder, combining graph-based tactical context with per-player visual features; (3) square-root frequency class weighting to address the 213:1 pass-to-tackle imbalance in the training data; and (4) a post processing pipeline comprising per-class logit gating, temporal frame refinement, jersey re-assignment, and a two-model ensemble. Our system achieves 0.548 Macro F1 on the test set and 0.446 on the challenge set (server evaluation).", "date": "2026-06-08", "categories": []}
{"id": "2606.07355", "title": "Spatial-Temporal Decoupled Adapter for Micro-gesture Online Recognition", "abstract": "Micro-gesture online recognition aims to temporally localize and classify subtle gestures in untrimmed videos. Owing to their extremely short duration, low motion amplitude, and ambiguous visual cues, capturing discriminative spatiotemporal representations remains highly challenging. Existing parameter-efficient adapters typically employ a single branch to model spatial and temporal cues jointly, which may fail to capture the fine-grained patterns of micro-gestures. To address this limitation, we propose a Spatial-Temporal Decoupled Adapter that decomposes video adaptation into independent temporal and spatial branches via lightweight depthwise convolutions. In addition, to address the long-tail distribution problem in the benchmark dataset, we introduce Adaptive Soft Balanced Augmentation, which dynamically allocates augmentation intensity based on class rarity and learning difficulty, without manual thresholds. Our method achieves an F1 score of 0.43808, ranking 1st in Track 2 of the 4th EI-MiGA-IJCAI Challenge.", "date": "2026-06-05", "categories": []}
{"id": "2605.24496", "title": "EgoAction: Egocentric Action Composition with Reliability-Aware Temporal Fusion for the EPIC-KITCHENS Action Detection Challenge at CVPR 2026", "abstract": "The EPIC-KITCHENS-100 Action Detection challenge evaluates whether a model can localize the start and end of each action in long untrimmed egocentric videos and assign the corresponding verb--noun action label. In this report, we formulate our submission as EgoAction (Egocentric Action Composition with Reliability-Aware Temporal Fusion), a unified decoupled detection and fusion pipeline. The pipeline uses EPIC-finetuned VideoMAE-L features, trains separate noun and verb temporal detectors with causal temporal modeling, composes action hypotheses from top noun--verb pairs, and introduces a confidence-adaptive boundary fusion rule at post-processing time. The key observation is that verb and noun streams often fail differently: verb scores are sensitive to motion transitions, whereas noun scores are sensitive to hand-object visibility and object clutter. A fixed arithmetic mean of their predicted boundaries can therefore amplify localization errors when one stream degenerates. We replace this hard-coded mean with Dynamic Weighted Fusion (DWF), which normalizes the maximum noun and verb classification confidences into proposal-wise boundary weights and linearly combines the two intervals. This lightweight tensor-only operator shifts boundary authority toward the more reliable stream while preserving the decoupled action scoring mechanism. Together with sliding-window inference, top-K noun--verb action composition, and class-wise Soft-NMS, EgoAction provides a compact and reproducible system for egocentric temporal action detection.", "date": "2026-05-23", "categories": []}
{"id": "2604.28173", "title": "Action Motifs: Self-Supervised Hierarchical Representation of Human Body Movements", "abstract": "Effective human behavior modeling requires a representation of the human body movement that capitalizes on its compositionality. We propose a hierarchical representation consisting of Action Atoms that capture the atomic joint movements and Action Motifs that are formed by their temporal compositions and encode similar body movements found across different overall human actions. We derive A4Mer, a nested latent Transformer to learn this hierarchical representation from human pose data in a fully self-supervised manner. A4Mer splits a 3D pose sequence into variable-length segments and represents each segment as a single latent token (Action Atoms). Through bottom-up representation learning, temporal patterns composed of these Action Atoms, which capture meaningful temporal spans of reusable, semantic segments of body movements, naturally emerge (Action Motifs). A4Mer achieves this with a unified pretext task of masked token prediction in their respective latent spaces. We also introduce Action Motif Dataset (AMD), a large-scale dataset of multi-view human behavior videos with full SMPL annotations. We introduce a novel use of cameras by mounting them on the feet to achieve their frame-wise annotations despite frequent and heavy body occlusions. Experimental results demonstrate the effectiveness of A4Mer for extracting meaningful Action Motifs, which significantly benefit human behavior modeling tasks including action recognition, motion prediction, and motion interpolation.", "date": "2026-04-30", "categories": []}
{"id": "2606.04019", "title": "Gravity-Aware Hierarchical Routing for Lightweight SensorLLM on Human Activity Recognition", "abstract": "Recent studies on sensor-language alignment have shown that two-stage frameworks can improve the semantic modeling ability of wearable-sensor human activity recognition (HAR), where SensorLLM-style methods first perform motion-to-language alignment and then fine-tune the model for downstream tasks. However, our experiments reveal a consistent failure mode when the Stage 2 backbone is compressed to a compact model such as TinyLlama: recognition of dynamic activities remains relatively strong, while the discrimination of low-motion static classes such as standing, sitting, and lying degrades substantially. To address this issue, we propose a gravity-aware hierarchical routing head as a lightweight post-alignment adaptation built on top of an already aligned model, rather than a new large-scale pretraining framework. The method uses the per-channel mean and std from the Chronos tokenizer state to extract statistical cues related to posture and gravity direction, and adaptively combines a static expert and a full expert through soft routing, together with a load-balancing loss for stable training. On the MHealth dataset, this design significantly improves macro-F1 with minimal parameter overhead, and the gains are concentrated mainly on static classes while preserving strong performance on dynamic activities. As a first arXiv disclosure, the current paper reports results on a single dataset only, with the goal of highlighting the core method and laying the groundwork for broader evaluation in future work.", "date": "2026-06-01", "categories": []}
{"id": "2606.04798", "title": "Uncertainty-Aware (Un)Supervised Few-Shot User Adaptation for On-Device Personalized Human Activity Recognition", "abstract": "Sensor-based Human Activity Recognition (HAR) models often degrade on unseen users due to domain shifts caused by individual movement patterns and sensor placement. Practical wearable HAR systems therefore require personalization methods that are lightweight, applicable whether calibration data is labeled, unlabeled, or unavailable, and robust under limited calibration. We present a gradient-free framework that repurposes pretrained HAR classifiers as Prototypical Networks using using prior prototypes, which preserve zero-shot performance and regularize adaptation. For labeled calibration, we introduce closed-form Bayesian prototype estimation and extend the same principle to unlabeled calibration. With only 3 seconds of calibration data per activity (one shot), supervised adaptation improves macro-F1 by +2.76 to +33.44 percentage points across four datasets, while unsupervised adaptation improves by +0.56 to +32.13 points. Since adaptation requires only closed-form prototype updates, the framework enables efficient and robust on-device personalization of preexisting HAR classifiers.", "date": "2026-06-03", "categories": []}
{"id": "2606.03160", "title": "SRENet: Spectral Re-Entry Network for Point Cloud Action Recognition", "abstract": "Recognizing human actions from point cloud sequences is critical for 3D perception driven applications such as autonomous driving and human-computer interaction. However, the irregular structure and temporal inconsistency of point clouds pose unique challenges for spatio-temporal representation learning, especially in capturing both global motion context and fine-grained temporal dynamics. We propose SRENet, a spectral-aware framework designed to explicitly learn both global context and fine-grained temporal dynamics of motion from a frequency perspective for action recognition. SRENet introduces a Spectral Decomposition Block (SDeBlock) that performs wavelet-based analysis along temporal and spatial axes, disentangling features into low- and high-frequency components with frequency-specific attention. To recover residual dynamics and re-align temporal frequency structures distorted during semantic fusion, a Spectral Re-entry Block (SReBlock) performs secondary temporal decomposition. Furthermore, a spectral-aware learning strategy is devised to enhance discriminability in both frequency subspaces via contrastive loss and a curriculum schedule that gradually shifts focus from low- to high-frequency spaces in line with coarse to detailed motion patterns. Extensive experiments on MSR-Action3D, NTU-RGBD and NTU-RGBD120 demonstrate that SRENet achieves state-of-the-art performance, validating the effectiveness of frequency modeling in point cloud-based action understanding.", "date": "2026-06-02", "categories": []}
{"id": "2606.03490", "title": "TrAction: Action Recognition with Sparse Trajectories", "abstract": "Modern action recognition models operate on memory- and compute-intensive dense RGB video volumes and frequently exploit appearance and background shortcuts, for example, predicting actions from objects or scenes instead of characteristic motion. We investigate an efficient alternative input modality that is largely free of such biases by construction: sparse point trajectories. To this end, we develop a simple transformer architecture for 2.5D trajectory-based recognition together with a masked-trajectory pretraining, which we show to substantially improve downstream action recognition accuracy. Despite using only a fraction of the dense RGB input, our method reaches 45% top-1 on Something-Something V2 and 54% on EPIC-Kitchens-100, and surpasses V-JEPA on time-reversal sensitivity. More importantly, we find trajectory features to be complementary to state-of-the-art appearance-based features. Fusing our pretrained model with DINOv2 and V-JEPA 2 improves top-1 accuracy on Something-Something V2 by 8.7 and 1.6 points, respectively. Code: https://github.com/ecker-lab/TrAction", "date": "2026-06-02", "categories": []}
{"id": "2606.03610", "title": "SkelHCC: A Hyperbolic CLIP-Driven Cache Adaptation Framework for Skeleton-based One-Shot Action Recognition", "abstract": "Skeleton-based action recognition aims to understand human behaviors from body joint sequences and is especially challenging in the one-shot setting, where only a single labeled exemplar is available for each novel action. A key challenge is learning representations that capture the hierarchical and compositional structure of human motion while aligning effectively with high-level action semantics under extreme data scarcity. Existing approaches, largely based on Euclidean embeddings and low-level motion cues, struggle to model the tree-like organization of skeleton data, limiting cross-modal alignment and generalization to unseen action categories. We propose SkelHCC, a unified skeleton hyperbolic CLIP-driven cache adaptation framework for one-shot skeleton-based action recognition. SkelHCC introduces an Explicitly Hierarchical Hyperbolic CLIP (EH-HCLIP) module that embeds skeleton sequences and action language into a shared hyperbolic space. By leveraging the negative curvature and exponential volume growth of hyperbolic geometry, EH-HCLIP naturally encodes the joint-part-body hierarchy of human anatomy and yields structurally consistent cross-modal representations. To support efficient one-shot adaptation, SkelHCC further integrates a training-free LLM-guided Multi-granularity Voting Cache (LMV-Cache) for context-aware inference. Experiments on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD demonstrate that SkelHCC consistently outperforms state-of-the-art methods.", "date": "2026-06-02", "categories": []}
{"id": "2604.09063", "title": "Frequency-Enhanced Diffusion Models: Curriculum-Guided Semantic Alignment for Zero-Shot Skeleton Action Recognition", "abstract": "Human action recognition is pivotal in computer vision, with applications ranging from surveillance to human-robot interaction. Despite the effectiveness of supervised skeleton-based methods, their reliance on exhaustive annotation limits generalization to novel actions. Zero-Shot Skeleton Action Recognition (ZSAR) emerges as a promising paradigm, yet it faces challenges due to the spectral bias of diffusion models, which oversmooth high-frequency dynamics. Here, we propose Frequency-Aware Diffusion for Skeleton-Text Matching (FDSM), integrating a Semantic-Guided Spectral Residual Module, a Timestep-Adaptive Spectral Loss, and Curriculum-based Semantic Abstraction to address these challenges. Our approach effectively recovers fine-grained motion details, achieving state-of-the-art performance on NTU RGB+D, PKU-MMD, and Kinetics-skeleton datasets. Code has been made available at https://github.com/yuzhi535/FDSM. Project homepage: https://yuzhi535.github.io/FDSM.github.io/", "date": "2026-04-10", "categories": []}
{"id": "2606.00662", "title": "TAP-JEPA: Frozen Future-Latent Probing and Two-Stage Score Fusion for EPIC-KITCHENS-100 Action Anticipation", "abstract": "This report presents TAP-JEPA, our runner-up submission to the EPIC-KITCHENS-100 (EK-100) Action Anticipation Challenge at EgoVis 2026. The task is to anticipate the next verb, noun, and verb-noun action from an egocentric clip that ends before the target action begins. Instead of fine-tuning a large video backbone, TAP-JEPA builds a compact anticipation model on frozen V-JEPA 2.1 features: a ViT-G/384 encoder extracts visible pre-action tokens, the pre-trained latent predictor estimates near-future tokens from the observed context, and both token groups are fused by attentive probes with task-specific queries for verbs, nouns, and action pairs. For the final submission, we expand supervised training with the official training split and most of the validation split, reserving a small subset for sanity checks and qualitative inspection, and adopt a two-stage score fusion that first averages eight independently initialized probe replicas within each epoch and then merges candidates from epochs 12-20 with field-dependent weights. On the official open-testing leaderboard, our sunshinesky entry achieves 27.91 percent overall action Mean Top-5 Recall (MT5R), ranking second and only 0.04 percentage points behind the top score.", "date": "2026-05-30", "categories": []}
{"id": "2606.01834", "title": "Physics-Guided Attention in a Lightweight TCN for Efficient WiFi CSI-Based Human Activity Recognition", "abstract": "Human Action Recognition (HAR) using WiFi Channel State Information (CSI) has gained increasing attention due to its non-contact, low-cost, and privacy-preserving nature. However, existing learning-based approaches largely rely on deep, computationally intensive architectures to implicitly capture motion dynamics from CSI measurements, thereby increasing model complexity and reducing efficiency. Instead, we argue that incorporating appropriate inductive biases tailored to the physical characteristics of CSI signals enables more efficient and effective learning. In this work, we propose a compact temporal convolutional network (TCN)-based framework that explicitly incorporates motion-aware inductive biases into feature learning. Specifically, we introduce a Doppler-energy-guided temporal attention mechanism in feature space to emphasize motion-salient time segments, and a variance-driven channel attention module to weight informative subcarriers based on temporal motion statistics adaptively. By integrating these domain-specific priors, the proposed model effectively captures motion dynamics without increasing architectural depth. Extensive experiments on multiple benchmark datasets demonstrate that our approach achieves superior performance compared to deeper baselines, while significantly reducing parameter count and computational cost.", "date": "2026-06-01", "categories": []}
{"id": "2606.02246", "title": "Ego-METAS: Egocentric online Multimodal Energy-efficient Temporal Action Segmentation benchmark", "abstract": "To operate in the physical world, embodied agents must perceive their environment in an \"always-on\" fashion, selectively accessing the most informative sensors to balance energy constraints and task accuracy. Despite its importance for resource-constrained devices, energy-aware perception remains under-explored, with most prior work assuming unlimited compute. To address this, we introduce Ego-METAS: the first Egocentric online Multimodal Energy-efficient Temporal Action Segmentation benchmark. Ego-METAS provides a unified testbed of more than 100 hours of untrimmed egocentric video from EgoExo4D, CMU-MMAC, and CaptainCook4D, spanning 5 modalities (RGB, audio, gaze, IMU, and monochrome camera). We formulate an online temporal action segmentation task where models must dynamically select which sensors to activate at each timestep while strictly adhering to hardware-representative energy budgets. Alongside the benchmark, we release unified splits, cleaned annotations, pre-extracted features, and a diverse suite of baseline routing policies. Our evaluations show that optimal routing is highly scenario-dependent, and that existing policy-learning methods, designed primarily for trimmed clips, struggle to adapt to continuous, untrimmed environments. However, even simple dynamic fusion of complementary modalities (e.g., via random routing) proves critical for balancing predictive accuracy against strict energy budgets. Ultimately, Ego-METAS provides a standardized foundation to develop robust, cost-aware policies for autonomous, always-on embodied AI.", "date": "2026-05-29", "categories": []}
{"id": "2605.30689", "title": "ConTrans: Learning Text-enhanced Local-global Temporal Representations for Zero-shot Temporal Action Localization", "abstract": "Zero-shot Temporal Action Localization (ZS-TAL) aims to detect and locate previously unseen actions in untrimmed videos. However, existing approaches primarily focus on modeling long-range contextual information, often neglecting the critical relative-offset-based local correlations between video frames. Furthermore, their performance is hindered by limited feature representation capabilities due to the shallow nature of their network architectures. In this paper, we address these limitations by introducing a novel local-global multi-scale feature representation module. We propose a novel multi-scale encoder architecture, termed ConTrans, that integrates convolutional (Conv) inductive biases with transformer Self-attention to jointly capture fine-grained local dependencies and long-range global context, leading to more comprehensive feature representations than existing methods. Experimental evaluations on the ActivityNet-1.3 and THUMOS14 datasets demonstrate that ConTrans significantly outperforms existing methods, establishing a new benchmark for ZS-TAL.", "date": "2026-05-29", "categories": []}
{"id": "2605.31115", "title": "Polyphony: Diffusion-based Dual-Hand Action Segmentation with Alternating Vision Transformer and Semantic Conditioning", "abstract": "Dual-hand action segmentation, densely predicting actions for both hands from untrimmed videos, is essential for understanding complex bimanual activities. However, it poses several unique challenges: complex inter-hand dependencies, visual asymmetry between hands, representation conflicts where the dominant hand monopolizes gradients, and semantic ambiguity in fine-grained actions. We propose Polyphony, a three-stage method to address these challenges through: (1) an Alternating Dual-Hand Vision Transformer that alternates training between left- and right-hand mini-batches to ensure balanced gradient contributions from both hands while sharing a spatio-temporal encoder; (2) Semantic Feature Conditioning that aligns visual features with structured, compositional action descriptions to enhance discrimination of semantically similar actions; and (3) Diffusion-Based Segmentation with cross-hand feature fusion for inter-hand coordination and adaptive loss weighting for balancing performance. Polyphony achieves state-of-the-art on both dual-hand datasets (HA-ViD, ATTACH) with improvements up to 16.8 points, and on the single-stream Breakfast dataset (82.5%), outperforming the prior best method that uses a 12x larger backbone. Notably, our unified model with a single shared backbone surpasses baselines requiring separate per-hand models. Code is at https://github.com/x-labs-xyz/Polyphony-Dual-hand-Action-Segmentation.", "date": "2026-05-29", "categories": []}
{"id": "2605.29858", "title": "Masked Diffusion Vision-Language Models for Temporal Action Localization", "abstract": "Temporal action localization (TAL) requires recognizing the target event and localizing its start and end times precisely in untrimmed videos. Recent vision-language formulations improve semantic reasoning and support language-conditioned outputs, but their autoregressive decoders still generate tokens from left to right, preventing later semantic evidence from revising earlier timestamp predictions.\n  We adapt masked diffusion vision-language models (MDVLMs) to TAL so that semantic tokens and boundary tokens remain editable throughout iterative denoising with bidirectional attention, allowing temporal boundaries and semantic content to be refined jointly.\n  Direct adaptation, however, creates two TAL-specific mismatches: standard masked diffusion training corrupts all positions uniformly at random, but the time tokens are more reliable when enough semantic context is available; and token-level cross-entropy does not reflect temporal IoU. To address these mismatches, we introduce a Planned Training Objective that uses boundary-aware masking and step-weighted reconstruction to rehearse the late recovery of time tokens, together with a Step-Level IoU Reward that provides overlap-aware supervision during denoising. A standard sequence-level cross-entropy term provides the base reconstruction signal.\n  Experiments on ActivityNet-RTL, ActivityNet-1.3, and THUMOS-14 show that MDVLM-TAL improves both temporal reasoning and boundary localization over autoregressive vision-language baselines, with especially strong gains under stricter temporal IoU criteria.", "date": "2026-05-28", "categories": []}
{"id": "2504.10079", "title": "Hierarchical Relation-augmented Representation Generalization for Few-shot Action Recognition", "abstract": "Few-shot action recognition (FSAR) aims to recognize novel action categories with few exemplars. Existing methods typically learn frame-level representations for each video by designing inter-frame temporal modeling strategies or inter-video interaction at the coarse video-level granularity. However, they treat each episode task in isolation and neglect fine-grained temporal relation modeling between videos, thus failing to capture shared fine-grained temporal patterns across videos and reuse temporal knowledge from historical tasks. In light of this, we propose HR2G-shot, a Hierarchical Relation-augmented Representation Generalization framework for FSAR, which unifies three types of relation modeling (inter-frame, inter-video, and inter-task) to learn task-specific temporal patterns from a holistic view. Going beyond conducting inter-frame temporal interactions, we further devise two components to respectively explore inter-video and inter-task relationships: i) Inter-video Semantic Correlation (ISC) performs cross-video frame-level interactions in a fine-grained manner, thereby capturing task-specific query features and enhancing both intra-class consistency and inter-class separability; ii) Inter-task Knowledge Transfer (IKT) retrieves and aggregates relevant temporal knowledge from the bank, which stores diverse temporal patterns from historical episode tasks. Extensive experiments on five benchmarks show that HR2G-shot outperforms current top-leading FSAR methods.", "date": "2025-04-14", "categories": []}
{"id": "2605.24302", "title": "Cross-Modal Action Recognition in Egocentric Video Using Mamba: Integrating RGB and Hand Skeleton Streams via CLS Token Fusion Strategies", "abstract": "Egocentric action recognition is a challenging task due to erratic camera motion, frequent hand occlusion, and the difficulty of maintaining consistent visual representations over time. In this work, we propose a cross-modal architecture that combines RGB video and temporal hand skeleton data within a unified Mamba-based framework, exploiting the linear time complexity of State Space Models (SSMs). Our architecture consists of three components: a VideoMamba module for visual feature extraction, a skeleton encoder built on a stack of Mamba blocks, and a fusion module that integrates both modalities into a single representation. A central contribution of this work is the design and evaluation of four Class (CLS) token mixing strategies for multimodal fusion: Naive, Average, Weighted and Context-based. These strategies differ in how the pretrained unimodal CLS tokens, which role is to act as information sinks concentrating learned representations, are leveraged to initialize the mixed CLS token used for final classification. We evaluate all strategies on the H2O dataset. Experimental results show that the Average strategy achieves the best performance, yielding gains of over 10% Top-1 accuracy in the Tiny configuration and 2% in the Small configuration over the VideoMamba baseline.", "date": "2026-05-23", "categories": []}
{"id": "2605.27464", "title": "Beyond Motion Primitives: Behavioral Activity Recognition from Head-Mounted IMU", "abstract": "AR smart glasses need continuous behavioral context to offer proactive assistance, yet their most practical always-on sensor, the head-mounted Inertial Measurement Unit (IMU), detects only motion primitives such as walking or standing. We push beyond motion primitives to behavioral-level recognition, defining five categories that balance AR application need with sensor observability. To this end, we construct a 160K-sample Ego4D dataset with a four-tier quality assurance framework spanning 8 activity scenarios, and propose HiT-HAR, a 703K-parameter hierarchical model that outperforms prior head-mounted IMU models on five-class action and eight-class scenario recognition. We further map the observability frontier of head-mounted IMU through per-class separability analysis, identifying which behavioral categories are reliably observable (Locomotion), which benefit from temporal context (Object Transfer, Task Operation), and where scenario-dependent signal overlap poses remaining challenges. Our results indicate that architectural choices exploiting temporal context and scenario structure outperform simply scaling model size. The code and dataset are publicly available at https://github.com/Harvard-AI-and-Robotics-Lab/HiT-HAR.", "date": "2026-05-26", "categories": []}
{"id": "2605.28387", "title": "CLANE: Continual Learning of Actions on Neuromorphic Hardware from Event Cameras", "abstract": "Recognizing and continuously learning novel human actions without forgetting prior classes is a requirement for emerging AR/VR and robotics applications. For these applications, both on-device processing and learning are essential for privacy and low-latency adaptation. Event cameras address the efficiency of visual sensing with sparse, asynchronous output that is naturally compatible with neuromorphic processing. Yet no prior system has deployed a continual on-device learning pipeline for event-based action recognition using neuromorphic hardware. We present CLANE, Continual Learning of Actions on Neuromorphic Hardware from Event Cameras, deployed end-to-end on Intel Loihi 2. CLANE combines a spiking 2D CNN for spatiotemporal feature extraction with CLP-SNN as its on-chip learning head, extended to action clips via a Temporal Aggregation Layer and a fixed-point Normalization Layer, both novel Loihi 2 modules. On THU E-ACT-50, a 50-class dataset captured under real-world conditions, CLANE achieves 70.4% accuracy in a continual learning task while delivering more than 100x energy reduction and 16x lower latency over a sequential CNN+GRU+CLP edge GPU baseline, validated through iso-algorithm cross-platform benchmarking across three evaluation levels.", "date": "2026-05-27", "categories": []}
{"id": "2409.09953", "title": "Uncertainty-Guided Appearance-Motion Association Network for Out-of-Distribution Action Detection", "abstract": "Out-of-distribution (OOD) detection targets to detect and reject test samples with semantic shifts, to prevent models trained on in-distribution (ID) dataset from producing unreliable predictions. Existing works only extract the appearance features on image datasets, and cannot handle dynamic multimedia scenarios with much motion information. Therefore, we target a more realistic and challenging OOD detection task: OOD action detection (ODAD). Given an untrimmed video, ODAD first classifies the ID actions and recognizes the OOD actions, and then localizes ID and OOD actions. To this end, in this paper, we propose a novel Uncertainty-Guided Appearance-Motion Association Network (UAAN), which explores both appearance features and motion contexts to reason spatial-temporal inter-object interaction for ODAD.Firstly, we design separate appearance and motion branches to extract corresponding appearance-oriented and motion-aspect object representations. In each branch, we construct a spatial-temporal graph to reason appearance-guided and motion-driven inter-object interaction. Then, we design an appearance-motion attention module to fuse the appearance and motion features for final action detection. Experimental results on two challenging datasets show that UAAN beats state-of-the-art methods by a significant margin, illustrating its effectiveness.", "date": "2024-09-16", "categories": []}
{"id": "2512.11941", "title": "DynaPURLS: Dynamic Refinement of Part-Aware Representations for Skeleton-Based Zero-Shot Action Recognition", "abstract": "Zero-shot skeleton-based action recognition (ZS-SAR) is fundamentally constrained by prevailing approaches that rely on aligning skeleton features with static, class-level semantics. This coarse-grained alignment fails to bridge the domain shift between seen and unseen classes, thereby impeding the effective transfer of fine-grained visual knowledge. To address these limitations, we introduce \\textbf{DynaPURLS}, a unified framework that establishes robust, multi-scale visual-semantic correspondences and dynamically refines them at inference time to enhance generalization. Our framework leverages a large language model to generate hierarchical textual descriptions that encompass both global movements and local body-part dynamics. Concurrently, an adaptive partitioning module produces fine-grained visual representations by semantically grouping skeleton joints. To fortify this fine-grained alignment against the train-test domain shift, DynaPURLS incorporates a dynamic refinement module. During inference, this module adapts textual features to the incoming visual stream via a lightweight learnable projection. This refinement process is stabilized by a confidence-aware, class-balanced memory bank, which mitigates error propagation from noisy pseudo-labels. Extensive experiments on three large-scale benchmark datasets, including NTU RGB+D 60/120 and PKU-MMD, demonstrate that DynaPURLS significantly outperforms prior art, setting new state-of-the-art records. The source code is made publicly available at https://github.com/Alchemist0754/DynaPURLS", "date": "2025-12-12", "categories": []}
{"id": "2605.25615", "title": "UAV-OVO: Out-of-Viewpoint Generalization in UAV Action Recognition", "abstract": "UAV action recognition faces a deployment shift that standard benchmarks often obscure: a model trained on UAV footage captured from low-depression viewpoints may be required to recognize the same action classes from high-depression viewpoints. While the action labels remain unchanged, this shift alters body visibility, motion projection, and scene context, encouraging models to rely on viewpoint-specific shortcuts. We introduce UAV-OVO, an Out-of-Viewpoint generalization benchmark for UAV action recognition. UAV-OVO derives view scores from uncalibrated videos, uses a view-isolation band to assign low-depression videos to the training and in-distribution test splits while reserving high-depression videos for out-of-distribution testing, and constructs ID/OOD test sets matched by class distribution so that performance differences reflect viewpoint shift rather than label imbalance. Across representative video recognizers, UAV-OVO reveals a substantial ID/OOD gap: models that fit the low-depression training distribution well often fail to transfer to held-out high-depression views, exposing viewpoint shortcuts hidden by aggregate accuracy. We further propose LATER, LoRA-Anchored Test-time Re-centering, which first adapts the recognizer with Low-Rank Adaptation (LoRA) and then uses the learned LoRA subspace as a semantic anchor for online feature re-centering. Specifically, LATER projects target-domain displacement onto the orthogonal complement of the LoRA subspace before re-centering features, reducing viewpoint-induced drift while preserving task-relevant semantics. Together, UAV-OVO and LATER provide a controlled testbed and a practical adaptation method for viewpoint-robust UAV video understanding.", "date": "2026-05-25", "categories": []}
{"id": "2605.23288", "title": "Spatio-Temporal Similarity Volume Aggregation for Open-Vocabulary Action Recognition", "abstract": "Recent Open-Vocabulary Action Recognition (OVAR) methods typically aggregate visual features into a global representation before computing text alignment, a process that obscures local patch information and fine-grained spatio-temporal cues. We propose Similarity Volume Aggregation (SimVA), a framework that constructs a dense 4D spatio-temporal similarity volume from patch-level visual-text similarities. SimVA constructs a spatio-temporal similarity volume over local video tokens and action classes, and employs class sampling to ensure similarity aggregation scalable to large vocabularies. The similarity volume is refined by spatial aggregation, which contextualizes local similarity patterns to improve intra-frame consistency. Motion-aware modulation further injects inter-frame variation cues, highlighting dynamically changing regions. Mamba-based temporal aggregation then models the evolution of class-conditioned similarity patterns across frames. By maintaining dense visual-text correspondence, SimVA effectively transfers CLIP to video action recognition, achieving competitive performance across zero-shot, few-shot, and base-to-novel benchmarks.", "date": "2026-05-22", "categories": []}
{"id": "2605.23355", "title": "Decoupling Spatio-Temporal Adapter for Fine-Grained Badminton Action Localization", "abstract": "Temporal Action Localization (TAL) has been extensively studied in generic video understanding, while fine-grained sports scenarios, such as professional badminton, remain underexplored due to their complex and subtle spatio-temporal dynamics. In this paper, we focus on fine-grained TAL in professional badminton videos and introduce a new benchmark dataset, Fine-Badminton, which consists of 31 matches with 29 fine-grained stroke categories, covering 2104 rallies and 27597 annotated actions. To effectively capture the intricate motion patterns in such scenarios, we propose a Decoupling Spatio-Temporal Adapter (DSTA), which enables efficient modeling of spatio-temporal features within a parameter-efficient framework. Specifically, DSTA decomposes motion representation into three parallel branches, capturing temporal dynamics as well as vertical and horizontal spatial variations. The design allows the model to better distinguish subtle differences among fine-grained actions. Extensive experiments on both the Fine-Badminton dataset and the ShuttleSet benchmark demonstrate that the proposed method achieves state-of-the-art performance while introducing only a marginal increase in computational and parameter cost. These results validate the effectiveness and efficiency of the proposed approach for fine-grained temporal action localization.", "date": "2026-05-22", "categories": []}
{"id": "2605.23409", "title": "Online Hand Gesture Recognition Using 3D Convolutional Neural Networks", "abstract": "In human computer interaction, real-time detection and classification of dynamic hand gestures is challenging as: 1) the system must run in a real-time video stream and there is no noticeable lag in response after performing a gesture; 2) there is a large difference in how people perform gestures, making recognition more difficult. In this paper, an online hand gesture recognition system is proposed, which is able to localize gestures in real-time video stream and recognize what these gestures are. To improve the robustness of the system, the sliding window approach is used to refine results from multiple windows. All of the models in my project are trained on Jester database, achieving 98+% accuracy for detector and 90+% accuracy for classifier. For the overall performance of the system, the best group can respond within three seconds and reach 37.5% Levenshtein accuracy on the homemade dataset. The project codes used in this work are publicly available.", "date": "2026-05-22", "categories": []}
{"id": "2602.23231", "title": "Skarimva: Skeleton-based Action Recognition is a Multi-view Application", "abstract": "Human action recognition plays an important role when developing intelligent interactions between humans and machines. While there is a lot of active research on improving the machine learning algorithms for skeleton-based action recognition, not much attention has been given to the quality of the input skeleton data itself. This work demonstrates that by making use of multiple camera views to triangulate more accurate 3D~skeletons, the performance of state-of-the-art action recognition models can be improved significantly. This suggests that the quality of the input data is currently a limiting factor for the performance of these models. Based on these results, it is argued that the cost-benefit ratio of using multiple cameras is very favorable in most practical use-cases, therefore future research in skeleton-based action recognition should consider multi-view applications as the standard setup.", "date": "2026-02-26", "categories": []}
{"id": "2605.19578", "title": "Lens Privacy Sealing: A New Benchmark and Method for Physical Privacy-Preserving Action Recognition", "abstract": "RGB camera-based surveillance systems enable human action recognition for public safety and healthcare, yet raise serious privacy concerns. Existing methods rely on post-capture algorithms, which fail to protect privacy during data acquisition. We propose Lens Privacy Sealing (LPS), a simple hardware solution that physically obscures camera lenses with adjustable laminating film, providing pre-sensor privacy protection at minimal cost. Unlike software methods or expensive engineered optics, LPS achieves strong privacy through stochastic multi-layer scattering that is physically irreversible. We introduce the P$^3$AR dataset for privacy-preserving action recognition, featuring both large-scale replay-captured (P$^3$AR-NTU, 114K videos) and real-world collected (P$^3$AR-PKU) subsets with privacy attribute annotations. To handle video degradation from LPS, we propose MSPNet, a single-stage framework incorporating Inter-Frame Noise Suppressor (IFNS) and Cross-Frame Semantic Aggregator (CFSA), enhanced by contrastive language-image pre-training for robust semantic extraction. Extensive experiments demonstrate that MSPNet with IFNS and CFSA nearly doubles action recognition accuracy compared to baseline methods while suppressing identity recognition to low levels. Comprehensive validation shows LPS achieves a superior privacy-utility trade-off compared to state-of-the-art hardware methods, resists reconstruction attacks including PSF inversion and data-driven recovery, and generalizes robustly across optical configurations and challenging environments. Code is available at https://github.com/wangzy01/MSPNet.", "date": "2026-05-19", "categories": []}
{"id": "2605.22086", "title": "GenHAR: Generalizing Cross-domain Human Activity Recognition for Last-mile Delivery", "abstract": "Human Activity Recognition (HAR) has shown remarkable effectiveness in various applications, such as smart healthcare and intelligent manufacturing. However, a major challenge faced by HAR is the distribution shift across different sensor data domains, which often leads to decreased performance when deployed for real-world applications. To address this issue, this paper introduces GenHAR, a novel framework designed to mitigate the domain gap by learning domain-invariant sensor representations. GenHAR aims to enhance the generalization capabilities of HAR on target domains purely with data from the source domain. The key novelty of GenHAR lies in two aspects. Firstly, GenHAR tokenizes sensor data and learns correlations among frequency sensor channel dimensions to improve the robustness of HAR models. Secondly, GenHAR improves the efficiency via selective masking and an efficient attention mechanism. We conduct a systematic analysis of GenHAR by comparing it with state-of-the-art HAR methods on real-world human activity datasets. Results show that GenHAR outperforms state-of-the-art methods by 9.97% in accuracy, and reduces Floating Point Operations by 6.4 times. Moreover, we deploy GenHAR at a leading logistics company in 4 cities, and have detected 2.15 billion real-time activities. We release our code at: https://github.com/Sensor-FoundationModel/GenHAR.", "date": "2026-05-21", "categories": []}
{"id": "2605.22201", "title": "Zero-Shot Temporal Action Localization Through Textual Guidance", "abstract": "Zero-shot temporal action localization (ZS-TAL) consists of classifying and localizing actions in untrimmed videos, where action classes are unseen at training time. Existing work uses Vision and Language Models (VLMs), taking advantage of their strong zero-shot transfer capabilities. Yet, these models face evident challenges with fine-grained action classification, making it difficult to directly use them to distinguish between the presence and absence of an action. Most current methods for ZS-TAL address these challenges by training models on large-scale video datasets, which require annotated data and often result in limited generalization performance. Recently, approaches discarding the use of labeled data have emerged as an alternative. Following this direction, we propose a novel approach, ``Textual Guidance for finer localization of actions in videos'' (TEGU), that compensates for the lack of supervision from training data by exploiting rich textual information derived from large language models and structured text extracted from captions. This additional linguistic context can improve fine-grained discrimination by providing richer cues about fine-grained action differences within videos. We validate the effectiveness of the proposed method by conducting experiments on the THUMOS14 and the ActivityNet-v1.3 datasets. Our results show that, by exploiting rich textual information for improved action localization, TEGU outperforms state-of-the-art ZS-TAL approaches that do not involve training", "date": "2026-05-21", "categories": []}
{"id": "2605.22695", "title": "Improving Viewpoint-Invariance and Temporal Consistency for Action Detection", "abstract": "Viewpoint change invariance and action temporal consistency are critical aspects for the effective deployment of human action detection of untrimmed videos. Existing appearance-based video detection methods often struggle with limited viewpoint diversity during training, while motion-based detection approaches frequently fail to model fine-grained temporal relationships across consecutive motion windows. This paper introduces a novel two-stage action detection approach designed to improve both view-invariance and global temporal coherence properties. In the first stage, we extract motion features from augmented virtual viewpoints, solely used at training. Then, the second stage introduces a new view-invariant, multi-scale temporal encoder based on selective state-space sequence modelling to aggregate information across viewpoints and time scales. Experiments on PKU-MMD and BABEL benchmarks demonstrate that this approach significantly outperforms state-of-the-art methods in all considered splits. Code and trained models are available at: https://icb-vision-ai.github.io/HydraView-TAD", "date": "2026-05-21", "categories": []}
{"id": "2605.22697", "title": "Cross-Domain Human Action Recognition from Multiview Motion and Textual Descriptions", "abstract": "Robustness to domain changes is a key capability for effective deployment of human action recognition systems in real-world scenarios, where action categories at inference can present important domain shifts or even unseen actions from training. In this context, improving the recognition capabilities of Zero-Shot Action Recognition models (ZSAR), without requiring strong annotation efforts, remains a central challenge. Most ZSAR approaches assume that actions are observed under geometric conditions similar to those seen during training. In practice, variations in human body orientation and camera viewpoint add a significant domain gap in ZSAR, substantially limiting generalization to novel action-motion combinations. In this context, this paper presents a novel orientation-aware action recognition approach with improved cross-domain capabilities. Our approach combines motion cues of multiple camera viewpoints and text descriptions of human actions in the training phase. We present a new orientation-aware motion encoding network to learn different motion features, and adapt a specific orientation-aware text prompt to match the corresponding features at inference. Extensive experiments demonstrate that the proposed method consistently improves ZSAR performance across different recognition benchmarks, outperforming recent state-of-the-art zero-shot approaches on NTU-RGB+D, BABEL, NW-UCLA, and on two surveillance datasets. In addition, the learned representations exhibit strong transfer learning capabilities, yielding competitive performance on both cross-domain and same-domain recognition of seen actions. Code and trained models are available at: https://icb-vision-ai.github.io/OrientationAware-HAR", "date": "2026-05-21", "categories": []}
{"id": "2605.20645", "title": "Seeing Through Fog: Towards Fog-Invariant Action Recognition", "abstract": "Foggy conditions are commonly encountered in real-world applications; however, existing action recognition approaches typically assume favorable weather and high-quality video inputs. On foggy days, unpredictable visibility degradation and reduced contrast obstruct the extraction of semantic cues, posing significant challenges for current action recognition methods. In this paper, we mitigate the issues faced in action recognition under foggy conditions by employing two strategies. First, we present FogAct, the first benchmark dataset for foggy action recognition, consisting of paired clean and foggy videos captured with a stereo camera system. The dataset spans 10 scenes and 55 action categories, comprising nearly 10,000 video clips. Second, we propose FogNet, a two-stream CLIP model that discovers fog-invariant semantic information hidden behind the degraded videos. FogNet learns robust representations of foggy videos with guidance from clean videos, effectively capturing shared structural and motion cues between clean and foggy videos. Extensive experiments on FogAct and three other popular datasets demonstrate that our method achieves competitive performance compared with state-of-the-art (SOTA) approaches. Our FogAct and FogNet are given in our project page.", "date": "2026-05-20", "categories": []}
{"id": "2605.20680", "title": "DarkShake-DVS: Event-based Human Action Recognition under Low-light andShaking Camera Conditions", "abstract": "Human Action Recognition (HAR) is a fundamental computer vision task with diverse real-world applications. Practical deployments often involve low-light environments and unconstrained 6-DoF camera motion, conditions that degrade visual quality, disrupt temporal coherence, and compromise reliability of existing methods. Event cameras, with high low-light sensitivity and microsecond-level temporal resolution, paired with an inertial measurement unit (IMU), present a promising solution. However, current research faces two key challenges: absence of a benchmark integrating low-light conditions, 6-DoF motion, and synchronized IMU data; and lack of effective motion compensation techniques. To address these, we propose Event-IMU Stabilized HAR (EIS-HAR), with two modules. The first is an EIS module that reduces motion blur via a non-linear warping function to reconstruct a motion-compensated input. The second is a HAR module with a four-stage hybrid architecture to efficiently extract spatiotemporal features for accurate action recognition. To alleviate data scarcity, we introduce DarkShake-DVS, the first large-scale event-based HAR benchmark that includes 18,041 realworld clips captured in low light and intense 6-DoF motion, supplemented by synchronized IMU data. Extensive experiments on three datasets demonstrate consistent superiority of EIS-HAR over state-of-the-art methods.", "date": "2026-05-20", "categories": []}
{"id": "2605.20904", "title": "JFAA: Technical Report for the EPIC-KITCHENS-100 Action Anticipation Challenge at EgoVis 2026", "abstract": "We propose JFAA, a JEPA-based Future Action Anticipation method for the EPIC-KITCHENS-100 (EK-100) Action Anticipation task. Inspired by the representation learning and future prediction ability of V-JEPA 2.1, JFAA uses a frozen encoder and predictor to extract observed context features and near-future latent tokens. A lightweight attentive probe is then trained to predict verb, noun, and action logits with separate task queries. To improve robustness, we further build a field-aware ensemble over selected epoch-level predictions, allowing each output field to benefit from its most reliable candidates. Experimental results on the official challenge server show that JFAA achieves first place in the EgoVis 2026 EK-100 Action Anticipation Challenge. Our code will be released at https://github.com/CorrineQiu/JFAA.", "date": "2026-05-20", "categories": []}
{"id": "2602.09872", "title": "BabyMamba-HAR: Lightweight Selective State Space Models for Efficient Human Activity Recognition on Resource Constrained Devices", "abstract": "Human activity recognition (HAR) on resource constrained devices requires high accuracy across diverse sensor setups. Selective state space models (SSMs) offer efficient linear time sequence processing, presenting a compelling alternative to attention mechanisms. However, their TinyML design space remains unexplored. This paper introduces BabyMamba-HAR, comprising two lightweight architectures: (1) CI-BabyMamba-HAR, utilizing a channel independent stem for noise robustness, and (2) Crossover-BiDir-BabyMamba-HAR, utilizing an early fusion stem for channel count independent complexity. Both integrate weight tied bidirectional scanning and gated temporal attention pooling. Across eight benchmarks, Crossover-BiDir-BabyMamba-HAR averages an 86.52% F1-score with 27K parameters and 2.21M MACs, matching TinyHAR (86.16%) while requiring 11x fewer MACs on high channel datasets. On-device deployment on the Raspberry Pi Pico 2 and ESP32 utilized a mixed precision C++ runtime (INT8 projections, float32 states). A fused computation strategy with lifetime aware memory management reduces peak memory footprint from O(B*dmodel*L*dstate) to O(B*dmodel*dstate), adapting to support weight-tied bidirectional and channel-streaming execution. Both architectures achieved full 8/8 dataset coverage with >99.2% PyTorch parity, whereas INT8 quantized TFLite baselines showed degraded coverage and parity (TinyHAR: 7/8 and 4/8 coverage at 60.4% and 88.6% parity, TinierHAR: 8/8 and 6/8 at 54.2% and 90.8%, DeepConvLSTM: 1/8 and 0/8 on Pico 2 and ESP32, respectively). Crossover-BiDir-BabyMamba-HAR averages 154.4 ms latency on ESP32 and 481.9 ms on Pico 2. Ablations confirm bidirectional scanning and gated attention improve F1-scores by up to 8.42% and 8.94%, respectively, establishing practical principles for TinyML SSM deployment.", "date": "2026-02-10", "categories": []}
{"id": "2501.13795", "title": "Training-Free Zero-Shot Temporal Action Detection with Vision-Language Models", "abstract": "Existing zero-shot temporal action detection (ZSTAD) methods predominantly use fully supervised or unsupervised strategies to recognize unseen activities. However, these training-based methods are prone to domain shifts and require high computational costs, which hinder their practical applicability in real-world scenarios. In this paper, unlike previous works, we propose a training-Free Zero-shot temporal Action Detection (FreeZAD) method, leveraging existing vision-language (ViL) models to directly classify and localize unseen activities within untrimmed videos without any additional fine-tuning or adaptation. We mitigate the need for explicit temporal modeling and reliance on pseudo-label quality by designing the LOGarithmic decay weighted Outer-Inner-Contrastive Score (LogOIC) and frequency-based Actionness Calibration. Furthermore, we introduce a test-time adaptation (TTA) strategy using Prototype-Centric Sampling (PCS) to expand FreeZAD, enabling ViL models to adapt more effectively for ZSTAD. Extensive experiments on the THUMOS14 and ActivityNet-1.3 datasets demonstrate that our training-free method outperforms state-of-the-art unsupervised methods while requiring only 1/13 of the runtime. When equipped with TTA, the enhanced method further narrows the gap with fully supervised methods.", "date": "2025-01-23", "categories": []}
{"id": "2602.06523", "title": "MicroBi-ConvLSTM: An Ultra-Lightweight Efficient Model for Human Activity Recognition on Resource Constrained Devices", "abstract": "Human Activity Recognition (HAR) on resource constrained wearables requires models that balance accuracy against strict memory and computational budgets. State of the art lightweight architectures such as TinierHAR (34K parameters) and TinyHAR (55K parameters) achieve strong accuracy, but exceed memory budgets of microcontrollers with limited SRAM once operating system overhead is considered.\n  We present MicroBi-ConvLSTM, an ultra-lightweight convolutional recurrent architecture achieving 11.4K parameters on average through two stage convolutional feature extraction with 4x temporal pooling, and a single bidirectional LSTM layer. This represents 2.9x parameter reduction versus TinierHAR and 11.9x versus DeepConvLSTM while preserving linear O(N) complexity.\n  Evaluation across eight diverse HAR benchmarks shows that MicroBi-ConvLSTM maintains competitive performance within the ultra-lightweight regime: 93.41% macro F1 on UCI-HAR, 94.46% on SKODA assembly gestures, and 88.98% on Daphnet gait freeze detection. Systematic ablation reveals task dependent component contributions where bidirectionality benefits episodic event detection, but provides marginal gains on periodic locomotion. On-device deployment on the Raspberry Pi Pico 2 and ESP32 validates hardware viability under both INT8 quantized and FP32 full-precision paths. Under INT8 quantization, MicroBi-ConvLSTM is the only architecture achieving full 8/8 dataset coverage on both platforms, with 72.8 ms average latency on Pico 2 and 97.9% PyTorch parity on ESP32. Under FP32 deployment, it achieves 100.0% parity on all successful configurations (8/8 Pico 2, 7/8 ESP32), confirming that all INT8 fidelity degradation is a quantization artifact rather than an architectural limitation.", "date": "2026-02-06", "categories": []}
{"id": "2504.05451", "title": "ViewBridge: Curriculum Knowledge Distillation for Activity View-Invariance Under Extreme Viewpoint Changes", "abstract": "Traditional methods for view-invariant learning rely on controlled multi-view training data with minimal scene clutter. However, they struggle with in-the-wild videos that exhibit extreme viewpoint differences and share little visual content. We introduce ViewBridge, a framework for learning rich video representations in the presence of severe view-occlusions. We introduce a knowledge distillation objective that preserves action-centric semantics, together with a novel curriculum learning procedure that pairs incrementally more challenging views over time, thereby allowing smooth adaptation to extreme viewpoint differences. To sort training video segments for the proposed curriculum, we define a geometry-based metric that reflects their likely occlusion level. While training leverages multi-view data, at inference time, the input is an uncalibrated, single-viewpoint video. Evaluating our approach on two tasks -- temporal keystep grounding and fine-grained keystep recognition -- we outperform SOTA approaches across three datasets (Ego-Exo4D, LEMMA, EPFL-Smart-Kitchen-30). Project page: https://vision.cs.utexas.edu/projects/learning_view_distill/ .", "date": "2025-04-07", "categories": []}
{"id": "2509.21261", "title": "Every Subtlety Counts: Fine-grained Person Independence Micro-Action Recognition via Distributionally Robust Optimization", "abstract": "Micro-action Recognition is vital for psychological assessment and human-computer interaction. However, existing methods often fail in real-world scenarios because inter-person variability causes the same action to manifest differently, hindering robust generalization. To address this, we propose the Person Independence Universal Micro-action Recognition Framework, which integrates Distributionally Robust Optimization principles to learn person-agnostic representations. Our framework contains two plug-and-play components operating at the feature and loss levels. At the feature level, the Temporal-Frequency Alignment Module normalizes person-specific motion characteristics with a dual-branch design: the temporal branch applies Wasserstein-regularized alignment to stabilize dynamic trajectories, while the frequency branch introduces variance-guided perturbations to enhance robustness against person-specific spectral differences. A consistency-driven fusion mechanism integrates both branches. At the loss level, the Group-Invariant Regularized Loss partitions samples into pseudo-groups to simulate unseen person-specific distributions. By up-weighting boundary cases and regularizing subgroup variance, it forces the model to generalize beyond easy or frequent samples, thus enhancing robustness to difficult variations. Experiments on the large-scale MA-52 dataset demonstrate that our framework outperforms existing methods in both accuracy and robustness, achieving stable generalization under fine-grained conditions.", "date": "2025-09-25", "categories": []}
{"id": "2605.13202", "title": "STAR: Semantic-Temporal Adaptive Representation Learning for Few-Shot Action Recognition", "abstract": "Few-shot action recognition (FSAR) requires models to generalize to novel action categories from only a handful of annotated samples. Despite progress with vision-language models, existing approaches still suffer from semantic-temporal misalignment, where static textual prompts fail to capture decisive visual cues that appear sparsely across sequences, and from inadequate modeling of multi-scale temporal dynamics, as short-term discriminative cues and long-range dependencies are often either oversmoothed or fragmented. To address these challenges, we propose Semantic Temporal Adaptive Representation Learning (STAR), a unified framework, consisting of a semantic-alignment component and a temporal-aware component, effectively bridging the semantic and temporal gaps and transferring the sequence modeling capability of Mamba into the FSAR. The semantic alignment module introduces a Temporal Semantic Attention (TSA) mechanism, which performs frame-level cross-modal alignment with textual cues, ensuring fine-grained semantic-temporal consistency. The temporal-aware module incorporates a Semantic Temporal Prototype Refiner (STPR) that integrates semantic-guided Mamba blocks with multi-frequency temporal sampling and bidirectional state-space refinement, yielding semantically aligned prototypes with enhanced discriminative fidelity and temporal consistency. Furthermore, temporally dependent class descriptors derived from large language models (LLMs) provide long-range semantic guidance. Extensive experiments on five FSAR benchmarks demonstrate the consistent superiority of STAR over state-of-the-art methods. For instance, STAR achieves up to 8.1% and 6.7% gains on the SSv2-Full and SSv2-Small datasets under the 1-shot setting, and 7.3% on HMDB51, validating its effectiveness under limited supervision. The code is available at https://github.com/HongliLiu1/STAR-main.", "date": "2026-05-13", "categories": []}
{"id": "2605.12019", "title": "Efficient and Adaptive Human Activity Recognition via LLM Backbones", "abstract": "Human Activity Recognition (HAR) is a core task in pervasive computing systems, where models must operate under strict computational constraints while remaining robust to heterogeneous and evolving deployment conditions. Recent advances based on Transformer architectures have significantly improved recognition performance, but typically rely on task-specific models trained from scratch, resulting in high training cost, large data requirements, and limited adaptability to domain shifts. In this paper, we propose a paradigm shift that reuses large pretrained language models (LLMs) as generic temporal backbones for sensor-based HAR, instead of designing domain-specific Transformers. To bridge the modality gap between inertial time series and language models, we introduce a structured convolutional projection that maps multivariate accelerometer and gyroscope signals into the latent space of the LLM. The pretrained backbone is kept frozen and adapted using parameter-efficient Low-Rank Adaptation (LoRA), drastically reducing the number of trainable parameters and the overall training cost. Through extensive experiments on standard HAR benchmarks, we show that this approach enables rapid convergence, strong data efficiency, and robust cross-dataset transfer, particularly in low-data and few-shot settings. At the same time, our results highlight the complementary roles of convolutional frontends and LLMs, where local invariances are handled at the signal level while long-range temporal dependencies are captured by the pretrained backbone. Overall, this work demonstrates that LLMs can serve as a practical, frugal, and scalable foundation for adaptive HAR systems, opening new directions for reusing foundation models beyond their original language domain.", "date": "2026-05-12", "categories": []}
{"id": "2512.08984", "title": "RAG-HAR: Retrieval Augmented Generation-based Human Activity Recognition", "abstract": "Human Activity Recognition (HAR) underpins applications in healthcare, rehabilitation, fitness tracking, and smart environments, yet existing deep learning approaches demand dataset-specific training, large labeled corpora, and significant computational resources.We introduce RAG-HAR, a training-free retrieval-augmented framework that leverages large language models (LLMs) for HAR. RAG-HAR computes lightweight statistical descriptors, retrieves semantically similar samples from a vector database, and uses this contextual evidence to make LLM-based activity identification. We further enhance RAG-HAR by first applying prompt optimization and introducing an LLM-based activity descriptor that generates context-enriched vector databases for delivering accurate and highly relevant contextual information. Along with these mechanisms, RAG-HAR achieves state-of-the-art performance across six diverse HAR benchmarks. Most importantly, RAG-HAR attains these improvements without requiring model training or fine-tuning, emphasizing its robustness and practical applicability. RAG-HAR moves beyond known behaviors, enabling the recognition and meaningful labelling of multiple unseen human activities.", "date": "2025-12-06", "categories": []}
{"id": "2605.02743", "title": "Triple Spectral Fusion for Sensor-based Human Activity Recognition", "abstract": "The field of sensor-based human activity recognition (HAR) mainly uses posture, motion and context data of Inertial Measurement Units (IMUs) to identify daily activities. Despite the advancements in learning-based methods, it is challenging to perform information fusion from the temporal perspective due to the complexities in fusing heterogeneous sensor data and establishing long-term context correlations. This paper proposes a novel triple spectral fusion framework tailored for HAR. First, we develop an adaptive complementary filtering technique for noise suppression and organize each IMU's sensors into posture and motion modality nodes. Given that IMU nodes form a dynamic heterogeneous graph, we then apply adaptive filtering within the graph Fourier domain to merge both homogeneous and heterogeneous node information. Furthermore, an adaptive wavelet frequency selection approach is implemented to suppress context redundancy and shorten the length of features. This approach enhances both timestamp-based graph aggregation and the correlation of long-term contexts. Our framework uses adaptive filtering in the Fourier, graph Fourier, and wavelet domains, enabling effective multi-sensor fusion and context correlation. Extensive experiments on ten benchmark datasets demonstrate the superior performance of our framework. Project page: https://github.com/crocodilegogogo/TSF-TPAMI2026.", "date": "2026-05-04", "categories": []}
{"id": "2605.04617", "title": "Temporal Structure Matters for Efficient Test-Time Adaptation in Wearable Human Activity Recognition", "abstract": "Wearable human activity recognition (WHAR) models often suffer from performance degradation under real-world cross-user distribution shifts. Test-time adaptation (TTA) mitigates this degradation by adapting models online using unlabeled test streams, yet existing methods largely inherit assumptions from vision tasks and underexploit the inherent inter-window temporal structure in WHAR streams. In this paper, we revisit such temporal structure as a feature-conditioned inference signal rather than merely an output-space smoothing prior. We derive the insight that temporal continuity and observation-induced feature deviations provide complementary cues for determining when to preserve or release temporal inertia and where to route prediction refinement during likely transitions. Building upon this insight, we propose SIGHT, a lightweight and backpropagation-free TTA framework for WHAR, enabling real-time edge deployment. SIGHT estimates predictive surprise by comparing the current feature with a prototype-based expected state, and then uses the resulting feature deviation to guide geometry-aware transition routing based on prototype alignment and stream-level marginal habit tracking. Evaluations on real-world datasets confirm that SIGHT outperforms existing TTA baselines while reducing computational and memory costs.", "date": "2026-05-06", "categories": []}
{"id": "2605.08117", "title": "Modular Retrieval-Augmented Generalization for Human Action Recognition", "abstract": "Inertial Measurement Unit (IMU)-based Human Activity Recognition (HAR) aims to interpret and classify user behaviors from temporal motion signals. Recently, deep learning frameworks have advanced this task by learning and extracting discriminative spatiotemporal representations, significantly improving recognition performance. However, IMU-based HAR still faces several critical challenges, particularly limited training samples and static knowledge utilization, both of which severely hinder its large-scale deployment. In this paper, we introduce MoRA, the first Retrieval-Augmented Module specifically designed for motion series. It can be flexibly integrated into any existing HAR model, enhancing recognition performance while maintaining inference efficiency. To address issues such as information redundancy in retrieval results and rigid fusion strategies, we propose an uncertainty-adaptive fusion unit within MoRA. This unit leverages previous physical knowledge from IMU signals to dynamically adjust the fusion strategy between original outputs and retrieved information, enabling more robust recognition. Extensive experiments on ten real-world datasets demonstrate that MoRA significantly improves the performance of existing IMU-based HAR models, consistently delivering stable and effective gains. The source code of MoRA is available at: https://github.com/liavonpenn/mora.", "date": "2026-04-28", "categories": []}
{"id": "2605.09976", "title": "OZ-TAL: Online Zero-Shot Temporal Action Localization", "abstract": "Online Temporal Action Localization (On-TAL) aims to detect the occurrence time and category of actions in untrimmed streaming videos immediately upon their completion. Recent advancements in this field focus on developing more sophisticated frameworks, shifting from Online Action Detection (OAD)-based aggregation paradigm to instance-level understanding. However, existing approaches are typically trained on specific domains and often exhibit limited generalization capabilities when applied to arbitrary videos, particularly in the presence of previously unseen actions. In this paper, we introduce a new task called Online Zero-shot Temporal Action Localization (OZ-TAL), which aims to detect previously unseen actions in an online fashion. Furthermore, we propose a training-free framework that leverages off-the-shelf Vision-Language Models (VLMs) while introducing additional mechanisms to enhance visual representations and mitigate their inherent biases. We establish new benchmarks and representative baselines for OZ-TAL on THUMOS14 and ActivityNet-1.3, and extensive experiments demonstrate that our method substantially outperforms existing state-of-the-art approaches under both offline and online zero-shot settings.", "date": "2026-05-11", "categories": []}
{"id": "2605.10149", "title": "Improving Temporal Action Segmentation via Constraint-Aware Decoding", "abstract": "Temporal action segmentation (TAS) divides untrimmed videos into labeled action segments. While fully supervised methods have advanced the field, challenges such as action variability, ambiguous boundaries, and high annotation costs remain, especially in new or low-resource domains. Grammar-based approaches improve segmentation with structural priors but rely on complex parsing limiting scalability. In this work, we propose a lightweight, constraint-based refinement framework that enhances TAS predictions by integrating statistical structural priors such as transition confidence, action boundary sets, and per-class duration, that can be directly extracted from annotated data. These constraints are integrated into a modified Viterbi decoding algorithm, allowing inference-time refinement without retraining or added model complexity. Our approach improves both fully and semi-supervised TAS models by correcting structural prediction errors while maintaining high efficiency. Code is available at https://github.com/LUNAProject22/CAD", "date": "2026-05-11", "categories": []}
{"id": "2605.10732", "title": "iPay: Integrated Payment Action Recognition via Multimodal Networks and Adaptive Spatial Prior Learning", "abstract": "Automated transit payment analysis is vital for scalable fare auditing and passenger analytics, yet practice still relies on limited manual inspection. Prior vision- and skeleton-based methods remain brittle under noisy onboard surveillance and often depend on poorly generalizable handcrafted features. Building on the success of graph convolutional networks in human action recognition, we observe that skeleton features excel at modeling global spatiotemporal dependencies but tend to underemphasize the subtle local relative motions that distinguish payment actions. In contrast, RGB features preserve fine-grained spatial details yet often lack reliable temporal continuity in surveillance footage. To bridge both system-level deployment needs and model-level design challenges, we present iPay, an integrated payment action recognition framework for onboard transit surveillance system. iPay adopts a multimodal mixture-of-experts architecture with four tightly coupled streams: (1) an RGB expert stream emphasizing local evidence via region-focused computation; (2) a skeleton expert stream modeling articulated motion with a graph convolutional backbone; (3) a dual-attention fusion stream enabling skeleton-to-RGB temporal transfer and RGB-to-skeleton spatial enhancement; and (4) a prior-driven Spatial Difference Discriminator (SDD) that explicitly models hand-to-anchor relative motion to improve task-specific discriminability. We also collaborate with local transit agencies to collect over 55 hours of real onboard surveillance footage, yielding 500+ payment clips. Experiments show that iPay outperforms prior methods and achieves 83.45\\% recognition accuracy with competitive computational efficiency, making it suitable for edge deployment. Code is available at https://github.com/ccoopq/iPay.", "date": "2026-05-11", "categories": []}
{"id": "2605.06809", "title": "LookWhen? Fast Video Recognition by Learning When, Where, and What to Compute", "abstract": "Transformers dominate video recognition. They split videos into tokens, and processing them has expensive superlinear computational cost. Yet videos are filled with redundancy, so we can question the need for this expense. We introduce LookWhen, a selector-extractor framework that factorizes video recognition into learning when, where, and what to compute. Our shallow selector gets a scaled-down video and quickly scores all tokens across space-time, while our deep extractor gets the top-K selected tokens to approximate full-video representations without actually processing all the tokens. A key challenge is defining effective supervision for selection and extraction. For selection pre-training, we introduce a score on representations that ranks tokens by uniqueness using a simple nearest-neighbor distance. For extraction pre-training, we distill both a video teacher and an image teacher, for which we normalize its frame-wise representations to learn what changes within videos. Through these strategies, our selector-extractor learns general and efficient representations for feature extraction or fine-tuning to a task. Through experiments on Kinetics-400, SSv2, Epic-Kitchens, Diving48, Jester, and Charades, we show that LookWhen achieves a better accuracy-computation trade-off than efficient models and upgraded baselines of similar size. LookWhen Pareto-dominates in accuracy-FLOPs on 9 of 12 cases (6 tasks x 2 settings) and roughly matches on 3. In accuracy-throughput, measuring time in practice, LookWhen is more efficient still at 6.7x faster than InternVideo2-B at equal accuracy.", "date": "2026-05-07", "categories": []}
{"id": "2605.07140", "title": "Neurosymbolic Framework for Concept-Driven Logical Reasoning in Skeleton-Based Human Action Recognition", "abstract": "Skeleton-based human activity recognition has achieved strong empirical performance, yet most existing models remain black boxes and difficult to interpret. In this work, we introduce a neurosymbolic formulation of skeleton-based HAR that reframes action recognition as concept-driven first-order logical reasoning over motion primitives. Our framework bridges representation learning and symbolic inference by grounding first-order logic predicates in learnable spatial and temporal motion concepts. Specifically, we employ a standard spatio-temporal skeleton encoder to extract latent motion representations, which are then mapped to interpretable concept predicates via a spatio-temporal concept decoder that explicitly separates pose-centric and dynamics-centric abstractions. These concept predicates are composed through differentiable first-order logic layers, enabling the model to learn human-readable logical rules that govern action semantics. To impose semantic structure on the learned concepts, we align skeleton representations with LLM-derived descriptions of atomic motion primitives, establishing a shared conceptual space for perception and reasoning. Extensive experiments on NTU RGB+D 60/120 and NW-UCLA demonstrate that our approach achieves competitive recognition performance while providing explicit, interpretable explanations grounded in logical structure. Our results highlight neurosymbolic reasoning as an effective paradigm for interpretable spatio-temporal action understanding. Code: https://github.com/Mr-TalhaIlyas/REASON", "date": "2026-05-08", "categories": []}
{"id": "2604.26227", "title": "HOI-aware Adaptive Network for Weakly-supervised Action Segmentation", "abstract": "In this paper, we propose an HOI-aware adaptive network named AdaAct for weakly-supervised action segmentation. Most existing methods learn a fixed network to predict the action of each frame with the neighboring frames. However, this would result in ambiguity when estimating similar actions, such as pouring juice and pouring coffee. To address this, we aim to exploit temporally global but spatially local human-object interactions (HOI) as video-level prior knowledge for action segmentation. The long-term HOI sequence provides crucial contextual information to distinguish ambiguous actions, where our network dynamically adapts to the given HOI sequence at test time. More specifically, we first design a video HOI encoder that extracts, selects, and integrates the most representative HOI throughout the video. Then, we propose a two-branch HyperNetwork to learn an adaptive temporal encoder, which automatically adjusts the parameters based on the HOI information of various videos on the fly. Extensive experiments on two widely-used datasets including Breakfast and 50Salads demonstrate the effectiveness of our method under different evaluation metrics.", "date": "2026-04-29", "categories": []}
{"id": "2605.05692", "title": "CFE-PPAR: Compression-friendly encryption for privacy-preserving action recognition leveraging video transformers", "abstract": "Privacy-preserving action recognition (PPAR) enables machines to understand human activities in videos without revealing sensitive visual content. Among the various strategies for PPAR, encryption-based methods achieve strong privacy protection while maintaining high recognition performance. However, these methods lead to a catastrophic decrease in recognition performance and visual quality when the encrypted videos are compressed. That is, the previous methods are not compression-friendly. To address these issues, in this paper, we propose the first compression-friendly encryption method for PPAR, called CFE-PPAR. In CFE-PPAR, videos encrypted with secret keys can be directly recognized by a video transformer, which uses parameters transformed by the same keys as those used for video encryption. In experiments, it is verified that CFE-PPAR outperforms previous methods on the UCF101 and HMDB51 datasets under Motion-JPEG and H.264 compression.", "date": "2026-05-07", "categories": []}
{"id": "2605.02558", "title": "TemPose-TF-ASF: Two-Stage Bidirectional Stroke Context Fusion for Badminton Stroke Classification", "abstract": "Accurate badminton stroke prediction is crucial for fine-grained sports analysis and tactical decision support. However, existing methods struggle to model rich temporal context. This paper introduces TemPose-TF-ASF (Adjacent-Stroke Fusion), a context-aware extension of TemPose. It enhances stroke recognition by incorporating stroke-type information from both preceding and subsequent strokes. A two-stage training and inference strategy is adopted. Preliminary predictions from the baseline model are reused as estimated temporal context. These predictions guide the joint optimization of the ASF module and the classifier. By explicitly modeling bidirectional temporal stroke dependencies, the proposed method can be seamlessly integrated into existing state-of-the-art models. Experiments on a large-scale badminton match dataset show consistent improvements over the baseline and its variants in terms of Accuracy and Macro-F1. Moreover, integrating ASF into other advanced methods yields notable performance gains. These results demonstrate strong transferability and generalization capability.", "date": "2026-05-04", "categories": []}
{"id": "2512.16504", "title": "Skeleton-Snippet Contrastive Learning with Multiscale Feature Fusion for Action Localization", "abstract": "The self-supervised pretraining paradigm has achieved great success in learning 3D action representations for skeleton-based action recognition using contrastive learning. However, learning effective representations for skeleton-based temporal action localization remains challenging and underexplored. Unlike video-level {action} recognition, detecting action boundaries requires temporally sensitive features that capture subtle differences between adjacent frames where labels change. To this end, we formulate a snippet discrimination pretext task for self-supervised pretraining, which densely projects skeleton sequences into non-overlapping segments and promotes features that distinguish them across videos via contrastive learning. Additionally, we build on strong backbones of skeleton-based action recognition models by fusing intermediate features with a U-shaped module to enhance feature resolution for frame-level localization. Our approach consistently improves existing skeleton-based contrastive learning methods for action localization on BABEL across diverse subsets and evaluation protocols. We also achieve state-of-the-art transfer learning performance on PKUMMD with pretraining on NTU RGB+D and BABEL.", "date": "2025-12-18", "categories": []}
{"id": "2605.02834", "title": "VideoNet: A Large-Scale Dataset for Domain-Specific Action Recognition", "abstract": "Videos are unique in their ability to capture actions which transcend multiple frames. Accordingly, for many years action recognition was the quintessential task for video understanding. Unfortunately, due to a lack of sufficiently diverse and challenging data, modern vision-language models (VLMs) are no longer evaluated on their action recognition capabilities. To revitalize action recognition in the era of VLMs, we advocate for a returned focus on domain-specific actions. To this end, we introduce VideoNet, a domain-specific action recognition benchmark covering 1,000 distinct actions from 37 domains. We begin with a multiple-choice evaluation setting, where the difference between closed and open models is stark: Gemini 3.1 Pro attains 69.9% accuracy while Qwen3-VL-8B gets a mere 45.0%. To understand why VLMs struggle on VideoNet, we relax the questions into a binary setting, where random chance is 50%. Still, Qwen achieves only 59.2% accuracy. Further relaxing the evaluation setup, we provide $k\\in\\{1,2,3\\}$ in-context examples of the action. Some models excel in the few-shot setting, while others falter; Qwen improves $+7.0\\%$, while Gemini declines $-4.8\\%$. Notably, these gains fall short of the $+13.6\\%$ improvement in non-expert humans when given few-shot examples. Finding that VLMs struggle to fully exploit in-context examples, we shift from test-time improvements to the training side. We collect the first large-scale training dataset for domain-specific actions, totaling nearly 500k video question-answer pairs. Fine-tuning a Molmo2-4B model on our data, we surpass all open-weight 8B models on the VideoNet benchmark.", "date": "2026-05-04", "categories": []}
{"id": "2604.13279", "title": "Explainable Fall Detection for Elderly Monitoring via Temporally Stable SHAP in Skeleton-Based Human Activity Recognition", "abstract": "Reliable fall detection in elderly care requires monitoring systems that are not only accurate but also capable of producing stable, interpretable explanations of motion dynamics, a requirement that existing post hoc explainability methods rarely satisfy when applied to sequential biosignals. This study introduces a lightweight framework for skeleton-based fall detection that combines a Long Short-Term Memory (LSTM) model with a temporally stabilized attribution mechanism. We propose Temporal SHAP (T-SHAP), which treats frame-wise SHAP attributions as a temporal signal and applies a linear smoothing operator to reduce high-frequency variance. From a signal processing perspective, this operation is analogous to low-pass filtering, enabling the extraction of consistent temporal patterns while preserving the theoretical properties of Shapley-based attributions. Experiments conducted on the NTU RGB+D dataset demonstrate that the proposed approach achieves 94.3% classification accuracy with an end-to-end latency below 25 ms, supporting real-time applicability. Quantitative evaluation using perturbation-based faithfulness metrics shows that T-SHAP improves attribution reliability compared to standard SHAP (AUP: 0.91 vs. 0.89) and Grad-CAM (0.82), while also reducing temporal variance in the attribution signals. The resulting explanations highlight biomechanically relevant motion patterns, such as lower-limb instability and changes in trunk posture, which are consistent with known characteristics of fall events. The resulting framework is computationally lightweight, requires no additional model training, and produces explanations that are both temporally stable and biomechanically meaningful, properties directly relevant to the reliability demands of AI-assisted clinical monitoring.", "date": "2026-04-14", "categories": []}
{"id": "2605.01165", "title": "CEZSAR: A Contrastive Embedding Method for Zero-Shot Action Recognition", "abstract": "This paper proposes a novel Zero-Shot Action Recognition~(ZSAR) method based on contrastive learning. In ZSAR, we aim to classify examples from classes that were missing during training. Two well-known problems remain in ZSAR: the semantic gap and the domain shift. A semantic gap occurs because label representations come from the textual domain (i.e., language models) and must be associated with visual representations (i.e., CNNs, RNNs, transformer-based). This multimodal nature implies that the semantic properties of the two spaces are not identical. On the other hand, the domain shift arises from differences between the training and test sets and is inherent to ZSAR once the test set is unknown. One of the most promising methods to address both issues is learning joint embedding spaces. Therefore, we propose a new model that encodes videos and sentences in a joint embedding space, trained by aligning videos with their natural-language descriptions. We design an automatic negative sampling procedure to augment the training dataset and generate unpaired data, i.e., visual appearance and unrelated descriptions. Our results are state-of-the-art on the UCF-101 and Kinetics-400 datasets under several split configurations. Our code is available at https://github.com/valterlej/cezsar.", "date": "2026-05-01", "categories": []}
{"id": "2605.01668", "title": "IMPACT-Scribe: Interactive Temporal Action Segmentation with Boundary Scribbles and Query Planning", "abstract": "Dense temporal annotation of procedural activity videos is vital for action understanding and embodied intelligence but remains labor-intensive due to reactive tools. Each correction is treated as an isolated edit, limiting reuse of information on annotator uncertainty and model reliability. We introduce IMPACT-Scribe, a correction-driven framework for dense labeling that uses each correction to improve future human-machine collaboration. IMPACT-Scribe combines uncertainty-aware boundary scribble supervision, local proposal modeling, cost-aware query planning, structured propagation, and correction-driven adaptation. Experiments and a human study show that this closed-loop design improves labeling quality per effort, enhances boundary accuracy, and fosters better human-machine interaction over time. The code will be made publicly available at https://github.com/BanzQians/IMPACT_AS.", "date": "2026-05-03", "categories": []}
{"id": "2605.02596", "title": "HARMES: A Multi-Modal Dataset for Wearable Human Activity Recognition with Motion, Environmental Sensing and Sound", "abstract": "With each sensing modality exhibiting inherent strengths and limitations, multi-modal approaches for wearable Human Activity Recognition (HAR) are becoming increasingly relevant -- particularly for recognizing Activities of Daily Living (ADLs), where individual modalities often produce ambiguous signals for similar or complex activities. This work introduces HARMES, a multi-modal wearable dataset combining three wrist-recorded modalities: motion sensing via an Inertial Measurement Unit (IMU), atmospheric environmental sensors (humidity, temperature, and pressure), and audio. Collected from 20 participants performing household activities in their own homes, HARMES totals over 80 hours of recorded data, with approximately three hours of labeled activity data per participant across 15 ADL classes. To the best of our knowledge, HARMES is the first dataset to combine this particular sensor trio, and it is nearly six times larger than the previously largest wrist-inertial-acoustic HAR dataset. In an extensive benchmark, we evaluate cross-subject generalization and conduct an ablation study revealing that modality contributions are activity-dependent and can provide complementary value, particularly for activities that are ambiguous from motion data alone. HARMES is freely available at Zenodo, alongside example code for loading the dataset and training models on GitHub.", "date": "2026-05-04", "categories": []}
{"id": "2605.02659", "title": "Human Activity Recognition Method for Moderate Violence Detection", "abstract": "Physical violence in public spaces is a significant public health concern, with minor incidents such as pushing often serving as precursors to more severe escalations. This research develops an automated system for the real-time detection of moderate physical violence, specifically pushing, in surveillance camera footage. The proposed solution integrates state-of-the-art computer vision models, utilizing YOLO11 and YOLO11-Pose for human detection and skeletal keypoint extraction. By calculating body inclination and joint angles between shoulders and hips, a Random Forest classifier was trained to distinguish between normal behavior and aggressive physical contact. The system's performance was evaluated through three progressive case studies representing increasing levels of difficulty. In controlled environments with frontal views, the model achieved a precision of 0.98. In the most challenging scenario, featuring high-altitude, steep-angle recordings from real-world surveillance infrastructure, the system maintained a precision of 0.72 despite significant perspective distortion and visual noise. These results demonstrate the feasibility of using skeletal analysis for early violence intervention in urban security contexts.", "date": "2026-05-04", "categories": []}
{"id": "2605.00496", "title": "High-Speed Vision Improves Zero-Shot Semantic Understanding of Human Actions", "abstract": "Understanding human actions from visual observations is essential for human--robot interaction, particularly when semantic interpretation of unfamiliar or hard-to-annotate actions is required. In scenarios such as rapid and less common activities, collecting sufficient labeled data for supervised learning is challenging, making zero-shot approaches a practical alternative for semantic understanding without task-specific training. While recent advances in large-scale pretrained models enable such zero-shot reasoning, the impact of temporal resolution, especially for rapid and fine-grained motions, remains underexplored.\n  In this study, we investigate how temporal resolution affects zero-shot semantic understanding of high-speed human actions. Using kendo as a representative case of rapid and subtle motion patterns, we propose a training-free pipeline that combines a pre-trained video-language model for semantic representation with large language model-based reasoning for pairwise action comparison. Through controlled experiments across multiple frame rates (120 Hz, 60 Hz, and 30 Hz), we show that higher temporal resolution significantly improves semantic separability in zero-shot settings. We further analyze the role of tracking-based human joint information under both full and partial observation scenarios. Quantitative evaluation using a nearest-class prototype strategy demonstrates that high-speed video provides more stable and interpretable semantic representations for fast actions. These findings highlight the importance of temporal resolution in training-free action recognition and suggest that high-speed perception can enhance semantic understanding capabilities.", "date": "2026-05-01", "categories": []}
{"id": "2504.18662", "title": "M2R2: MultiModal Robotic Representation for Temporal Action Segmentation", "abstract": "Temporal action segmentation (TAS) has long been a key area of research in both robotics and computer vision. In robotics, algorithms have primarily focused on leveraging proprioceptive information to determine skill boundaries, with recent approaches in surgical robotics incorporating vision. In contrast, computer vision typically relies on exteroceptive sensors, such as cameras. Existing multimodal TAS models in robotics integrate feature fusion within the model, making it difficult to reuse learned features across different models. Meanwhile, pretrained vision-only feature extractors commonly used in computer vision struggle in scenarios with limited object visibility. In this work, we address these challenges by proposing M2R2, a multimodal feature extractor tailored for TAS, which combines information from both proprioceptive and exteroceptive sensors. We introduce a novel training strategy that enables the reuse of learned features across multiple TAS models. Our method sets a new state-of-the-art performance on three robotic datasets REASSEMBLE, (Im)PerfectPour, and JIGSAWS. Additionally, we conduct an extensive ablation study to evaluate the contribution of different modalities in robotic TAS tasks.", "date": "2025-04-25", "categories": []}
{"id": "2604.18274", "title": "LiquidTAD: Efficient Temporal Action Detection via Parallel Liquid-Inspired Temporal Relaxation", "abstract": "Temporal Action Detection (TAD) requires precise localization of action boundaries within long, untrimmed video sequences. While current high-performing methods achieve strong accuracy, they are often characterized by excessive parameter counts, substantial computational overhead, and a reliance on specialized operators that hinder deployment across diverse hardware platforms. This paper presents LiquidTAD, a framework that distills the exponential relaxation prior of liquid neural dynamics into a parallel temporal operator, rather than reproducing full Liquid Neural Network (LNN) dynamics. By introducing a Parallel Liquid-inspired Relaxation mechanism, sequential ODE solving is avoided through a fully vectorized, non-recursive formulation built entirely upon standard neural operations, enabling hardware-agnostic deployment with linear complexity with respect to the temporal length. A complementary Hierarchical Decay-Rate Sharing Strategy further adapts this relaxation prior across feature pyramid levels, stabilizing optimization and implicitly compensating for temporal compression in deeper layers. Experimental evaluations on THUMOS-14 and ActivityNet-1.3 demonstrate that LiquidTAD achieves accuracy competitive with strong baselines while substantially lowering the model footprint. Specifically, on THUMOS-14, LiquidTAD achieves 69.46\\% average mAP with only 10.82M parameters and 27.17G FLOPs, reducing the parameter count by over 60\\% compared with ActionFormer.", "date": "2026-04-20", "categories": []}
{"id": "2604.23281", "title": "Contrastive Learning for Multimodal Human Activity Recognition with Limited Labeled Data", "abstract": "Human activity recognition serves as the foundation for various emerging applications. In recent years, researchers have used collaborative sensing of multi-source sensors to capture complex and dynamic human activities. However, multimodal human activity sensing typically encounters highly heterogeneous data across modalities and label scarcity, resulting in an application gap between existing solutions and real-world needs. In this paper, we propose CLMM, a general contrastive learning framework for human activity recognition that achieves effective multimodal recognition with limited labeled data.\n  CLMM employs a novel two-stage training strategy. In the first stage, CLMM employs a CNN-DiffTransformer encoder to capture cross-modal shared information by extracting local and global features. Meanwhile, a hard-positive samples weighting algorithm enhances gradient propagation to reinforce shared learning. In the second stage, a dual-branch architecture combining quality-guided attention and bidirectional gated units captures modality-specific information, while a primary-auxiliary collaborative training strategy fuses both shared and modality-specific information. Experimental results on three public datasets demonstrate that CLMM significantly improves state-of-the-art baselines in both recognition accuracy and convergence performance.", "date": "2026-04-25", "categories": []}
{"id": "2604.23415", "title": "A Heterogeneous Two-Stream Framework for Video Action Recognition with Comparative Fusion Analysis", "abstract": "Most two-stream action recognition networks apply the same convolutional backbone to both RGB and optical flow streams, ignoring the fact that the two modalities have fundamentally different structural properties. Optical flow captures fine-grained motion patterns, while RGB frames carry rich appearance and scene context - treating them identically discards this distinction. We propose DualStreamHybrid, a heterogeneous two-stream architecture that assigns each stream a backbone suited to its input: a pretrained ViT-Tiny/16 for RGB frames, and a MobileNetV2 trained from scratch on a 20-channel stacked optical flow representation. A learned projection layer maps the two differently-sized feature vectors to a common dimensionality before fusion, enabling the two streams to interact without forcing architectural symmetry. We design five fusion strategies within a unified framework - late fusion, concatenation, cross-attention, weighted fusion, and gated fusion - and evaluate them on UCF11 (1,600 videos, 11 classes) and UCF50 (6,681 videos, 50 classes) to study how fusion behaviour scales with dataset size. On UCF11, cross-attention achieves 98.12% test accuracy, outperforming the RGB-only ViT-Tiny baseline of 95.94%, which suggests that explicit inter-modal attention is particularly effective on smaller, less complex datasets. On UCF50, weighted fusion reaches 96.86% and proves the most consistent strategy across both benchmarks. The learned stream weights reveal an interesting pattern: UCF11 sees near-equal modality contribution (RGB: 0.507, flow: 0.493), while UCF50 favours the RGB stream slightly more (RGB: 0.554, flow: 0.446) - arguably reflecting the larger and more visually diverse action space. Taken together, these results suggest that even a lightweight motion stream meaningfully complements a strong appearance encoder, and that the optimal fusion strategy depends on dataset scale.", "date": "2026-04-25", "categories": []}
{"id": "2604.22036", "title": "EgoMAGIC- An Egocentric Video Field Medicine Dataset for Training Perception Algorithms", "abstract": "This paper introduces EgoMAGIC (Medical Assistance, Guidance, Instruction, and Correction), an egocentric medical activity dataset collected as part of DARPA's Perceptually-enabled Task Guidance (PTG) program. This dataset comprises 3,355 videos of 50 medical tasks, with at least 50 labeled videos per task. The primary objective of the PTG program was to develop virtual assistants integrated into augmented reality headsets to assist users in performing complex tasks.\n  To encourage exploration and research using this dataset, the medical training data has been released along with an action detection challenge focused on eight medical tasks. The majority of the videos were recorded using a head-mounted stereo camera with integrated audio. From this dataset, 40 YOLO models were trained using 1.95 million labels to detect 124 medical objects, providing a robust starting point for developers working on medical AI applications.\n  In addition to introducing the dataset, this paper presents baseline results on action detection for the eight selected medical tasks across three models, with the best-performing method achieving average mAP 0.526. Although this paper primarily addresses action detection as the benchmark, the EgoMAGIC dataset is equally suitable for action recognition, object identification and detection, error detection, and other challenging computer vision tasks.\n  The dataset is accessible via zenodo.org (DOI: 10.5281/zenodo.19239154).", "date": "2026-04-23", "categories": []}
{"id": "2604.22595", "title": "EV-CLIP: Efficient Visual Prompt Adaptation for CLIP in Few-shot Action Recognition under Visual Challenges", "abstract": "CLIP has demonstrated strong generalization in visual domains through natural language supervision, even for video action recognition. However, most existing approaches that adapt CLIP for action recognition have primarily focused on temporal modeling, often overlooking spatial perception. In real-world scenarios, visual challenges such as low-light environments or egocentric viewpoints can severely impair spatial understanding, an essential precursor for effective temporal reasoning. To address this limitation, we propose Efficient Visual Prompting for CLIP (EV-CLIP), an efficient adaptation framework designed for few-shot video action recognition across diverse scenes and viewpoints. EV-CLIP introduces two visual prompts: mask prompts, which guide the model's attention to action-relevant regions by reweighting pixels, and context prompts, which perform lightweight temporal modeling by compressing frame-wise features into a compact representation. For a comprehensive evaluation, we curate five benchmark datasets and analyze domain shifts to quantify the influence of diverse visual and semantic factors on action recognition. Experimental results demonstrate that EV-CLIP outperforms existing parameter-efficient methods in overall performance. Moreover, its efficiency remains independent of the backbone scale, making it well-suited for deployment in real-world, resource-constrained scenarios. The code is available at https://github.com/AI-CV-Lab/EV-CLIP.", "date": "2026-04-24", "categories": []}
{"id": "2604.21011", "title": "Micro-DualNet: Dual-Path Spatio-Temporal Network for Micro-Action Recognition", "abstract": "Micro-actions are subtle, localized movements lasting 1-3 seconds such as scratching one's head or tapping fingers. Such subtle actions are essential for social communication, ubiquitously used in natural interactions, and thus critical for fine-grained video understanding, yet remain poorly understood by current computer vision systems. We identify a fundamental challenge: micro-actions exhibit diverse spatio-temporal characteristics where some are defined by spatial configurations while others manifest through temporal dynamics. Existing methods that commit to a single spatio-temporal decomposition cannot accommodate this diversity. We propose a dual-path network that processes anatomically-grounded spatial entities through parallel Spatial-Temporal (ST) and Temporal-Spatial (TS) pathways. The ST path captures spatial configurations before modeling temporal dynamics, while the TS path inverts this order to prioritize temporal dynamics. Rather than fixed fusion, we introduce entity-level adaptive routing where each body part learns its optimal processing preference, complemented by Mutual Action Consistency (MAC) loss that enforces cross-path coherence. Extensive experiments demonstrate competitive performance on MA-52 dataset and state-of-the-art results on iMiGUE dataset. Our work reveals that architectural adaptation to the inherent complexity of micro-actions is essential for advancing fine-grained video understanding.", "date": "2026-04-22", "categories": []}
{"id": "2604.21572", "title": "Deep kernel video approximation for unsupervised action segmentation", "abstract": "This work focuses on per-video unsupervised action segmentation, which is of interest to applications where storing large datasets is either not possible, or nor permitted. We propose to segment videos by learning in deep kernel space, to approximate the underlying frame distribution, as closely as possible. To define this closeness metric between the original video distribution and its approximation, we rely on maximum mean discrepancy (MMD) which is a geometry-preserving metric in distribution space, and thus gives more reliable estimates. Moreover, unlike the commonly used optimal transport metric, MMD is both easier to optimize, and faster. We choose to use neural tangent kernels (NTKs) to define the kernel space where MMD operates, because of their improved descriptive power as opposed to fixed kernels. And, also, because NTKs sidestep the trivial solution, when jointly learning the inputs (video approximation) and the kernel function. Finally, we show competitive results when compared to state-of-the-art per-video methods, on six standard benchmarks. Additionally, our method has higher F1 scores than prior agglomerative work, when the number of segments is unknown.", "date": "2026-04-23", "categories": []}
{"id": "2604.20760", "title": "Exploring High-Order Self-Similarity for Video Understanding", "abstract": "Space-time self-similarity (STSS), which captures visual correspondences across frames, provides an effective way to represent temporal dynamics for video understanding. In this work, we explore higher-order STSS and demonstrate how STSSs at different orders reveal distinct aspects of these dynamics. We then introduce the Multi-Order Self-Similarity (MOSS) module, a lightweight neural module designed to learn and integrate multi-order STSS features. It can be applied to diverse video tasks to enhance motion modeling capabilities while consuming only marginal computational cost and memory usage. Extensive experiments on video action recognition, motion-centric video VQA, and real-world robotic tasks consistently demonstrate substantial improvements, validating the broad applicability of MOSS as a general temporal modeling module. The source code and checkpoints will be publicly available.", "date": "2026-04-22", "categories": []}
{"id": "2503.07259", "title": "COMODO: Cross-Modal Video-to-IMU Distillation for Efficient Egocentric Human Activity Recognition", "abstract": "The goal of creating intelligent, human-centered wearable systems for continuous activity understanding faces a fundamental trade-off: Egocentric video-based models capture rich semantic information and have demonstrated strong performance in human activity recognition (HAR), but their high power consumption, privacy concerns, and dependence on lighting limit their feasibility for continuous on-device recognition. In contrast, inertial measurement unit (IMU) sensors offer an energy-efficient, privacy-preserving alternative, yet lack large-scale annotated datasets, leading to weaker generalization. To bridge this gap, we propose COMODO, a cross-modal self-supervised distillation framework that transfers semantic knowledge from video to IMU without requiring labels. COMODO leverages a pretrained and frozen video encoder to construct a dynamic instance queue to align the feature distributions of video and IMU embeddings. This enables the IMU encoder to inherit rich semantic structure from video while maintaining its efficiency for real-world applications. Experiments on multiple egocentric HAR datasets show that COMODO consistently improves downstream performance, matching or surpassing fully supervised models, and demonstrating strong cross-dataset generalization. Benefiting from its simplicity and flexibility, COMODO is compatible with diverse pretrained video and time-series models, offering the potential to leverage more powerful teacher and student foundation models in future ubiquitous computing research. The code is available at this repository: https://github.com/cruiseresearchgroup/COMODO.", "date": "2025-03-10", "categories": []}
{"id": "2511.12606", "title": "Pixels or Positions? Benchmarking Modalities in Group Activity Recognition", "abstract": "Group Activity Recognition (GAR) is well studied on the video modality for surveillance and indoor team sports (e.g., volleyball, basketball). Yet, other modalities such as agent positions and trajectories over time, i.e. tracking, remain comparatively under-explored despite being compact, agent-centric signals that explicitly encode spatial interactions. Understanding whether pixel (video) or position (tracking) modalities leads to better group activity recognition is therefore important to drive further research on the topic. However, no standardized benchmark currently exists that aligns broadcast video and tracking data for the same group activities, leading to a lack of apples-to-apples comparison between these modalities for GAR. In this work, we introduce SoccerNet-GAR, a multimodal dataset built from the $64$ matches of the football World Cup 2022. Specifically, the broadcast videos and player tracking modalities for $87{,}939$ group activities are synchronized and annotated with $10$ categories. Furthermore, we define a unified evaluation protocol to benchmark two strong unimodal approaches: (i) competitive video-based classifiers and (ii) tracking-based classifiers leveraging graph neural networks. In particular, our novel role-aware graph architecture for tracking-based GAR directly encodes tactical structure through positional edges connecting players by their on-pitch roles. Our tracking model achieves $77.8\\%$ balanced accuracy compared to $60.9\\%$ for the best video baseline, while training with $7 \\times$ less GPU hours and $479 \\times$ fewer parameters ($180K$ vs. $86.3M$). This study provides new insights into the relative strengths of pixels and positions for group activity recognition in sports.", "date": "2025-11-16", "categories": []}
{"id": "2604.16577", "title": "Multilevel neural networks with dual-stage feature fusion for human activity recognition", "abstract": "Human activity recognition (HAR) refers to the process of identifying human actions and activities using data collected from sensors. Neural networks, such as convolutional neural networks (CNNs), long short-term memory (LSTM) networks, convolutional LSTM, and their hybrid combinations, have demonstrated exceptional performance in various research domains. Developing a multilevel individual or hybrid model for HAR involves strategically integrating multiple networks to capitalize on their complementary strengths. The structural arrangement of these components is a critical factor influencing the overall performance. This study explores a novel framework of a two-level network architecture with dual-stage feature fusion: late fusion, which combines the outputs from the first network level, and intermediate fusion, which integrates the features from both the first and second levels. We evaluated $15$ different network architectures of CNNs, LSTMs, and convolutional LSTMs, incorporating late fusion with and without intermediate fusion, to identify the optimal configuration. Experimental evaluation on two public benchmark datasets demonstrates that architectures incorporating both late and intermediate fusion achieve higher accuracy than those relying on late fusion alone. Moreover, the optimal configuration outperforms baseline models, thereby validating its effectiveness for HAR.", "date": "2026-04-17", "categories": []}
{"id": "2604.16588", "title": "MambaKick: Early Penalty Direction Prediction from HAR Embeddings", "abstract": "Penalty kicks in soccer are decided under extreme time constraints, where goalkeepers benefit from anticipating shot direction from the kickers motion before or around ball contact. In this paper, MambaKick is presented as a learning-based framework for penalty direction prediction that leverages pretrained human action recognition (HAR) embeddings extracted from contact-centered short video segments and combines them with a lightweight temporal predictor. Rather than relying on explicit kinematic reconstruction or handcrafted biomechanical features, the approach reuses transferable spatiotemporal representations and utilizes selective state-spare models (Mamba) for efficient sequence aggregation. Simple contextual metadata (e.g., field side and footedness) are also considered as complementary cues that may reduce ambiguity in real-world footage. Across a range of HAR backbones, MambaKick consistently improves or matches strong embedding baselines, achieving up to 53.1% accuracy for three classes and 64.5% for two classes under the proposed methodology. Overall, the results indicate that combining pretrained HAR representations with efficient state-space temporal modeling is a practical direction for low-latency intention prediction in real-world sports video. The code will be available at GitHub: https://github.com/hvelesaca/MambaKick/", "date": "2026-04-17", "categories": []}
{"id": "2604.16675", "title": "Appearance-free Action Recognition: Zero-shot Generalization in Humans and a Two-Pathway Model", "abstract": "Action recognition is a fundamental ability for social species. Yet, its underlying computations are not well understood. Classical psychophysical studies using simplified stimuli have shown that humans can perceive body motion even under degradation of relevant shape cues. Recent work using real-world action videos and their appearance-free counterparts (that preserve motion but lack static shape cues) included explicit training of humans and models on the appearance-free videos. Whether humans and vision models generalize in a zero-shot manner to appearance-free transformations of real-world action videos is not yet known. To measure this generalization in humans, we conducted a laboratory-based psychophysics experiment. 22 participants were trained to recognize five action categories using naturalistic videos (UCF5 dataset), and tested zero-shot on two types of appearance-free transformations: (i) dense-noise motion videos from an existing dataset (AFD5) and (ii) random-dot appearance-free videos. We find that participants recognize actions in both types of appearance-free videos well above chance, albeit with reduced accuracy compared to naturalistic videos. To model this behavior, we developed a two-pathway 3D CNN-based model combining an RGB (form) stream and an optical flow (motion) stream, including a coherence-gating mechanism inspired by Gestalt common-fate grouping. Our model generalizes to both appearance-free datasets and outperforms contemporary video classification models, narrowing the gap to human performance. We find that the motion pathway is critical for generalization to appearance-free videos, while the form pathway improves performance on naturalistic videos. Our findings highlight the importance of motion-based representations for generalization to appearance-free videos, and support the use of multi-stream architectures to model video-based action recognition.", "date": "2026-04-17", "categories": []}
{"id": "2604.17013", "title": "Towards Universal Skeleton-Based Action Recognition", "abstract": "With the development of robotics, skeleton-based action recognition has become increasingly important, as human-robot interaction requires understanding the actions of humans and humanoid robots. Due to different sources of human skeletons and structures of humanoid robots, skeleton data naturally exhibit heterogeneity. However, previous works overlook the data heterogeneity of skeletons and solely construct models using homogeneous skeletons. Moreover, open-vocabulary action recognition is also essential for real-world applications. To this end, this work studies the challenging problem of heterogeneous skeleton-based action recognition with open vocabularies. We construct a large-scale Heterogeneous Open-Vocabulary (HOV) Skeleton dataset by integrating and refining multiple representative large-scale skeleton-based action datasets. To address universal skeleton-based action recognition, we propose a Transformer-based model that comprises three key components: unified skeleton representation, motion encoder for skeletons, and multi-grained motion-text alignment. The motion encoder feeds multi-modal skeleton embeddings into a two-stream Transformer-based encoder to learn spatio-temporal action representations, which are then mapped to a semantic space to align with text embeddings. Multi-grained motion-text alignment incorporates contrastive learning at three levels: global instance alignment, stream-specific alignment, and fine-grained alignment. Extensive experiments on popular benchmarks with heterogeneous skeleton data demonstrate both the effectiveness and the generalization ability of the proposed method. Code is available at https://github.com/jidongkuang/Universal-Skeleton.", "date": "2026-04-18", "categories": []}
{"id": "2604.17062", "title": "Motion-Guided Semantic Alignment with Negative Prompts for Zero-Shot Video Action Recognition", "abstract": "Zero-shot action recognition is challenging due to the semantic gap between seen and unseen classes. We present a novel framework that enhances CLIP with disentangled embeddings and semantic-guided interaction. A Motion Separation Module (MSM) separates motion-sensitive and global-static features, while a Motion Aggregation Block (MAB) employs gated cross-attention to refine motion representation without re-coupling redundant information. To facilitate generalization to unseen categories, we enforce semantic alignment between video features and textual representations by aligning projected embeddings with positive textual prompts, while leveraging negative prompts to explicitly model \"non-class\" semantics. Experiments on standard benchmarks demonstrate that our method consistently outperforms prior CLIP-based approaches, achieving robust zero-shot action recognition across both coarse and fine-grained datasets.", "date": "2026-04-18", "categories": []}
{"id": "2604.17914", "title": "Beyond Binary Contrast: Modeling Continuous Skeleton Action Spaces with Transitional Anchors", "abstract": "Self-supervised contrastive learning has emerged as a powerful paradigm for skeleton-based action recognition by enforcing consistency in the embedding space. However, existing methods rely on binary contrastive objectives that overlook the intrinsic continuity of human motion, resulting in fragmented feature clusters and rigid class boundaries. To address these limitations, we propose TranCLR, a Transitional anchor-based Contrastive Learning framework that captures the continuous geometry of the action space. Specifically, the proposed Action Transitional Anchor Construction (ATAC) explicitly models the geometric structure of transitional states to enhance the model's perception of motion continuity. Building upon these anchors, a Multi-Level Geometric Manifold Calibration (MGMC) mechanism is introduced to adaptively calibrate the action manifold across multiple levels of continuity, yielding a smoother and more discriminative representation space. Extensive experiments on the NTU RGB+D, NTU RGB+D 120 and PKU-MMD datasets demonstrate that TranCLR achieves superior accuracy and calibration performance, effectively learning continuous and uncertainty-aware skeleton representations. The code is available at https://github.com/Philchieh/TranCLR.", "date": "2026-04-20", "categories": []}
{"id": "2604.17971", "title": "Identifying Ethical Biases in Action Recognition Models", "abstract": "Human Action Recognition (HAR) models are increasingly deployed in high-stakes environments, yet their fairness across different human appearances has not been analyzed. We introduce a framework for auditing bias in HAR models using synthetic video data, generated with full control over visual identity attributes such as skin color. Unlike prior work that focuses on static images or pose estimation, our approach preserves temporal consistency, allowing us to isolate and test how changes to a single attribute affect model predictions. Through controlled interventions using the BEDLAM simulation platform, we show whether some popular HAR models exhibit statistically significant biases on the skin color even when the motion remains identical. Our results highlight how models may encode unwanted visual associations, and we provide evidence of systematic errors across groups. This work contributes a framework for auditing HAR models and supports the development of more transparent, accountable systems in light of upcoming regulatory standards.", "date": "2026-04-20", "categories": []}
{"id": "2604.18064", "title": "Understanding Human Actions through the Lens of Executable Models", "abstract": "Human-centred systems require an understanding of human actions in the physical world. Temporally extended sequences of actions are intentional and structured, yet existing methods for recognising what actions are performed often do not attempt to capture their structure, particularly how the actions are executed. This, however, is crucial for assessing the quality of the action's execution and its differences from other actions. To capture the internal mechanics of actions, we introduce a domain-specific language EXACT that represents human motions as underspecified motion programs, interpreted as reward-generating functions for zero-shot policy inference using forward-backwards representations. By leveraging the compositional nature of EXACT motion programs, we combine individual policies into an executable neuro-symbolic model that uses program structure for compositional modelling. We evaluate the utility of the proposed pipeline for creating executable action models by analysing motion-capture data to understand human actions, for the tasks of human action segmentation and action anomaly detection. Our results show that the use of executable action models improves data efficiency and captures intuitive relationships between actions compared with monolithic, task-specific approaches.", "date": "2026-04-20", "categories": []}
{"id": "2604.18313", "title": "Denoise and Align: Diffusion-Driven Foreground Knowledge Prompting for Open-Vocabulary Temporal Action Detection", "abstract": "Open-Vocabulary Temporal Action Detection (OV-TAD) aims to localize and classify action segments of unseen categories in untrimmed videos, where effective alignment between action semantics and video representations is critical for accurate detection. However, existing methods struggle to mitigate the semantic imbalance between concise, abstract action labels and rich, complex video contents, inevitably introducing semantic noise and misleading cross-modal alignment. To address this challenge, we propose DFAlign, the first framework that leverages diffusion-based denoising to generate foreground knowledge for the guidance of action-video alignment. Following the 'conditioning, denoising and aligning' manner, we first introduce the Semantic-Unify Conditioning (SUC) module, which unifies action-shared and action-specific semantics as conditions for diffusion denoising. Then, the Background-Suppress Denoising (BSD) module generates foreground knowledge by progressively removing background redundancy from videos through denoising process. This foreground knowledge serves as effective intermediate semantic anchor between video and text representations, mitigating the semantic gap and enhancing the discriminability of action-relevant segments. Furthermore, we introduce the Foreground-Prompt Alignment (FPA) module to inject extracted foreground knowledge as prompt tokens into text representations, guiding model's attention towards action-relevant segments and enabling precise cross-modal alignment. Extensive experiments demonstrate that our method achieves state-of-the-art performance on two OV-TAD benchmarks. The code repository is provided as follows: https://anonymous.4open.science/r/Code-2114/.", "date": "2026-04-20", "categories": []}
{"id": "2604.18367", "title": "EAST: Early Action Prediction Sampling Strategy with Token Masking", "abstract": "Early action prediction seeks to anticipate an action before it fully unfolds, but limited visual evidence makes this task especially challenging. We introduce EAST, a simple and efficient framework that enables a model to reason about incomplete observations. In our empirical study, we identify key components when training early action prediction models. Our key contribution is a randomized training strategy that samples a time step separating observed and unobserved video frames, enabling a single model to generalize seamlessly across all test-time observation ratios. We further show that joint learning on both observed and future (oracle) representations significantly boosts performance, even allowing an encoder-only model to excel. To improve scalability, we propose a token masking procedure that cuts memory usage in half and accelerates training by 2x with negligible accuracy loss. Combined with a forecasting decoder, EAST sets a new state of the art on NTU60, SSv2, and UCF101, surpassing previous best work by 10.1, 7.7, and 3.9 percentage points, respectively.", "date": "2026-04-20", "categories": []}
{"id": "2411.18328", "title": "EventCrab: Harnessing Frame and Point Synergy for Event-based Action Recognition and Beyond", "abstract": "Event-based Action Recognition (EAR) possesses the advantages of high-temporal resolution capturing and privacy preservation compared with traditional action recognition. Current leading EAR solutions typically follow two regimes: project unconstructed event streams into dense constructed event frames and adopt powerful frame-specific networks, or employ lightweight point-specific networks to handle sparse unconstructed event points directly. However, such two regimes are blind to a fundamental issue: failing to accommodate the unique dense temporal and sparse spatial properties of asynchronous event data. In this article, we present a synergy-aware framework, i.e., EventCrab, that adeptly integrates the \"lighter\" frame-specific networks for dense event frames with the \"heavier\" point-specific networks for sparse event points, balancing accuracy and efficiency. Furthermore, we establish a joint frame-text-point representation space to bridge distinct event frames and points. In specific, to better exploit the unique spatiotemporal relationships inherent in asynchronous event points, we devise two strategies for the \"heavier\" point-specific embedding: i) a Spiking-like Context Learner (SCL) that extracts contextualized event points from raw event streams. ii) an Event Point Encoder (EPE) that further explores event-point long spatiotemporal features in a Hilbert-scan way. Experiments on four datasets demonstrate the significant performance of our proposed EventCrab, particularly gaining improvements of 5.17% on SeAct and 7.01% on HARDVS.", "date": "2024-11-27", "categories": []}
{"id": "2510.08480", "title": "Video-STAR: Reinforcing Open-Vocabulary Action Recognition with Tools", "abstract": "Multimodal large language models (MLLMs) have demonstrated remarkable potential in bridging visual and textual reasoning, yet their reliance on text-centric priors often limits their ability to disentangle semantically similar actions in open-vocabulary scenarios. To address this, we propose Video-STAR, a framework that harmonizes contextual sub-motion decomposition with tool-augmented reinforcement learning for open-vocabulary action recognition (OVAR). Unlike prior methods that treat actions as monolithic entities, our approach innovatively decomposes actions into discriminative sub-motions for fine-grained matching while dynamically invoking domain-specific tools for cross-modal interleaving, thereby enabling category-specific reasoning capacity and reducing cross-modal hallucination. Moreover, by designing a hierarchical reward that balances tool-usage efficiency, sub-motion relevance, and structural coherence in reasoning, our method autonomously leverages external tools to prioritize sub-motion patterns without explicit supervision, transmitting from text-centric reasoning to visually grounded inference. Extensive evaluations on HMDB-51, UCF-101, SSv2, Kinetics-400, and Kinetics-600 datasets demonstrate our state-of-the-art performance, outperforming existing methods in distinguishing fine-grained actions and handling cross-modal hallucination, validating our excellent robustness and generalization.", "date": "2025-10-09", "categories": []}
{"id": "2604.14329", "title": "Interpretable Human Activity Recognition for Subtle Robbery Detection in Surveillance Videos", "abstract": "Non-violent street robberies (snatch-and-run) are difficult to detect automatically because they are brief, subtle, and often indistinguishable from benign human interactions in unconstrained surveillance footage. This paper presents a hybrid, pose-driven approach for detecting snatch-and-run events that combines real-time perception with an interpretable classification stage suitable for edge deployment. The system uses a YOLO-based pose estimator to extract body keypoints for each tracked person and computes kinematic and interaction features describing hand speed, arm extension, proximity, and relative motion between an aggressor-victim pair. A Random Forest classifier is trained on these descriptors, and a temporal hysteresis filter is applied to stabilize frame-level predictions and reduce spurious alarms. We evaluate the method on a staged dataset and on a disjoint test set collected from internet videos, demonstrating promising generalization across different scenes and camera viewpoints. Finally, we implement the complete pipeline on an NVIDIA Jetson Nano and report real-time performance, supporting the feasibility of proactive, on-device robbery detection.", "date": "2026-04-15", "categories": []}
{"id": "2604.14933", "title": "Generative Data Augmentation for Skeleton Action Recognition", "abstract": "Skeleton-based human action recognition is a powerful approach for understanding human behaviour from pose data, but collecting large-scale, diverse, and well-annotated 3D skeleton datasets is both expensive and labor-intensive. To address this challenge, we propose a conditional generative pipeline for data augmentation in skeleton action recognition. Our method learns the distribution of real skeleton sequences under the constraint of action labels, enabling the synthesis of diverse and high-fidelity data. Even with limited training samples, it can effectively generate skeleton sequences and achieve competitive recognition performance in low-data scenarios, demonstrating strong generalisation in downstream tasks. Specifically, we introduce a Transformer-based encoder-decoder architecture, combined with a generative refinement module and a dropout mechanism, to balance fidelity and diversity during sampling. Experiments on HumanAct12 and the refined NTU-RGBD (NTU-VIBE) dataset show that our approach consistently improves the accuracy of multiple skeleton-based action recognition models, validating its effectiveness in both few-shot and full-data settings. The source code can be found at here.", "date": "2026-04-16", "categories": []}
{"id": "2604.15196", "title": "Unsupervised Skeleton-Based Action Segmentation via Hierarchical Spatiotemporal Vector Quantization", "abstract": "We propose a novel hierarchical spatiotemporal vector quantization framework for unsupervised skeleton-based temporal action segmentation. We first introduce a hierarchical approach, which includes two consecutive levels of vector quantization. Specifically, the lower level associates skeletons with fine-grained subactions, while the higher level further aggregates subactions into action-level representations. Our hierarchical approach outperforms the non-hierarchical baseline, while primarily exploiting spatial cues by reconstructing input skeletons. Next, we extend our approach by leveraging both spatial and temporal information, yielding a hierarchical spatiotemporal vector quantization scheme. In particular, our hierarchical spatiotemporal approach performs multi-level clustering, while simultaneously recovering input skeletons and their corresponding timestamps. Lastly, extensive experiments on multiple benchmarks, including HuGaDB, LARa, and BABEL, demonstrate that our approach establishes a new state-of-the-art performance and reduces segment length bias in unsupervised skeleton-based temporal action segmentation.", "date": "2026-04-16", "categories": []}
{"id": "2508.04038", "title": "ZARA: Training-Free Motion Time-Series Reasoning via Evidence-Grounded LLM Agents", "abstract": "Motion sensor time-series are central to Human Activity Recognition (HAR), yet conventional approaches are constrained to fixed activity sets and typically require costly parameter retraining to adapt to new behaviors. While Large Language Models (LLMs) offer promising open-set reasoning capabilities, applying them directly to numerical time-series often leads to hallucinations and weak grounding. To address this challenge, we propose ZARA (Zero-training Activity Reasoning Agents), a knowledge- and retrieval-augmented agentic framework for motion time-series reasoning in a training-free inference setting. Rather than relying on black-box projections, ZARA distills reference data into a statistically grounded textual knowledge base that transforms implicit signal patterns into verifiable natural-language priors. Guided by retrieved evidence, ZARA iteratively selects discriminative cues and performs grounded reasoning over candidate activities. Extensive experiments on eight benchmarks show that ZARA generalizes robustly to unseen subjects and across datasets, demonstrating strong transferability across heterogeneous sensor domains. These results mark a step toward trustworthy, plug-and-play motion understanding beyond dataset-specific artifacts. Our code is available at https://github.com/zechenli03/ZARA.", "date": "2025-08-06", "categories": []}
{"id": "2601.21078", "title": "Towards Mitigating Modality Bias in Vision-Language Models for Temporal Action Localization", "abstract": "Temporal Action Localization (TAL) requires identifying both the boundaries and categories of actions in untrimmed videos. While vision-language models (VLMs) offer rich semantics to complement visual evidence, existing approaches tend to overemphasize linguistic priors at the expense of visual performance, leading to a pronounced modality bias. We propose ActionVLM, a vision-language aggregation framework that systematically mitigates modality bias in TAL. Our key insight is to preserve vision as the dominant signal while adaptively exploiting language only when beneficial. To this end, we introduce (i) a debiasing reweighting module that estimates the language advantage-the incremental benefit of language over vision-only predictions-and dynamically reweights language modality accordingly, and (ii) a residual aggregation strategy that treats language as a complementary refinement rather than the primary driver. This combination alleviates modality bias, reduces overconfidence from linguistic priors, and strengthens temporal reasoning. Experiments on THUMOS14 show that our model outperforms state-of-the-art by up to 3.2% mAP. Our code is available at https://github.com/JiaqiLi404/ActionVLM", "date": "2026-01-28", "categories": []}
{"id": "2604.10078", "title": "Attention-Guided Dual-Stream Learning for Group Engagement Recognition: Fusing Transformer-Encoded Motion Dynamics with Scene Context via Adaptive Gating", "abstract": "Student engagement is crucial for improving learning outcomes in group activities. Highly engaged students perform better both individually and contribute to overall group success. However, most existing automated engagement recognition methods are designed for online classrooms or estimate engagement at the individual level. Addressing this gap, we propose DualEngage, a novel two-stream framework for group-level engagement recognition from in-classroom videos. It models engagement as a joint function of both individual and group-level behaviors. The primary stream models person-level motion dynamics by detecting and tracking students, extracting dense optical flow with the Recurrent All-Pairs Field Transforms network, encoding temporal motion patterns using a transformer encoder, and finally aggregating per-student representations through attention pooling into a unified representation. The secondary stream captures scene-level spatiotemporal information from the full video clip, leveraging a pretrained three-dimensional Residual Network. The two-stream representations are combined via softmax-gated fusion, which dynamically weights each stream's contribution based on the joint context of both features. DualEngage learns a joint representation of individual actions with overarching group dynamics. We evaluate the proposed approach using fivefold cross-validation on the Classroom Group Engagement Dataset developed by Ocean University of China, achieving an average classification accuracy of 0.9621+/-0.0161 with a macro-averaged F1 of 0.9530+/-0.0204. To understand the contribution of each branch, we further conduct an ablation study comparing single-stream variants against the two-stream model. This work is among the first in classroom engagement recognition to adopt a dual-stream design that explicitly leverages motion cues as an estimator.", "date": "2026-04-11", "categories": []}
{"id": "2604.10409", "title": "IMPACT: A Dataset for Multi-Granularity Human Procedural Action Understanding in Industrial Assembly", "abstract": "We introduce IMPACT, a synchronized five-view RGB-D dataset for deployment-oriented industrial procedural understanding, built around real assembly and disassembly of a commercial angle grinder with professional-grade tools. To our knowledge, IMPACT is the first real industrial assembly benchmark that jointly provides synchronized ego-exo RGB-D capture, decoupled bimanual annotation, compliance-aware state tracking, and explicit anomaly--recovery supervision within a single real industrial workflow. It comprises 112 trials from 13 participants totaling 39.5 hours, with multi-route execution governed by a partial-order prerequisite graph, a six-category anomaly taxonomy, and operator cognitive load measured via NASA-TLX. The annotation hierarchy links hand-specific atomic actions to coarse procedural steps, component assembly states, and per-hand compliance phases, with synchronized null spans across views to decouple perceptual limitations from algorithmic failure. Systematic baselines reveal fundamental limitations that remain invisible to single-task benchmarks, particularly under realistic deployment conditions that involve incomplete observations, flexible execution paths, and corrective behavior. The full dataset, annotations, and evaluation code are available at https://github.com/Kratos-Wen/IMPACT.", "date": "2026-04-12", "categories": []}
{"id": "2604.11498", "title": "TAG-Head: Time-Aligned Graph Head for Plug-and-Play Fine-grained Action Recognition", "abstract": "Fine-grained human action recognition (FHAR) is challenging because visually similar actions differ by subtle spatio-temporal cues. Many recent systems enhance discriminability with extra modalities (e.g., pose, text, optical flow), but this increases annotation burden and computational cost. We introduce TAG-Head, a lightweight spatio-temporal graph head that upgrades standard 3D backbones (SlowFast, R(2+1)D-34, I3D, etc.) for FHAR using RGB only. Our pipeline first applies a Transformer encoder with learnable 3D positional encodings to the backbone tokens, capturing long-range dependencies across space and time. The resulting features are then refined by a graph in which (i) fully-connected intra-frame edges to resolve subtle appearance differences within frames, and (ii) time-aligned temporal edges that connect features at the same spatial location across frames to stabilise motion cues without over-smoothing. The head is compact (little parameter/FLOP overhead), plug-and-play across backbones, and trained end-to-end with the backbone. Extensive evaluations on FineGym (Gym99 and Gym288) and HAA500 show that TAG-Head sets a new state-of-the-art among RGB-only models and surpasses many recent multimodal approaches (video + pose + text) that rely on privileged information. Ablations disentangle the contributions of the Transformer and the graph topology, and complexity analyses confirm low latency. TAG-Head advances FHAR by explicitly coupling global context with high-resolution spatial interactions and low-variance temporal continuity inside a slim, composable graph head. The simplicity of the design enables straightforward adoption in practical systems that favour RGB-only sensors, while delivering performance gains typically associated with heavier or multimodal models. Code will be released on GitHub.", "date": "2026-04-13", "categories": []}
{"id": "2603.24245", "title": "B-MoE: A Body-Part-Aware Mixture-of-Experts \"All Parts Matter\" Approach to Micro-Action Recognition", "abstract": "Micro-actions, fleeting and low-amplitude motions, such as glances, nods, or minor posture shifts, carry rich social meaning but remain difficult for current action recognition models to recognize due to their subtlety, short duration, and high inter-class ambiguity. In this paper, we introduce B-MoE, a Body-part-aware Mixture-of-Experts framework designed to explicitly model the structured nature of human motion. In B-MoE, each expert specializes in a distinct body region (head, body, upper limbs, lower limbs), and is based on the lightweight Macro-Micro Motion Encoder (M3E) that captures long-range contextual structure and fine-grained local motion. A cross-attention routing mechanism learns inter-region relationships and dynamically selects the most informative regions for each micro-action. B-MoE uses a dual-stream encoder that fuses these region-specific semantic cues with global motion features to jointly capture spatially localized cues and temporally subtle variations that characterize micro-actions. Experiments on three challenging benchmarks (MA-52, SocialGesture, and MPII-GroupInteraction) show consistent state-of-theart gains, with improvements in ambiguous, underrepresented, and low amplitude classes.", "date": "2026-03-25", "categories": []}
{"id": "2604.08762", "title": "InstrAct: Towards Action-Centric Understanding in Instructional Videos", "abstract": "Understanding instructional videos requires recognizing fine-grained actions and modeling their temporal relations, which remains challenging for current Video Foundation Models (VFMs). This difficulty stems from noisy web supervision and a pervasive \"static bias\", where models rely on objects rather than motion cues. To address this, we propose InstrAction, a pretraining framework for instructional videos' action-centric representations. We first introduce a data-driven strategy, which filters noisy captions and generates action-centric hard negatives to disentangle actions from objects during contrastive learning. At the visual feature level, an Action Perceiver extracts motion-relevant tokens from redundant video encodings. Beyond contrastive learning, we introduce two auxiliary objectives: Dynamic Time Warping alignment (DTW-Align) for modeling sequential temporal structure, and Masked Action Modeling (MAM) for strengthening cross-modal grounding. Finally, we introduce the InstrAct Bench to evaluate action-centric understanding, where our method consistently outperforms state-of-the-art VFMs on semantic reasoning, procedural logic, and fine-grained retrieval tasks.", "date": "2026-04-09", "categories": []}
{"id": "2604.09051", "title": "Fine-Grained Action Segmentation for Renorrhaphy in Robot-Assisted Partial Nephrectomy", "abstract": "Fine-grained action segmentation during renorrhaphy in robot-assisted partial nephrectomy requires frame-level recognition of visually similar suturing gestures with variable duration and substantial class imbalance. The SIA-RAPN benchmark defines this problem on 50 clinical videos acquired with the da Vinci Xi system and annotated with 12 frame-level labels. The benchmark compares four temporal models built on I3D features: MS-TCN++, AsFormer, TUT, and DiffAct. Evaluation uses balanced accuracy, edit score, segmental F1 at overlap thresholds of 10, 25, and 50, frame-wise accuracy, and frame-wise mean average precision. In addition to the primary evaluation across five released split configurations on SIA-RAPN, the benchmark reports cross-domain results on a separate single-port RAPN dataset. Across the strongest reported values over those five runs on the primary dataset, DiffAct achieves the highest F1, frame-wise accuracy, edit score, and frame mAP, while MS-TCN++ attains the highest balanced accuracy.", "date": "2026-04-10", "categories": []}
{"id": "2604.09164", "title": "Efficient Spatial-Temporal Focal Adapter with SSM for Temporal Action Detection", "abstract": "Temporal human action detection aims to identify and localize action segments within untrimmed videos, serving as a pivotal task in video understanding. Despite the progress achieved by prior architectures like CNN and Transformer models, these continue to struggle with feature redundancy and degraded global dependency modeling capabilities when applied to long video sequences. These limitations severely constrain their scalability in real-world video analysis. State Space Models (SSMs) offer a promising alternative with linear long-term modeling and robust global temporal reasoning capabilities. Rethinking the application of SSMs in temporal modeling, this research constructs a novel framework for video human action detection. Specifically, we introduce the Efficient Spatial-Temporal Focal (ESTF) Adapter into the pre-trained layers. This module integrates the advantages of our proposed Temporal Boundary-aware SSM(TB-SSM) for temporal feature modeling with efficient processing of spatial features. We perform comprehensive and quantitative analyses across multiple benchmarks, comparing our proposed method against previous SSM-based and other structural methods. Extensive experiments demonstrate that our improved strategy significantly enhances both localization performance and robustness, validating the effectiveness of our proposed method.", "date": "2026-04-10", "categories": []}
{"id": "2403.14922", "title": "CODA: A Continuous Online Evolve Framework for Deploying HAR Sensing Systems", "abstract": "In always-on HAR deployments, model accuracy erodes silently as domain shift accumulates over time. Addressing this challenge requires moving beyond one-off updates toward instance-driven adaptation from streaming data. However, continuous adaptation exposes a fundamental tension: systems must selectively learn from informative instances while actively forgetting obsolete ones under long-term, non-stationary drift. To address them, we propose CODA, a continuous online adaptation framework for mobile sensing. CODA introduces two synergistic components: (i) Cache-based Selective Assimilation, which prioritizes informative instances likely to enhance system performance under sparse supervision, and (ii) an Adaptive Temporal Retention Strategy, which enables the system to gradually forget obsolete instances as sensing conditions evolve. By treating adaptation as a principled cache evolution rather than parameter-heavy retraining, CODA maintains high accuracy without model reconfiguration. We conduct extensive evaluations on four heterogeneous datasets spanning phone, watch, and multi-sensor configurations. Results demonstrate that CODA consistently outperforms one-off adaptation under non-stationary drift, remains robust against imperfect feedback, and incurs negligible on-device latency.", "date": "2024-03-22", "categories": []}
{"id": "2604.06783", "title": "Insights from Visual Cognition: Understanding Human Action Dynamics with Overall Glance and Refined Gaze Transformer", "abstract": "Recently, Transformer has made significant progress in various vision tasks. To balance computation and efficiency in video tasks, recent works heavily rely on factorized or window-based self-attention. However, these approaches split spatiotemporal correlations between regions of interest in videos, limiting the models' ability to capture motion and long-range dependencies. In this paper, we argue that, similar to the human visual system, the importance of temporal and spatial information varies across different time scales, and attention is allocated sparsely over time through glance and gaze behavior. Is equal consideration of time and space crucial for success in video tasks? Motivated by this understanding, we propose a dual-path network called the Overall Glance and Refined Gaze (OG-ReG) Transformer. The Glance path extracts coarse-grained overall spatiotemporal information, while the Gaze path supplements the Glance path by providing local details. Our model achieves state-of-the-art results on the Kinetics-400, Something-Something v2, and Diving-48, demonstrating its competitive performance. The code will be available at https://github.com/linuxsino/OG-ReG.", "date": "2026-04-08", "categories": []}
{"id": "2601.16211", "title": "Why Can't I Open My Drawer? Mitigating Object-Driven Shortcuts in Zero-Shot Compositional Action Recognition", "abstract": "Zero-Shot Compositional Action Recognition (ZS-CAR) requires recognizing novel verb-object combinations composed of previously observed primitives. In this work, we tackle a key failure mode: models predict verbs via object-driven shortcuts (i.e., relying on the labeled object class) rather than temporal evidence. We argue that sparse compositional supervision and verb-object learning asymmetry can promote object-driven shortcut learning. Our analysis with proposed diagnostic metrics shows that existing methods overfit to training co-occurrence patterns and underuse temporal verb cues, resulting in weak generalization to unseen compositions. To address object-driven shortcuts, we propose Robust COmpositional REpresentations (RCORE) with two components. Co-occurrence Prior Regularization (CPR) adds explicit supervision for unseen compositions and regularizes the model against frequent co-occurrence priors by treating them as hard negatives. Temporal Order Regularization for Composition (TORC) enforces temporal-order sensitivity to learn temporally grounded verb representations. Across Sth-com and EK100-com, RCORE reduces shortcut diagnostics and consequently improves compositional generalization.", "date": "2026-01-22", "categories": []}
{"id": "2604.05947", "title": "Mixture-of-Modality-Experts with Holistic Token Learning for Fine-Grained Multimodal Visual Analytics in Driver Action Recognition", "abstract": "Robust multimodal visual analytics remains challenging when heterogeneous modalities provide complementary but input-dependent evidence for decision-making.Existing multimodal learning methods mainly rely on fixed fusion modules or predefined cross-modal interactions, which are often insufficient to adapt to changing modality reliability and to capture fine-grained action cues. To address this issue, we propose a Mixture-of-Modality-Experts (MoME) framework with a Holistic Token Learning (HTL) strategy. MoME enables adaptive collaboration among modality-specific experts, while HTL improves both intra-expert refinement and inter-expert knowledge transfer through class tokens and spatio-temporal tokens. In this way, our method forms a knowledge-centric multimodal learning framework that improves expert specialization while reducing ambiguity in multimodal fusion.We validate the proposed framework on driver action recognition as a representative multimodal understanding taskThe experimental results on the public benchmark show that the proposed MoME framework and the HTL strategy jointly outperform representative single-modal and multimodal baselines. Additional ablation, validation, and visualization results further verify that the proposed HTL strategy improves subtle multimodal understanding and offers better interpretability.", "date": "2026-04-07", "categories": []}
{"id": "2604.03590", "title": "SBF: An Effective Representation to Augment Skeleton for Video-based Human Action Recognition", "abstract": "Many modern video-based human action recognition (HAR) approaches use 2D skeleton as the intermediate representation in their prediction pipelines. Despite overall encouraging results, these approaches still struggle in many common scenes, mainly because the skeleton does not capture critical action-related information pertaining to the depth of the joints, contour of the human body, and interaction between the human and objects. To address this, we propose an effective approach to augment skeleton with a representation capturing action-related information in the pipeline of HAR. The representation, termed Scale-Body-Flow (SBF), consists of three distinct components, namely a scale map volume given by the scale (and hence depth information) of each joint, a body map outlining the human subject, and a flow map indicating human-object interaction given by pixel-wise optical flow values. To predict SBF, we further present SFSNet, a novel segmentation network supervised by the skeleton and optical flow without extra annotation overhead beyond the existing skeleton extraction. Extensive experiments across different datasets demonstrate that our pipeline based on SBF and SFSNet achieves significantly higher HAR accuracy with similar compactness and efficiency as compared with the state-of-the-art skeleton-only approaches.", "date": "2026-04-04", "categories": []}
{"id": "2604.04467", "title": "Group-DINOmics: Incorporating People Dynamics into DINO for Self-supervised Group Activity Feature Learning", "abstract": "This paper proposes Group Activity Feature (GAF) learning without group activity annotations. Unlike prior work, which uses low-level static local features to learn GAFs, we propose leveraging dynamics-aware and group-aware pretext tasks, along with local and global features provided by DINO, for group-dynamics-aware GAF learning. To adapt DINO and GAF learning to local dynamics and global group features, our pretext tasks use person flow estimation and group-relevant object location estimation, respectively. Person flow estimation is used to represent the local motion of each person, which is an important cue for understanding group activities. In contrast, group-relevant object location estimation encourages GAFs to learn scene context (e.g., spatial relations of people and objects) as global features. Comprehensive experiments on public datasets demonstrate the state-of-the-art performance of our method in group activity retrieval and recognition. Our ablation studies verify the effectiveness of each component in our method. Code: https://github.com/tezuka0001/Group-DINOmics.", "date": "2026-04-06", "categories": []}
{"id": "2604.01318", "title": "ViTs for Action Classification in Videos: An Approach to Risky Tackle Detection in American Football Practice Videos", "abstract": "Early identification of hazardous actions in contact sports enables timely intervention and improves player safety. We present a method for detecting risky tackles in American football practice videos and introduce a substantially expanded dataset for this task. Our work contains 733 single-athlete-dummy tackle clips, each temporally localized around first point contact and labeled with a strike zone component of the standardized Assessment for Tackling Technique (SATT-3), extending prior work that reported 178 annotated videos. Using a Vision transformer-based model with imbalance-aware training, we obtain risky recall of 0.67 and Risky F1 of 0.59 under crossvalidation. Relative to the previous baseline in a smaller subset (risky recall of 0.58; Risky F1 0.56 ), our approach improves risky recall by more than 8% points on a much larger dataset. These results indicate that the vision transformer-based video analysis, coupled with careful handling of class imbalance, can reliably detect rare but safety-critical tackling patterns, offering a practical pathway toward coach-centered injury prevention tools.", "date": "2026-04-01", "categories": []}
{"id": "2604.01859", "title": "Combining Boundary Supervision and Segment-Level Regularization for Fine-Grained Action Segmentation", "abstract": "Recent progress in Temporal Action Segmentation (TAS) has increasingly relied on complex architectures, which can hinder practical deployment. We present a lightweight dual-loss training framework that improves fine-grained segmentation quality with only one additional output channel and two auxiliary loss terms, requiring minimal architectural modification. Our approach combines a boundary-regression loss that promotes accurate temporal localization via a single-channel boundary prediction and a CDF-based segment-level regularization loss that encourages coherent within-segment structure by matching cumulative distributions over predicted and ground-truth segments. The framework is architecture-agnostic and can be integrated into existing TAS models (e.g., MS-TCN, C2F-TCN, FACT) as a training-time loss function. Across three benchmark datasets, the proposed method improves segment-level consistency and boundary quality, yielding higher F1 and Edit scores across three different models. Frame-wise accuracy remains largely unchanged, highlighting that precise segmentation can be achieved through simple loss design rather than heavier architectures or inference-time refinements.", "date": "2026-04-02", "categories": []}
{"id": "2604.02222", "title": "SCALE: Semantic- and Confidence-Aware Conditional Variational Autoencoder for Zero-shot Skeleton-based Action Recognition", "abstract": "Zero-shot skeleton-based action recognition (ZSAR) aims to recognize action classes without any training skeletons from those classes, relying instead on auxiliary semantics from text. Existing approaches frequently depend on explicit skeleton-text alignment, which can be brittle when action names underspecify fine-grained dynamics and when unseen classes are semantically confusable. We propose SCALE, a lightweight and deterministic Semantic- and Confidence-Aware Listwise Energy-based framework that formulates ZSAR as class-conditional energy ranking. SCALE builds a text-conditioned Conditional Variational Autoencoder where frozen text representations parameterize both the latent prior and the decoder, enabling likelihood-based evaluation for unseen classes without generating samples at test time. To separate competing hypotheses, we introduce a semantic- and confidence-aware listwise energy loss that emphasizes semantically similar hard negatives and incorporates posterior uncertainty to adapt decision margins and reweight ambiguous training instances. Additionally, we utilize a latent prototype contrast objective to align posterior means with text-derived latent prototypes, improving semantic organization and class separability without direct feature matching. Experiments on NTU-60 and NTU-120 datasets show that SCALE consistently improves over prior VAE- and alignment-based baselines while remaining competitive with diffusion-based methods.", "date": "2026-04-02", "categories": []}
{"id": "2511.09388", "title": "Learning by Neighbor-Aware Semantics, Deciding by Open-form Flows: Towards Robust Zero-Shot Skeleton Action Recognition", "abstract": "Recognizing unseen skeleton action categories remains highly challenging due to the absence of corresponding skeletal priors. Existing approaches generally follow an ``align-then-classify'' paradigm but face two fundamental issues, \\textit{i.e.}, (i) fragile point-to-point alignment arising from imperfect semantics, and (ii) rigid classifiers restricted by static decision boundaries and coarse-grained anchors. To address these issues, we propose a novel method for zero-shot skeleton action recognition, termed \\texttt{\\textbf{Flora}}, which builds upon \\textbf{F}lexib\\textbf{L}e neighb\\textbf{O}r-aware semantic attunement and open-form dist\\textbf{R}ibution-aware flow cl\\textbf{A}ssifier. Specifically, we flexibly attune textual semantics by incorporating neighboring inter-class contextual cues to form direction-aware regional semantics, coupled with a cross-modal geometric consistency objective that ensures stable and robust point-to-region alignment. Furthermore, we employ noise-free flow matching to bridge the modality distribution gap between semantic and skeleton latent embeddings, while a condition-free contrastive regularization enhances discriminability, leading to a distribution-aware classifier with fine-grained decision boundaries achieved through token-level velocity predictions. Extensive experiments on three benchmark datasets validate the effectiveness of our method, showing particularly impressive performance even when trained with only 10% of the seen data. Code is available at https://github.com/cseeyangchen/Flora.", "date": "2025-11-12", "categories": []}
{"id": "2604.00517", "title": "Toward Optimal Sampling Rate Selection and Unbiased Classification for Precise Animal Activity Recognition", "abstract": "With the rapid advancements in deep learning techniques, wearable sensor-aided animal activity recognition (AAR) has demonstrated promising performance, thereby improving livestock management efficiency as well as animal health and welfare monitoring. However, existing research often prioritizes overall performance, overlooking the fact that classification accuracies for specific animal behavioral categories may remain unsatisfactory. This issue typically stems from suboptimal sampling rates or class imbalance problems. To address these challenges and achieve high classification accuracy across all individual behaviors in farm animals, we propose a novel Individual-Behavior-Aware Network (IBA-Net). This network enhances the recognition of each specific behavior by simultaneously customizing features and calibrating the classifier. Specifically, considering that different behaviors require varying sampling rates to achieve optimal performance, we design a Mixture-of-Experts (MoE)-based Feature Customization (MFC) module. This module adaptively fuses data from multiple sampling rates, capturing customized features tailored to various animal behaviors. Additionally, to mitigate classifier bias toward majority classes caused by class imbalance, we develop a Neural Collapse-driven Classifier Calibration (NC3) module. This module introduces a fixed equiangular tight frame (ETF) classifier during the classification stage, maximizing the angles between pair-wise classifier vectors and thereby improving the classification performance for minority classes. To validate the effectiveness of IBA-Net, we conducted experiments on three public datasets covering goat, cattle, and horse activity recognition. The results demonstrate that our method consistently outperforms existing approaches across all datasets.", "date": "2026-04-01", "categories": []}
{"id": "2604.00761", "title": "PrivHAR-Bench: A Graduated Privacy Benchmark Dataset for Video-Based Action Recognition", "abstract": "Existing research on privacy-preserving Human Activity Recognition (HAR) typically evaluates methods against a binary paradigm: clear video versus a single privacy transformation. This limits cross-method comparability and obscures the nuanced relationship between privacy strength and recognition utility. We introduce \\textit{PrivHAR-Bench}, a multi-tier benchmark dataset designed to standardize the evaluation of the \\textit{Privacy-Utility Trade-off} in video-based action recognition. PrivHAR-Bench applies a graduated spectrum of visual privacy transformations: from lightweight spatial obfuscation to cryptographic block permutation, to a curated subset of 15 activity classes selected for human articulation diversity. Each of the 1,932 source videos is distributed across 9 parallel tiers of increasing privacy strength, with additional background-removed variants to isolate the contribution of human motion features from contextual scene bias. We provide lossless frame sequences, per-frame bounding boxes, estimated pose keypoints with joint-level confidence scores, standardized group-based train/test splits, and an evaluation toolkit computing recognition accuracy and privacy metrics. Empirical validation using R3D-18 demonstrates a measurable and interpretable degradation curve across tiers, with within-tier accuracy declining from 88.8\\% (clear) to 53.5\\% (encrypted, background-removed) and cross-domain accuracy collapsing to 4.8\\%, establishing PrivHAR-Bench as a controlled benchmark for comparing privacy-preserving HAR methods under standardized conditions. The dataset, generation pipeline, and evaluation code are publicly available.", "date": "2026-04-01", "categories": []}
{"id": "2603.29692", "title": "SkeletonContext: Skeleton-side Context Prompt Learning for Zero-Shot Skeleton-based Action Recognition", "abstract": "Zero-shot skeleton-based action recognition aims to recognize unseen actions by transferring knowledge from seen categories through semantic descriptions. Most existing methods typically align skeleton features with textual embeddings within a shared latent space. However, the absence of contextual cues, such as objects involved in the action, introduces an inherent gap between skeleton and semantic representations, making it difficult to distinguish visually similar actions. To address this, we propose SkeletonContext, a prompt-based framework that enriches skeletal motion representations with language-driven contextual semantics. Specifically, we introduce a Cross-Modal Context Prompt Module, which leverages a pretrained language model to reconstruct masked contextual prompts under guidance derived from LLMs. This design effectively transfers linguistic context to the skeleton encoder for instance-level semantic grounding and improved cross-modal alignment. In addition, a Key-Part Decoupling Module is incorporated to decouple motion-relevant joint features, ensuring robust action understanding even in the absence of explicit object interactions. Extensive experiments on multiple benchmarks demonstrate that SkeletonContext achieves state-of-the-art performance under both conventional and generalized zero-shot settings, validating its effectiveness in reasoning about context and distinguishing fine-grained, visually similar actions.", "date": "2026-03-31", "categories": []}
{"id": "2508.09428", "title": "What-Meets-Where: Unified Learning of Action and Contact Localization in Images", "abstract": "People control their bodies to establish contact with the environment. To comprehensively understand actions across diverse visual contexts, it is essential to simultaneously consider \\textbf{what} action is occurring and \\textbf{where} it is happening. Current methodologies, however, often inadequately capture this duality, typically failing to jointly model both action semantics and their spatial contextualization within scenes. To bridge this gap, we introduce a novel vision task that simultaneously predicts high-level action semantics and fine-grained body-part contact regions. Our proposed framework, PaIR-Net, comprises three key components: the Contact Prior Aware Module (CPAM) for identifying contact-relevant body parts, the Prior-Guided Concat Segmenter (PGCS) for pixel-wise contact segmentation, and the Interaction Inference Module (IIM) responsible for integrating global interaction relationships. To facilitate this task, we present PaIR (Part-aware Interaction Representation), a comprehensive dataset containing 13,979 images that encompass 654 actions, 80 object categories, and 17 body parts. Experimental evaluation demonstrates that PaIR-Net significantly outperforms baseline approaches, while ablation studies confirm the efficacy of each architectural component. The code and dataset will be released upon publication.", "date": "2025-08-13", "categories": []}
{"id": "2603.27103", "title": "LLM Enhanced Action Recognition via Hierarchical Global-Local Skeleton-Language Model", "abstract": "Skeleton-based human action recognition has achieved remarkable progress in recent years. However, most existing GCN-based methods rely on short-range motion topologies, which not only struggle to capture long-range joint dependencies and complex temporal dynamics but also limit cross-modal semantic alignment and understanding due to insufficient modeling of action semantics. To address these challenges, we propose a hierarchical global-local skeleton-language model (HocSLM), enabling the large action model be more representative of action semantics. First, we design a hierarchical global-local network (HGLNet) that consists of a composite-topology spatial module and a dual-path hierarchical temporal module. By synergistically integrating multi-level global and local modules, HGLNet achieves dynamically collaborative modeling at both global and local scales while preserving prior knowledge of human physical structure, significantly enhancing the model's representation of complex spatio-temporal relationships. Then, a large vision-language model (VLM) is employed to generate textual descriptions by passing the original RGB video sequences to this model, providing the rich action semantics for further training the skeleton-language model. Furthermore, we introduce a skeleton-language sequential fusion module by combining the features from HGLNet and the generated descriptions, which utilizes a skeleton-language model (SLM) for aligning skeletal spatio-temporal features and textual action descriptions precisely within a unified semantic space. The SLM model could significantly enhance the HGLNet's semantic discrimination capabilities and cross-modal understanding abilities. Extensive experiments demonstrate that the proposed HocSLM achieves the state-of-the-art performance on three mainstream benchmark datasets: NTU RGB+D 60, NTU RGB+D 120, and Northwestern-UCLA.", "date": "2026-03-28", "categories": []}
{"id": "2504.17696", "title": "Hierarchical and Multimodal Data for Daily Activity Understanding", "abstract": "Daily Activity Recordings for Artificial Intelligence (DARai, pronounced \"Dahr-ree\") is a multimodal, hierarchically annotated dataset constructed to understand human activities in real-world settings. DARai consists of continuous scripted and unscripted recordings of 50 participants in 10 different environments, totaling over 200 hours of data from 20 sensors including multiple camera views, depth and radar sensors, wearable inertial measurement units (IMUs), electromyography (EMG), insole pressure sensors, biomonitor sensors, and gaze tracker.\n  To capture the complexity in human activities, DARai is annotated at three levels of hierarchy: (i) high-level activities (L1) that are independent tasks, (ii) lower-level actions (L2) that are patterns shared between activities, and (iii) fine-grained procedures (L3) that detail the exact execution steps for actions. The dataset annotations and recordings are designed so that 22.7% of L2 actions are shared between L1 activities and 14.2% of L3 procedures are shared between L2 actions. The overlap and unscripted nature of DARai allows counterfactual activities in the dataset.\n  Experiments with various machine learning models showcase the value of DARai in uncovering important challenges in human-centered applications. Specifically, we conduct unimodal and multimodal sensor fusion experiments for recognition, temporal localization, and future action anticipation across all hierarchical annotation levels. To highlight the limitations of individual sensors, we also conduct domain-variant experiments that are enabled by DARai's multi-sensor and counterfactual activity design setup.\n  The code, documentation, and dataset are available at the dedicated DARai website: https://alregib.ece.gatech.edu/software-and-datasets/darai-daily-activity-recordings-for-artificial-intelligence-and-machine-learning/", "date": "2025-04-24", "categories": []}
{"id": "2603.26033", "title": "Knowledge is Power: Advancing Few-shot Action Recognition with Multimodal Semantics from MLLMs", "abstract": "Multimodal Large Language Models (MLLMs) have propelled the field of few-shot action recognition (FSAR). However, preliminary explorations in this area primarily focus on generating captions to form a suboptimal feature->caption->feature pipeline and adopt metric learning solely within the visual space. In this paper, we propose FSAR-LLaVA, the first end-to-end method to leverage MLLMs (such as Video-LLaVA) as a multimodal knowledge base for directly enhancing FSAR. First, at the feature level, we leverage the MLLM's multimodal decoder to extract spatiotemporally and semantically enriched representations, which are then decoupled and enhanced by our Multimodal Feature-Enhanced Module into distinct visual and textual features that fully exploit their semantic knowledge for FSAR. Next, we leverage the versatility of MLLMs to craft input prompts that flexibly adapt to diverse scenarios, and use their aligned outputs to drive our designed Composite Task-Oriented Prototype Construction, effectively bridging the distribution gap between meta-train and meta-test sets. Finally, to enable multimodal features to guide metric learning jointly, we introduce a training-free Multimodal Prototype Matching Metric that adaptively selects the most decisive cues and efficiently leverages the decoupled feature representations produced by MLLMs. Extensive experiments demonstrate superior performance across various tasks with minimal trainable parameters.", "date": "2026-03-27", "categories": []}
{"id": "2603.23988", "title": "CAKE: Real-time Action Detection via Motion Distillation and Background-aware Contrastive Learning", "abstract": "Online Action Detection (OAD) systems face two primary challenges: high computational cost and insufficient modeling of discriminative temporal dynamics against background motion. Adding optical flow could provides strong motion cues but it incurs significant computational overhead. We propose CAKE, a OAD Flow-based distillation framework to transfer motion knowledge into RGB models. We propose Dynamic Motion Adapter (DMA) to suppress static background noise and emphasize pixel changes, effectively approximating optical flow without explicit computation. The framework also integrates a Floating Contrastive Learning strategy to distinguish informative motion dynamics from temporal background. Various experiments conducted on the TVSeries, THUMOS'14, Kinetics-400 datasets show effectiveness of our model. CAKE achieves a standout mAP compared with SOTA while using the same backbone. Our model operates at over 72 FPS on a single CPU, making it highly suitable for resource-constrained systems.", "date": "2026-03-25", "categories": []}
{"id": "2603.24030", "title": "Decompose and Transfer: CoT-Prompting Enhanced Alignment for Open-Vocabulary Temporal Action Detection", "abstract": "Open-Vocabulary Temporal Action Detection (OV-TAD) aims to classify and localize action segments in untrimmed videos for unseen categories. Previous methods rely solely on global alignment between label-level semantics and visual features, which is insufficient to transfer temporal consistent visual knowledge from seen to unseen classes. To address this, we propose a Phase-wise Decomposition and Alignment (PDA) framework, which enables fine-grained action pattern learning for effective prior knowledge transfer. Specifically, we first introduce the CoT-Prompting Semantic Decomposition (CSD) module, which leverages the chain-of-thought (CoT) reasoning ability of large language models to automatically decompose action labels into coherent phase-level descriptions, emulating human cognitive processes. Then, Text-infused Foreground Filtering (TIF) module is introduced to adaptively filter action-relevant segments for each phase leveraging phase-wise semantic cues, producing semantically aligned visual representations. Furthermore, we propose the Adaptive Phase-wise Alignment (APA) module to perform phase-level visual-textual matching, and adaptively aggregates alignment results across phases for final prediction. This adaptive phase-wise alignment facilitates the capture of transferable action patterns and significantly enhances generalization to unseen actions. Extensive experiments on two OV-TAD benchmarks demonstrated the superiority of the proposed method.", "date": "2026-03-25", "categories": []}
{"id": "2603.24097", "title": "LaDy: Lagrangian-Dynamic Informed Network for Skeleton-based Action Segmentation via Spatial-Temporal Modulation", "abstract": "Skeleton-based Temporal Action Segmentation (STAS) aims to densely parse untrimmed skeletal sequences into frame-level action categories. However, existing methods, while proficient at capturing spatio-temporal kinematics, neglect the underlying physical dynamics that govern human motion. This oversight limits inter-class discriminability between actions with similar kinematics but distinct dynamic intents, and hinders precise boundary localization where dynamic force profiles shift. To address these, we propose the Lagrangian-Dynamic Informed Network (LaDy), a framework integrating principles of Lagrangian dynamics into the segmentation process. Specifically, LaDy first computes generalized coordinates from joint positions and then estimates Lagrangian terms under physical constraints to explicitly synthesize the generalized forces. To further ensure physical coherence, our Energy Consistency Loss enforces the work-energy theorem, aligning kinetic energy change with the work done by the net force. The learned dynamics then drive a Spatio-Temporal Modulation module: Spatially, generalized forces are fused with spatial representations to provide more discriminative semantics. Temporally, salient dynamic signals are constructed for temporal gating, thereby significantly enhancing boundary awareness. Experiments on challenging datasets show that LaDy achieves state-of-the-art performance, validating the integration of physical dynamics for action segmentation. Code is available at https://github.com/HaoyuJi/LaDy.", "date": "2026-03-25", "categories": []}
{"id": "2407.16344", "title": "SOAP: Enhancing Spatio-Temporal Relation and Motion Information Capturing for Few-Shot Action Recognition", "abstract": "High frame-rate (HFR) videos of action recognition improve fine-grained expression while reducing the spatio-temporal relation and motion information density. Thus, large amounts of video samples are continuously required for traditional data-driven training. However, samples are not always sufficient in real-world scenarios, promoting few-shot action recognition (FSAR) research. We observe that most recent FSAR works build spatio-temporal relation of video samples via temporal alignment after spatial feature extraction, cutting apart spatial and temporal features within samples. They also capture motion information via narrow perspectives between adjacent frames without considering density, leading to insufficient motion information capturing. Therefore, we propose a novel plug-and-play architecture for FSAR called Spatio-tempOral frAme tuPle enhancer (SOAP) in this paper. The model we designed with such architecture refers to SOAP-Net. Temporal connections between different feature channels and spatio-temporal relation of features are considered instead of simple feature extraction. Comprehensive motion information is also captured, using frame tuples with multiple frames containing more motion information than adjacent frames. Combining frame tuples of diverse frame counts further provides a broader perspective. SOAP-Net achieves new state-of-the-art performance across well-known benchmarks such as SthSthV2, Kinetics, UCF101, and HMDB51. Extensive empirical evaluations underscore the competitiveness, pluggability, generalization, and robustness of SOAP. The code is released at https://github.com/wenbohuang1002/SOAP.", "date": "2024-07-23", "categories": []}
{"id": "2412.07481", "title": "Manta: Enhancing Mamba for Few-Shot Action Recognition of Long Sub-Sequence", "abstract": "In few-shot action recognition (FSAR), long sub-sequences of video naturally express entire actions more effectively. However, the high computational complexity of mainstream Transformer-based methods limits their application. Recent Mamba demonstrates efficiency in modeling long sequences, but directly applying Mamba to FSAR overlooks the importance of local feature modeling and alignment. Moreover, long sub-sequences within the same class accumulate intra-class variance, which adversely impacts FSAR performance. To solve these challenges, we propose a Matryoshka MAmba and CoNtrasTive LeArning framework (Manta). Firstly, the Matryoshka Mamba introduces multiple Inner Modules to enhance local feature representation, rather than directly modeling global features. An Outer Module captures dependencies of timeline between these local features for implicit temporal alignment. Secondly, a hybrid contrastive learning paradigm, combining both supervised and unsupervised methods, is designed to mitigate the negative effects of intra-class variance accumulation. The Matryoshka Mamba and the hybrid contrastive learning paradigm operate in two parallel branches within Manta, enhancing Mamba for FSAR of long sub-sequence. Manta achieves new state-of-the-art performance on prominent benchmarks, including SSv2, Kinetics, UCF101, and HMDB51. Extensive empirical studies prove that Manta significantly improves FSAR of long sub-sequence from multiple perspectives.", "date": "2024-12-10", "categories": []}
{"id": "2603.09798", "title": "Test-time Ego-Exo-centric Adaptation for Action Anticipation via Multi-Label Prototype Growing and Dual-Clue Consistency", "abstract": "Efficient adaptation between Egocentric (Ego) and Exocentric (Exo) views is crucial for applications such as human-robot cooperation. However, the success of most existing Ego-Exo adaptation methods relies heavily on target-view data for training, thereby increasing computational and data collection costs. In this paper, we make the first exploration of a Test-time Ego-Exo Adaptation for Action Anticipation (TE$^{2}$A$^{3}$) task, which aims to adjust the source-view-trained model online during test time to anticipate target-view actions. It is challenging for existing Test-Time Adaptation (TTA) methods to address this task due to the multi-action candidates and significant temporal-spatial inter-view gap. Hence, we propose a novel Dual-Clue enhanced Prototype Growing Network (DCPGN), which accumulates multi-label knowledge and integrates cross-modality clues for effective test-time Ego-Exo adaptation and action anticipation. Specifically, we propose a Multi-Label Prototype Growing Module (ML-PGM) to balance multiple positive classes via multi-label assignment and confidence-based reweighting for class-wise memory banks, which are updated by an entropy priority queue strategy. Then, the Dual-Clue Consistency Module (DCCM) introduces a lightweight narrator to generate textual clues indicating action progressions, which complement the visual clues containing various objects. Moreover, we constrain the inferred textual and visual logits to construct dual-clue consistency for temporally and spatially bridging Ego and Exo views. Extensive experiments on the newly proposed EgoMe-anti and the existing EgoExoLearn benchmarks show the effectiveness of our method, which outperforms related state-of-the-art methods by a large margin. Code is available at \\href{https://github.com/ZhaofengSHI/DCPGN}{https://github.com/ZhaofengSHI/DCPGN}.", "date": "2026-03-10", "categories": []}
{"id": "2506.02845", "title": "Go Beyond Earth: Understanding Human Actions and Scenes in Microgravity Environments", "abstract": "Despite substantial progress in video understanding, most existing datasets are limited to Earth's gravitational conditions. However, microgravity alters human motion, interactions, and visual semantics, revealing a critical gap for real-world vision systems. This presents a challenge for domain-robust video understanding in safety-critical space applications. To address this, we introduce MicroG-4M, the first benchmark for spatio-temporal and semantic understanding of human activities in microgravity. Constructed from real-world space missions and cinematic simulations, the dataset includes 4,759 clips covering 50 actions, 1,238 context-rich captions, and over 7,000 question-answer pairs on astronaut activities and scene understanding. MicroG-4M supports three core tasks: fine-grained multi-label action recognition, temporal video captioning, and visual question answering, enabling a comprehensive evaluation of both spatial localization and semantic reasoning in microgravity contexts. We establish baselines using state-of-the-art models. All data, annotations, and code are available at https://github.com/LEI-QI-233/HAR-in-Space.", "date": "2025-06-03", "categories": []}
{"id": "2603.21048", "title": "A Two-stage Transformer Framework for Temporal Localization of Distracted Driver Behaviors", "abstract": "The identification of hazardous driving behaviors from in-cabin video streams is essential for enhancing road safety and supporting the detection of traffic violations and unsafe driver actions. However, current temporal action localization techniques often struggle to balance accuracy with computational efficiency. In this work, we develop and evaluate a temporal action localization framework tailored for driver monitoring scenarios, particularly suitable for periodic inspection settings such as transportation safety checkpoints or fleet management assessment systems. Our approach follows a two-stage pipeline that combines VideoMAE-based feature extraction with an Augmented Self-Mask Attention (AMA) detector, enhanced by a Spatial Pyramid Pooling-Fast (SPPF) module to capture multi-scale temporal features. Experimental results reveal a distinct trade-off between model capacity and efficiency. At the feature extraction stage, the ViT-Giant backbone delivers higher representations with 88.09% Top-1 test accuracy, while the ViT-based variant proves to be a practical alternative, achieving 82.55% accuracy with significantly lower computational fine-tuning costs (101.85 GFLOPs/segment compared to 1584.06 GFLOPs/segment for Giant). In the downstream localization task, the integration of SPPF consistently improves performance across all configurations. Notably, the ViT-Giant + SPPF model achieves a peak mAP of 92.67%, while the lightweight ViT-based configuration maintains robust results.", "date": "2026-03-22", "categories": []}
{"id": "2603.21305", "title": "Privacy-Preserving Federated Action Recognition via Differentially Private Selective Tuning and Efficient Communication", "abstract": "Federated video action recognition enables collaborative model training without sharing raw video data, yet remains vulnerable to two key challenges: \\textit{model exposure} and \\textit{communication overhead}. Gradients exchanged between clients and the server can leak private motion patterns, while full-model synchronization of high-dimensional video networks causes significant bandwidth and communication costs. To address these issues, we propose \\textit{Federated Differential Privacy with Selective Tuning and Efficient Communication for Action Recognition}, namely \\textit{FedDP-STECAR}. Our \\textit{FedDP-STECAR} framework selectively fine-tunes and perturbs only a small subset of task-relevant layers under Differential Privacy (DP), reducing the surface of information leakage while preserving temporal coherence in video features. By transmitting only the tuned layers during aggregation, communication traffic is reduced by over 99\\% compared to full-model updates. Experiments on the UCF-101 dataset using the MViT-B-16x4 transformer show that \\textit{FedDP-STECAR} achieves up to \\textbf{70.2\\% higher accuracy} under strict privacy ($\\epsilon=0.65$) in centralized settings and \\textbf{48\\% faster training} with \\textbf{73.1\\% accuracy} in federated setups, enabling scalable and privacy-preserving video action recognition. Code available at https://github.com/izakariyya/mvit-federated-videodp", "date": "2026-03-22", "categories": []}
{"id": "2511.06741", "title": "Otter: Mitigating Background Distractions of Wide-Angle Few-Shot Action Recognition with Enhanced RWKV", "abstract": "Wide-angle videos in few-shot action recognition (FSAR) effectively express actions within specific scenarios. However, without a global understanding of both subjects and background, recognizing actions in such samples remains challenging because of the background distractions. Receptance Weighted Key Value (RWKV), which learns interaction between various dimensions, shows promise for global modeling. While directly applying RWKV to wide-angle FSAR may fail to highlight subjects due to excessive background information. Additionally, temporal relation degraded by frames with similar backgrounds is difficult to reconstruct, further impacting performance. Therefore, we design the CompOund SegmenTation and Temporal REconstructing RWKV (Otter). Specifically, the Compound Segmentation Module~(CSM) is devised to segment and emphasize key patches in each frame, effectively highlighting subjects against background information. The Temporal Reconstruction Module (TRM) is incorporated into the temporal-enhanced prototype construction to enable bidirectional scanning, allowing better reconstruct temporal relation. Furthermore, a regular prototype is combined with the temporal-enhanced prototype to simultaneously enhance subject emphasis and temporal modeling, improving wide-angle FSAR performance. Extensive experiments on benchmarks such as SSv2, Kinetics, UCF101, and HMDB51 demonstrate that Otter achieves state-of-the-art performance. Extra evaluation on the VideoBadminton dataset further validates the superiority of Otter in wide-angle FSAR.", "date": "2025-11-10", "categories": []}
{"id": "2510.16988", "title": "CARE: Contrastive Alignment for ADL Recognition from Event-Triggered Sensor Streams", "abstract": "The recognition of Activities of Daily Living (ADLs) from event-triggered ambient sensors is an essential task in Ambient Assisted Living, yet existing methods remain constrained by representation-level limitations. Sequence-based approaches preserve temporal order of sensor activations but are sensitive to noise and lack spatial awareness, while image-based approaches capture global patterns and implicit spatial correlations but compress fine-grained temporal dynamics and distort sensor layouts. Naive fusion (e.g., feature concatenation) fails to enforce alignment between sequence- and image-based representation views, underutilizing their complementary strengths. We propose Contrastive Alignment for ADL Recognition from Event-Triggered Sensor Streams (CARE), an end-to-end framework that jointly optimizes representation learning via Sequence-Image Contrastive Alignment (SICA) and classification via cross-entropy, ensuring both cross-representation alignment and task-specific discriminability. CARE integrates (i) time-aware, noise-resilient sequence encoding with (ii) spatially-informed and frequency-sensitive image representations, and employs (iii) a joint contrastive-classification objective for end-to-end learning of aligned and discriminative embeddings. Evaluated on three CASAS datasets, CARE achieves state-of-the-art performance (89.8% on Milan, 88.9% on Cairo, and 73.3% on Kyoto7) and demonstrates robustness to sensor malfunctions and layout variability, highlighting its potential for reliable ADL recognition in smart homes. We release our code at https://github.com/Jhziiiig/CARE.", "date": "2025-10-19", "categories": []}
{"id": "2603.13410", "title": "Bridging the Visual-to-Physical Gap: Physically Aligned Representations for Fall Risk Analysis", "abstract": "Vision-based fall analysis has advanced rapidly, but a key bottleneck remains: visually similarmotions can correspond to very different physical outcomes because small differences in contactmechanics and protective responses are hard to infer from appearance alone. Most existingapproaches handle this by supervised injury prediction, which depends on reliable injury labels.In practice, such labels are difficult to obtain: video evidence is often ambiguous (occlusion,viewpoint limits), and true injury events are rare and cannot be safely staged, leading to noisysupervision. We address this problem with PHARL (PHysics-aware Alignment RepresentationLearning), which learns physically meaningful fall representations without requiring clinicaloutcome labels. PHARL regularizes motion embeddings with two complementary constraints:(1) trajectory-level temporal consistency for stable representation learning, and (2) multi-classphysics alignment, where simulation-derived contact outcomes shape embedding geometry. Bypairing video windows with temporally aligned simulation descriptors, PHARL captures localimpact-relevant dynamics while keeping inference purely feed-forward. Experiments on fourpublic datasets show that PHARL consistently improves risk-aligned representation quality overvisual-only baselines while maintaining strong fall-detection performance. Notably, PHARL alsoexhibits zero-shot ordinality: an interpretable severity structure (Head > Trunk > Supported)emerges without explicit ordinal supervision.", "date": "2026-03-12", "categories": []}
{"id": "2603.12764", "title": "SAVA-X: Ego-to-Exo Imitation Error Detection via Scene-Adaptive View Alignment and Bidirectional Cross View Fusion", "abstract": "Error detection is crucial in industrial training, healthcare, and assembly quality control. Most existing work assumes a single-view setting and cannot handle the practical case where a third-person (exo) demonstration is used to assess a first-person (ego) imitation. We formalize Ego$\\rightarrow$Exo Imitation Error Detection: given asynchronous, length-mismatched ego and exo videos, the model must localize procedural steps on the ego timeline and decide whether each is erroneous. This setting introduces cross-view domain shift, temporal misalignment, and heavy redundancy. Under a unified protocol, we adapt strong baselines from dense video captioning and temporal action detection and show that they struggle in this cross-view regime. We then propose SAVA-X, an Align-Fuse-Detect framework with (i) view-conditioned adaptive sampling, (ii) scene-adaptive view embeddings, and (iii) bidirectional cross-attention fusion. On the EgoMe benchmark, SAVA-X consistently improves AUPRC and mean tIoU over all baselines, and ablations confirm the complementary benefits of its components. Code is available at https://github.com/jack1ee/SAVAX.", "date": "2026-03-13", "categories": []}
{"id": "2603.12787", "title": "Generalized Recognition of Basic Surgical Actions Enables Skill Assessment and Vision-Language-Model-based Surgical Planning", "abstract": "Artificial intelligence, imaging, and large language models have the potential to transform surgical practice, training, and automation. Understanding and modeling of basic surgical actions (BSA), the fundamental unit of operation in any surgery, is important to drive the evolution of this field. In this paper, we present a BSA dataset comprising 10 basic actions across 6 surgical specialties with over 11,000 video clips, which is the largest to date. Based on the BSA dataset, we developed a new foundation model that conducts general-purpose recognition of basic actions. Our approach demonstrates robust cross-specialist performance in experiments validated on datasets from different procedural types and various body parts. Furthermore, we demonstrate downstream applications enabled by the BAS foundation model through surgical skill assessment in prostatectomy using domain-specific knowledge, and action planning in cholecystectomy and nephrectomy using large vision-language models. Multinational surgeons' evaluation of the language model's output of the action planning explainable texts demonstrated clinical relevance. These findings indicate that basic surgical actions can be robustly recognized across scenarios, and an accurate BSA understanding model can essentially facilitate complex applications and speed up the realization of surgical superintelligence.", "date": "2026-03-13", "categories": []}
{"id": "2511.22433", "title": "SkeletonAgent: An Agentic Interaction Framework for Skeleton-based Action Recognition", "abstract": "Recent advances in skeleton-based action recognition increasingly leverage semantic priors from Large Language Models (LLMs) to enrich skeletal representations. However, the LLM is typically queried in isolation from the recognition model and receives no performance feedback. As a result, it often fails to deliver the targeted discriminative cues critical to distinguish similar actions. To overcome these limitations, we propose SkeletonAgent, a novel framework that bridges the recognition model and the LLM through two cooperative agents, i.e., Questioner and Selector. Specifically, the Questioner identifies the most frequently confused classes and supplies them to the LLM as context for more targeted guidance. Conversely, the Selector parses the LLM's response to extract precise joint-level constraints and feeds them back to the recognizer, enabling finer-grained cross-modal alignment. Comprehensive evaluations on five benchmarks, including NTU RGB+D, NTU RGB+D 120, Kinetics-Skeleton, FineGYM, and UAV-Human, demonstrate that SkeletonAgent consistently outperforms state-of-the-art benchmark methods. The code is available at https://github.com/firework8/SkeletonAgent.", "date": "2025-11-27", "categories": []}
{"id": "2603.10648", "title": "Less is More: Decoder-Free Masked Modeling for Efficient Skeleton Representation Learning", "abstract": "The landscape of skeleton-based action representation learning has evolved from Contrastive Learning (CL) to Masked Auto-Encoder (MAE) architectures. However, each paradigm faces inherent limitations: CL often overlooks fine-grained local details, while MAE is burdened by computationally heavy decoders. Moreover, MAE suffers from severe computational asymmetry -- benefiting from efficient masking during pre-training but requiring exhaustive full-sequence processing for downstream tasks. To resolve these bottlenecks, we propose SLiM (Skeleton Less is More), a novel unified framework that harmonizes masked modeling with contrastive learning via a shared encoder. By eschewing the reconstruction decoder, SLiM not only eliminates computational redundancy but also compels the encoder to capture discriminative features directly. SLiM is the first framework with decoder-free masked modeling of representative learning. Crucially, to prevent trivial reconstruction arising from high skeletal-temporal correlation, we introduce semantic tube masking, alongside skeletal-aware augmentations designed to ensure anatomical consistency across diverse temporal granularities. Extensive experiments demonstrate that SLiM consistently achieves state-of-the-art performance across all downstream protocols. Notably, our method delivers this superior accuracy with exceptional efficiency, reducing inference computational cost by 7.89x compared to existing MAE methods.", "date": "2026-03-11", "categories": []}
{"id": "2512.21064", "title": "Multimodal Skeleton-Based Action Representation Learning via Decomposition and Composition", "abstract": "Multimodal human action understanding is a significant problem in computer vision, with the central challenge being the effective utilization of the complementarity among diverse modalities while maintaining model efficiency. However, most existing methods rely on simple late fusion to enhance performance, which results in substantial computational overhead. Although early fusion with a shared backbone for all modalities is efficient, it struggles to achieve excellent performance. To address the dilemma of balancing efficiency and effectiveness, we introduce a self-supervised multimodal skeleton-based action representation learning framework, named Decomposition and Composition. The Decomposition strategy meticulously decomposes the fused multimodal features into distinct unimodal features, subsequently aligning them with their respective ground truth unimodal counterparts. On the other hand, the Composition strategy integrates multiple unimodal features, leveraging them as self-supervised guidance to enhance the learning of multimodal representations. Extensive experiments on the NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD II datasets demonstrate that the proposed method strikes an excellent balance between computational cost and model performance.", "date": "2025-12-24", "categories": []}
{"id": "2603.08289", "title": "Novel Semantic Prompting for Zero-Shot Action Recognition", "abstract": "Zero-shot action recognition relies on transferring knowledge from vision-language models to unseen actions using semantic descriptions. While recent methods focus on temporal modeling or architectural adaptations to handle video data, we argue that semantic prompting alone provides a strong and underexplored signal for zero-shot action understanding. We introduce SP-CLIP, a lightweight framework that augments frozen vision-language models with structured semantic prompts describing actions at multiple levels of abstraction, such as intent, motion, and object interaction. Without modifying the visual encoder or learning additional parameters, SP-CLIP aligns video representations with enriched textual semantics through prompt aggregation and consistency scoring. Experiments across standard benchmarks show that semantic prompting substantially improves zero-shot action recognition, particularly for fine-grained and compositional actions, while preserving the efficiency and generalization of pretrained models.", "date": "2026-03-09", "categories": []}
{"id": "2603.08317", "title": "Human-AI Divergence in Ego-centric Action Recognition under Spatial and Spatiotemporal Manipulations", "abstract": "Humans consistently outperform state-of-the-art AI models in action recognition, particularly in challenging real-world conditions involving low resolution, occlusion, and visual clutter. Understanding the sources of this performance gap is essential for developing more robust and human-aligned models. In this paper, we present a large-scale human-AI comparative study of egocentric action recognition using Minimal Identifiable Recognition Crops (MIRCs), defined as the smallest spatial or spatiotemporal regions sufficient for reliable human recognition. We used our previously introduced, Epic ReduAct, a systematically spatially reduced and temporally scrambled dataset derived from 36 EPIC KITCHENS videos, spanning multiple spatial reduction levels and temporal conditions. Recognition performance is evaluated using over 3,000 human participants and the Side4Video model. Our analysis combines quantitative metrics, Average Reduction Rate and Recognition Gap, with qualitative analyses of spatial (high-, mid-, and low-level visual features) and spatiotemporal factors, including a categorisation of actions into Low Temporal Actions (LTA) and High Temporal Actions (HTA). Results show that human performance exhibits sharp declines when transitioning from MIRCs to subMIRCs, reflecting a strong reliance on sparse, semantically critical cues such as hand-object interactions. In contrast, the model degrades more gradually and often relies on contextual and mid- to low-level features, sometimes even exhibiting increased confidence under spatial reduction. Temporally, humans remain robust to scrambling when key spatial cues are preserved, whereas the model often shows insensitivity to temporal disruption, revealing class-dependent temporal sensitivities.", "date": "2026-03-09", "categories": []}
{"id": "2409.18300", "title": "FALCON: Future-Aware Learning with Contextual Object-Centric Pretraining for UAV Action Recognition", "abstract": "We introduce FALCON, a unified self-supervised video pretraining approach for UAV action recognition from raw RGB aerial footage, requiring no additional preprocessing at inference. UAV videos exhibit severe spatial imbalance: large, cluttered backgrounds dominate the field of view, causing reconstruction-based pretraining to waste capacity on uninformative regions and under-learn action-relevant human/object cues. FALCON addresses this by integrating object-aware masked autoencoding with object-centric dual-horizon future reconstruction. Using detections only during pretraining, we construct objectness priors that (i) enforce balanced token visibility during masking and (ii) concentrate reconstruction supervision on action-relevant regions, preventing learning from being dominated by background appearance. To promote temporal dynamics learning, we further reconstruct short- and long-horizon future content within an object-centric supervision region, injecting anticipatory temporal supervision that is robust to noisy aerial context. Across UAV benchmarks, FALCON improves top-1 accuracy by 2.9\\% on NEC-Drone and 5.8\\% on UAV-Human with a ViT-B backbone, while achieving 2$\\times$--5$\\times$ faster inference than supervised approaches that rely on heavy test-time augmentation.", "date": "2024-09-26", "categories": []}
{"id": "2511.17929", "title": "MambaTAD: When State-Space Models Meet Long-Range Temporal Action Detection", "abstract": "Temporal Action Detection (TAD) aims to identify and localize actions by determining their starting and ending frames within untrimmed videos. Recent Structured State-Space Models such as Mamba have demonstrated potential in TAD due to their long-range modeling capability and linear computational complexity. On the other hand, structured state-space models often face two key challenges in TAD, namely, decay of temporal context due to recursive processing and self-element conflict during global visual context modeling, which become more severe while handling long-span action instances. Additionally, traditional methods for TAD struggle with detecting long-span action instances due to a lack of global awareness and inefficient detection heads. This paper presents MambaTAD, a new state-space TAD model that introduces long-range modeling and global feature detection capabilities for accurate temporal action detection. MambaTAD comprises two novel designs that complement each other with superior TAD performance. First, it introduces a Diagonal-Masked Bidirectional State-Space (DMBSS) module which effectively facilitates global feature fusion and temporal action detection. Second, it introduces a global feature fusion head that refines the detection progressively with multi-granularity features and global awareness. In addition, MambaTAD tackles TAD in an end-to-end one-stage manner using a new state-space temporal adapter(SSTA) which reduces network parameters and computation cost with linear complexity. Extensive experiments show that MambaTAD achieves superior TAD performance consistently across multiple public benchmarks.", "date": "2025-11-22", "categories": []}
{"id": "2603.04509", "title": "Recognition of Daily Activities through Multi-Modal Deep Learning: A Video, Pose, and Object-Aware Approach for Ambient Assisted Living", "abstract": "Recognition of daily activities is a critical element for effective Ambient Assisted Living (AAL) systems, particularly to monitor the well-being and support the independence of older adults in indoor environments. However, developing robust activity recognition systems faces significant challenges, including intra-class variability, inter-class similarity, environmental variability, camera perspectives, and scene complexity. This paper presents a multi-modal approach for the recognition of activities of daily living tailored for older adults within AAL settings. The proposed system integrates visual information processed by a 3D Convolutional Neural Network (CNN) with 3D human pose data analyzed by a Graph Convolutional Network. Contextual information, derived from an object detection module, is fused with the 3D CNN features using a cross-attention mechanism to enhance recognition accuracy. This method is evaluated using the Toyota SmartHome dataset, which consists of real-world indoor activities. The results indicate that the proposed system achieves competitive classification accuracy for a range of daily activities, highlighting its potential as an essential component for advanced AAL monitoring solutions. This advancement supports the broader goal of developing intelligent systems that promote safety and autonomy among older adults.", "date": "2026-03-04", "categories": []}
{"id": "2603.02390", "title": "OpenMarcie: Dataset for Multimodal Action Recognition in Industrial Environments", "abstract": "Smart factories use advanced technologies to optimize production and increase efficiency. To this end, the recognition of worker activity allows for accurate quantification of performance metrics, improving efficiency holistically while contributing to worker safety. OpenMarcie is, to the best of our knowledge, the biggest multimodal dataset designed for human action monitoring in manufacturing environments. It includes data from wearables sensing modalities and cameras distributed in the surroundings. The dataset is structured around two experimental settings, involving a total of 36 participants. In the first setting, twelve participants perform a bicycle assembly and disassembly task under semi-realistic conditions without a fixed protocol, promoting divergent and goal-oriented problem-solving. The second experiment involves twenty-five volunteers (24 valid data) engaged in a 3D printer assembly task, with the 3D printer manufacturer's instructions provided to guide the volunteers in acquiring procedural knowledge. This setting also includes sequential collaborative assembly, where participants assess and correct each other's progress, reflecting real-world manufacturing dynamics. OpenMarcie includes over 37 hours of egocentric and exocentric, multimodal, and multipositional data, featuring eight distinct data types and more than 200 independent information channels. The dataset is benchmarked across three human activity recognition tasks: activity classification, open vocabulary captioning, and cross-modal alignment.", "date": "2026-03-02", "categories": []}
{"id": "2603.02477", "title": "E2E-GNet: An End-to-End Skeleton-based Geometric Deep Neural Network for Human Motion Recognition", "abstract": "Geometric deep learning has recently gained significant attention in the computer vision community for its ability to capture meaningful representations of data lying in a non-Euclidean space. To this end, we propose E2E-GNet, an end-to-end geometric deep neural network for skeleton-based human motion recognition. To enhance the discriminative power between different motions in the non-Euclidean space, E2E-GNet introduces a geometric transformation layer that jointly optimizes skeleton motion sequences on this space and applies a differentiable logarithm map activation to project them onto a linear space. Building on this, we further design a distortion-aware optimization layer that limits skeleton shape distortions caused by this projection, enabling the network to retain discriminative geometric cues and achieve a higher motion recognition rate. We demonstrate the impact of each layer through ablation studies and extensive experiments across five datasets spanning three domains show that E2E-GNet outperforms other methods with lower cost.", "date": "2026-03-03", "categories": []}
{"id": "2603.02546", "title": "On Discriminative vs. Generative classifiers: Rethinking MLLMs for Action Understanding", "abstract": "Multimodal Large Language Models (MLLMs) have advanced open-world action understanding and can be adapted as generative classifiers for closed-set settings by autoregressively generating action labels as text. However, this approach is inefficient, and shared subwords across action labels introduce semantic overlap, leading to ambiguity in generation. In contrast, discriminative classifiers learn task-specific representations with clear decision boundaries, enabling efficient one-step classification without autoregressive decoding. We first compare generative and discriminative classifiers with MLLMs for closed-set action understanding, revealing the superior accuracy and efficiency of the latter. To bridge the performance gap, we design strategies that elevate generative classifiers toward performance comparable with discriminative ones. Furthermore, we show that generative modeling can complement discriminative classifiers, leading to better performance while preserving efficiency. To this end, we propose Generation-Assisted Discriminative~(GAD) classifier for closed-set action understanding. GAD operates only during fine-tuning, preserving full compatibility with MLLM pretraining. Extensive experiments on temporal action understanding benchmarks demonstrate that GAD improves both accuracy and efficiency over generative methods, achieving state-of-the-art results on four tasks across five datasets, including an average 2.5% accuracy gain and 3x faster inference on our largest COIN benchmark.", "date": "2026-03-03", "categories": []}
{"id": "2407.13750", "title": "PO-GUISE+: Pose and object guided transformer token selection for efficient driver action recognition", "abstract": "We address the task of identifying distracted driving by analyzing in-car videos using efficient transformers. Although transformer models have achieved outstanding performance in human action recognition tasks, their high computational costs limit their application onboard a vehicle. We introduce POGUISE+, a multi-task video transformer that, given an input clip, predicts the distracted driving action, the driver's pose, and the interacting object. Our enhanced features for token selection are specifically adapted to driver actions by leveraging information about object interaction and the driver's pose. With POGUISE+, we significantly reduce the model's computational demands while maintaining or improving baseline accuracy across various computational budgets. Additionally, to evaluate our model's performance in real-world scenarios, we have developed benchmarks on a Jetson computing platform, demonstrating its effectiveness across different configurations and computational budgets. Our model outperforms current state-of-the-art results on the Drive&Act, 100-Driver, and 3MDAD datasets, while having superior efficiency compared to existing video transformer-based methods.", "date": "2024-07-18", "categories": []}
{"id": "2603.00878", "title": "MMTA: Multi Membership Temporal Attention for Fine-Grained Stroke Rehabilitation Assessment", "abstract": "To empower the iterative assessments involved during a person's rehabilitation, automated assessment of a person's abilities during daily activities requires temporally precise segmentation of fine-grained actions in therapy videos. Existing temporal action segmentation (TAS) models struggle to capture sub-second micro-movements while retaining exercise context, blurring rapid phase transitions and limiting reliable downstream assessment of motor recovery. We introduce Multi-Membership Temporal Attention (MMTA), a high-resolution temporal transformer for fine-grained rehabilitation assessment. Unlike standard temporal attention, which assigns each frame a single attention context per layer, MMTA lets each frame attend to multiple locally normalized temporal attention windows within the same layer. We fuse these concurrent temporal views via feature-space overlap resolution, preserving competing local contexts near transitions while enabling longer-range reasoning through layer-wise propagation. This increases boundary sensitivity without additional depth or multi-stage refinement. MMTA supports both video and wearable IMU inputs within a unified single-stage architecture, making it applicable to both clinical and home settings. MMTA consistently improves over the Global Attention transformer, boosting Edit Score by +1.3 (Video) and +1.6 (IMU) on StrokeRehab while further improving 50Salads by +3.3. Ablations confirm that performance gains stem from multi-membership temporal views rather than architectural complexity, offering a practical solution for resource-constrained rehabilitation assessment.", "date": "2026-03-01", "categories": []}
{"id": "2504.08578", "title": "Multimodal Knowledge Distillation for Egocentric Action Recognition Robust to Missing Modalities", "abstract": "Egocentric action recognition enables robots to facilitate human-robot interactions and monitor task progress. Existing methods often rely solely on RGB videos, although additional modalities, such as audio, can improve accuracy under challenging conditions. However, most multimodal approaches assume that all modalities are available at inference time, leading to significant accuracy drops, or even failure, when inputs are missing. To address this limitation, we introduce KARMMA, a multimodal Knowledge distillation framework for egocentric Action Recognition robust to Missing ModAlities that does not require modality alignment across all samples during training or inference. KARMMA distills knowledge from a multimodal teacher into a multimodal student that leverages all available modalities while remaining robust to missing ones, enabling deployment across diverse sensor configurations without retraining. Our student uses approximately 50% fewer computational resources than the teacher, resulting in a lightweight and fast model that is well suited for on-robot deployment. Experiments on Epic-Kitchens and Something-Something demonstrate that our student achieves competitive accuracy while significantly reducing performance degradation under missing modality conditions.", "date": "2025-04-11", "categories": []}
{"id": "2602.24275", "title": "Hierarchical Action Learning for Weakly-Supervised Action Segmentation", "abstract": "Humans perceive actions through key transitions that structure actions across multiple abstraction levels, whereas machines, relying on visual features, tend to over-segment. This highlights the difficulty of enabling hierarchical reasoning in video understanding. Interestingly, we observe that lower-level visual and high-level action latent variables evolve at different rates, with low-level visual variables changing rapidly, while high-level action variables evolve more slowly, making them easier to identify. Building on this insight, we propose the Hierarchical Action Learning (\\textbf{HAL}) model for weakly-supervised action segmentation. Our approach introduces a hierarchical causal data generation process, where high-level latent action governs the dynamics of low-level visual features. To model these varying timescales effectively, we introduce deterministic processes to align these latent variables over time. The \\textbf{HAL} model employs a hierarchical pyramid transformer to capture both visual features and latent variables, and a sparse transition constraint is applied to enforce the slower dynamics of high-level action variables. This mechanism enhances the identification of these latent variables over time. Under mild assumptions, we prove that these latent action variables are strictly identifiable. Experimental results on several benchmarks show that the \\textbf{HAL} model significantly outperforms existing methods for weakly-supervised action segmentation, confirming its practical effectiveness in real-world applications.", "date": "2026-02-27", "categories": []}
{"id": "2602.21406", "title": "Exploring Vision-Language Models for Open-Vocabulary Zero-Shot Action Segmentation", "abstract": "Temporal Action Segmentation (TAS) requires dividing videos into action segments, yet the vast space of activities and alternative breakdowns makes collecting comprehensive datasets infeasible. Existing methods remain limited to closed vocabularies and fixed label sets. In this work, we explore the largely unexplored problem of Open-Vocabulary Zero-Shot Temporal Action Segmentation (OVTAS) by leveraging the strong zero-shot capabilities of Vision-Language Models (VLMs). We introduce a training-free pipeline that follows a segmentation-by-classification design: Frame-Action Embedding Similarity (FAES) matches video frames to candidate action labels, and Similarity-Matrix Temporal Segmentation (SMTS) enforces temporal consistency. Beyond proposing OVTAS, we present a systematic study across 14 diverse VLMs, providing the first broad analysis of their suitability for open-vocabulary action segmentation. Experiments on standard benchmarks show that OVTAS achieves strong results without task-specific supervision, underscoring the potential of VLMs for structured temporal understanding.", "date": "2026-02-24", "categories": []}
{"id": "2601.16694", "title": "Affinity Contrastive Learning for Skeleton-based Human Activity Understanding", "abstract": "In skeleton-based human activity understanding, existing methods often adopt the contrastive learning paradigm to construct a discriminative feature space. However, many of these approaches fail to exploit the structural inter-class similarities and overlook the impact of anomalous positive samples. In this study, we introduce ACLNet, an Affinity Contrastive Learning Network that explores the intricate clustering relationships among human activity classes to improve feature discrimination. Specifically, we propose an affinity metric to refine similarity measurements, thereby forming activity superclasses that provide more informative contrastive signals. A dynamic temperature schedule is also introduced to adaptively adjust the penalty strength for various superclasses. In addition, we employ a margin-based contrastive strategy to improve the separation of hard positive and negative samples within classes. Extensive experiments on NTU RGB+D 60, NTU RGB+D 120, Kinetics-Skeleton, PKU-MMD, FineGYM, and CASIA-B demonstrate the superiority of our method in skeleton-based action recognition, gait recognition, and person re-identification. The source code is available at https://github.com/firework8/ACLNet.", "date": "2026-01-23", "categories": []}
{"id": "2503.24298", "title": "Order Matters: On Parameter-Efficient Image-to-Video Probing for Recognizing Nearly Symmetric Actions", "abstract": "Fine-grained understanding of human actions is essential for safe and intuitive human--robot interaction. We study the challenge of recognizing nearly symmetric actions, such as picking up vs. placing down a tool or opening vs. closing a drawer. These actions are common in close human-robot collaboration, yet they are rare and largely overlooked in mainstream vision frameworks. Pretrained vision foundation models (VFMs) are often adapted using probing, valued in robotics for its efficiency and low data needs, or parameter-efficient fine-tuning (PEFT), which adds temporal modeling through adapters or prompts. However, our analysis shows that probing is permutation-invariant and blind to frame order, while PEFT is prone to overfitting on smaller HRI datasets, and less practical in real-world robotics due to compute constraints.\n  To address this, we introduce STEP (Self-attentive Temporal Embedding Probing), a lightweight extension to probing that models temporal order via frame-wise positional encodings, a global CLS token, and a simplified attention block. Compared to conventional probing, STEP improves accuracy by 4--10% on nearly symmetric actions and 6--15% overall across action recognition benchmarks in human-robot-interaction, industrial assembly, and driver assistance. Beyond probing, STEP surpasses heavier PEFT methods and even outperforms fully fine-tuned models on all three benchmarks, establishing a new state-of-the-art. Code and models will be made publicly available: https://github.com/th-nesh/STEP.", "date": "2025-03-31", "categories": []}
{"id": "2510.10682", "title": "Action-Dynamics Modeling and Cross-Temporal Interaction for Online Action Understanding", "abstract": "Action understanding, encompassing action detection and anticipation, plays a crucial role in numerous practical applications. However, untrimmed videos are often characterized by substantial redundant information and noise. Moreover, in modeling action understanding, the influence of the agent's intention on the action is often overlooked. Motivated by these issues, we propose a novel framework called the State-Specific Model (SSM), designed to unify and enhance both action detection and anticipation tasks. In the proposed framework, the Critical State-Based Memory Compression module compresses frame sequences into critical states, reducing information redundancy. The Action Pattern Learning module constructs a state-transition graph with multi-dimensional edges to model action dynamics in complex scenarios, on the basis of which potential future cues can be generated to represent intention. Furthermore, our Cross-Temporal Interaction module models the mutual influence between intentions and past as well as current information through cross-temporal interactions, thereby refining present and future features and ultimately realizing simultaneous action detection and anticipation. Extensive experiments on multiple benchmark datasets -- including EPIC-Kitchens-100, THUMOS'14, TVSeries, and the introduced Parkinson's Disease Mouse Behaviour (PDMB) dataset -- demonstrate the superior performance of our proposed framework compared to other state-of-the-art approaches. These results highlight the importance of action dynamics learning and cross-temporal interactions, laying a foundation for future action understanding research.", "date": "2025-10-12", "categories": []}
{"id": "2511.03665", "title": "A Lightweight 3D-CNN for Event-Based Human Action Recognition with Privacy-Preserving Potential", "abstract": "This paper presents a lightweight three-dimensional convolutional neural network (3DCNN) for human activity recognition (HAR) using event-based vision data. Privacy preservation is a key challenge in human monitoring systems, as conventional frame-based cameras capture identifiable personal information. In contrast, event cameras record only changes in pixel intensity, providing an inherently privacy-preserving sensing modality. The proposed network effectively models both spatial and temporal dynamics while maintaining a compact design suitable for edge deployment. To address class imbalance and enhance generalization, focal loss with class reweighting and targeted data augmentation strategies are employed. The model is trained and evaluated on a composite dataset derived from the Toyota Smart Home and ETRI datasets. Experimental results demonstrate an F1-score of 0.9415 and an overall accuracy of 94.17%, outperforming benchmark 3D-CNN architectures such as C3D, ResNet3D, and MC3_18 by up to 3%. These results highlight the potential of event-based deep learning for developing accurate, efficient, and privacy-aware human action recognition systems suitable for real-world edge applications.", "date": "2025-11-05", "categories": []}
{"id": "2602.18977", "title": "Frame2Freq: Spectral Adapters for Fine-Grained Video Understanding", "abstract": "Adapting image-pretrained backbones to video typically relies on time-domain adapters tuned to a single temporal scale. Our experiments show that these modules pick up static image cues and very fast flicker changes, while overlooking medium-speed motion. Capturing dynamics across multiple time-scales is, however, crucial for fine-grained temporal analysis (i.e., opening vs. closing bottle).\n  To address this, we introduce Frame2Freq -- a family of frequency-aware adapters that perform spectral encoding during image-to-video adaptation of pretrained Vision Foundation Models (VFMs), improving fine-grained action recognition. Frame2Freq uses Fast Fourier Transform (FFT) along time and learns frequency-band specific embeddings that adaptively highlight the most discriminative frequency ranges. Across five fine-grained activity recognition datasets, Frame2Freq outperforms prior PEFT methods and even surpasses fully fine-tuned models on four of them. These results provide encouraging evidence that frequency analysis methods are a powerful tool for modeling temporal dynamics in image-to-video transfer. Code is available at https://github.com/th-nesh/Frame2Freq.", "date": "2026-02-21", "categories": []}
{"id": "2602.18043", "title": "Spatio-temporal Decoupled Knowledge Compensator for Few-Shot Action Recognition", "abstract": "Few-Shot Action Recognition (FSAR) is a challenging task that requires recognizing novel action categories with a few labeled videos. Recent works typically apply semantically coarse category names as auxiliary contexts to guide the learning of discriminative visual features. However, such context provided by the action names is too limited to provide sufficient background knowledge for capturing novel spatial and temporal concepts in actions. In this paper, we propose DiST, an innovative Decomposition-incorporation framework for FSAR that makes use of decoupled Spatial and Temporal knowledge provided by large language models to learn expressive multi-granularity prototypes. In the decomposition stage, we decouple vanilla action names into diverse spatio-temporal attribute descriptions (action-related knowledge). Such commonsense knowledge complements semantic contexts from spatial and temporal perspectives. In the incorporation stage, we propose Spatial/Temporal Knowledge Compensators (SKC/TKC) to discover discriminative object-level and frame-level prototypes, respectively. In SKC, object-level prototypes adaptively aggregate important patch tokens under the guidance of spatial knowledge. Moreover, in TKC, frame-level prototypes utilize temporal attributes to assist in inter-frame temporal relation modeling. These learned prototypes thus provide transparency in capturing fine-grained spatial details and diverse temporal patterns. Experimental results show DiST achieves state-of-the-art results on five standard FSAR datasets.", "date": "2026-02-20", "categories": []}
{"id": "2602.08755", "title": "Align and Adapt: Multimodal Multiview Human Activity Recognition under Arbitrary View Combinations", "abstract": "Multimodal multiview learning seeks to integrate information from diverse sources to enhance task performance. Existing approaches often struggle with flexible view configurations, including arbitrary view combinations, numbers of views, and heterogeneous modalities. Focusing on the context of human activity recognition, we propose AliAd, a model that combines multiview contrastive learning with a mixture-of-experts module to support arbitrary view availability during both training and inference. Instead of trying to reconstruct missing views, an adjusted center contrastive loss is used for self-supervised representation learning and view alignment, mitigating the impact of missing views on multiview fusion. This loss formulation allows for the integration of view weights to account for view quality. Additionally, it reduces computational complexity from $O(V^2)$ to $O(V)$, where $V$ is the number of views. To address residual discrepancies not captured by contrastive learning, we employ a mixture-of-experts module with a specialized load balancing strategy, tasked with adapting to arbitrary view combinations. We highlight the geometric relationship among components in our model and how they combine well in the latent space. AliAd is validated on four datasets encompassing inertial and human pose modalities, with the number of views ranging from three to nine, demonstrating its performance and flexibility.", "date": "2026-02-09", "categories": []}
{"id": "2602.11810", "title": "How to Sample High Quality 3D Fractals for Action Recognition Pre-Training?", "abstract": "Synthetic datasets are being recognized in the deep learning realm as a valuable alternative to exhaustively labeled real data. One such synthetic data generation method is Formula Driven Supervised Learning (FDSL), which can provide an infinite number of perfectly labeled data through a formula driven approach, such as fractals or contours. FDSL does not have common drawbacks like manual labor, privacy and other ethical concerns. In this work we generate 3D fractals using 3D Iterated Function Systems (IFS) for pre-training an action recognition model. The fractals are temporally transformed to form a video that is used as a pre-training dataset for downstream task of action recognition. We find that standard methods of generating fractals are slow and produce degenerate 3D fractals. Therefore, we systematically explore alternative ways of generating fractals and finds that overly-restrictive approaches, while generating aesthetically pleasing fractals, are detrimental for downstream task performance. We propose a novel method, Targeted Smart Filtering, to address both the generation speed and fractal diversity issue. The method reports roughly 100 times faster sampling speed and achieves superior downstream performance against other 3D fractal filtering methods.", "date": "2026-02-12", "categories": []}
{"id": "2602.12002", "title": "Can Local Vision-Language Models improve Activity Recognition over Vision Transformers? -- Case Study on Newborn Resuscitation", "abstract": "Accurate documentation of newborn resuscitation is essential for quality improvement and adherence to clinical guidelines, yet remains underutilized in practice. Previous work using 3D-CNNs and Vision Transformers (ViT) has shown promising results in detecting key activities from newborn resuscitation videos, but also highlighted the challenges in recognizing such fine-grained activities. This work investigates the potential of generative AI (GenAI) methods to improve activity recognition from such videos. Specifically, we explore the use of local vision-language models (VLMs), combined with large language models (LLMs), and compare them to a supervised TimeSFormer baseline. Using a simulated dataset comprising 13.26 hours of newborn resuscitation videos, we evaluate several zero-shot VLM-based strategies and fine-tuned VLMs with classification heads, including Low-Rank Adaptation (LoRA). Our results suggest that small (local) VLMs struggle with hallucinations, but when fine-tuned with LoRA, the results reach F1 score at 0.91, surpassing the TimeSformer results of 0.70.", "date": "2026-02-12", "categories": []}
{"id": "2602.10818", "title": "Resource-Efficient RGB-Only Action Recognition for Edge Deployment", "abstract": "Action recognition on edge devices poses stringent constraints on latency, memory, storage, and power consumption. While auxiliary modalities such as skeleton and depth information can enhance recognition performance, they often require additional sensors or computationally expensive pose-estimation pipelines, limiting practicality for edge use. In this work, we propose a compact RGB-only network tailored for efficient on-device inference. Our approach builds upon an X3D-style backbone augmented with Temporal Shift, and further introduces selective temporal adaptation and parameter-free attention. Extensive experiments on the NTU RGB+D 60 and 120 benchmarks demonstrate a strong accuracy-efficiency balance. Moreover, deployment-level profiling on the Jetson Orin Nano verifies a smaller on-device footprint and practical resource utilization compared to existing RGB-based action recognition techniques.", "date": "2026-02-11", "categories": []}
{"id": "2506.01102", "title": "Keystep Recognition using Graph Neural Networks", "abstract": "We pose keystep recognition as a node classification task, and propose a flexible graph-learning framework for fine-grained keystep recognition that is able to effectively leverage long-term dependencies in egocentric videos. Our approach, termed GLEVR, consists of constructing a graph where each video clip of the egocentric video corresponds to a node. The constructed graphs are sparse and computationally efficient, outperforming existing larger models substantially. We further leverage alignment between egocentric and exocentric videos during training for improved inference on egocentric videos, as well as adding automatic captioning as an additional modality. We consider each clip of each exocentric video (if available) or video captions as additional nodes during training. We examine several strategies to define connections across these nodes. We perform extensive experiments on the Ego-Exo4D dataset and show that our proposed flexible graph-based framework notably outperforms existing methods.", "date": "2025-06-01", "categories": []}
{"id": "2602.09518", "title": "A Universal Action Space for General Behavior Analysis", "abstract": "Analyzing animal and human behavior has long been a challenging task in computer vision. Early approaches from the 1970s to the 1990s relied on hand-crafted edge detection, segmentation, and low-level features such as color, shape, and texture to locate objects and infer their identities-an inherently ill-posed problem. Behavior analysis in this era typically proceeded by tracking identified objects over time and modeling their trajectories using sparse feature points, which further limited robustness and generalization. A major shift occurred with the introduction of ImageNet by Deng and Li in 2010, which enabled large-scale visual recognition through deep neural networks and effectively served as a comprehensive visual dictionary. This development allowed object recognition to move beyond complex low-level processing toward learned high-level representations. In this work, we follow this paradigm to build a large-scale Universal Action Space (UAS) using existing labeled human-action datasets. We then use this UAS as the foundation for analyzing and categorizing mammalian and chimpanzee behavior datasets. The source code is released on GitHub at https://github.com/franktpmvu/Universal-Action-Space.", "date": "2026-02-10", "categories": []}
{"id": "2602.07024", "title": "A Distributed Multi-Modal Sensing Approach for Human Activity Recognition in Real-Time Human-Robot Collaboration", "abstract": "Human activity recognition (HAR) is fundamental in human-robot collaboration (HRC), enabling robots to respond to and dynamically adapt to human intentions. This paper introduces a HAR system combining a modular data glove equipped with Inertial Measurement Units and a vision-based tactile sensor to capture hand activities in contact with a robot. We tested our activity recognition approach under different conditions, including offline classification of segmented sequences, real-time classification under static conditions, and a realistic HRC scenario. The experimental results show a high accuracy for all the tasks, suggesting that multiple collaborative settings could benefit from this multi-modal approach.", "date": "2026-02-02", "categories": []}
{"id": "2602.06251", "title": "ASMa: Asymmetric Spatio-temporal Masking for Skeleton Action Representation Learning", "abstract": "Self-supervised learning (SSL) has shown remarkable success in skeleton-based action recognition by leveraging data augmentations to learn meaningful representations. However, existing SSL methods rely on data augmentations that predominantly focus on masking high-motion frames and high-degree joints such as joints with degree 3 or 4. This results in biased and incomplete feature representations that struggle to generalize across varied motion patterns. To address this, we propose Asymmetric Spatio-temporal Masking (ASMa) for Skeleton Action Representation Learning, a novel combination of masking to learn a full spectrum of spatio-temporal dynamics inherent in human actions. ASMa employs two complementary masking strategies: one that selectively masks high-degree joints and low-motion, and another that masks low-degree joints and high-motion frames. These masking strategies ensure a more balanced and comprehensive skeleton representation learning. Furthermore, we introduce a learnable feature alignment module to effectively align the representations learned from both masked views. To facilitate deployment in resource-constrained settings and on low-resource devices, we compress the learned and aligned representation into a lightweight model using knowledge distillation. Extensive experiments on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD datasets demonstrate that our approach outperforms existing SSL methods with an average improvement of 2.7-4.4% in fine-tuning and up to 5.9% in transfer learning to noisy datasets and achieves competitive performance compared to fully supervised baselines. Our distilled model achieves 91.4% parameter reduction and 3x faster inference on edge devices while maintaining competitive accuracy, enabling practical deployment in resource-constrained scenarios.", "date": "2026-02-05", "categories": []}
{"id": "2503.21843", "title": "CMD-HAR: Cross-Modal Disentanglement for Wearable Human Activity Recognition", "abstract": "Human Activity Recognition (HAR) is a fundamental technology for numerous human - centered intelligent applications. Although deep learning methods have been utilized to accelerate feature extraction, issues such as multimodal data mixing, activity heterogeneity, and complex model deployment remain largely unresolved. The aim of this paper is to address issues such as multimodal data mixing, activity heterogeneity, and complex model deployment in sensor-based human activity recognition. We propose a spatiotemporal attention modal decomposition alignment fusion strategy to tackle the problem of the mixed distribution of sensor data. Key discriminative features of activities are captured through cross-modal spatio-temporal disentangled representation, and gradient modulation is combined to alleviate data heterogeneity. In addition, a wearable deployment simulation system is constructed. We conducted experiments on a large number of public datasets, demonstrating the effectiveness of the model.", "date": "2025-03-27", "categories": []}
{"id": "2602.05718", "title": "Exploring the Temporal Consistency for Point-Level Weakly-Supervised Temporal Action Localization", "abstract": "Point-supervised Temporal Action Localization (PTAL) adopts a lightly frame-annotated paradigm (\\textit{i.e.}, labeling only a single frame per action instance) to train a model to effectively locate action instances within untrimmed videos. Most existing approaches design the task head of models with only a point-supervised snippet-level classification, without explicit modeling of understanding temporal relationships among frames of an action. However, understanding the temporal relationships of frames is crucial because it can help a model understand how an action is defined and therefore benefits localizing the full frames of an action. To this end, in this paper, we design a multi-task learning framework that fully utilizes point supervision to boost the model's temporal understanding capability for action localization. Specifically, we design three self-supervised temporal understanding tasks: (i) Action Completion, (ii) Action Order Understanding, and (iii) Action Regularity Understanding. These tasks help a model understand the temporal consistency of actions across videos. To the best of our knowledge, this is the first attempt to explicitly explore temporal consistency for point supervision action localization. Extensive experimental results on four benchmark datasets demonstrate the effectiveness of the proposed method compared to several state-of-the-art approaches.", "date": "2026-02-05", "categories": []}
{"id": "2602.04182", "title": "HoloEv-Net: Efficient Event-based Action Recognition via Holographic Spatial Embedding and Global Spectral Gating", "abstract": "Event-based Action Recognition (EAR) has attracted significant attention due to the high temporal resolution and high dynamic range of event cameras. However, existing methods typically suffer from (i) the computational redundancy of dense voxel representations, (ii) structural redundancy inherent in multi-branch architectures, and (iii) the under-utilization of spectral information in capturing global motion patterns. To address these challenges, we propose an efficient EAR framework named HoloEv-Net. First, to simultaneously tackle representation and structural redundancies, we introduce a Compact Holographic Spatiotemporal Representation (CHSR). Departing from computationally expensive voxel grids, CHSR implicitly embeds horizontal spatial cues into the Time-Height (T-H) view, effectively preserving 3D spatiotemporal contexts within a 2D representation. Second, to exploit the neglected spectral cues, we design a Global Spectral Gating (GSG) module. By leveraging the Fast Fourier Transform (FFT) for global token mixing in the frequency domain, GSG enhances the representation capability with negligible parameter overhead. Extensive experiments demonstrate the scalability and effectiveness of our framework. Specifically, HoloEv-Net-Base achieves state-of-the-art performance on THU-EACT-50-CHL, HARDVS and DailyDVS-200, outperforming existing methods by 10.29%, 1.71% and 6.25%, respectively. Furthermore, our lightweight variant, HoloEv-Net-Small, delivers highly competitive accuracy while offering extreme efficiency, reducing parameters by 5.4 times, FLOPs by 300times, and latency by 2.4times compared to heavy baselines, demonstrating its potential for edge deployment.", "date": "2026-02-04", "categories": []}
{"id": "2602.02542", "title": "Auto-Augmentation Contrastive Learning for Wearable-based Human Activity Recognition", "abstract": "For low-semantic sensor signals from human activity recognition (HAR), contrastive learning (CL) is essential to implement novel applications or generic models without manual annotation, which is a high-performance self-supervised learning (SSL) method. However, CL relies heavily on data augmentation for pairwise comparisons. Especially for low semantic data in the HAR area, conducting good performance augmentation strategies in pretext tasks still rely on manual attempts lacking generalizability and flexibility. To reduce the augmentation burden, we propose an end-to-end auto-augmentation contrastive learning (AutoCL) method for wearable-based HAR. AutoCL is based on a Siamese network architecture that shares the parameters of the backbone and with a generator embedded to learn auto-augmentation. AutoCL trains the generator based on the representation in the latent space to overcome the disturbances caused by noise and redundant information in raw sensor data. The architecture empirical study indicates the effectiveness of this design. Furthermore, we propose a stop-gradient design and correlation reduction strategy in AutoCL to enhance encoder representation learning. Extensive experiments based on four wide-used HAR datasets demonstrate that the proposed AutoCL method significantly improves recognition accuracy compared with other SOTA methods.", "date": "2026-01-30", "categories": []}
{"id": "2509.18473", "title": "MoCrop: Training Free Motion Guided Cropping for Efficient Video Action Recognition", "abstract": "Standard video action recognition models often process typically resized full frames, suffering from spatial redundancy and high computational costs. To address this, we introduce MoCrop, a motion-aware adaptive cropping module designed for efficient video action recognition in the compressed domain. Leveraging Motion Vectors (MVs) naturally available in H.264 video, MoCrop localizes motion-dense regions to produce adaptive crops at inference without requiring any training or parameter updates. Our lightweight pipeline synergizes three key components: Merge & Denoise (MD) for outlier filtering, Monte Carlo Sampling (MCS) for efficient importance sampling, and Motion Grid Search (MGS) for optimal region localization. This design allows MoCrop to serve as a versatile \"plug-and-play\" module for diverse backbones. Extensive experiments on UCF101 demonstrate that MoCrop serves as both an accelerator and an enhancer. With ResNet-50, it achieves a +3.5% boost in Top-1 accuracy at equivalent FLOPs (Attention Setting), or a +2.4% accuracy gain with 26.5% fewer FLOPs (Efficiency Setting). When applied to CoViAR, it improves accuracy to 89.2% or reduces computation by roughly 27% (from 11.6 to 8.5 GFLOPs). Consistent gains across MobileNet-V3, EfficientNet-B1, and Swin-B confirm its strong generality and suitability for real-time deployment. Our code and models are available at https://github.com/microa/MoCrop.", "date": "2025-09-22", "categories": []}
{"id": "2602.00131", "title": "PovNet+: A Deep Learning Architecture for Socially Assistive Robots to Learn and Assist with Multiple Activities of Daily Living", "abstract": "A significant barrier to the long-term deployment of autonomous socially assistive robots is their inability to both perceive and assist with multiple activities of daily living (ADLs). In this paper, we present the first multimodal deep learning architecture, POVNet+, for multi-activity recognition for socially assistive robots to proactively initiate assistive behaviors. Our novel architecture introduces the use of both ADL and motion embedding spaces to uniquely distinguish between a known ADL being performed, a new unseen ADL, or a known ADL being performed atypically in order to assist people in real scenarios. Furthermore, we apply a novel user state estimation method to the motion embedding space to recognize new ADLs while monitoring user performance. This ADL perception information is used to proactively initiate robot assistive interactions. Comparison experiments with state-of-the-art human activity recognition methods show our POVNet+ method has higher ADL classification accuracy. Human-robot interaction experiments in a cluttered living environment with multiple users and the socially assistive robot Leia using POVNet+ demonstrate the ability of our multi-modal ADL architecture in successfully identifying different seen and unseen ADLs, and ADLs being performed atypically, while initiating appropriate assistive human-robot interactions.", "date": "2026-01-28", "categories": []}
{"id": "2602.01257", "title": "Boosting Point-supervised Temporal Action Localization via Text Refinement and Alignment", "abstract": "Recently, point-supervised temporal action localization has gained significant attention for its effective balance between labeling costs and localization accuracy. However, current methods only consider features from visual inputs, neglecting helpful semantic information from the text side. To address this issue, we propose a Text Refinement and Alignment (TRA) framework that effectively utilizes textual features from visual descriptions to complement the visual features as they are semantically rich. This is achieved by designing two new modules for the original point-supervised framework: a Point-based Text Refinement module (PTR) and a Point-based Multimodal Alignment module (PMA). Specifically, we first generate descriptions for video frames using a pre-trained multimodal model. Next, PTR refines the initial descriptions by leveraging point annotations together with multiple pre-trained models. PMA then projects all features into a unified semantic space and leverages a point-level multimodal feature contrastive learning to reduce the gap between visual and linguistic modalities. Last, the enhanced multi-modal features are fed into the action detector for precise localization. Extensive experimental results on five widely used benchmarks demonstrate the favorable performance of our proposed framework compared to several state-of-the-art methods. Moreover, our computational overhead analysis shows that the framework can run on a single 24 GB RTX 3090 GPU, indicating its practicality and scalability.", "date": "2026-02-01", "categories": []}
{"id": "2602.01850", "title": "WS-IMUBench: Can Weakly Supervised Methods from Audio, Image, and Video Be Adapted for IMU-based Temporal Action Localization?", "abstract": "IMU-based Human Activity Recognition (HAR) has enabled a wide range of ubiquitous computing applications, yet its dominant clip classification paradigm cannot capture the rich temporal structure of real-world behaviors. This motivates a shift toward IMU Temporal Action Localization (IMU-TAL), which predicts both action categories and their start/end times in continuous streams. However, current progress is strongly bottlenecked by the need for dense, frame-level boundary annotations, which are costly and difficult to scale. To address this bottleneck, we introduce WS-IMUBench, a systematic benchmark study of weakly supervised IMU-TAL (WS-IMU-TAL) under only sequence-level labels. Rather than proposing a new localization algorithm, we evaluate how well established weakly supervised localization paradigms from audio, image, and video transfer to IMU-TAL under only sequence-level labels. We benchmark seven representative weakly supervised methods on seven public IMU datasets, resulting in over 3,540 model training runs and 7,080 inference evaluations. Guided by three research questions on transferability, effectiveness, and insights, our findings show that (i) transfer is modality-dependent, with temporal-domain methods generally more stable than image-derived proposal-based approaches; (ii) weak supervision can be competitive on favorable datasets (e.g., with longer actions and higher-dimensional sensing); and (iii) dominant failure modes arise from short actions, temporal ambiguity, and proposal quality. Finally, we outline concrete directions for advancing WS-IMU-TAL (e.g., IMU-specific proposal generation, boundary-aware objectives, and stronger temporal reasoning). Beyond individual results, WS-IMUBench establishes a reproducible benchmarking template, datasets, protocols, and analyses, to accelerate community-wide progress toward scalable WS-IMU-TAL.", "date": "2026-02-02", "categories": []}
{"id": "2505.09455", "title": "Beyond Pixels: Leveraging the Language of Soccer to Improve Spatio-Temporal Action Detection in Broadcast Videos", "abstract": "State-of-the-art spatio-temporal action detection (STAD) methods show promising results for extracting soccer events from broadcast videos. However, when operated in the high-recall, low-precision regime required for exhaustive event coverage in soccer analytics, their lack of contextual understanding becomes apparent: many false positives could be resolved by considering a broader sequence of actions and game-state information. In this work, we address this limitation by reasoning at the game level and improving STAD through the addition of a denoising sequence transduction task. Sequences of noisy, context-free player-centric predictions are processed alongside clean game state information using a Transformer-based encoder-decoder model. By modeling extended temporal context and reasoning jointly over team-level dynamics, our method leverages the \"language of soccer\" - its tactical regularities and inter-player dependencies - to generate \"denoised\" sequences of actions. This approach improves both precision and recall in low-confidence regimes, enabling more reliable event extraction from broadcast video and complementing existing pixel-based methods.", "date": "2025-05-14", "categories": []}
{"id": "2601.22675", "title": "Fire on Motion: Optimizing Video Pass-bands for Efficient Spiking Action Recognition", "abstract": "Spiking neural networks (SNNs) have gained traction in vision due to their energy efficiency, bio-plausibility, and inherent temporal processing. Yet, despite this temporal capacity, most progress concentrates on static image benchmarks, and SNNs still underperform on dynamic video tasks compared to artificial neural networks (ANNs). In this work, we diagnose a fundamental pass-band mismatch: Standard spiking dynamics behave as a temporal low pass that emphasizes static content while attenuating motion bearing bands, where task relevant information concentrates in dynamic tasks. This phenomenon explains why SNNs can approach ANNs on static tasks yet fall behind on tasks that demand richer temporal understanding.To remedy this, we propose the Pass-Bands Optimizer (PBO), a plug-and-play module that optimizes the temporal pass-band toward task-relevant motion bands. PBO introduces only two learnable parameters, and a lightweight consistency constraint that preserves semantics and boundaries, incurring negligible computational overhead and requires no architectural changes. PBO deliberately suppresses static components that contribute little to discrimination, effectively high passing the stream so that spiking activity concentrates on motion bearing content. On UCF101, PBO yields over ten percentage points improvement. On more complex multi-modal action recognition and weakly supervised video anomaly detection, PBO delivers consistent and significant gains, offering a new perspective for SNN based video processing and understanding.", "date": "2026-01-30", "categories": []}
{"id": "2501.13518", "title": "Text-driven Online Action Detection", "abstract": "Detecting actions as they occur is essential for applications like video surveillance, autonomous driving, and human-robot interaction. Known as online action detection, this task requires classifying actions in streaming videos, handling background noise, and coping with incomplete actions. Transformer architectures are the current state-of-the-art, yet the potential of recent advancements in computer vision, particularly vision-language models (VLMs), remains largely untapped for this problem, partly due to high computational costs. In this paper, we introduce TOAD: a Text-driven Online Action Detection architecture that supports zero-shot and few-shot learning. TOAD leverages CLIP (Contrastive Language-Image Pretraining) textual embeddings, enabling efficient use of VLMs without significant computational overhead. Our model achieves 82.46% mAP on the THUMOS14 dataset, outperforming existing methods, and sets new baselines for zero-shot and few-shot performance on the THUMOS14 and TVSeries datasets.", "date": "2025-01-23", "categories": []}
{"id": "2601.22039", "title": "Understanding Multimodal Complementarity for Single-Frame Action Anticipation", "abstract": "Human action anticipation is commonly treated as a video understanding problem, implicitly assuming that dense temporal information is required to reason about future actions. In this work, we challenge this assumption by investigating what can be achieved when action anticipation is constrained to a single visual observation. We ask a fundamental question: how much information about the future is already encoded in a single frame, and how can it be effectively exploited? Building on our prior work on Action Anticipation at a Glimpse (AAG), we conduct a systematic investigation of single-frame action anticipation enriched with complementary sources of information. We analyze the contribution of RGB appearance, depth-based geometric cues, and semantic representations of past actions, and investigate how different multimodal fusion strategies, keyframe selection policies and past-action history sources influence anticipation performance. Guided by these findings, we consolidate the most effective design choices into AAG+, a refined single-frame anticipation framework. Despite operating on a single frame, AAG+ consistently improves upon the original AAG and achieves performance comparable to, or exceeding, that of state-of-the-art video-based methods on challenging anticipation benchmarks including IKEA-ASM, Meccano and Assembly101. Our results offer new insights into the limits and potential of single-frame action anticipation, and clarify when dense temporal modeling is necessary and when a carefully selected glimpse is sufficient.", "date": "2026-01-29", "categories": []}
{"id": "2503.18712", "title": "LLaVAction: evaluating and training multi-modal large language models for action understanding", "abstract": "Understanding human behavior requires measuring behavioral actions. Due to its complexity, behavior is best mapped onto a rich, semantic structure such as language. Emerging multimodal large language models (MLLMs) are promising candidates, but their fine-grained action understanding ability has not been fully examined. In this work, we reformulate EPIC-KITCHENS-100, one of the largest and most challenging egocentric action recognition datasets, into a MLLM benchmark (EPIC-KITCHENS-100-MQA). We demonstrate that when we sample difficult answers based on specialist models as distractors, leading MLLMs struggle to recognize the correct actions. How can we increase the performance of MLLMs? We curated a supervised finetuning dataset that includes `hard' action recognition, temporal detection, captioning, and free-form question answering to improve models' diverse action understanding capabilities. We introduce a new model called LLaVAction that adds an action token to boost models' attention on visual tokens and a two-stage pipeline to obtain structured actions. LLaVAction greatly improves the MLLMs' ability of action understanding, achieving strong improvements on both MLLM benchmarks (21 points in accuracy over GPT-4o on EPIC-KITCHENS-100-MQA) and established action recognition benchmarks, suggesting that our methods prepare MLLMs to be a promising path forward for complex action tasks. Code, data, the benchmark, and models are available at https://github.com/AdaptiveMotorControlLab/LLaVAction.", "date": "2025-03-24", "categories": []}
{"id": "2508.04945", "title": "Towards Robust Evaluation of Visual Activity Recognition: Resolving Verb Ambiguity with Sense Clustering", "abstract": "Evaluating visual activity recognition systems is challenging due to inherent ambiguities in verb semantics and image interpretation. When describing actions in images, synonymous verbs can refer to the same event (e.g., brushing vs. grooming), while different perspectives can lead to equally valid but distinct verb choices (e.g., piloting vs. operating). Standard exact-match evaluation, which relies on a single gold answer, fails to capture these ambiguities, resulting in an incomplete assessment of model performance. To address this, we propose a vision-language clustering framework that constructs verb sense clusters, providing a more robust evaluation. Our analysis of the imSitu dataset shows that each image maps to around four sense clusters, with each cluster representing a distinct perspective of the image. We evaluate multiple activity recognition models and compare our cluster-based evaluation with standard evaluation methods. Additionally, our human alignment analysis suggests that the cluster-based evaluation better aligns with human judgments, offering a more nuanced assessment of model performance.", "date": "2025-08-07", "categories": []}
{"id": "2512.11654", "title": "Kinetic Mining in Context: Few-Shot Action Synthesis via Text-to-Motion Distillation", "abstract": "The acquisition cost for large, annotated motion datasets remains a critical bottleneck for skeletal-based Human Activity Recognition (HAR). Although Text-to-Motion (T2M) generative models offer a compelling, scalable source of synthetic data, their training objectives, which emphasize general artistic motion, and dataset structures fundamentally differ from HAR's requirements for kinematically precise, class-discriminative actions. This disparity creates a significant domain gap, making generalist T2M models ill-equipped for generating motions suitable for HAR classifiers. To address this challenge, we propose KineMIC (Kinetic Mining In Context), a transfer learning framework for few-shot action synthesis. KineMIC adapts a T2M diffusion model to an HAR domain by hypothesizing that semantic correspondences in the text encoding space can provide soft supervision for kinematic distillation. We operationalize this via a kinetic mining strategy that leverages CLIP text embeddings to establish correspondences between sparse HAR labels and T2M source data. This process guides fine-tuning, transforming the generalist T2M backbone into a specialized few-shot Action-to-Motion generator. We validate KineMIC using HumanML3D as the source T2M dataset and a subset of NTU RGB+D 120 as the target HAR domain, randomly selecting just 10 samples per action class. Our approach generates significantly more coherent motions, providing a robust data augmentation source that delivers a +23.1% accuracy points improvement. Animated illustrations and supplementary materials are available at https://lucazzola.github.io/publications/kinemic.", "date": "2025-12-12", "categories": []}
{"id": "2601.17049", "title": "Summary of the Unusual Activity Recognition Challenge for Developmental Disability Support", "abstract": "This paper presents an overview of the Recognize the Unseen: Unusual Behavior Recognition from Pose Data Challenge, hosted at ISAS 2025. The challenge aims to address the critical need for automated recognition of unusual behaviors in facilities for individuals with developmental disabilities using non-invasive pose estimation data. Participating teams were tasked with distinguishing between normal and unusual activities based on skeleton keypoints extracted from video recordings of simulated scenarios. The dataset reflects real-world imbalance and temporal irregularities in behavior, and the evaluation adopted a Leave-One-Subject-Out (LOSO) strategy to ensure subject-agnostic generalization. The challenge attracted broad participation from 40 teams applying diverse approaches ranging from classical machine learning to deep learning architectures. Submissions were assessed primarily using macro-averaged F1 scores to account for class imbalance. The results highlight the difficulty of modeling rare, abrupt actions in noisy, low-dimensional data, and emphasize the importance of capturing both temporal and contextual nuances in behavior modeling. Insights from this challenge may contribute to future developments in socially responsible AI applications for healthcare and behavior monitoring.", "date": "2026-01-21", "categories": []}
{"id": "2601.17056", "title": "Ego4OOD: Rethinking Egocentric Video Domain Generalization via Covariate Shift Scoring", "abstract": "Egocentric video action recognition under domain shifts remains challenging due to large intra-class spatio-temporal variability, long-tailed feature distributions, and strong correlations between actions and environments. Existing benchmarks for egocentric domain generalization often conflate covariate shifts with concept shifts, making it difficult to reliably evaluate a model's ability to generalize across input distributions. To address this limitation, we introduce Ego4OOD, a domain generalization benchmark derived from Ego4D that emphasizes measurable covariate diversity while reducing concept shift through semantically coherent, moment-level action categories. Ego4OOD spans eight geographically distinct domains and is accompanied by a clustering-based covariate shift metric that provides a quantitative proxy for domain difficulty. We further leverage a one-vs-all binary training objective that decomposes multi-class action recognition into independent binary classification tasks. This formulation is particularly well-suited for covariate shift by reducing interference between visually similar classes under feature distribution shift. Using this formulation, we show that a lightweight two-layer fully connected network achieves performance competitive with state-of-the-art egocentric domain generalization methods on both Argo1M and Ego4OOD, despite using fewer parameters and no additional modalities. Our empirical analysis demonstrates a clear relationship between measured covariate shift and recognition performance, highlighting the importance of controlled benchmarks and quantitative domain characterization for studying out-of-distribution generalization in egocentric video.", "date": "2026-01-21", "categories": []}
{"id": "2601.17391", "title": "SMV-EAR: Bring Spatiotemporal Multi-View Representation Learning into Efficient Event-Based Action Recognition", "abstract": "Event cameras action recognition (EAR) offers compelling privacy-protecting and efficiency advantages, where temporal motion dynamics is of great importance. Existing spatiotemporal multi-view representation learning (SMVRL) methods for event-based object recognition (EOR) offer promising solutions by projecting H-W-T events along spatial axis H and W, yet are limited by its translation-variant spatial binning representation and naive early concatenation fusion architecture. This paper reexamines the key SMVRL design stages for EAR and propose: (i) a principled spatiotemporal multi-view representation through translation-invariant dense conversion of sparse events, (ii) a dual-branch, dynamic fusion architecture that models sample-wise complementarity between motion features from different views, and (iii) a bio-inspired temporal warping augmentation that mimics speed variability of real-world human actions. On three challenging EAR datasets of HARDVS, DailyDVS-200 and THU-EACT-50-CHL, we show +7.0%, +10.7%, and +10.2% Top-1 accuracy gains over existing SMVRL EOR method with surprising 30.1% reduced parameters and 35.7% lower computations, establishing our framework as a novel and powerful EAR paradigm.", "date": "2026-01-24", "categories": []}
{"id": "2508.18695", "title": "A Novel Deep Hybrid Framework with Ensemble-Based Feature Optimization for Robust Real-Time Human Activity Recognition", "abstract": "Real-time Human Activity Recognition (HAR) has wide-ranging applications in areas such as context-aware environments, public safety, assistive technologies, and autonomous monitoring and surveillance systems. However, existing real-time HAR systems face significant challenges, including limited scalability and high computational costs arising from redundant features. To address these issues, the Inception-V3 model was customized with region-based and boundary-aware operations, using average pooling and max pooling, respectively, to enhance region homogeneity, suppress noise, and capture discriminative local features, while improving robustness through down-sampling. Furthermore, to effectively encode motion dynamics, an Attention-Augmented Long Short-Term Memory (AA-LSTM) network was employed to learn temporal dependencies across video frames. Features are extracted from video dataset and are then optimized through a novel proposed dynamic composite feature selection method called Adaptive Dynamic Fitness Sharing and Attention (ADFSA). This ADFSA mechanism is embedded within a genetic algorithm to select a compact, optimized subset of features by dynamically balancing multiple objectives, accuracy, redundancy reduction, feature uniqueness, and complexity minimization. As a result, the selected subset of diverse and discriminative features enables lightweight machine learning classifiers to achieve accurate and robust HAR in heterogeneous environments. Experimental results demonstrate up to 99.65\\% accuracy using as few as seven selected features, with improved inference time on the challenging UCF-YouTube dataset, which includes factors such as occlusion, cluttered backgrounds, complex motion dynamics, and poor illumination conditions.", "date": "2025-08-26", "categories": []}
{"id": "2601.10521", "title": "BikeActions: An Open Platform and Benchmark for Cyclist-Centric VRU Action Recognition", "abstract": "Anticipating the intentions of Vulnerable Road Users (VRUs) is a critical challenge for safe autonomous driving (AD) and mobile robotics. While current research predominantly focuses on pedestrian crossing behaviors from a vehicle's perspective, interactions within dense shared spaces remain underexplored. To bridge this gap, we introduce FUSE-Bike, the first fully open perception platform of its kind. Equipped with two LiDARs, a camera, and GNSS, it facilitates high-fidelity, close-range data capture directly from a cyclist's viewpoint. Leveraging this platform, we present BikeActions, a novel multi-modal dataset comprising 852 annotated samples across 5 distinct action classes, specifically tailored to improve VRU behavior modeling. We establish a rigorous benchmark by evaluating state-of-the-art graph convolution and transformer-based models on our publicly released data splits, establishing the first performance baselines for this challenging task. We release the full dataset together with data curation tools, the open hardware design, and the benchmark code to foster future research in VRU action understanding under https://iv.ee.hm.edu/bikeactions/.", "date": "2026-01-15", "categories": []}
{"id": "2601.11990", "title": "DAOS: A Multimodal In-cabin Behavior Monitoring with Driver Action-Object Synergy Dataset", "abstract": "In driver activity monitoring, movements are mostly limited to the upper body, which makes many actions look similar. To tell these actions apart, human often rely on the objects the driver is using, such as holding a phone compared with gripping the steering wheel. However, most existing driver-monitoring datasets lack accurate object-location annotations or do not link objects to their associated actions, leaving a critical gap for reliable action recognition. To address this, we introduce the Driver Action with Object Synergy (DAOS) dataset, comprising 9,787 video clips annotated with 36 fine-grained driver actions and 15 object classes, totaling more than 2.5 million corresponding object instances. DAOS offers multi-modal, multi-view data (RGB, IR, and depth) from front, face, left, and right perspectives. Although DAOS captures a wide range of cabin objects, only a few are directly relevant to each action for prediction, so focusing on task-specific human-object relations is essential. To tackle this challenge, we propose the Action-Object-Relation Network (AOR-Net). AOR-Net comprehends complex driver actions through multi-level reasoning and a chain-of-action prompting mechanism that models the logical relationships among actions, objects, and their relations. Additionally, the Mixture of Thoughts module is introduced to dynamically select essential knowledge at each stage, enhancing robustness in object-rich and object-scarce conditions. Extensive experiments demonstrate that our model outperforms other state-of-the-art methods on various datasets.", "date": "2026-01-17", "categories": []}
{"id": "2601.12432", "title": "SkeFi: Cross-Modal Knowledge Transfer for Wireless Skeleton-Based Action Recognition", "abstract": "Skeleton-based action recognition leverages human pose keypoints to categorize human actions, which shows superior generalization and interoperability compared to regular end-to-end action recognition. Existing solutions use RGB cameras to annotate skeletal keypoints, but their performance declines in dark environments and raises privacy concerns, limiting their use in smart homes and hospitals. This paper explores non-invasive wireless sensors, i.e., LiDAR and mmWave, to mitigate these challenges as a feasible alternative. Two problems are addressed: (1) insufficient data on wireless sensor modality to train an accurate skeleton estimation model, and (2) skeletal keypoints derived from wireless sensors are noisier than RGB, causing great difficulties for subsequent action recognition models. Our work, SkeFi, overcomes these gaps through a novel cross-modal knowledge transfer method acquired from the data-rich RGB modality. We propose the enhanced Temporal Correlation Adaptive Graph Convolution (TC-AGC) with frame interactive enhancement to overcome the noise from missing or inconsecutive frames. Additionally, our research underscores the effectiveness of enhancing multiscale temporal modeling through dual temporal convolution. By integrating TC-AGC with temporal modeling for cross-modal transfer, our framework can extract accurate poses and actions from noisy wireless sensors. Experiments demonstrate that SkeFi realizes state-of-the-art performances on mmWave and LiDAR. The code is available at https://github.com/Huang0035/Skefi.", "date": "2026-01-18", "categories": []}
{"id": "2601.14069", "title": "Unsupervised Video Class-Incremental Learning via Deep Embedded Clustering Management", "abstract": "Unsupervised video class incremental learning (uVCIL) represents an important learning paradigm for learning video information without forgetting, and without considering any data labels. Prior approaches have focused on supervised class-incremental learning, relying on using the knowledge of labels and task boundaries, which is costly, requires human annotation, or is simply not a realistic option. In this paper, we propose a simple yet effective approach to address the uVCIL. We first consider a deep feature extractor network, providing a set of representative video features during each task without assuming any class or task information. We then progressively build a series of deep clusters from the extracted features. During the successive task learning, the model updated from the previous task is used as an initial state in order to transfer knowledge to the current learning task. We perform in-depth evaluations on three standard video action recognition datasets, including UCF101, HMDB51, and Something-to-Something V2, by ignoring the labels from the supervised setting. Our approach significantly outperforms other baselines on all datasets.", "date": "2026-01-20", "categories": []}
{"id": "2601.14086", "title": "Two-Stream temporal transformer for video action classification", "abstract": "Motion representation plays an important role in video understanding and has many applications including action recognition, robot and autonomous guidance or others. Lately, transformer networks, through their self-attention mechanism capabilities, have proved their efficiency in many applications. In this study, we introduce a new two-stream transformer video classifier, which extracts spatio-temporal information from content and optical flow representing movement information. The proposed model identifies self-attention features across the joint optical flow and temporal frame domain and represents their relationships within the transformer encoder mechanism. The experimental results show that our proposed methodology provides excellent classification results on three well-known video datasets of human activities.", "date": "2026-01-20", "categories": []}
{"id": "2601.14101", "title": "Curriculum-Based Strategies for Efficient Cross-Domain Action Recognition", "abstract": "Despite significant progress in human action recognition, generalizing to diverse viewpoints remains a challenge. Most existing datasets are captured from ground-level perspectives, and models trained on them often struggle to transfer to drastically different domains such as aerial views. This paper examines how curriculum-based training strategies can improve generalization to unseen real aerial-view data without using any real aerial data during training.\n  We explore curriculum learning for cross-view action recognition using two out-of-domain sources: synthetic aerial-view data and real ground-view data. Our results on the evaluation on order of training (fine-tuning on synthetic aerial data vs. real ground data) shows that fine-tuning on real ground data but differ in how they transition from synthetic to real. The first uses a two-stage curriculum with direct fine-tuning, while the second applies a progressive curriculum that expands the dataset in multiple stages before fine-tuning. We evaluate both methods on the REMAG dataset using SlowFast (CNN-based) and MViTv2 (Transformer-based) architectures.\n  Results show that combining the two out-of-domain datasets clearly outperforms training on a single domain, whether real ground-view or synthetic aerial-view. Both curriculum strategies match the top-1 accuracy of simple dataset combination while offering efficiency gains. With the two-step fine-tuning method, SlowFast achieves up to a 37% reduction in iterations and MViTv2 up to a 30% reduction compared to simple combination. The multi-step progressive approach further reduces iterations, by up to 9% for SlowFast and 30% for MViTv2, relative to the two-step method. These findings demonstrate that curriculum-based training can maintain comparable performance (top-1 accuracy within 3% range) while improving training efficiency in cross-view action recognition.", "date": "2026-01-20", "categories": []}
{"id": "2601.10854", "title": "Effects of Different Attention Mechanisms Applied on 3D Models in Video Classification", "abstract": "Human action recognition has become an important research focus in computer vision due to the wide range of applications where it is used. 3D Resnet-based CNN models, particularly MC3, R3D, and R(2+1)D, have different convolutional filters to extract spatiotemporal features. This paper investigates the impact of reducing the captured knowledge from temporal data, while increasing the resolution of the frames. To establish this experiment, we created similar designs to the three originals, but with a dropout layer added before the final classifier. Secondly, we then developed ten new versions for each one of these three designs. The variants include special attention blocks within their architecture, such as convolutional block attention module (CBAM), temporal convolution networks (TCN), in addition to multi-headed and channel attention mechanisms. The purpose behind that is to observe the extent of the influence each of these blocks has on performance for the restricted-temporal models. The results of testing all the models on UCF101 have shown accuracy of 88.98% for the variant with multiheaded attention added to the modified R(2+1)D. This paper concludes the significance of missing temporal features in the performance of the newly created increased resolution models. The variants had different behavior on class-level accuracy, despite the similarity of their enhancements to the overall performance.", "date": "2026-01-15", "categories": []}
{"id": "2601.10592", "title": "Action100M: A Large-scale Video Action Dataset", "abstract": "Inferring physical actions from visual observations is a fundamental capability for advancing machine intelligence in the physical world. Achieving this requires large-scale, open-vocabulary video action datasets that span broad domains. We introduce Action100M, a large-scale dataset constructed from 1.2M Internet instructional videos (14.6 years of duration), yielding O(100 million) temporally localized segments with open-vocabulary action supervision and rich captions. Action100M is generated by a fully automated pipeline that (i) performs hierarchical temporal segmentation using V-JEPA 2 embeddings, (ii) produces multi-level frame and segment captions organized as a Tree-of-Captions, and (iii) aggregates evidence with a reasoning model (GPT-OSS-120B) under a multi-round Self-Refine procedure to output structured annotations (brief/detailed action, actor, brief/detailed caption). Training VL-JEPA on Action100M demonstrates consistent data-scaling improvements and strong zero-shot performance across diverse action recognition benchmarks, establishing Action100M as a new foundation for scalable research in video understanding and world modeling.", "date": "2026-01-15", "categories": []}
{"id": "2601.08241", "title": "Improving Zero-shot ADL Recognition with Large Language Models through Event-based Context and Confidence", "abstract": "Unobtrusive sensor-based recognition of Activities of Daily Living (ADLs) in smart homes by processing data collected from IoT sensing devices supports applications such as healthcare, safety, and energy management. Recent zero-shot methods based on Large Language Models (LLMs) have the advantage of removing the reliance on labeled ADL sensor data. However, existing approaches rely on time-based segmentation, which is poorly aligned with the contextual reasoning capabilities of LLMs. Moreover, existing approaches lack methods for estimating prediction confidence. This paper proposes to improve zero-shot ADL recognition with event-based segmentation and a novel method for estimating prediction confidence. Our experimental evaluation shows that event-based segmentation consistently outperforms time-based LLM approaches on complex, realistic datasets and surpasses supervised data-driven methods, even with relatively small LLMs (e.g., Gemma 3 27B). The proposed confidence measure effectively distinguishes correct from incorrect predictions.", "date": "2026-01-13", "categories": []}
{"id": "2511.20928", "title": "Smooth regularization for efficient video recognition", "abstract": "We propose a smooth regularization technique that instills a strong temporal inductive bias in video recognition models, particularly benefiting lightweight architectures. Our method encourages smoothness in the intermediate-layer embeddings of consecutive frames by modeling their changes as a Gaussian Random Walk (GRW). This penalizes abrupt representational shifts, thereby promoting low-acceleration solutions that better align with the natural temporal coherence inherent in videos. By leveraging this enforced smoothness, lightweight models can more effectively capture complex temporal dynamics. Applied to such models, our technique yields a 3.8% to 6.4% accuracy improvement on Kinetics-600. Notably, the MoViNets model family trained with our smooth regularization improves the current state of the art by 3.8% to 6.1% within their respective FLOP constraints, while MobileNetV3 and the MoViNets-Stream family achieve gains of 4.9% to 6.4% over prior state-of-the-art models with comparable memory footprints. Our code and models are available at https://github.com/cmusatyalab/grw-smoothing.", "date": "2025-11-25", "categories": []}
{"id": "2601.07469", "title": "Knowledge Distillation for LLM-Based Human Activity Recognition in Homes", "abstract": "Human Activity Recognition (HAR) is a central problem for context-aware applications, especially for smart homes and assisted living. A few very recent studies have shown that Large Language Models (LLMs) can be used for HAR at home, reaching high performance and addressing key challenges. In this paper, we provide new experimental results regarding the use of LLMs for HAR, on two state-of-the-art datasets. More specifically, we show how recognition performance evolves depending on the size of the LLM used. Moreover, we experiment on the use of knowledge distillation techniques to fine-tune smaller LLMs with HAR reasoning examples generated by larger LLMs. We show that such fine-tuned models can perform almost as well as the largest LLMs, while having 50 times less parameters.", "date": "2026-01-12", "categories": []}
{"id": "2601.07666", "title": "Variational Contrastive Learning for Skeleton-based Action Recognition", "abstract": "In recent years, self-supervised representation learning for skeleton-based action recognition has advanced with the development of contrastive learning methods. However, most of contrastive paradigms are inherently discriminative and often struggle to capture the variability and uncertainty intrinsic to human motion. To address this issue, we propose a variational contrastive learning framework that integrates probabilistic latent modeling with contrastive self-supervised learning. This formulation enables the learning of structured and semantically meaningful representations that generalize across different datasets and supervision levels. Extensive experiments on three widely used skeleton-based action recognition benchmarks show that our proposed method consistently outperforms existing approaches, particularly in low-label regimes. Moreover, qualitative analyses show that the features provided by our method are more relevant given the motion and sample characteristics, with more focus on important skeleton joints, when compared to the other methods.", "date": "2026-01-12", "categories": []}
{"id": "2504.11749", "title": "SkeletonX: Data-Efficient Skeleton-based Action Recognition via Cross-sample Feature Aggregation", "abstract": "While current skeleton action recognition models demonstrate impressive performance on large-scale datasets, their adaptation to new application scenarios remains challenging. These challenges are particularly pronounced when facing new action categories, diverse performers, and varied skeleton layouts, leading to significant performance degeneration. Additionally, the high cost and difficulty of collecting skeleton data make large-scale data collection impractical. This paper studies one-shot and limited-scale learning settings to enable efficient adaptation with minimal data. Existing approaches often overlook the rich mutual information between labeled samples, resulting in sub-optimal performance in low-data scenarios. To boost the utility of labeled data, we identify the variability among performers and the commonality within each action as two key attributes. We present SkeletonX, a lightweight training pipeline that integrates seamlessly with existing GCN-based skeleton action recognizers, promoting effective training under limited labeled data. First, we propose a tailored sample pair construction strategy on two key attributes to form and aggregate sample pairs. Next, we develop a concise and effective feature aggregation module to process these pairs. Extensive experiments are conducted on NTU RGB+D, NTU RGB+D 120, and PKU-MMD with various GCN backbones, demonstrating that the pipeline effectively improves performance when trained from scratch with limited data. Moreover, it surpasses previous state-of-the-art methods in the one-shot setting, with only 1/10 of the parameters and much fewer FLOPs. The code and data are available at: https://github.com/zzysteve/SkeletonX", "date": "2025-04-16", "categories": []}
{"id": "2601.01914", "title": "Learning Action Hierarchies via Hybrid Geometric Diffusion", "abstract": "Temporal action segmentation is a critical task in video understanding, where the goal is to assign action labels to each frame in a video. While recent advances leverage iterative refinement-based strategies, they fail to explicitly utilize the hierarchical nature of human actions. In this work, we propose HybridTAS - a novel framework that incorporates a hybrid of Euclidean and hyperbolic geometries into the denoising process of diffusion models to exploit the hierarchical structure of actions. Hyperbolic geometry naturally provides tree-like relationships between embeddings, enabling us to guide the action label denoising process in a coarse-to-fine manner: higher diffusion timesteps are influenced by abstract, high-level action categories (root nodes), while lower timesteps are refined using fine-grained action classes (leaf nodes). Extensive experiments on three benchmark datasets, GTEA, 50Salads, and Breakfast, demonstrate that our method achieves state-of-the-art performance, validating the effectiveness of hyperbolic-guided denoising for the temporal action segmentation task.", "date": "2026-01-05", "categories": []}
{"id": "2503.14960", "title": "Body-Hand Modality Expertized Networks with Cross-attention for Fine-grained Skeleton Action Recognition", "abstract": "Skeleton-based Human Action Recognition (HAR) is a vital technology in robotics and human-robot interaction. However, most existing methods concentrate primarily on full-body movements and often overlook subtle hand motions that are critical for distinguishing fine-grained actions. Recent work leverages a unified graph representation that combines body, hand, and foot keypoints to capture detailed body dynamics. Yet, these models often blur fine hand details due to the disparity between body and hand action characteristics and the loss of subtle features during the spatial-pooling. In this paper, we propose BHaRNet (Body-Hand action Recognition Network), a novel framework that augments a typical body-expert model with a hand-expert model. Our model jointly trains both streams with an ensemble loss that fosters cooperative specialization, functioning in a manner reminiscent of a Mixture-of-Experts (MoE). Moreover, cross-attention is employed via an expertized branch method and a pooling-attention module to enable feature-level interactions and selectively fuse complementary information. Inspired by MMNet, we also demonstrate the applicability of our approach to multi-modal tasks by leveraging RGB information, where body features guide RGB learning to capture richer contextual cues. Experiments on large-scale benchmarks (NTU RGB+D 60, NTU RGB+D 120, PKU-MMD, and Northwestern-UCLA) demonstrate that BHaRNet achieves SOTA accuracies -- improving from 86.4\\% to 93.0\\% in hand-intensive actions -- while maintaining fewer GFLOPs and parameters than the relevant unified methods.", "date": "2025-03-19", "categories": []}
{"id": "2512.22385", "title": "LLM-Guided Exemplar Selection for Few-Shot Wearable-Sensor Human Activity Recognition", "abstract": "In this paper, we propose an LLM-Guided Exemplar Selection framework to address a key limitation in state-of-the-art Human Activity Recognition (HAR) methods: their reliance on large labeled datasets and purely geometric exemplar selection, which often fail to distinguish similar wearable sensor activities such as walking, walking upstairs, and walking downstairs. Our method incorporates semantic reasoning via an LLM-generated knowledge prior that captures feature importance, inter-class confusability, and exemplar budget multipliers, and uses it to guide exemplar scoring and selection. These priors are combined with margin-based validation cues, PageRank centrality, hubness penalization, and facility-location optimization to obtain a compact and informative set of exemplars. Evaluated on the UCI-HAR dataset under strict few-shot conditions, the framework achieves a macro F1-score of 88.78%, outperforming classical approaches such as random sampling, herding, and k-center. The results show that LLM-derived semantic priors, when integrated with structural and geometric cues, provide a stronger foundation for selecting representative sensor exemplars in few-shot wearable-sensor HAR.", "date": "2025-12-26", "categories": []}
{"id": "2601.00311", "title": "ReMA: A Training-Free Plug-and-Play Mixing Augmentation for Video Behavior Recognition", "abstract": "Video behavior recognition demands stable and discriminative representations under complex spatiotemporal variations. However, prevailing data augmentation strategies for videos remain largely perturbation-driven, often introducing uncontrolled variations that amplify non-discriminative factors, which finally weaken intra-class distributional structure and representation drift with inconsistent gains across temporal scales. To address these problems, we propose Representation-aware Mixing Augmentation (ReMA), a plug-and-play augmentation strategy that formulates mixing as a controlled replacement process to expand representations while preserving class-conditional stability. ReMA integrates two complementary mechanisms. Firstly, the Representation Alignment Mechanism (RAM) performs structured intra-class mixing under distributional alignment constraints, suppressing irrelevant intra-class drift while enhancing statistical reliability. Then, the Dynamic Selection Mechanism (DSM) generates motion-aware spatiotemporal masks to localize perturbations, guiding them away from discrimination-sensitive regions and promoting temporal coherence. By jointly controlling how and where mixing is applied, ReMA improves representation robustness without additional supervision or trainable parameters. Extensive experiments on diverse video behavior benchmarks demonstrate that ReMA consistently enhances generalization and robustness across different spatiotemporal granularities.", "date": "2026-01-01", "categories": []}
{"id": "2601.00369", "title": "BHaRNet: Reliability-Aware Body-Hand Modality Expertized Networks for Fine-grained Skeleton Action Recognition", "abstract": "Skeleton-based human action recognition (HAR) has achieved remarkable progress with graph-based architectures. However, most existing methods remain body-centric, focusing on large-scale motions while neglecting subtle hand articulations that are crucial for fine-grained recognition. This work presents a probabilistic dual-stream framework that unifies reliability modeling and multi-modal integration, generalizing expertized learning under uncertainty across both intra-skeleton and cross-modal domains. The framework comprises three key components: (1) a calibration-free preprocessing pipeline that removes canonical-space transformations and learns directly from native coordinates; (2) a probabilistic Noisy-OR fusion that stabilizes reliability-aware dual-stream learning without requiring explicit confidence supervision; and (3) an intra- to cross-modal ensemble that couples four skeleton modalities (Joint, Bone, Joint Motion, and Bone Motion) to RGB representations, bridging structural and visual motion cues in a unified cross-modal formulation. Comprehensive evaluations across multiple benchmarks (NTU RGB+D~60/120, PKU-MMD, N-UCLA) and a newly defined hand-centric benchmark exhibit consistent improvements and robustness under noisy and heterogeneous conditions.", "date": "2026-01-01", "categories": []}
{"id": "2512.25067", "title": "FineTec: Fine-Grained Action Recognition Under Temporal Corruption via Skeleton Decomposition and Sequence Completion", "abstract": "Recognizing fine-grained actions from temporally corrupted skeleton sequences remains a significant challenge, particularly in real-world scenarios where online pose estimation often yields substantial missing data. Existing methods often struggle to accurately recover temporal dynamics and fine-grained spatial structures, resulting in the loss of subtle motion cues crucial for distinguishing similar actions. To address this, we propose FineTec, a unified framework for Fine-grained action recognition under Temporal Corruption. FineTec first restores a base skeleton sequence from corrupted input using context-aware completion with diverse temporal masking. Next, a skeleton-based spatial decomposition module partitions the skeleton into five semantic regions, further divides them into dynamic and static subgroups based on motion variance, and generates two augmented skeleton sequences via targeted perturbation. These, along with the base sequence, are then processed by a physics-driven estimation module, which utilizes Lagrangian dynamics to estimate joint accelerations. Finally, both the fused skeleton position sequence and the fused acceleration sequence are jointly fed into a GCN-based action recognition head. Extensive experiments on both coarse-grained (NTU-60, NTU-120) and fine-grained (Gym99, Gym288) benchmarks show that FineTec significantly outperforms previous methods under various levels of temporal corruption. Specifically, FineTec achieves top-1 accuracies of 89.1% and 78.1% on the challenging Gym99-severe and Gym288-severe settings, respectively, demonstrating its robustness and generalizability. Code and datasets could be found at https://smartdianlab.github.io/projects-FineTec/.", "date": "2025-12-31", "categories": []}
{"id": "2504.12039", "title": "RadMamba: Efficient Human Activity Recognition through Radar-based Micro-Doppler-Oriented Mamba State-Space Model", "abstract": "Radar-based Human Activity Recognition (HAR) is an attractive alternative to wearables and cameras because it preserves privacy, and is contactless and robust to occlusions. However, dominant Convolutional Neural Network (CNN)- and Recurrent Neural Network (RNN)-based solutions are computationally intensive at deployment, and recent lightweight Vision Transformer (ViT) and State Space Model (SSM) variants still exhibit substantial complexity. In this paper, we present RadMamba, a parameter-efficient, micro-Doppler-oriented Mamba SSM tailored to radar HAR under on-sensor compute, latency, and energy constraints typical of distributed radar systems. RadMamba combines (i) channel fusion with downsampling, (ii) Doppler-aligned segmentation that preserves the physical continuity of Doppler over time, and (iii) convolutional token projections that better capture Doppler-span variations, thereby retaining temporal-Doppler structure while reducing the number of Floating-point Operations per Inference (#FLOP/Inf.). Evaluated across three datasets with different radars and types of activities, RadMamba matches the prior best 99.8% accuracy of a recent SSM-based model on the Continuous Wave (CW) radar dataset, while requiring only 1/400 of its parameters. On a dataset of non-continuous activities with Frequency Modulated Continuous Wave (FMCW) radar, RadMamba remains competitive with leading 92.0% results using about 1/10 of the parameters, and on a continuous FMCW radar dataset it surpasses methods with far more parameters by at least 3%, using only 6.7k parameters. Code: https://github.com/lab-emi/AIRHAR.", "date": "2025-04-16", "categories": []}
{"id": "2406.02978", "title": "Self-Supervised Skeleton-Based Action Representation Learning: A Benchmark and Beyond", "abstract": "Self-supervised learning (SSL), which aims to learn meaningful prior representations from unlabeled data, has been proven effective for skeleton-based action understanding. Different from the image domain, skeleton data possesses sparser spatial structures and diverse representation forms, with the absence of background clues and the additional temporal dimension, presenting new challenges for spatial-temporal motion pretext task design. Recently, many endeavors have been made for skeleton-based SSL, achieving remarkable progress. However, a systematic and thorough review is still lacking. In this paper, we conduct, for the first time, a comprehensive survey on self-supervised skeleton-based action representation learning. Following the taxonomy of context-based, generative learning, and contrastive learning approaches, we make a thorough review and benchmark of existing works and shed light on the future possible directions. Remarkably, our investigation demonstrates that most SSL works rely on the single paradigm, learning representations of a single level, and are evaluated on the action recognition task solely, which leaves the generalization power of skeleton SSL models under-explored. To this end, a novel and effective SSL method for skeleton is further proposed, which integrates versatile representation learning objectives of different granularity, substantially boosting the generalization capacity for multiple skeleton downstream tasks. Extensive experiments under three large-scale datasets demonstrate our method achieves superior generalization performance on various downstream tasks, including recognition, retrieval, detection, and few-shot learning.", "date": "2024-06-05", "categories": []}
{"id": "2504.13460", "title": "Chain-of-Evidence Multimodal Reasoning for Few-shot Temporal Action Localization", "abstract": "Traditional temporal action localization (TAL) methods rely on large amounts of detailed annotated data, whereas few-shot TAL reduces this dependence by using only a few training samples to identify unseen action categories. However, existing few-shot TAL methods typically focus solely on video-level information, neglecting textual information, which can provide valuable semantic support for the action localization task. To address these issues, in this work, we propose a new few-shot temporal action localization method by Chain-of-Evidence multimodal reasoning to improve localization performance. Specifically, we design a novel few-shot learning framework to capture action commonalities and variations, which includes a semantic-aware text-visual alignment module designed to align the query and support videos at different levels. Meanwhile, to better express the temporal dependencies and causal relationships between actions at the textual level, we design a Chain-of-Evidence (CoE) reasoning method that progressively guides the Vision Language Model (VLM) and Large Language Model (LLM) to generate CoE text descriptions for videos. The generated texts can capture more variance of action than visual features. We conduct extensive experiments on the publicly available ActivityNet1.3, THUMOS14 and our newly collected Human-related Anomaly Localization Dataset. The experimental results demonstrate that our proposed method significantly outperforms existing methods in single-instance and multi-instance scenarios. Our source code and data are available at https://github.com/MICLAB-BUPT/VAL-VLM.", "date": "2025-04-18", "categories": []}
{"id": "2512.21916", "title": "Patch as Node: Human-Centric Graph Representation Learning for Multimodal Action Recognition", "abstract": "While human action recognition has witnessed notable achievements, multimodal methods fusing RGB and skeleton modalities still suffer from their inherent heterogeneity and fail to fully exploit the complementary potential between them. In this paper, we propose PAN, the first human-centric graph representation learning framework for multimodal action recognition, in which token embeddings of RGB patches containing human joints are represented as spatiotemporal graphs. The human-centric graph modeling paradigm suppresses the redundancy in RGB frames and aligns well with skeleton-based methods, thus enabling a more effective and semantically coherent fusion of multimodal features. Since the sampling of token embeddings heavily relies on 2D skeletal data, we further propose attention-based post calibration to reduce the dependency on high-quality skeletal data at a minimal cost interms of model performance. To explore the potential of PAN in integrating with skeleton-based methods, we present two variants: PAN-Ensemble, which employs dual-path graph convolution networks followed by late fusion, and PAN-Unified, which performs unified graph representation learning within a single network. On three widely used multimodal action recognition datasets, both PAN-Ensemble and PAN-Unified achieve state-of-the-art (SOTA) performance in their respective settings of multimodal fusion: separate and unified modeling, respectively.", "date": "2025-12-26", "categories": []}
{"id": "2511.12870", "title": "View-aware Cross-modal Distillation for Multi-view Action Recognition", "abstract": "The widespread use of multi-sensor systems has increased research in multi-view action recognition. While existing approaches in multi-view setups with fully overlapping sensors benefit from consistent view coverage, partially overlapping settings where actions are visible in only a subset of views remain underexplored. This challenge becomes more severe in real-world scenarios, as many systems provide only limited input modalities and rely on sequence-level annotations instead of dense frame-level labels. In this study, we propose View-aware Cross-modal Knowledge Distillation (ViCoKD), a framework that distills knowledge from a fully supervised multi-modal teacher to a modality- and annotation-limited student. ViCoKD employs a cross-modal adapter with cross-modal attention, allowing the student to exploit multi-modal correlations while operating with incomplete modalities. Moreover, we propose a View-aware Consistency module to address view misalignment, where the same action may appear differently or only partially across viewpoints. It enforces prediction alignment when the action is co-visible across views, guided by human-detection masks and confidence-weighted Jensen-Shannon divergence between their predicted class distributions. Experiments on the real-world MultiSensor-Home dataset show that ViCoKD consistently outperforms competitive distillation methods across multiple backbones and environments, delivering significant gains and surpassing the teacher model under limited conditions.", "date": "2025-11-17", "categories": []}
{"id": "2512.19742", "title": "On-device Large Multi-modal Agent for Human Activity Recognition", "abstract": "Human Activity Recognition (HAR) has been an active area of research, with applications ranging from healthcare to smart environments. The recent advancements in Large Language Models (LLMs) have opened new possibilities to leverage their capabilities in HAR, enabling not just activity classification but also interpretability and human-like interaction. In this paper, we present a Large Multi-Modal Agent designed for HAR, which integrates the power of LLMs to enhance both performance and user engagement. The proposed framework not only delivers activity classification but also bridges the gap between technical outputs and user-friendly insights through its reasoning and question-answering capabilities. We conduct extensive evaluations using widely adopted HAR datasets, including HHAR, Shoaib, Motionsense to assess the performance of our framework. The results demonstrate that our model achieves high classification accuracy comparable to state-of-the-art methods while significantly improving interpretability through its reasoning and Q&A capabilities.", "date": "2025-12-17", "categories": []}
{"id": "2512.20104", "title": "Effect of Activation Function and Model Optimizer on the Performance of Human Activity Recognition System Using Various Deep Learning Models", "abstract": "Human Activity Recognition (HAR) plays a vital role in healthcare, surveillance, and innovative environments, where reliable action recognition supports timely decision-making and automation. Although deep learning-based HAR systems are widely adopted, the impact of Activation Functions (AFs) and Model Optimizers (MOs) on performance has not been sufficiently analyzed, particularly regarding how their combinations influence model behavior in practical scenarios. Most existing studies focus on architecture design, while the interaction between AF and MO choices remains relatively unexplored. In this work, we investigate the effect of three commonly used activation functions (ReLU, Sigmoid, and Tanh) combined with four optimization algorithms (SGD, Adam, RMSprop, and Adagrad) using two recurrent deep learning architectures, namely BiLSTM and ConvLSTM. Experiments are conducted on six medically relevant activity classes selected from the HMDB51 and UCF101 datasets, considering their suitability for healthcare-oriented HAR applications. Our experimental results show that ConvLSTM consistently outperforms BiLSTM across both datasets. ConvLSTM, combined with Adam or RMSprop, achieves an accuracy of up to 99.00%, demonstrating strong spatio-temporal learning capabilities and stable performance. While BiLSTM performs reasonably well on UCF101, with accuracy approaching 98.00%, its performance drops to approximately 60.00% on HMDB51, indicating limited robustness across datasets and weaker sensitivity to AF and MO variations. This study provides practical insights for optimizing HAR systems, particularly for real-world healthcare environments where fast and precise activity detection is critical.", "date": "2025-12-23", "categories": []}
{"id": "2512.20409", "title": "DETACH : Decomposed Spatio-Temporal Alignment for Exocentric Video and Ambient Sensors with Staged Learning", "abstract": "Aligning egocentric video with wearable sensors have shown promise for human action recognition, but face practical limitations in user discomfort, privacy concerns, and scalability. We explore exocentric video with ambient sensors as a non-intrusive, scalable alternative. While prior egocentric-wearable works predominantly adopt Global Alignment by encoding entire sequences into unified representations, this approach fails in exocentric-ambient settings due to two problems: (P1) inability to capture local details such as subtle motions, and (P2) over-reliance on modality-invariant temporal patterns, causing misalignment between actions sharing similar temporal patterns with different spatio-semantic contexts. To resolve these problems, we propose DETACH, a decomposed spatio-temporal framework. This explicit decomposition preserves local details, while our novel sensor-spatial features discovered via online clustering provide semantic grounding for context-aware alignment. To align the decomposed features, our two-stage approach establishes spatial correspondence through mutual supervision, then performs temporal alignment via a spatial-temporal weighted contrastive loss that adaptively handles easy negatives, hard negatives, and false negatives. Comprehensive experiments with downstream tasks on Opportunity++ and HWU-USP datasets demonstrate substantial improvements over adapted egocentric-wearable baselines.", "date": "2025-12-23", "categories": []}
{"id": "2512.17953", "title": "Seeing Beyond the Scene: Analyzing and Mitigating Background Bias in Action Recognition", "abstract": "Human action recognition models often rely on background cues rather than human movement and pose to make predictions, a behavior known as background bias. We present a systematic analysis of background bias across classification models, contrastive text-image pretrained models, and Video Large Language Models (VLLM) and find that all exhibit a strong tendency to default to background reasoning. Next, we propose mitigation strategies for classification models and show that incorporating segmented human input effectively decreases background bias by 3.78%. Finally, we explore manual and automated prompt tuning for VLLMs, demonstrating that prompt design can steer predictions towards human-focused reasoning by 9.85%.", "date": "2025-12-17", "categories": []}
{"id": "2512.17983", "title": "Parameter-Efficient Fine-Tuning for HAR: Integrating LoRA and QLoRA into Transformer Models", "abstract": "Human Activity Recognition is a foundational task in pervasive computing. While recent advances in self-supervised learning and transformer-based architectures have significantly improved HAR performance, adapting large pretrained models to new domains remains a practical challenge due to limited computational resources on target devices. This papers investigates parameter-efficient fine-tuning techniques, specifically Low-Rank Adaptation (LoRA) and Quantized LoRA, as scalable alternatives to full model fine-tuning for HAR. We propose an adaptation framework built upon a Masked Autoencoder backbone and evaluate its performance under a Leave-One-Dataset-Out validation protocol across five open HAR datasets. Our experiments demonstrate that both LoRA and QLoRA can match the recognition performance of full fine-tuning while significantly reducing the number of trainable parameters, memory usage, and training time. Further analyses reveal that LoRA maintains robust performance even under limited supervision and that the adapter rank provides a controllable trade-off between accuracy and efficiency. QLoRA extends these benefits by reducing the memory footprint of frozen weights through quantization, with minimal impact on classification quality.", "date": "2025-12-19", "categories": []}
{"id": "2512.18750", "title": "Context-Aware Network Based on Multi-scale Spatio-temporal Attention for Action Recognition in Videos", "abstract": "Action recognition is a critical task in video understanding, requiring the comprehensive capture of spatio-temporal cues across various scales. However, existing methods often overlook the multi-granularity nature of actions. To address this limitation, we introduce the Context-Aware Network (CAN). CAN consists of two core modules: the Multi-scale Temporal Cue Module (MTCM) and the Group Spatial Cue Module (GSCM). MTCM effectively extracts temporal cues at multiple scales, capturing both fast-changing motion details and overall action flow. GSCM, on the other hand, extracts spatial cues at different scales by grouping feature maps and applying specialized extraction methods to each group. Experiments conducted on five benchmark datasets (Something-Something V1 and V2, Diving48, Kinetics-400, and UCF101) demonstrate the effectiveness of CAN. Our approach achieves competitive performance, outperforming most mainstream methods, with accuracies of 50.4% on Something-Something V1, 63.9% on Something-Something V2, 88.4% on Diving48, 74.9% on Kinetics-400, and 86.9% on UCF101. These results highlight the importance of capturing multi-scale spatio-temporal cues for robust action recognition.", "date": "2025-12-21", "categories": []}
{"id": "2512.19036", "title": "Distinguishing Visually Similar Actions: Prompt-Guided Semantic Prototype Modulation for Few-Shot Action Recognition", "abstract": "Few-shot action recognition aims to enable models to quickly learn new action categories from limited labeled samples, addressing the challenge of data scarcity in real-world applications. Current research primarily addresses three core challenges: (1) temporal modeling, where models are prone to interference from irrelevant static background information and struggle to capture the essence of dynamic action features; (2) visual similarity, where categories with subtle visual differences are difficult to distinguish; and (3) the modality gap between visual-textual support prototypes and visual-only queries, which complicates alignment within a shared embedding space. To address these challenges, this paper proposes a CLIP-SPM framework, which includes three components: (1) the Hierarchical Synergistic Motion Refinement (HSMR) module, which aligns deep and shallow motion features to improve temporal modeling by reducing static background interference; (2) the Semantic Prototype Modulation (SPM) strategy, which generates query-relevant text prompts to bridge the modality gap and integrates them with visual features, enhancing the discriminability between similar actions; and (3) the Prototype-Anchor Dual Modulation (PADM) method, which refines support prototypes and aligns query features with a global semantic anchor, improving consistency across support and query samples. Comprehensive experiments across standard benchmarks, including Kinetics, SSv2-Full, SSv2-Small, UCF101, and HMDB51, demonstrate that our CLIP-SPM achieves competitive performance under 1-shot, 3-shot, and 5-shot settings. Extensive ablation studies and visual analyses further validate the effectiveness of each component and its contributions to addressing the core challenges. The source code and models are publicly available at GitHub.", "date": "2025-12-22", "categories": []}
{"id": "2209.14711", "title": "Low-Resolution Action Recognition for Tiny Actions Challenge", "abstract": "Tiny Actions Challenge focuses on understanding human activities in real-world surveillance. Basically, there are two main difficulties for activity recognition in this scenario. First, human activities are often recorded at a distance, and appear in a small resolution without much discriminative clue. Second, these activities are naturally distributed in a long-tailed way. It is hard to alleviate data bias for such heavy category imbalance. To tackle these problems, we propose a comprehensive recognition solution in this paper. First, we train video backbones with data balance, in order to alleviate overfitting in the challenge benchmark. Second, we design a dual-resolution distillation framework, which can effectively guide low-resolution action recognition by super-resolution knowledge. Finally, we apply model en-semble with post-processing, which can further boost per-formance on the long-tailed categories. Our solution ranks Top-1 on the leaderboard.", "date": "2022-09-28", "categories": []}
{"id": "2402.18951", "title": "Percept, Chat, and then Adapt: Multimodal Knowledge Transfer of Foundation Models for Open-World Video Recognition", "abstract": "Open-world video recognition is challenging since traditional networks are not generalized well on complex environment variations. Alternatively, foundation models with rich knowledge have recently shown their generalization power. However, how to apply such knowledge has not been fully explored for open-world video recognition. To this end, we propose a generic knowledge transfer pipeline, which progressively exploits and integrates external multimodal knowledge from foundation models to boost open-world video recognition. We name it PCA, based on three stages of Percept, Chat, and Adapt. First, we perform Percept process to reduce the video domain gap and obtain external visual knowledge. Second, we generate rich linguistic semantics as external textual knowledge in Chat stage. Finally, we blend external multimodal knowledge in Adapt stage, by inserting multimodal knowledge adaptation modules into networks. We conduct extensive experiments on three challenging open-world video benchmarks, i.e., TinyVIRAT, ARID, and QV-Pipe. Our approach achieves state-of-the-art performance on all three datasets.", "date": "2024-02-29", "categories": []}
{"id": "2501.06138", "title": "MS-Temba: Multi-Scale Temporal Mamba for Understanding Long Untrimmed Videos", "abstract": "Temporal Action Detection (TAD) in untrimmed videos poses significant challenges, particularly for Activities of Daily Living (ADL) requiring models to (1) process long-duration videos, (2) capture temporal variations in actions, and (3) simultaneously detect dense overlapping actions. Existing CNN and Transformer-based approaches, struggle to jointly capture fine-grained detail and long-range structure at scale. State-space Model (SSM) based Mamba offers powerful long-range modeling, but naive application to TAD collapses fine-grained temporal structure and fails to account for the challenges inherent to TAD. To this end, we propose Multi-Scale Temporal Mamba (MS-Temba), which extends Mamba to TAD with newly introduced dilated SSMs. Each Temba block, comprising dilated SSMs coupled with our proposed additional losses, enables the learning of discriminative representations across temporal scales. A lightweight Multi-scale Mamba Fuser then unifies these multi-scale features via SSM-based aggregation, yielding precise action-boundary localization. With only 17M parameters, MS-Temba achieves state-of-the-art performance on densely labeled ADL benchmarks TSU & Charades, and further generalizes to long-form video summarization, setting new state-of-the-art results on TVSum & SumMe.", "date": "2025-01-10", "categories": []}
{"id": "2512.11340", "title": "Task-Specific Distance Correlation Matching for Few-Shot Action Recognition", "abstract": "Few-shot action recognition (FSAR) has recently made notable progress through set matching and efficient adaptation of large-scale pre-trained models. However, two key limitations persist. First, existing set matching metrics typically rely on cosine similarity to measure inter-frame linear dependencies and then perform matching with only instance-level information, thus failing to capture more complex patterns such as nonlinear relationships and overlooking task-specific cues. Second, for efficient adaptation of CLIP to FSAR, recent work performing fine-tuning via skip-fusion layers (which we refer to as side layers) has significantly reduced memory cost. However, the newly introduced side layers are often difficult to optimize under limited data conditions. To address these limitations, we propose TS-FSAR, a framework comprising three components: (1) a visual Ladder Side Network (LSN) for efficient CLIP fine-tuning; (2) a metric called Task-Specific Distance Correlation Matching (TS-DCM), which uses $\\alpha$-distance correlation to model both linear and nonlinear inter-frame dependencies and leverages a task prototype to enable task-specific matching; and (3) a Guiding LSN with Adapted CLIP (GLAC) module, which regularizes LSN using the adapted frozen CLIP to improve training for better $\\alpha$-distance correlation estimation under limited supervision. Extensive experiments on five widely-used benchmarks demonstrate that our TS-FSAR yields superior performance compared to prior state-of-the-arts.", "date": "2025-12-12", "categories": []}
{"id": "2512.12013", "title": "Exploring Spatial-Temporal Representation via Star Graph for mmWave Radar-based Human Activity Recognition", "abstract": "Human activity recognition (HAR) requires extracting accurate spatial-temporal features with human movements. A mmWave radar point cloud-based HAR system suffers from sparsity and variable-size problems due to the physical features of the mmWave signal. Existing works usually borrow the preprocessing algorithms for the vision-based systems with dense point clouds, which may not be optimal for mmWave radar systems. In this work, we proposed a graph representation with a discrete dynamic graph neural network (DDGNN) to explore the spatial-temporal representation of human movement-related features. Specifically, we designed a star graph to describe the high-dimensional relative relationship between a manually added static center point and the dynamic mmWave radar points in the same and consecutive frames. We then adopted DDGNN to learn the features residing in the star graph with variable sizes. Experimental results demonstrated that our approach outperformed other baseline methods using real-world HAR datasets. Our system achieved an overall classification accuracy of 94.27\\%, which gets the near-optimal performance with a vision-based skeleton data accuracy of 97.25\\%. We also conducted an inference test on Raspberry Pi~4 to demonstrate its effectiveness on resource-constraint platforms. \\sh{ We provided a comprehensive ablation study for variable DDGNN structures to validate our model design. Our system also outperformed three recent radar-specific methods without requiring resampling or frame aggregators.", "date": "2025-12-12", "categories": []}
{"id": "2512.12586", "title": "StegaVAR: Privacy-Preserving Video Action Recognition via Steganographic Domain Analysis", "abstract": "Despite the rapid progress of deep learning in video action recognition (VAR) in recent years, privacy leakage in videos remains a critical concern. Current state-of-the-art privacy-preserving methods often rely on anonymization. These methods suffer from (1) low concealment, where producing visually distorted videos that attract attackers' attention during transmission, and (2) spatiotemporal disruption, where degrading essential spatiotemporal features for accurate VAR. To address these issues, we propose StegaVAR, a novel framework that embeds action videos into ordinary cover videos and directly performs VAR in the steganographic domain for the first time. Throughout both data transmission and action analysis, the spatiotemporal information of hidden secret video remains complete, while the natural appearance of cover videos ensures the concealment of transmission. Considering the difficulty of steganographic domain analysis, we propose Secret Spatio-Temporal Promotion (STeP) and Cross-Band Difference Attention (CroDA) for analysis within the steganographic domain. STeP uses the secret video to guide spatiotemporal feature extraction in the steganographic domain during training. CroDA suppresses cover interference by capturing cross-band semantic differences. Experiments demonstrate that StegaVAR achieves superior VAR and privacy-preserving performance on widely used datasets. Moreover, our framework is effective for multiple steganographic models.", "date": "2025-12-14", "categories": []}
{"id": "2403.01543", "title": "Efficient Action Counting with Dynamic Queries", "abstract": "Temporal repetition counting aims to quantify the repeated action cycles within a video. The majority of existing methods rely on the similarity correlation matrix to characterize the repetitiveness of actions, but their scalability is hindered due to the quadratic computational complexity. In this work, we introduce a novel approach that employs an action query representation to localize repeated action cycles with linear computational complexity. Based on this representation, we further develop two key components to tackle the essential challenges of temporal repetition counting. Firstly, to facilitate open-set action counting, we propose the dynamic update scheme on action queries. Unlike static action queries, this approach dynamically embeds video features into action queries, offering a more flexible and generalizable representation. Secondly, to distinguish between actions of interest and background noise actions, we incorporate inter-query contrastive learning to regularize the video representations corresponding to different action queries. As a result, our method significantly outperforms previous works, particularly in terms of long video sequences, unseen actions, and actions at various speeds. On the challenging RepCountA benchmark, we outperform the state-of-the-art method TransRAC by 26.5% in OBO accuracy, with a 22.7% mean error decrease and 94.1% computational burden reduction. Code is available at https://github.com/lizishi/DeTRC.", "date": "2024-03-03", "categories": []}
{"id": "2503.15126", "title": "Text-Derived Relational Graph-Enhanced Network for Skeleton-Based Action Segmentation", "abstract": "Skeleton-based Temporal Action Segmentation (STAS) aims to segment and recognize various actions from long, untrimmed sequences of human skeletal movements. Current STAS methods typically employ spatio-temporal modeling to establish dependencies among joints as well as frames, and utilize one-hot encoding with cross-entropy loss for frame-wise classification supervision. However, these methods overlook the intrinsic correlations among joints and actions within skeletal features, leading to a limited understanding of human movements. To address this, we propose a Text-Derived Relational Graph-Enhanced Network (TRG-Net) that leverages prior graphs generated by Large Language Models (LLM) to enhance both modeling and supervision. For modeling, the Dynamic Spatio-Temporal Fusion Modeling (DSFM) method incorporates Text-Derived Joint Graphs (TJG) with channel- and frame-level dynamic adaptation to effectively model spatial relations, while integrating spatio-temporal core features during temporal modeling. For supervision, the Absolute-Relative Inter-Class Supervision (ARIS) method employs contrastive learning between action features and text embeddings to regularize the absolute class distributions, and utilizes Text-Derived Action Graphs (TAG) to capture the relative inter-class relationships among action features. Additionally, we propose a Spatial-Aware Enhancement Processing (SAEP) method, which incorporates random joint occlusion and axial rotation to enhance spatial generalization. Performance evaluations on four public datasets demonstrate that TRG-Net achieves state-of-the-art results.", "date": "2025-03-19", "categories": []}
{"id": "2512.11189", "title": "Multi-task Learning with Extended Temporal Shift Module for Temporal Action Localization", "abstract": "We present our solution to the BinEgo-360 Challenge at ICCV 2025, which focuses on temporal action localization (TAL) in multi-perspective and multi-modal video settings. The challenge provides a dataset containing panoramic, third-person, and egocentric recordings, annotated with fine-grained action classes. Our approach is built on the Temporal Shift Module (TSM), which we extend to handle TAL by introducing a background class and classifying fixed-length non-overlapping intervals. We employ a multi-task learning framework that jointly optimizes for scene classification and TAL, leveraging contextual cues between actions and environments. Finally, we integrate multiple models through a weighted ensemble strategy, which improves robustness and consistency of predictions. Our method is ranked first in both the initial and extended rounds of the competition, demonstrating the effectiveness of combining multi-task learning, an efficient backbone, and ensemble learning for TAL.", "date": "2025-12-12", "categories": []}
{"id": "2512.11350", "title": "Surveillance Video-Based Traffic Accident Detection Using Transformer Architecture", "abstract": "Road traffic accidents represent a leading cause of mortality globally, with incidence rates rising due to increasing population, urbanization, and motorization. Rising accident rates raise concerns about traffic surveillance effectiveness. Traditional computer vision methods for accident detection struggle with limited spatiotemporal understanding and poor cross-domain generalization. Recent advances in transformer architectures excel at modeling global spatial-temporal dependencies and parallel computation. However, applying these models to automated traffic accident detection is limited by small, non-diverse datasets, hindering the development of robust, generalizable systems. To address this gap, we curated a comprehensive and balanced dataset that captures a wide spectrum of traffic environments, accident types, and contextual variations. Utilizing the curated dataset, we propose an accident detection model based on a transformer architecture using pre-extracted spatial video features. The architecture employs convolutional layers to extract local correlations across diverse patterns within a frame, while leveraging transformers to capture sequential-temporal dependencies among the retrieved features. Moreover, most existing studies neglect the integration of motion cues, which are essential for understanding dynamic scenes, especially during accidents. These approaches typically rely on static features or coarse temporal information. In this study, multiple methods for incorporating motion cues were evaluated to identify the most effective strategy. Among the tested input approaches, concatenating RGB features with optical flow achieved the highest accuracy at 88.3%. The results were further compared with vision language models (VLM) such as GPT, Gemini, and LLaVA-NeXT-Video to assess the effectiveness of the proposed method.", "date": "2025-12-12", "categories": []}
{"id": "2512.11458", "title": "Boosting Skeleton-based Zero-Shot Action Recognition with Training-Free Test-Time Adaptation", "abstract": "We introduce Skeleton-Cache, the first training-free test-time adaptation framework for skeleton-based zero-shot action recognition (SZAR), aimed at improving model generalization to unseen actions during inference. Skeleton-Cache reformulates inference as a lightweight retrieval process over a non-parametric cache that stores structured skeleton representations, combining both global and fine-grained local descriptors. To guide the fusion of descriptor-wise predictions, we leverage the semantic reasoning capabilities of large language models (LLMs) to assign class-specific importance weights. By integrating these structured descriptors with LLM-guided semantic priors, Skeleton-Cache dynamically adapts to unseen actions without any additional training or access to training data. Extensive experiments on NTU RGB+D 60/120 and PKU-MMD II demonstrate that Skeleton-Cache consistently boosts the performance of various SZAR backbones under both zero-shot and generalized zero-shot settings. The code is publicly available at https://github.com/Alchemist0754/Skeleton-Cache.", "date": "2025-12-12", "categories": []}
{"id": "2512.11503", "title": "TSkel-Mamba: Temporal Dynamic Modeling via State Space Model for Human Skeleton-based Action Recognition", "abstract": "Skeleton-based action recognition has garnered significant attention in the computer vision community. Inspired by the recent success of the selective state-space model (SSM) Mamba in modeling 1D temporal sequences, we propose TSkel-Mamba, a hybrid Transformer-Mamba framework that effectively captures both spatial and temporal dynamics. In particular, our approach leverages Spatial Transformer for spatial feature learning while utilizing Mamba for temporal modeling. Mamba, however, employs separate SSM blocks for individual channels, which inherently limits its ability to model inter-channel dependencies. To better adapt Mamba for skeleton data and enhance Mamba`s ability to model temporal dependencies, we introduce a Temporal Dynamic Modeling (TDM) block, which is a versatile plug-and-play component that integrates a novel Multi-scale Temporal Interaction (MTI) module. The MTI module employs multi-scale Cycle operators to capture cross-channel temporal interactions, a critical factor in action recognition. Extensive experiments on NTU-RGB+D 60, NTU-RGB+D 120, NW-UCLA and UAV-Human datasets demonstrate that TSkel-Mamba achieves state-of-the-art performance while maintaining low inference time, making it both efficient and highly effective.", "date": "2025-12-12", "categories": []}
{"id": "2507.00752", "title": "Multi-Modal Graph Convolutional Network with Sinusoidal Encoding for Robust Human Action Segmentation", "abstract": "Accurate temporal segmentation of human actions is critical for intelligent robots in collaborative settings, where a precise understanding of sub-activity labels and their temporal structure is essential. However, the inherent noise in both human pose estimation and object detection often leads to over-segmentation errors, disrupting the coherence of action sequences. To address this, we propose a Multi-Modal Graph Convolutional Network (MMGCN) that integrates low-frame-rate (e.g., 1 fps) visual data with high-frame-rate (e.g., 30 fps) motion data (skeleton and object detections) to mitigate fragmentation. Our framework introduces three key contributions. First, a sinusoidal encoding strategy that maps 3D skeleton coordinates into a continuous sin-cos space to enhance spatial representation robustness. Second, a temporal graph fusion module that aligns multi-modal inputs with differing resolutions via hierarchical feature aggregation, Third, inspired by the smooth transitions inherent to human actions, we design SmoothLabelMix, a data augmentation technique that mixes input sequences and labels to generate synthetic training examples with gradual action transitions, enhancing temporal consistency in predictions and reducing over-segmentation artifacts.\n  Extensive experiments on the Bimanual Actions Dataset, a public benchmark for human-object interaction understanding, demonstrate that our approach outperforms state-of-the-art methods, especially in action segmentation accuracy, achieving F1@10: 94.5% and F1@25: 92.8%.", "date": "2025-07-01", "categories": []}
{"id": "2507.00756", "title": "Towards Open-World Human Action Segmentation Using Graph Convolutional Networks", "abstract": "Human-object interaction segmentation is a fundamental task of daily activity understanding, which plays a crucial role in applications such as assistive robotics, healthcare, and autonomous systems. Most existing learning-based methods excel in closed-world action segmentation, they struggle to generalize to open-world scenarios where novel actions emerge. Collecting exhaustive action categories for training is impractical due to the dynamic diversity of human activities, necessitating models that detect and segment out-of-distribution actions without manual annotation. To address this issue, we formally define the open-world action segmentation problem and propose a structured framework for detecting and segmenting unseen actions. Our framework introduces three key innovations: 1) an Enhanced Pyramid Graph Convolutional Network (EPGCN) with a novel decoder module for robust spatiotemporal feature upsampling. 2) Mixup-based training to synthesize out-of-distribution data, eliminating reliance on manual annotations. 3) A novel Temporal Clustering loss that groups in-distribution actions while distancing out-of-distribution samples.\n  We evaluate our framework on two challenging human-object interaction recognition datasets: Bimanual Actions and 2 Hands and Object (H2O) datasets. Experimental results demonstrate significant improvements over state-of-the-art action segmentation models across multiple open-set evaluation metrics, achieving 16.9% and 34.6% relative gains in open-set segmentation (F1@50) and out-of-distribution detection performances (AUROC), respectively. Additionally, we conduct an in-depth ablation study to assess the impact of each proposed component, identifying the optimal framework configuration for open-world action segmentation.", "date": "2025-07-01", "categories": []}
{"id": "2512.07136", "title": "A Large-Scale Multimodal Dataset and Benchmarks for Human Activity Scene Understanding and Reasoning", "abstract": "Multimodal human action recognition (HAR) leverages complementary sensors for activity classification. Beyond recognition, recent advances in large language models (LLMs) enable detailed descriptions and causal reasoning, motivating new tasks: human action understanding (HAU) and human action reasoning (HARn). However, most LLMs, especially large vision language models (LVLMs), struggle with non-RGB modalities such as depth, IMU, and mmWave due to the lack of large-scale data-caption resources. Existing HAR datasets mainly provide coarse data-label annotations, which are insufficient to capture fine-grained action dynamics needed for HAU and HARn. We consider two ground-truth pair types: (1) data label (discrete category) and (2) data caption (textual description). Naively generating captions from labels often lacks logical and spatiotemporal consistency. We introduce CUHK-X, a large-scale multimodal dataset and benchmark suite for HAR, HAU, and HARn. CUHK-X contains 58,445 samples covering 40 actions performed by 30 participants across two indoor environments. To improve caption consistency, we propose a prompt-based scene creation method that leverages LLMs to generate logically connected activity sequences, followed by human validation. CUHK-X includes three benchmarks with six evaluation tasks. Experiments report average accuracies of 76.52% (HAR), 40.76% (HAU), and 70.25% (HARn). CUHK-X aims to enable the community to apply and develop data-intensive learning methods for robust, multimodal human activity analysis. Project page and code: https://openaiotlab.github.io/CUHK-X/ and https://github.com/openaiotlab/CUHK-X.", "date": "2025-12-08", "categories": []}
{"id": "2512.07729", "title": "Improving action classification with brain-inspired deep networks", "abstract": "Action recognition is also key for applications ranging from robotics to healthcare monitoring. Action information can be extracted from the body pose and movements, as well as from the background scene. However, the extent to which deep neural networks (DNNs) make use of information about the body and information about the background remains unclear. Since these two sources of information may be correlated within a training dataset, DNNs might learn to rely predominantly on one of them, without taking full advantage of the other. Unlike DNNs, humans have domain-specific brain regions selective for perceiving bodies, and regions selective for perceiving scenes. The present work tests whether humans are thus more effective at extracting information from both body and background, and whether building brain-inspired deep network architectures with separate domain-specific streams for body and scene perception endows them with more human-like performance. We first demonstrate that DNNs trained using the HAA500 dataset perform almost as accurately on versions of the stimuli that show both body and background and on versions of the stimuli from which the body was removed, but are at chance-level for versions of the stimuli from which the background was removed. Conversely, human participants (N=28) can recognize the same set of actions accurately with all three versions of the stimuli, and perform significantly better on stimuli that show only the body than on stimuli that show only the background. Finally, we implement and test a novel architecture patterned after domain specificity in the brain with separate streams to process body and background information. We show that 1) this architecture improves action recognition performance, and 2) its accuracy across different versions of the stimuli follows a pattern that matches more closely the pattern of accuracy observed in human participants.", "date": "2025-12-08", "categories": []}
{"id": "2504.06580", "title": "Exploring Ordinal Bias in Action Recognition for Instructional Videos", "abstract": "Action recognition models have achieved promising results in understanding instructional videos. However, they often rely on dominant, dataset-specific action sequences rather than true video comprehension, a problem that we define as ordinal bias. To address this issue, we propose two effective video manipulation methods: Action Masking, which masks frames of frequently co-occurring actions, and Sequence Shuffling, which randomizes the order of action segments. Through comprehensive experiments, we demonstrate that current models exhibit significant performance drops when confronted with nonstandard action sequences, underscoring their vulnerability to ordinal bias. Our findings emphasize the importance of rethinking evaluation strategies and developing models capable of generalizing beyond fixed action patterns in diverse instructional videos.", "date": "2025-04-09", "categories": []}
{"id": "2507.07381", "title": "Multi-Focus Temporal Shifting for Precise Event Spotting in Sports Videos", "abstract": "Precise Event Spotting (PES) in sports videos requires frame-level recognition of fine-grained actions from single-camera footage. Existing PES models typically incorporate lightweight temporal modules such as the Gate Shift Module (GSM) or the Gate Shift Fuse to enrich 2D CNN feature extractors with temporal context. However, these modules are limited in both temporal receptive field and spatial adaptability. We propose Multi-Focus Temporal Shifting Module (MFS) that enhances GSM with multi-scale temporal shifts and Group Focus Module, enabling efficient modeling of both short and long-term dependencies while focusing on salient regions. MFS is a lightweight, plug-and-play module that integrates seamlessly with diverse 2D backbones. To further advance the field, we introduce the Table Tennis Australia dataset, the first PES benchmark for table tennis containing over 4,800 precisely annotated events. Extensive experiments across five PES benchmarks demonstrate that MFS consistently improves performance with minimal overhead, achieving leading results among lightweight methods (+4.09 mAP, 45 GFLOPs).", "date": "2025-07-10", "categories": []}
{"id": "2512.04943", "title": "Towards Adaptive Fusion of Multimodal Deep Networks for Human Action Recognition", "abstract": "This study introduces a pioneering methodology for human action recognition by harnessing deep neural network techniques and adaptive fusion strategies across multiple modalities, including RGB, optical flows, audio, and depth information. Employing gating mechanisms for multimodal fusion, we aim to surpass limitations inherent in traditional unimodal recognition methods while exploring novel possibilities for diverse applications. Through an exhaustive investigation of gating mechanisms and adaptive weighting-based fusion architectures, our methodology enables the selective integration of relevant information from various modalities, thereby bolstering both accuracy and robustness in action recognition tasks. We meticulously examine various gated fusion strategies to pinpoint the most effective approach for multimodal action recognition, showcasing its superiority over conventional unimodal methods. Gating mechanisms facilitate the extraction of pivotal features, resulting in a more holistic representation of actions and substantial enhancements in recognition performance. Our evaluations across human action recognition, violence action detection, and multiple self-supervised learning tasks on benchmark datasets demonstrate promising advancements in accuracy. The significance of this research lies in its potential to revolutionize action recognition systems across diverse fields. The fusion of multimodal information promises sophisticated applications in surveillance and human-computer interaction, especially in contexts related to active assisted living.", "date": "2025-12-04", "categories": []}
{"id": "2510.08635", "title": "Hi-OSCAR: Hierarchical Open-set Classifier for Human Activity Recognition", "abstract": "Within Human Activity Recognition (HAR), there is an insurmountable gap between the range of activities performed in life and those that can be captured in an annotated sensor dataset used in training. Failure to properly handle unseen activities seriously undermines any HAR classifier's reliability. Additionally within HAR, not all classes are equally dissimilar, some significantly overlap or encompass other sub-activities. Based on these observations, we arrange activity classes into a structured hierarchy. From there, we propose Hi-OSCAR: a Hierarchical Open-set Classifier for Activity Recognition, that can identify known activities at state-of-the-art accuracy while simultaneously rejecting unknown activities. This not only enables open-set classification, but also allows for unknown classes to be localized to the nearest internal node, providing insight beyond a binary \"known/unknown\" classification. To facilitate this and future open-set HAR research, we collected a new dataset: NFI_FARED. NFI_FARED contains data from multiple subjects performing nineteen activities from a range of contexts, including daily living, commuting, and rapid movements, which is fully public and available for download.", "date": "2025-10-08", "categories": []}
{"id": "2512.03837", "title": "Heatmap Pooling Network for Action Recognition from RGB Videos", "abstract": "Human action recognition (HAR) in videos has garnered widespread attention due to the rich information in RGB videos. Nevertheless, existing methods for extracting deep features from RGB videos face challenges such as information redundancy, susceptibility to noise and high storage costs. To address these issues and fully harness the useful information in videos, we propose a novel heatmap pooling network (HP-Net) for action recognition from videos, which extracts information-rich, robust and concise pooled features of the human body in videos through a feedback pooling module. The extracted pooled features demonstrate obvious performance advantages over the previously obtained pose data and heatmap features from videos. In addition, we design a spatial-motion co-learning module and a text refinement modulation module to integrate the extracted pooled features with other multimodal data, enabling more robust action recognition. Extensive experiments on several benchmarks namely NTU RGB+D 60, NTU RGB+D 120, Toyota-Smarthome and UAV-Human consistently verify the effectiveness of our HP-Net, which outperforms the existing human action recognition methods. Our code is publicly available at: https://github.com/liujf69/HPNet-Action.", "date": "2025-12-03", "categories": []}
{"id": "2305.19478", "title": "Permutation-Aware Action Segmentation via Unsupervised Frame-to-Segment Alignment", "abstract": "This paper presents an unsupervised transformer-based framework for temporal activity segmentation which leverages not only frame-level cues but also segment-level cues. This is in contrast with previous methods which often rely on frame-level information only. Our approach begins with a frame-level prediction module which estimates framewise action classes via a transformer encoder. The frame-level prediction module is trained in an unsupervised manner via temporal optimal transport. To exploit segment-level information, we utilize a segment-level prediction module and a frame-to-segment alignment module. The former includes a transformer decoder for estimating video transcripts, while the latter matches frame-level features with segment-level features, yielding permutation-aware segmentation results. Moreover, inspired by temporal optimal transport, we introduce simple-yet-effective pseudo labels for unsupervised training of the above modules. Our experiments on four public datasets, i.e., 50 Salads, YouTube Instructions, Breakfast, and Desktop Assembly show that our approach achieves comparable or better performance than previous methods in unsupervised activity segmentation. Our code and dataset are available on our research website: https://retrocausal.ai/research/.", "date": "2023-05-31", "categories": []}
{"id": "2512.02846", "title": "Action Anticipation at a Glimpse: To What Extent Can Multimodal Cues Replace Video?", "abstract": "Anticipating actions before they occur is a core challenge in action understanding research. While conventional methods rely on extracting and aggregating temporal information from videos, as humans we can often predict upcoming actions by observing a single moment from a scene, when given sufficient context. Can a model achieve this competence? The short answer is yes, although its effectiveness depends on the complexity of the task. In this work, we investigate to what extent video aggregation can be replaced with alternative modalities. To this end, based on recent advances in visual feature extraction and language-based reasoning, we introduce AAG, a method for Action Anticipation at a Glimpse. AAG combines RGB features with depth cues from a single frame for enhanced spatial reasoning, and incorporates prior action information to provide long-term context. This context is obtained either through textual summaries from Vision-Language Models, or from predictions generated by a single-frame action recognizer. Our results demonstrate that multimodal single-frame action anticipation using AAG can perform competitively compared to both temporally aggregated video baselines and state-of-the-art methods across three instructional activity datasets: IKEA-ASM, Meccano, and Assembly101.", "date": "2025-12-02", "categories": []}
{"id": "2305.19480", "title": "Learning by Aligning 2D Skeleton Sequences and Multi-Modality Fusion", "abstract": "This paper presents a self-supervised temporal video alignment framework which is useful for several fine-grained human activity understanding applications. In contrast with the state-of-the-art method of CASA, where sequences of 3D skeleton coordinates are taken directly as input, our key idea is to use sequences of 2D skeleton heatmaps as input. Unlike CASA which performs self-attention in the temporal domain only, we feed 2D skeleton heatmaps to a video transformer which performs self-attention both in the spatial and temporal domains for extracting effective spatiotemporal and contextual features. In addition, we introduce simple heatmap augmentation techniques based on 2D skeletons for self-supervised learning. Despite the lack of 3D information, our approach achieves not only higher accuracy but also better robustness against missing and noisy keypoints than CASA. Furthermore, extensive evaluations on three public datasets, i.e., Penn Action, IKEA ASM, and H2O, demonstrate that our approach outperforms previous methods in different fine-grained human activity understanding tasks. Finally, fusing 2D skeleton heatmaps with RGB videos yields the state-of-the-art on all metrics and datasets. To our best knowledge, our work is the first to utilize 2D skeleton heatmap inputs and the first to explore multi-modality fusion for temporal video alignment. Our code and dataset are available on our research website: https://retrocausal.ai/research/.", "date": "2023-05-31", "categories": []}
{"id": "2511.10060", "title": "Multivariate Gaussian Representation Learning for Medical Action Evaluation", "abstract": "Fine-grained action evaluation in medical vision faces unique challenges due to the unavailability of comprehensive datasets, stringent precision requirements, and insufficient spatiotemporal dynamic modeling of very rapid actions. To support development and evaluation, we introduce CPREval-6k, a multi-view, multi-label medical action benchmark containing 6,372 expert-annotated videos with 22 clinical labels. Using this dataset, we present GaussMedAct, a multivariate Gaussian encoding framework, to advance medical motion analysis through adaptive spatiotemporal representation learning. Multivariate Gaussian Representation projects the joint motions to a temporally scaled multi-dimensional space, and decomposes actions into adaptive 3D Gaussians that serve as tokens. These tokens preserve motion semantics through anisotropic covariance modeling while maintaining robustness to spatiotemporal noise. Hybrid Spatial Encoding, employing a Cartesian and Vector dual-stream strategy, effectively utilizes skeletal information in the form of joint and bone features. The proposed method achieves 92.1% Top-1 accuracy with real-time inference on the benchmark, outperforming baseline by +5.9% accuracy with only 10% FLOPs. Cross-dataset experiments confirm the superiority of our method in robustness.", "date": "2025-11-13", "categories": []}
{"id": "2512.01298", "title": "TBT-Former: Learning Temporal Boundary Distributions for Action Localization", "abstract": "Temporal Action Localization (TAL) remains a fundamental challenge in video understanding, aiming to identify the start time, end time, and category of all action instances within untrimmed videos. While recent single-stage, anchor-free models like ActionFormer have set a high standard by leveraging Transformers for temporal reasoning, they often struggle with two persistent issues: the precise localization of actions with ambiguous or \"fuzzy\" temporal boundaries and the effective fusion of multi-scale contextual information. In this paper, we introduce the Temporal Boundary Transformer (TBT-Former), a new architecture that directly addresses these limitations. TBT-Former enhances the strong ActionFormer baseline with three core contributions: (1) a higher-capacity scaled Transformer backbone with an increased number of attention heads and an expanded Multi-Layer Perceptron (MLP) dimension for more powerful temporal feature extraction; (2) a cross-scale feature pyramid network (FPN) that integrates a top-down pathway with lateral connections, enabling richer fusion of high-level semantics and low-level temporal details; and (3) a novel boundary distribution regression head. Inspired by the principles of Generalized Focal Loss (GFL), this new head recasts the challenging task of boundary regression as a more flexible probability distribution learning problem, allowing the model to explicitly represent and reason about boundary uncertainty. Within the paradigm of Transformer-based architectures, TBT-Former advances the formidable benchmark set by its predecessors, establishing a new level of performance on the highly competitive THUMOS14 and EPIC-Kitchens 100 datasets, while remaining competitive on the large-scale ActivityNet-1.3. Our code is available at https://github.com/aaivu/In21-S7-CS4681-AML-Research-Projects/tree/main/projects/210536K-Multi-Modal-Learning_Video-Understanding", "date": "2025-12-01", "categories": []}
{"id": "2507.21977", "title": "Motion Matters: Motion-guided Modulation Network for Skeleton-based Micro-Action Recognition", "abstract": "Micro-Actions (MAs) are an important form of non-verbal communication in social interactions, with potential applications in human emotional analysis. However, existing methods in Micro-Action Recognition often overlook the inherent subtle changes in MAs, which limits the accuracy of distinguishing MAs with subtle changes. To address this issue, we present a novel Motion-guided Modulation Network (MMN) that implicitly captures and modulates subtle motion cues to enhance spatial-temporal representation learning. Specifically, we introduce a Motion-guided Skeletal Modulation module (MSM) to inject motion cues at the skeletal level, acting as a control signal to guide spatial representation modeling. In parallel, we design a Motion-guided Temporal Modulation module (MTM) to incorporate motion information at the frame level, facilitating the modeling of holistic motion patterns in micro-actions. Finally, we propose a motion consistency learning strategy to aggregate the motion cues from multi-scale features for micro-action classification. Experimental results on the Micro-Action 52 and iMiGUE datasets demonstrate that MMN achieves state-of-the-art performance in skeleton-based micro-action recognition, underscoring the importance of explicitly modeling subtle motion cues. The code will be available at https://github.com/momiji-bit/MMN.", "date": "2025-07-29", "categories": []}
{"id": "2511.21202", "title": "Towards an Effective Action-Region Tracking Framework for Fine-grained Video Action Recognition", "abstract": "Fine-grained action recognition (FGAR) aims to identify subtle and distinctive differences among fine-grained action categories. However, current recognition methods often capture coarse-grained motion patterns but struggle to identify subtle details in local regions evolving over time. In this work, we introduce the Action-Region Tracking (ART) framework, a novel solution leveraging a query-response mechanism to discover and track the dynamics of distinctive local details, enabling effective distinction of similar actions. Specifically, we propose a region-specific semantic activation module that employs discriminative and text-constrained semantics as queries to capture the most action-related region responses in each video frame, facilitating interaction among spatial and temporal dimensions with corresponding video features. The captured region responses are organized into action tracklets, which characterize region-based action dynamics by linking related responses across video frames in a coherent sequence. The text-constrained queries encode nuanced semantic representations derived from textual descriptions of action labels extracted by language branches within Visual Language Models (VLMs). To optimize the action tracklets, we design a multi-level tracklet contrastive constraint among region responses at spatial and temporal levels, enabling effective discrimination within each frame and correlation between adjacent frames. Additionally, a task-specific fine-tuning mechanism refines textual semantics such that semantic representations encoded by VLMs are preserved while optimized for task preferences. Comprehensive experiments on widely used action recognition benchmarks demonstrate the superiority to previous state-of-the-art baselines.", "date": "2025-11-26", "categories": []}
{"id": "2509.21783", "title": "Prompt-guided Disentangled Representation for Action Recognition", "abstract": "Action recognition is a fundamental task in video understanding. Existing methods typically extract unified features to process all actions in one video, which makes it challenging to model the interactions between different objects in multi-action scenarios. To alleviate this issue, we explore disentangling any specified actions from complex scenes as an effective solution. In this paper, we propose Prompt-guided Disentangled Representation for Action Recognition (ProDA), a novel framework that disentangles any specified actions from a multi-action scene. ProDA leverages Spatio-temporal Scene Graphs (SSGs) and introduces Dynamic Prompt Module (DPM) to guide a Graph Parsing Neural Network (GPNN) in generating action-specific representations. Furthermore, we design a video-adapted GPNN that aggregates information using dynamic weights. Experiments in video action recognition demonstrate the effectiveness of our approach when compared with the state-of-the-art methods. Our code can be found in https://github.com/iamsnaping/ProDA.git", "date": "2025-09-26", "categories": []}
{"id": "2511.03725", "title": "Disentangled Concepts Speak Louder Than Words: Explainable Video Action Recognition", "abstract": "Effective explanations of video action recognition models should disentangle how movements unfold over time from the surrounding spatial context. However, existing methods based on saliency produce entangled explanations, making it unclear whether predictions rely on motion or spatial context. Language-based approaches offer structure but often fail to explain motions due to their tacit nature -- intuitively understood but difficult to verbalize. To address these challenges, we propose Disentangled Action aNd Context concept-based Explainable (DANCE) video action recognition, a framework that predicts actions through disentangled concept types: motion dynamics, objects, and scenes. We define motion dynamics concepts as human pose sequences. We employ a large language model to automatically extract object and scene concepts. Built on an ante-hoc concept bottleneck design, DANCE enforces prediction through these concepts. Experiments on four datasets -- KTH, Penn Action, HAA500, and UCF-101 -- demonstrate that DANCE significantly improves explanation clarity with competitive performance. We validate the superior interpretability of DANCE through a user study. Experimental results also show that DANCE is beneficial for model debugging, editing, and failure analysis.", "date": "2025-11-05", "categories": []}
{"id": "2511.16860", "title": "Parts-Mamba: Augmenting Joint Context with Part-Level Scanning for Occluded Human Skeleton", "abstract": "Skeleton action recognition involves recognizing human action from human skeletons. The use of graph convolutional networks (GCNs) has driven major advances in this recognition task. In real-world scenarios, the captured skeletons are not always perfect or complete because of occlusions of parts of the human body or poor communication quality, leading to missing parts in skeletons or videos with missing frames. In the presence of such non-idealities, existing GCN models perform poorly due to missing local context. To address this limitation, we propose Parts-Mamba, a hybrid GCN-Mamba model designed to enhance the ability to capture and maintain contextual information from distant joints. The proposed Parts-Mamba model effectively captures part-specific information through its parts-specific scanning feature and preserves non-neighboring joint context via a parts-body fusion module. Our proposed model is evaluated on the NTU RGB+D 60 and NTU RGB+D 120 datasets under different occlusion settings, achieving up to 12.9% improvement in accuracy.", "date": "2025-11-21", "categories": []}
{"id": "2511.17345", "title": "Label-Efficient Skeleton-based Recognition with Stable-Invertible Graph Convolutional Networks", "abstract": "Skeleton-based action recognition is a hotspot in image processing. A key challenge of this task lies in its dependence on large, manually labeled datasets whose acquisition is costly and time-consuming. This paper devises a novel, label-efficient method for skeleton-based action recognition using graph convolutional networks (GCNs). The contribution of the proposed method resides in learning a novel acquisition function -- scoring the most informative subsets for labeling -- as the optimum of an objective function mixing data representativity, diversity and uncertainty. We also extend this approach by learning the most informative subsets using an invertible GCN which allows mapping data from ambient to latent spaces where the inherent distribution of the data is more easily captured. Extensive experiments, conducted on two challenging skeleton-based recognition datasets, show the effectiveness and the outperformance of our label-frugal GCNs against the related work.", "date": "2025-11-21", "categories": []}
{"id": "2511.16183", "title": "FOOTPASS: A Multi-Modal Multi-Agent Tactical Context Dataset for Play-by-Play Action Spotting in Soccer Broadcast Videos", "abstract": "Soccer video understanding has motivated the creation of datasets for tasks such as temporal action localization, spatiotemporal action detection (STAD), or multiobject tracking (MOT). The annotation of structured sequences of events (who does what, when, and where) used for soccer analytics requires a holistic approach that integrates both STAD and MOT. However, current action recognition methods remain insufficient for constructing reliable play-by-play data and are typically used to assist rather than fully automate annotation. Parallel research has advanced tactical modeling, trajectory forecasting, and performance analysis, all grounded in game-state and play-by-play data. This motivates leveraging tactical knowledge as a prior to support computer-vision-based predictions, enabling more automated and reliable extraction of play-by-play data. We introduce Footovision Play-by-Play Action Spotting in Soccer Dataset (FOOTPASS), the first benchmark for play-by-play action spotting over entire soccer matches in a multi-modal, multi-agent tactical context. It enables the development of methods for player-centric action spotting that exploit both outputs from computer-vision tasks (e.g., tracking, identification) and prior knowledge of soccer, including its tactical regularities over long time horizons, to generate reliable play-by-play data streams. These streams form an essential input for data-driven sports analytics.", "date": "2025-11-20", "categories": []}
{"id": "2511.16524", "title": "BoxingVI: A Multi-Modal Benchmark for Boxing Action Recognition and Localization", "abstract": "Accurate analysis of combat sports using computer vision has gained traction in recent years, yet the development of robust datasets remains a major bottleneck due to the dynamic, unstructured nature of actions and variations in recording environments. In this work, we present a comprehensive, well-annotated video dataset tailored for punch detection and classification in boxing. The dataset comprises 6,915 high-quality punch clips categorized into six distinct punch types, extracted from 20 publicly available YouTube sparring sessions and involving 18 different athletes. Each clip is manually segmented and labeled to ensure precise temporal boundaries and class consistency, capturing a wide range of motion styles, camera angles, and athlete physiques. This dataset is specifically curated to support research in real-time vision-based action recognition, especially in low-resource and unconstrained environments. By providing a rich benchmark with diverse punch examples, this contribution aims to accelerate progress in movement analysis, automated coaching, and performance assessment within boxing and related domains.", "date": "2025-11-20", "categories": []}
{"id": "2511.15379", "title": "Zero-Shot Open-Vocabulary Human Motion Grounding with Test-Time Training", "abstract": "Understanding complex human activities demands the ability to decompose motion into fine-grained, semantic-aligned sub-actions. This motion grounding process is crucial for behavior analysis, embodied AI and virtual reality. Yet, most existing methods rely on dense supervision with predefined action classes, which are infeasible in open-vocabulary, real-world settings. In this paper, we propose ZOMG, a zero-shot, open-vocabulary framework that segments motion sequences into semantically meaningful sub-actions without requiring any annotations or fine-tuning. Technically, ZOMG integrates (1) language semantic partition, which leverages large language models to decompose instructions into ordered sub-action units, and (2) soft masking optimization, which learns instance-specific temporal masks to focus on frames critical to sub-actions, while maintaining intra-segment continuity and enforcing inter-segment separation, all without altering the pretrained encoder. Experiments on three motion-language datasets demonstrate state-of-the-art effectiveness and efficiency of motion grounding performance, outperforming prior methods by +8.7\\% mAP on HumanML3D benchmark. Meanwhile, significant improvements also exist in downstream retrieval, establishing a new paradigm for annotation-free motion understanding.", "date": "2025-11-19", "categories": []}
{"id": "2502.15462", "title": "Game State and Spatio-temporal Action Detection in Soccer using Graph Neural Networks and 3D Convolutional Networks", "abstract": "Soccer analytics rely on two data sources: the player positions on the pitch and the sequences of events they perform. With around 2000 ball events per game, their precise and exhaustive annotation based on a monocular video stream remains a tedious and costly manual task. While state-of-the-art spatio-temporal action detection methods show promise for automating this task, they lack contextual understanding of the game. Assuming professional players' behaviors are interdependent, we hypothesize that incorporating surrounding players' information such as positions, velocity and team membership can enhance purely visual predictions. We propose a spatio-temporal action detection approach that combines visual and game state information via Graph Neural Networks trained end-to-end with state-of-the-art 3D CNNs, demonstrating improved metrics through game state integration.", "date": "2025-02-21", "categories": []}
{"id": "2503.23447", "title": "CA^2ST: Cross-Attention in Audio, Space, and Time for Holistic Video Recognition", "abstract": "We propose Cross-Attention in Audio, Space, and Time (CA^2ST), a transformer-based method for holistic video recognition. Recognizing actions in videos requires both spatial and temporal understanding, yet most existing models lack a balanced spatio-temporal understanding of videos. To address this, we propose a novel two-stream architecture, called Cross-Attention in Space and Time (CAST), using only RGB input. In each layer of CAST, Bottleneck Cross-Attention (B-CA) enables spatial and temporal experts to exchange information and make synergistic predictions. For holistic video understanding, we extend CAST by integrating an audio expert, forming Cross-Attention in Visual and Audio (CAVA). We validate the CAST on benchmarks with different characteristics, EPIC-KITCHENS-100, Something-Something-V2, and Kinetics-400, consistently showing balanced performance. We also validate the CAVA on audio-visual action recognition benchmarks, including UCF-101, VGG-Sound, KineticsSound, and EPIC-SOUNDS. With a favorable performance of CAVA across these datasets, we demonstrate the effective information exchange among multiple experts within the B-CA module. In summary, CA^2ST combines CAST and CAVA by employing spatial, temporal, and audio experts through cross-attention, achieving balanced and holistic video understanding.", "date": "2025-03-30", "categories": []}
{"id": "2508.01742", "title": "Intention-Guided Cognitive Reasoning for Egocentric Long-Term Action Anticipation", "abstract": "Long-term action anticipation from egocentric video is critical for applications such as human-computer interaction and assistive technologies, where anticipating user intent enables proactive and context-aware AI assistance. However, existing approaches suffer from three key limitations: 1) underutilization of fine-grained visual cues from hand-object interactions, 2) neglect of semantic dependencies between verbs and nouns, and 3) lack of explicit cognitive reasoning, limiting generalization and long-term forecasting ability. To overcome these challenges, we propose INSIGHT, a unified two-stage framework for egocentric action anticipation. In the first stage, INSIGHT focuses on extracting semantically rich features from hand-object interaction regions and enhances action representations using a verb-noun co-occurrence matrix. In the second stage, it introduces a reinforcement learning-based module that simulates explicit cognitive reasoning through a structured process: visual perception (think) -> intention inference (reason) -> action anticipation (answer). Extensive experiments on Ego4D, EPIC-Kitchens-55, and EGTEA Gaze+ benchmarks show that INSIGHT achieves state-of-the-art performance, demonstrating its effectiveness and strong generalization capability.", "date": "2025-08-03", "categories": []}
{"id": "2511.13039", "title": "MGCA-Net: Multi-Grained Category-Aware Network for Open-Vocabulary Temporal Action Localization", "abstract": "Open-Vocabulary Temporal Action Localization (OV-TAL) aims to recognize and localize instances of any desired action categories in videos without explicitly curating training data for all categories. Existing methods mostly recognize action categories at a single granularity, which degrades the recognition accuracy of both base and novel action categories. To address these issues, we propose a Multi-Grained Category-Aware Network (MGCA-Net) comprising a localizer, an action presence predictor, a conventional classifier, and a coarse-to-fine classifier. Specifically, the localizer localizes category-agnostic action proposals. For these action proposals, the action presence predictor estimates the probability that they belong to an action instance. At the same time, the conventional classifier predicts the probability of each action proposal over base action categories at the snippet granularity. Novel action categories are recognized by the coarse-to-fine classifier, which first identifies action presence at the video granularity. Finally, it assigns each action proposal to one category from the coarse categories at the proposal granularity. Through coarse-to-fine category awareness for novel actions and the conventional classifier's awareness of base actions, multi-grained category awareness is achieved, effectively enhancing localization performance. Comprehensive evaluations on the THUMOS'14 and ActivityNet-1.3 benchmarks demonstrate that our method achieves state-of-the-art performance. Furthermore, our MGCA-Net achieves state-of-the-art results under the Zero-Shot Temporal Action Localization setting.", "date": "2025-11-17", "categories": []}
{"id": "2511.13315", "title": "Computer Vision based group activity detection and action spotting", "abstract": "Group activity detection in multi-person scenes is challenging due to complex human interactions, occlusions, and variations in appearance over time. This work presents a computer vision based framework for group activity recognition and action spotting using a combination of deep learning models and graph based relational reasoning. The system first applies Mask R-CNN to obtain accurate actor localization through bounding boxes and instance masks. Multiple backbone networks, including Inception V3, MobileNet, and VGG16, are used to extract feature maps, and RoIAlign is applied to preserve spatial alignment when generating actor specific features. The mask information is then fused with the feature maps to obtain refined masked feature representations for each actor. To model interactions between individuals, we construct Actor Relation Graphs that encode appearance similarity and positional relations using methods such as normalized cross correlation, sum of absolute differences, and dot product. Graph Convolutional Networks operate on these graphs to reason about relationships and predict both individual actions and group level activities. Experiments on the Collective Activity dataset demonstrate that the combination of mask based feature refinement, robust similarity search, and graph neural network reasoning leads to improved recognition performance across both crowded and non crowded scenarios. This approach highlights the potential of integrating segmentation, feature extraction, and relational graph reasoning for complex video understanding tasks.", "date": "2025-11-17", "categories": []}
{"id": "2309.12067", "title": "Survey of Action Recognition, Spotting and Spatio-Temporal Localization in Soccer -- Current Trends and Research Perspectives", "abstract": "Action scene understanding in soccer is a challenging task due to the complex and dynamic nature of the game, as well as the interactions between players. This article provides a comprehensive overview of this task divided into action recognition, spotting, and spatio-temporal action localization, with a particular emphasis on the modalities used and multimodal methods. We explore the publicly available data sources and metrics used to evaluate models' performance. The article reviews recent state-of-the-art methods that leverage deep learning techniques and traditional methods. We focus on multimodal methods, which integrate information from multiple sources, such as video and audio data, and also those that represent one source in various ways. The advantages and limitations of methods are discussed, along with their potential for improving the accuracy and robustness of models. Finally, the article highlights some of the open research questions and future directions in the field of soccer action recognition, including the potential for multimodal methods to advance this field. Overall, this survey provides a valuable resource for researchers interested in the field of action scene understanding in soccer.", "date": "2023-09-21", "categories": []}
{"id": "2511.10866", "title": "Short-Window Sliding Learning for Real-Time Violence Detection via LLM-based Auto-Labeling", "abstract": "This paper proposes a Short-Window Sliding Learning framework for real-time violence detection in CCTV footages. Unlike conventional long-video training approaches, the proposed method divides videos into 1-2 second clips and applies Large Language Model (LLM)-based auto-caption labeling to construct fine-grained datasets. Each short clip fully utilizes all frames to preserve temporal continuity, enabling precise recognition of rapid violent events. Experiments demonstrate that the proposed method achieves 95.25\\% accuracy on RWF-2000 and significantly improves performance on long videos (UCF-Crime: 83.25\\%), confirming its strong generalization and real-time applicability in intelligent surveillance systems.", "date": "2025-11-14", "categories": []}
{"id": "2511.03943", "title": "Temporal Zoom Networks: Distance Regression and Continuous Depth for Efficient Action Localization", "abstract": "Temporal action localization requires both precise boundary detection and computational efficiency. Current methods apply uniform computation across all temporal positions, wasting resources on easy boundaries while struggling with ambiguous ones. We address this through two complementary innovations: Boundary Distance Regression (BDR), which replaces classification-based boundary detection with signed-distance regression achieving 3.3--16.7$\\times$ lower variance; and Adaptive Temporal Refinement (ATR), which allocates transformer depth continuously ($\\tau\\in[0,1]$) to concentrate computation near difficult boundaries. On THUMOS14, our method achieves 56.5\\% mAP@0.7 and 58.2\\% average mAP@[0.3:0.7] with 151G FLOPs, using 36\\% fewer FLOPs than ActionFormer++ (55.7\\% mAP@0.7 at 235G). Compared to uniform baselines, we achieve +2.9\\% mAP@0.7 (+1.8\\% avg mAP, 5.4\\% relative) with 24\\% fewer FLOPs and 29\\% lower latency, with particularly strong gains on short actions (+4.2\\%, 8.6\\% relative). Training requires 1.29$\\times$ baseline FLOPs, but this one-time cost is amortized over many inference runs; knowledge distillation further reduces this to 1.1$\\times$ while retaining 99.5\\% accuracy. Our contributions include: (i) a theoretically-grounded distance formulation with information-theoretic analysis showing optimal variance scaling; (ii) a continuous depth allocation mechanism avoiding discrete routing complexity; and (iii) consistent improvements across four datasets with gains correlating with boundary heterogeneity.", "date": "2025-11-06", "categories": []}
{"id": "2511.10091", "title": "SUGAR: Learning Skeleton Representation with Visual-Motion Knowledge for Action Recognition", "abstract": "Large Language Models (LLMs) hold rich implicit knowledge and powerful transferability. In this paper, we explore the combination of LLMs with the human skeleton to perform action classification and description. However, when treating LLM as a recognizer, two questions arise: 1) How can LLMs understand skeleton? 2) How can LLMs distinguish among actions? To address these problems, we introduce a novel paradigm named learning Skeleton representation with visUal-motion knowledGe for Action Recognition (SUGAR). In our pipeline, we first utilize off-the-shelf large-scale video models as a knowledge base to generate visual, motion information related to actions. Then, we propose to supervise skeleton learning through this prior knowledge to yield discrete representations. Finally, we use the LLM with untouched pre-training weights to understand these representations and generate the desired action targets and descriptions. Notably, we present a Temporal Query Projection (TQP) module to continuously model the skeleton signals with long sequences. Experiments on several skeleton-based action classification benchmarks demonstrate the efficacy of our SUGAR. Moreover, experiments on zero-shot scenarios show that SUGAR is more versatile than linear-based methods.", "date": "2025-11-13", "categories": []}
{"id": "2511.08910", "title": "OG-PCL: Efficient Sparse Point Cloud Processing for Human Activity Recognition", "abstract": "Human activity recognition (HAR) with millimeter-wave (mmWave) radar offers a privacy-preserving and robust alternative to camera- and wearable-based approaches. In this work, we propose the Occupancy-Gated Parallel-CNN Bi-LSTM (OG-PCL) network to process sparse 3D radar point clouds produced by mmWave sensing. Designed for lightweight deployment, the parameter size of the proposed OG-PCL is only 0.83M and achieves 91.75 accuracy on the RadHAR dataset, outperforming those existing baselines such as 2D CNN, PointNet, and 3D CNN methods. We validate the advantages of the tri-view parallel structure in preserving spatial information across three dimensions while maintaining efficiency through ablation studies. We further introduce the Occupancy-Gated Convolution (OGConv) block and demonstrate the necessity of its occupancy compensation mechanism for handling sparse point clouds. The proposed OG-PCL thus offers a compact yet accurate framework for real-time radar-based HAR on lightweight platforms.", "date": "2025-11-12", "categories": []}
{"id": "2511.09469", "title": "Revisiting Cross-Architecture Distillation: Adaptive Dual-Teacher Transfer for Lightweight Video Models", "abstract": "Vision Transformers (ViTs) have achieved strong performance in video action recognition, but their high computational cost limits their practicality. Lightweight CNNs are more efficient but suffer from accuracy gaps. Cross-Architecture Knowledge Distillation (CAKD) addresses this by transferring knowledge from ViTs to CNNs, yet existing methods often struggle with architectural mismatch and overlook the value of stronger homogeneous CNN teachers. To tackle these challenges, we propose a Dual-Teacher Knowledge Distillation framework that leverages both a heterogeneous ViT teacher and a homogeneous CNN teacher to collaboratively guide a lightweight CNN student. We introduce two key components: (1) Discrepancy-Aware Teacher Weighting, which dynamically fuses the predictions from ViT and CNN teachers by assigning adaptive weights based on teacher confidence and prediction discrepancy with the student, enabling more informative and effective supervision; and (2) a Structure Discrepancy-Aware Distillation strategy, where the student learns the residual features between ViT and CNN teachers via a lightweight auxiliary branch, focusing on transferable architectural differences without mimicking all of ViT's high-dimensional patterns. Extensive experiments on benchmarks including HMDB51, EPIC-KITCHENS-100, and Kinetics-400 demonstrate that our method consistently outperforms state-of-the-art distillation approaches, achieving notable performance improvements with a maximum accuracy gain of 5.95% on HMDB51.", "date": "2025-11-12", "categories": []}
{"id": "2407.10935", "title": "STARS: Self-supervised Tuning for 3D Action Recognition in Skeleton Sequences", "abstract": "Self-supervised pretraining methods with masked prediction demonstrate remarkable within-dataset performance in skeleton-based action recognition. However, we show that, unlike contrastive learning approaches, they do not produce well-separated clusters. Additionally, these methods struggle with generalization in few-shot settings. To address these issues, we propose Self-supervised Tuning for 3D Action Recognition in Skeleton sequences (STARS). Specifically, STARS first uses a masked prediction stage using an encoder-decoder architecture. It then employs nearest-neighbor contrastive learning to partially tune the weights of the encoder, enhancing the formation of semantic clusters for different actions. By tuning the encoder for a few epochs, and without using hand-crafted data augmentations, STARS achieves state-of-the-art self-supervised results in various benchmarks, including NTU-60, NTU-120, and PKU-MMD. In addition, STARS exhibits significantly better results than masked prediction models in few-shot settings, where the model has not seen the actions throughout pretraining. Project page: https://soroushmehraban.github.io/stars/", "date": "2024-07-15", "categories": []}
{"id": "2511.05622", "title": "Grounding Foundational Vision Models with 3D Human Poses for Robust Action Recognition", "abstract": "For embodied agents to effectively understand and interact within the world around them, they require a nuanced comprehension of human actions grounded in physical space. Current action recognition models, often relying on RGB video, learn superficial correlations between patterns and action labels, so they struggle to capture underlying physical interaction dynamics and human poses in complex scenes. We propose a model architecture that grounds action recognition in physical space by fusing two powerful, complementary representations: V-JEPA 2's contextual, predictive world dynamics and CoMotion's explicit, occlusion-tolerant human pose data. Our model is validated on both the InHARD and UCF-19-Y-OCC benchmarks for general action recognition and high-occlusion action recognition, respectively. Our model outperforms three other baselines, especially within complex, occlusive scenes. Our findings emphasize a need for action recognition to be supported by spatial understanding instead of statistical pattern recognition.", "date": "2025-11-06", "categories": []}
{"id": "2511.05773", "title": "MARAuder's Map: Motion-Aware Real-time Activity Recognition with Layout-Based Trajectories", "abstract": "Ambient sensor-based human activity recognition (HAR) in smart homes remains challenging due to the need for real-time inference, spatially grounded reasoning, and context-aware temporal modeling. Existing approaches often rely on pre-segmented, within-activity data and overlook the physical layout of the environment, limiting their robustness in continuous, real-world deployments. In this paper, we propose MARAuder's Map, a novel framework for real-time activity recognition from raw, unsegmented sensor streams. Our method projects sensor activations onto the physical floorplan to generate trajectory-aware, image-like sequences that capture the spatial flow of human movement. These representations are processed by a hybrid deep learning model that jointly captures spatial structure and temporal dependencies. To enhance temporal awareness, we introduce a learnable time embedding module that encodes contextual cues such as hour-of-day and day-of-week. Additionally, an attention-based encoder selectively focuses on informative segments within each observation window, enabling accurate recognition even under cross-activity transitions and temporal ambiguity. Extensive experiments on multiple real-world smart home datasets demonstrate that our method outperforms strong baselines, offering a practical solution for real-time HAR in ambient sensor environments.", "date": "2025-11-08", "categories": []}
{"id": "2407.12805", "title": "Dark Transformer: A Video Transformer for Action Recognition in the Dark", "abstract": "Recognizing human actions in adverse lighting conditions presents significant challenges in computer vision, with wide-ranging applications in visual surveillance and nighttime driving. Existing methods tackle action recognition and dark enhancement separately, limiting the potential for end-to-end learning of spatiotemporal representations for video action classification. This paper introduces Dark Transformer, a novel video transformer-based approach for action recognition in low-light environments. Dark Transformer leverages spatiotemporal self-attention mechanisms in cross-domain settings to enhance cross-domain action recognition. By extending video transformers to learn cross-domain knowledge, Dark Transformer achieves state-of-the-art performance on benchmark action recognition datasets, including InFAR, XD145, and ARID. The proposed approach demonstrates significant promise in addressing the challenges of action recognition in adverse lighting conditions, offering practical implications for real-world applications.", "date": "2024-06-25", "categories": []}
{"id": "2511.05039", "title": "PECL: A Heterogeneous Parallel Multi-Domain Network for Radar-Based Human Activity Recognition", "abstract": "Radar systems are increasingly favored for medical applications because they provide non-intrusive monitoring with high privacy and robustness to lighting conditions. However, existing research typically relies on single-domain radar signals and overlooks the temporal dependencies inherent in human activity, which complicates the classification of similar actions. To address this issue, we designed the Parallel-EfficientNet-CBAM-LSTM (PECL) network to process data in three complementary domains: Range-Time, Doppler-Time, and Range-Doppler. PECL combines a channel-spatial attention module and temporal units to capture more features and dynamic dependencies during action sequences, improving both accuracy and robustness. The experimental results show that PECL achieves an accuracy of 96.16% on the same dataset, outperforming existing methods by at least 4.78%. PECL also performs best in distinguishing between easily confused actions. Despite its strong performance, PECL maintains moderate model complexity, with 23.42M parameters and 1324.82M FLOPs. Its parameter-efficient design further reduces computational cost.", "date": "2025-11-07", "categories": []}
{"id": "2511.05250", "title": "Accurate online action and gesture recognition system using detectors and Deep SPD Siamese Networks", "abstract": "Online continuous motion recognition is a hot topic of research since it is more practical in real life application cases. Recently, Skeleton-based approaches have become increasingly popular, demonstrating the power of using such 3D temporal data. However, most of these works have focused on segment-based recognition and are not suitable for the online scenarios. In this paper, we propose an online recognition system for skeleton sequence streaming composed from two main components: a detector and a classifier, which use a Semi-Positive Definite (SPD) matrix representation and a Siamese network. The powerful statistical representations for the skeletal data given by the SPD matrices and the learning of their semantic similarity by the Siamese network enable the detector to predict time intervals of the motions throughout an unsegmented sequence. In addition, they ensure the classifier capability to recognize the motion in each predicted interval. The proposed detector is flexible and able to identify the kinetic state continuously. We conduct extensive experiments on both hand gesture and body action recognition benchmarks to prove the accuracy of our online recognition system which in most cases outperforms state-of-the-art performances.", "date": "2025-11-07", "categories": []}
{"id": "2511.02473", "title": "MVAFormer: RGB-based Multi-View Spatio-Temporal Action Recognition with Transformer", "abstract": "Multi-view action recognition aims to recognize human actions using multiple camera views and deals with occlusion caused by obstacles or crowds. In this task, cooperation among views, which generates a joint representation by combining multiple views, is vital. Previous studies have explored promising cooperation methods for improving performance. However, since their methods focus only on the task setting of recognizing a single action from an entire video, they are not applicable to the recently popular spatio-temporal action recognition~(STAR) setting, in which each person's action is recognized sequentially. To address this problem, this paper proposes a multi-view action recognition method for the STAR setting, called MVAFormer. In MVAFormer, we introduce a novel transformer-based cooperation module among views. In contrast to previous studies, which utilize embedding vectors with lost spatial information, our module utilizes the feature map for effective cooperation in the STAR setting, which preserves the spatial information. Furthermore, in our module, we divide the self-attention for the same and different views to model the relationship between multiple views effectively. The results of experiments using a newly collected dataset demonstrate that MVAFormer outperforms the comparison baselines by approximately $4.4$ points on the F-measure.", "date": "2025-11-04", "categories": []}
{"id": "2410.21275", "title": "Enhancing Action Recognition by Leveraging the Hierarchical Structure of Actions and Textual Context", "abstract": "We propose a novel approach to improve action recognition by exploiting the hierarchical organization of actions and by incorporating contextualized textual information, including location and previous actions, to reflect the action's temporal context. To achieve this, we introduce a transformer architecture tailored for action recognition that employs both visual and textual features. Visual features are obtained from RGB and optical flow data, while text embeddings represent contextual information. Furthermore, we define a joint loss function to simultaneously train the model for both coarse- and fine-grained action recognition, effectively exploiting the hierarchical nature of actions. To demonstrate the effectiveness of our method, we extend the Toyota Smarthome Untrimmed (TSU) dataset by incorporating action hierarchies, resulting in the Hierarchical TSU dataset, a hierarchical dataset designed for monitoring activities of the elderly in home environments. An ablation study assesses the performance impact of different strategies for integrating contextual and hierarchical data. Experimental results demonstrate that the proposed method consistently outperforms SOTA methods on the Hierarchical TSU dataset, Assembly101 and IkeaASM, achieving over a 17% improvement in top-1 accuracy.", "date": "2024-10-28", "categories": []}
{"id": "2510.27255", "title": "Enhancing Spatio-Temporal Zero-shot Action Recognition with Language-driven Description Attributes", "abstract": "Vision-Language Models (VLMs) have demonstrated impressive capabilities in zero-shot action recognition by learning to associate video embeddings with class embeddings. However, a significant challenge arises when relying solely on action classes to provide semantic context, particularly due to the presence of multi-semantic words, which can introduce ambiguity in understanding the intended concepts of actions. To address this issue, we propose an innovative approach that harnesses web-crawled descriptions, leveraging a large-language model to extract relevant keywords. This method reduces the need for human annotators and eliminates the laborious manual process of attribute data creation. Additionally, we introduce a spatio-temporal interaction module designed to focus on objects and action units, facilitating alignment between description attributes and video content. In our zero-shot experiments, our model achieves impressive results, attaining accuracies of 81.0%, 53.1%, and 68.9% on UCF-101, HMDB-51, and Kinetics-600, respectively, underscoring the model's adaptability and effectiveness across various downstream tasks.", "date": "2025-10-31", "categories": []}
{"id": "2511.00900", "title": "Learning with Category-Equivariant Representations for Human Activity Recognition", "abstract": "Human activity recognition is challenging because sensor signals shift with context, motion, and environment; effective models must therefore remain stable as the world around them changes. We introduce a categorical symmetry-aware learning framework that captures how signals vary over time, scale, and sensor hierarchy. We build these factors into the structure of feature representations, yielding models that automatically preserve the relationships between sensors and remain stable under realistic distortions such as time shifts, amplitude drift, and device orientation changes. On the UCI Human Activity Recognition benchmark, this categorical symmetry-driven design improves out-of-distribution accuracy by approx. 46 percentage points (approx. 3.6x over the baseline), demonstrating that abstract symmetry principles can translate into concrete performance gains in everyday sensing tasks via category-equivariant representation theory.", "date": "2025-11-02", "categories": []}
{"id": "2408.01701", "title": "Signal-SGN: A Spiking Graph Convolutional Network for Skeletal Action Recognition via Learning Temporal-Frequency Dynamics", "abstract": "For multimodal skeleton-based action recognition, Graph Convolutional Networks (GCNs) are effective models. Still, their reliance on floating-point computations leads to high energy consumption, limiting their applicability in battery-powered devices. While energy-efficient, Spiking Neural Networks (SNNs) struggle to model skeleton dynamics, leading to suboptimal solutions. We propose Signal-SGN (Spiking Graph Convolutional Network), which utilizes the temporal dimension of skeleton sequences as the spike time steps and represents features as multi-dimensional discrete stochastic signals for temporal-frequency domain feature extraction. It combines the 1D Spiking Graph Convolution (1D-SGC) module and the Frequency Spiking Convolution (FSC) module to extract features from the skeleton represented as spiking form. Additionally, the Multi-Scale Wavelet Transform Feature Fusion (MWTF) module is proposed to extract dynamic spiking features and capture frequency-specific characteristics, enhancing classification performance. Experiments across three large-scale datasets reveal Signal-SGN exceeding state-of-the-art SNN-based methods in accuracy and computational efficiency while attaining comparable performance with GCN methods and significantly reducing theoretical energy consumption.", "date": "2024-08-03", "categories": []}
{"id": "2510.25345", "title": "Informative Sample Selection Model for Skeleton-based Action Recognition with Limited Training Samples", "abstract": "Skeleton-based human action recognition aims to classify human skeletal sequences, which are spatiotemporal representations of actions, into predefined categories. To reduce the reliance on costly annotations of skeletal sequences while maintaining competitive recognition accuracy, the task of 3D Action Recognition with Limited Training Samples, also known as semi-supervised 3D Action Recognition, has been proposed. In addition, active learning, which aims to proactively select the most informative unlabeled samples for annotation, has been explored in semi-supervised 3D Action Recognition for training sample selection. Specifically, researchers adopt an encoder-decoder framework to embed skeleton sequences into a latent space, where clustering information, combined with a margin-based selection strategy using a multi-head mechanism, is utilized to identify the most informative sequences in the unlabeled set for annotation. However, the most representative skeleton sequences may not necessarily be the most informative for the action recognizer, as the model may have already acquired similar knowledge from previously seen skeleton samples. To solve it, we reformulate Semi-supervised 3D action recognition via active learning from a novel perspective by casting it as a Markov Decision Process (MDP). Built upon the MDP framework and its training paradigm, we train an informative sample selection model to intelligently guide the selection of skeleton sequences for annotation. To enhance the representational capacity of the factors in the state-action pairs within our method, we project them from Euclidean space to hyperbolic space. Furthermore, we introduce a meta tuning strategy to accelerate the deployment of our method in real-world scenarios. Extensive experiments on three 3D action recognition benchmarks demonstrate the effectiveness of our method.", "date": "2025-10-29", "categories": []}
{"id": "2510.21282", "title": "Sensor-Specific Transformer (PatchTST) Ensembles with Test-Matched Augmentation", "abstract": "We present a noise-aware, sensor-specific ensemble approach for robust human activity recognition on the 2nd WEAR Dataset Challenge. Our method leverages the PatchTST transformer architecture, training four independent models-one per inertial sensor location-on a tampered training set whose 1-second sliding windows are augmented to mimic the test-time noise. By aligning the train and test data schemas (JSON-encoded 50-sample windows) and applying randomized jitter, scaling, rotation, and channel dropout, each PatchTST model learns to generalize across real-world sensor perturbations. At inference, we compute softmax probabilities from all four sensor models on the Kaggle test set and average them to produce final labels. On the private leaderboard, this pipeline achieves a macro-F1 substantially above the baseline, demonstrating that test-matched augmentation combined with transformer-based ensembling is an effective strategy for robust HAR under noisy conditions.", "date": "2025-10-24", "categories": []}
{"id": "2506.03885", "title": "Video, How Do Your Tokens Merge?", "abstract": "Video transformer models require huge amounts of compute resources due to the spatio-temporal scaling of the input. Tackling this, recent methods have proposed to drop or merge tokens for image models, whether randomly or via learned methods. Merging tokens has many benefits: it can be plugged into any vision transformer, does not require model re-training, and it propagates information that would otherwise be dropped through the model. Before now, video token merging has not been evaluated on temporally complex datasets for video understanding. In this work, we explore training-free token merging for video to provide comprehensive experiments and find best practices across four video transformers on three datasets that exhibit coarse and fine-grained action recognition. Our results showcase the benefits of video token merging with a speedup of around $2.5$X while maintaining accuracy (avg. $-0.55\\%$ for ViViT). Code available at https://github.com/sjpollard/video-how-do-your-tokens-merge.", "date": "2025-06-04", "categories": []}
{"id": "2510.18705", "title": "A Renaissance of Explicit Motion Information Mining from Transformers for Action Recognition", "abstract": "Recently, action recognition has been dominated by transformer-based methods, thanks to their spatiotemporal contextual aggregation capacities. However, despite the significant progress achieved on scene-related datasets, they do not perform well on motion-sensitive datasets due to the lack of elaborate motion modeling designs. Meanwhile, we observe that the widely-used cost volume in traditional action recognition is highly similar to the affinity matrix defined in self-attention, but equipped with powerful motion modeling capacities. In light of this, we propose to integrate those effective motion modeling properties into the existing transformer in a unified and neat way, with the proposal of the Explicit Motion Information Mining module (EMIM). In EMIM, we propose to construct the desirable affinity matrix in a cost volume style, where the set of key candidate tokens is sampled from the query-based neighboring area in the next frame in a sliding-window manner. Then, the constructed affinity matrix is used to aggregate contextual information for appearance modeling and is converted into motion features for motion modeling as well. We validate the motion modeling capacities of our method on four widely-used datasets, and our method performs better than existing state-of-the-art approaches, especially on motion-sensitive datasets, i.e., Something-Something V1 & V2. Our project is available at https://github.com/PeiqinZhuang/EMIM .", "date": "2025-10-21", "categories": []}
{"id": "2510.20287", "title": "Breakdance Video classification in the age of Generative AI", "abstract": "Large Vision Language models have seen huge application in several sports use-cases recently. Most of these works have been targeted towards a limited subset of popular sports like soccer, cricket, basketball etc; focusing on generative tasks like visual question answering, highlight generation. This work analyzes the applicability of the modern video foundation models (both encoder and decoder) for a very niche but hugely popular dance sports - breakdance. Our results show that Video Encoder models continue to outperform state-of-the-art Video Language Models for prediction tasks. We provide insights on how to choose the encoder model and provide a thorough analysis into the workings of a finetuned decoder model for breakdance video classification.", "date": "2025-10-23", "categories": []}
{"id": "2510.19273", "title": "MobiAct: Efficient MAV Action Recognition Using MobileNetV4 with Contrastive Learning and Knowledge Distillation", "abstract": "Accurate and efficient recognition of Micro Air Vehicle (MAV) motion is essential for enabling real-time perception and coordination in autonomous aerial swarm. However, most existing approaches rely on large, computationally intensive models that are unsuitable for resource-limited MAV platforms, which results in a trade-off between recognition accuracy and inference speed. To address these challenges, this paper proposes a lightweight MAV action recognition framework, MobiAct, designed to achieve high accuracy with low computational cost. Specifically, MobiAct adopts MobileNetV4 as the backbone network and introduces a Stage-wise Orthogonal Knowledge Distillation (SOKD) strategy to effectively transfer MAV motion features from a teacher network (ResNet18) to a student network, thereby enhancing knowledge transfer efficiency. Furthermore, a parameter-free attention mechanism is integrated into the architecture to improve recognition accuracy without increasing model complexity. In addition, a hybrid loss training strategy is developed to combine multiple loss objectives, which ensures stable and robust optimization during training. Experimental results demonstrate that the proposed MobiAct achieves low-energy and low-computation MAV action recognition, while maintaining the fastest action decoding speed among compared methods. Across all three self-collected datasets, MobiAct achieves an average recognition accuracy of 92.12%, while consuming only 136.16 pJ of energy and processing recognition at a rate of 8.84 actions per second. Notably, MobiAct decodes actions up to 2 times faster than the leading method, with highly comparable recognition accuracy, highlighting its superior efficiency in MAV action recognition.", "date": "2025-10-22", "categories": []}
{"id": "2510.17863", "title": "Robotic Classification of Divers' Swimming States using Visual Pose Keypoints as IMUs", "abstract": "Traditional human activity recognition uses either direct image analysis or data from wearable inertial measurement units (IMUs), but can be ineffective in challenging underwater environments. We introduce a novel hybrid approach that bridges this gap to monitor scuba diver safety. Our method leverages computer vision to generate high-fidelity motion data, effectively creating a ``pseudo-IMU'' from a stream of 3D human joint keypoints. This technique circumvents the critical problem of wireless signal attenuation in water, which plagues conventional diver-worn sensors communicating with an Autonomous Underwater Vehicle (AUV). We apply this system to the vital task of identifying anomalous scuba diver behavior that signals the onset of a medical emergency such as cardiac arrest -- a leading cause of scuba diving fatalities. By integrating our classifier onboard an AUV and conducting experiments with simulated distress scenarios, we demonstrate the utility and effectiveness of our method for advancing robotic monitoring and diver safety.", "date": "2025-10-15", "categories": []}
{"id": "2502.03724", "title": "Seeing in the Dark: A Teacher-Student Framework for Dark Video Action Recognition via Knowledge Distillation and Contrastive Learning", "abstract": "Action recognition in dark or low-light (under-exposed) videos is a challenging task due to visibility degradation, which can hinder critical spatiotemporal details. This paper proposes ActLumos, a teacher-student framework that attains single-stream inference while retaining multi-stream level accuracy. The teacher consumes dual stream inputs, which include original dark frames and retinex-enhanced frames, processed by weight-shared R(2+1)D-34 backbones and fused by a Dynamic Feature Fusion (DFF) module, which dynamically re-weights the two streams at each time step, emphasising the most informative temporal segments. The teacher is also included with a supervised contrastive loss (SupCon) that sharpens class margins. The student shares the R(2+1)D-34 backbone but uses only dark frames and no fusion at test time. The student is first pre-trained with self-supervision on dark clips of both datasets without their labels and then fine-tuned with knowledge distillation from the teacher, transferring the teacher's multi-stream knowledge into a single-stream model. Under single-stream inference, the distilled student attains state-of-the-art accuracy of 96.92% (Top-1) on ARID V1.0, 88.27% on ARID V1.5, and 48.96% on Dark48. Ablation studies further highlight the individual contributions of each component, i.e., DFF in the teacher outperforms single or static fusion, knowledge distillation (KD) transfers these gains to the single-stream student, and two-view spatio-temporal SSL surpasses spatial-only or temporal-only variants without increasing inference cost. The official website of this work is available at: https://github.com/HrishavBakulBarua/ActLumos", "date": "2025-02-06", "categories": []}
{"id": "2510.16209", "title": "StretchySnake: Flexible SSM Training Unlocks Action Recognition Across Spatio-Temporal Scales", "abstract": "State space models (SSMs) have emerged as a competitive alternative to transformers in various tasks. Their linear complexity and hidden-state recurrence make them particularly attractive for modeling long sequences, whereas attention becomes quadratically expensive. However, current training methods for video understanding are tailored towards transformers and fail to fully leverage the unique attributes of SSMs. For example, video models are often trained at a fixed resolution and video length to balance the quadratic scaling of attention cost against performance. Consequently, these models suffer from degraded performance when evaluated on videos with spatial and temporal resolutions unseen during training; a property we call spatio-temporal inflexibility. In the context of action recognition, this severely limits a model's ability to retain performance across both short- and long-form videos. Therefore, we propose a flexible training method that leverages and improves the inherent adaptability of SSMs. Our method samples videos at varying temporal and spatial resolutions during training and dynamically interpolates model weights to accommodate any spatio-temporal scale. This instills our SSM, which we call StretchySnake, with spatio-temporal flexibility and enables it to seamlessly handle videos ranging from short, fine-grained clips to long, complex activities. We introduce and compare five different variants of flexible training, and identify the most effective strategy for video SSMs. On short-action (UCF-101, HMDB-51) and long-action (COIN, Breakfast) benchmarks, StretchySnake outperforms transformer and SSM baselines alike by up to 28%, with strong adaptability to fine-grained actions (SSV2, Diving-48). Therefore, our method provides a simple drop-in training recipe that makes video SSMs more robust, resolution-agnostic, and efficient across diverse action recognition scenarios.", "date": "2025-10-17", "categories": []}
{"id": "2510.16444", "title": "RefAtomNet++: Advancing Referring Atomic Video Action Recognition using Semantic Retrieval based Multi-Trajectory Mamba", "abstract": "Referring Atomic Video Action Recognition (RAVAR) aims to recognize fine-grained, atomic-level actions of a specific person of interest conditioned on natural language descriptions. Distinct from conventional action recognition and detection tasks, RAVAR emphasizes precise language-guided action understanding, which is particularly critical for interactive human action analysis in complex multi-person scenarios. In this work, we extend our previously introduced RefAVA dataset to RefAVA++, which comprises >2.9 million frames and >75.1k annotated persons in total. We benchmark this dataset using baselines from multiple related domains, including atomic action localization, video question answering, and text-video retrieval, as well as our earlier model, RefAtomNet. Although RefAtomNet surpasses other baselines by incorporating agent attention to highlight salient features, its ability to align and retrieve cross-modal information remains limited, leading to suboptimal performance in localizing the target person and predicting fine-grained actions. To overcome the aforementioned limitations, we introduce RefAtomNet++, a novel framework that advances cross-modal token aggregation through a multi-hierarchical semantic-aligned cross-attention mechanism combined with multi-trajectory Mamba modeling at the partial-keyword, scene-attribute, and holistic-sentence levels. In particular, scanning trajectories are constructed by dynamically selecting the nearest visual spatial tokens at each timestep for both partial-keyword and scene-attribute levels. Moreover, we design a multi-hierarchical semantic-aligned cross-attention strategy, enabling more effective aggregation of spatial and temporal tokens across different semantic hierarchies. Experiments show that RefAtomNet++ establishes new state-of-the-art results. The dataset and code are released at https://github.com/KPeng9510/refAVA2.", "date": "2025-10-18", "categories": []}
{"id": "2510.15448", "title": "MAVR-Net: Robust Multi-View Learning for MAV Action Recognition with Cross-View Attention", "abstract": "Recognizing the motion of Micro Aerial Vehicles (MAVs) is crucial for enabling cooperative perception and control in autonomous aerial swarms. Yet, vision-based recognition models relying only on RGB data often fail to capture the complex spatial temporal characteristics of MAV motion, which limits their ability to distinguish different actions. To overcome this problem, this paper presents MAVR-Net, a multi-view learning-based MAV action recognition framework. Unlike traditional single-view methods, the proposed approach combines three complementary types of data, including raw RGB frames, optical flow, and segmentation masks, to improve the robustness and accuracy of MAV motion recognition. Specifically, ResNet-based encoders are used to extract discriminative features from each view, and a multi-scale feature pyramid is adopted to preserve the spatiotemporal details of MAV motion patterns. To enhance the interaction between different views, a cross-view attention module is introduced to model the dependencies among various modalities and feature scales. In addition, a multi-view alignment loss is designed to ensure semantic consistency and strengthen cross-view feature representations. Experimental results on benchmark MAV action datasets show that our method clearly outperforms existing approaches, achieving 97.8\\%, 96.5\\%, and 92.8\\% accuracy on the Short MAV, Medium MAV, and Long MAV datasets, respectively.", "date": "2025-10-17", "categories": []}
{"id": "2510.15725", "title": "DGME-T: Directional Grid Motion Encoding for Transformer-Based Historical Camera Movement Classification", "abstract": "Camera movement classification (CMC) models trained on contemporary, high-quality footage often degrade when applied to archival film, where noise, missing frames, and low contrast obscure motion cues. We bridge this gap by assembling a unified benchmark that consolidates two modern corpora into four canonical classes and restructures the HISTORIAN collection into five balanced categories. Building on this benchmark, we introduce DGME-T, a lightweight extension to the Video Swin Transformer that injects directional grid motion encoding, derived from optical flow, via a learnable and normalised late-fusion layer. DGME-T raises the backbone's top-1 accuracy from 81.78% to 86.14% and its macro F1 from 82.08% to 87.81% on modern clips, while still improving the demanding World-War-II footage from 83.43% to 84.62% accuracy and from 81.72% to 82.63% macro F1. A cross-domain study further shows that an intermediate fine-tuning stage on modern data increases historical performance by more than five percentage points. These results demonstrate that structured motion priors and transformer representations are complementary and that even a small, carefully calibrated motion head can substantially enhance robustness in degraded film analysis. Related resources are available at https://github.com/linty5/DGME-T.", "date": "2025-10-17", "categories": []}
{"id": "2510.10478", "title": "MSF-Mamba: Motion-aware State Fusion Mamba for Efficient Micro-Gesture Recognition", "abstract": "Micro-gesture recognition (MGR) targets the identification of subtle and fine-grained human motions and requires accurate modeling of both long-range and local spatiotemporal dependencies. While CNNs are effective at capturing local patterns, they struggle with long-range dependencies due to their limited receptive fields. Transformer-based models address this limitation through self-attention mechanisms but suffer from high computational costs. Recently, Mamba has shown promise as an efficient model, leveraging state space models (SSMs) to enable linear-time processing However, directly applying the vanilla Mamba to MGR may not be optimal. This is because Mamba processes inputs as 1D sequences, with state updates relying solely on the previous state, and thus lacks the ability to model local spatiotemporal dependencies. In addition, previous methods lack a design of motion-awareness, which is crucial in MGR. To overcome these limitations, we propose motion-aware state fusion mamba (MSF-Mamba), which enhances Mamba with local spatiotemporal modeling by fusing local contextual neighboring states. Our design introduces a motion-aware state fusion module based on central frame difference (CFD). Furthermore, a multiscale version named MSF-Mamba+ has been proposed. Specifically, MSF-Mamba supports multiscale motion-aware state fusion, as well as an adaptive scale weighting module that dynamically weighs the fused states across different scales. These enhancements explicitly address the limitations of vanilla Mamba by enabling motion-aware local spatiotemporal modeling, allowing MSF-Mamba and MSF-Mamba to effectively capture subtle motion cues for MGR. Experiments on two public MGR datasets demonstrate that even the lightweight version, namely, MSF-Mamba, achieves SoTA performance, outperforming existing CNN-, Transformer-, and SSM-based models while maintaining high efficiency.", "date": "2025-10-12", "categories": []}
{"id": "2506.01353", "title": "EgoBrain: Synergizing Minds and Eyes For Human Action Understanding", "abstract": "The integration of brain-computer interfaces (BCIs), in particular electroencephalography (EEG), with artificial intelligence (AI) has shown tremendous promise in decoding human cognition and behavior from neural signals. In particular, the rise of multimodal AI models have brought new possibilities that have never been imagined before. Here, we present EgoBrain --the world's first large-scale, temporally aligned multimodal dataset that synchronizes egocentric vision and EEG of human brain over extended periods of time, establishing a new paradigm for human-centered behavior analysis. This dataset comprises 61 hours of synchronized 32-channel EEG recordings and first-person video from 40 participants engaged in 29 categories of daily activities. We then developed a muiltimodal learning framework to fuse EEG and vision for action understanding, validated across both cross-subject and cross-environment challenges, achieving an action recognition accuracy of 66.70%. EgoBrain paves the way for a unified framework for brain-computer interface with multiple modalities. All data, tools, and acquisition protocols are openly shared to foster open science in cognitive computing.", "date": "2025-06-02", "categories": []}
{"id": "2508.17025", "title": "Probabilistic Temporal Masked Attention for Cross-view Online Action Detection", "abstract": "As a critical task in video sequence classification within computer vision, Online Action Detection (OAD) has garnered significant attention. The sensitivity of mainstream OAD models to varying video viewpoints often hampers their generalization when confronted with unseen sources. To address this limitation, we propose a novel Probabilistic Temporal Masked Attention (PTMA) model, which leverages probabilistic modeling to derive latent compressed representations of video frames in a cross-view setting. The PTMA model incorporates a GRU-based temporal masked attention (TMA) cell, which leverages these representations to effectively query the input video sequence, thereby enhancing information interaction and facilitating autoregressive frame-level video analysis. Additionally, multi-view information can be integrated into the probabilistic modeling to facilitate the extraction of view-invariant features. Experiments conducted under three evaluation protocols: cross-subject (cs), cross-view (cv), and cross-subject-view (csv) show that PTMA achieves state-of-the-art performance on the DAHLIA, IKEA ASM, and Breakfast datasets.", "date": "2025-08-23", "categories": []}
{"id": "2504.03948", "title": "ProbRes: Probabilistic Jump Diffusion for Open-World Egocentric Activity Recognition", "abstract": "Open-world egocentric activity recognition poses a fundamental challenge due to its unconstrained nature, requiring models to infer unseen activities from an expansive, partially observed search space. We introduce ProbRes, a Probabilistic Residual search framework based on jump-diffusion that efficiently navigates this space by balancing prior-guided exploration with likelihood-driven exploitation. Our approach integrates structured commonsense priors to construct a semantically coherent search space, adaptively refines predictions using Vision-Language Models (VLMs) and employs a stochastic search mechanism to locate high-likelihood activity labels while minimizing exhaustive enumeration efficiently. We systematically evaluate ProbRes across multiple openness levels (L0-L3), demonstrating its adaptability to increasing search space complexity. In addition to achieving state-of-the-art performance on benchmark datasets (GTEA Gaze, GTEA Gaze+, EPIC-Kitchens, and Charades-Ego), we establish a clear taxonomy for open-world recognition, delineating the challenges and methodological advancements necessary for egocentric activity understanding. Our results highlight the importance of structured search strategies, paving the way for scalable and efficient open-world activity recognition.", "date": "2025-04-04", "categories": []}
{"id": "2505.19495", "title": "The Role of Video Generation in Enhancing Data-Limited Action Understanding", "abstract": "Video action understanding tasks in real-world scenarios always suffer data limitations. In this paper, we address the data-limited action understanding problem by bridging data scarcity. We propose a novel method that employs a text-to-video diffusion transformer to generate annotated data for model training. This paradigm enables the generation of realistic annotated data on an infinite scale without human intervention. We proposed the information enhancement strategy and the uncertainty-based label smoothing tailored to generate sample training. Through quantitative and qualitative analysis, we observed that real samples generally contain a richer level of information than generated samples. Based on this observation, the information enhancement strategy is proposed to enhance the informative content of the generated samples from two aspects: the environments and the characters. Furthermore, we observed that some low-quality generated samples might negatively affect model training. To address this, we devised the uncertainty-based label smoothing strategy to increase the smoothing of these samples, thus reducing their impact. We demonstrate the effectiveness of the proposed method on four datasets across five tasks and achieve state-of-the-art performance for zero-shot action recognition.", "date": "2025-05-26", "categories": []}
{"id": "2510.05506", "title": "Human Action Recognition from Point Clouds over Time", "abstract": "Recent research into human action recognition (HAR) has focused predominantly on skeletal action recognition and video-based methods. With the increasing availability of consumer-grade depth sensors and Lidar instruments, there is a growing opportunity to leverage dense 3D data for action recognition, to develop a third way. This paper presents a novel approach for recognizing actions from 3D videos by introducing a pipeline that segments human point clouds from the background of a scene, tracks individuals over time, and performs body part segmentation. The method supports point clouds from both depth sensors and monocular depth estimation. At the core of the proposed HAR framework is a novel backbone for 3D action recognition, which combines point-based techniques with sparse convolutional networks applied to voxel-mapped point cloud sequences. Experiments incorporate auxiliary point features including surface normals, color, infrared intensity, and body part parsing labels, to enhance recognition accuracy. Evaluation on the NTU RGB- D 120 dataset demonstrates that the method is competitive with existing skeletal action recognition algorithms. Moreover, combining both sensor-based and estimated depth inputs in an ensemble setup, this approach achieves 89.3% accuracy when different human subjects are considered for training and testing, outperforming previous point cloud action recognition methods.", "date": "2025-10-07", "categories": []}
{"id": "2510.07652", "title": "Dual-Stream Alignment for Action Segmentation", "abstract": "Action segmentation is a challenging yet active research area that involves identifying when and where specific actions occur in continuous video streams. Most existing work has focused on single-stream approaches that model the spatio-temporal aspects of frame sequences. However, recent research has shifted toward two-stream methods that learn action-wise features to enhance action segmentation performance. In this work, we propose the Dual-Stream Alignment Network (DSA Net) and investigate the impact of incorporating a second stream of learned action features to guide segmentation by capturing both action and action-transition cues. Communication between the two streams is facilitated by a Temporal Context (TC) block, which fuses complementary information using cross-attention and Quantum-based Action-Guided Modulation (Q-ActGM), enhancing the expressive power of the fused features. To the best of our knowledge, this is the first study to introduce a hybrid quantum-classical machine learning framework for action segmentation. Our primary objective is for the two streams (frame-wise and action-wise) to learn a shared feature space through feature alignment. This is encouraged by the proposed Dual-Stream Alignment Loss, which comprises three components: relational consistency, cross-level contrastive, and cycle-consistency reconstruction losses. Following prior work, we evaluate DSA Net on several diverse benchmark datasets: GTEA, Breakfast, 50Salads, and EgoProcel. We further demonstrate the effectiveness of each component through extensive ablation studies. Notably, DSA Net achieves state-of-the-art performance, significantly outperforming existing", "date": "2025-10-09", "categories": []}
{"id": "2505.15192", "title": "Leveraging Foundation Models for Multimodal Graph-Based Action Recognition", "abstract": "Foundation models have ushered in a new era for multimodal video understanding by enabling the extraction of rich spatiotemporal and semantic representations. In this work, we introduce a novel graph-based framework that integrates a vision-language foundation, leveraging VideoMAE for dynamic visual encoding and BERT for contextual textual embedding, to address the challenge of recognizing fine-grained bimanual manipulation actions. Departing from conventional static graph architectures, our approach constructs an adaptive multimodal graph where nodes represent frames, objects, and textual annotations, and edges encode spatial, temporal, and semantic relationships. These graph structures evolve dynamically based on learned interactions, allowing for flexible and context-aware reasoning. A task-specific attention mechanism within a Graph Attention Network further enhances this reasoning by modulating edge importance based on action semantics. Through extensive evaluations on diverse benchmark datasets, we demonstrate that our method consistently outperforms state-of-the-art baselines, underscoring the strength of combining foundation models with dynamic graph-based reasoning for robust and generalizable action recognition.", "date": "2025-05-21", "categories": []}
{"id": "2509.02511", "title": "Enhancing Fitness Movement Recognition with Attention Mechanism and Pre-Trained Feature Extractors", "abstract": "Fitness movement recognition, a focused subdomain of human activity recognition (HAR), plays a vital role in health monitoring, rehabilitation, and personalized fitness training by enabling automated exercise classification from video data. However, many existing deep learning approaches rely on computationally intensive 3D models, limiting their feasibility in real-time or resource-constrained settings. In this paper, we present a lightweight and effective framework that integrates pre-trained 2D Convolutional Neural Networks (CNNs) such as ResNet50, EfficientNet, and Vision Transformers (ViT) with a Long Short-Term Memory (LSTM) network enhanced by spatial attention. These models efficiently extract spatial features while the LSTM captures temporal dependencies, and the attention mechanism emphasizes informative segments. We evaluate the framework on a curated subset of the UCF101 dataset, achieving a peak accuracy of 93.34\\% with the ResNet50-based configuration. Comparative results demonstrate the superiority of our approach over several state-of-the-art HAR systems. The proposed method offers a scalable and real-time-capable solution for fitness activity recognition with broader applications in vision-based health and activity monitoring.", "date": "2025-09-02", "categories": []}
{"id": "2503.04470", "title": "Gate-Shift-Pose: Enhancing Action Recognition in Sports with Skeleton Information", "abstract": "This paper introduces Gate-Shift-Pose, an enhanced version of Gate-Shift-Fuse networks, designed for athlete fall classification in figure skating by integrating skeleton pose data alongside RGB frames. We evaluate two fusion strategies: early-fusion, which combines RGB frames with Gaussian heatmaps of pose keypoints at the input stage, and late-fusion, which employs a multi-stream architecture with attention mechanisms to combine RGB and pose features. Experiments on the FR-FS dataset demonstrate that Gate-Shift-Pose significantly outperforms the RGB-only baseline, improving accuracy by up to 40% with ResNet18 and 20% with ResNet50. Early-fusion achieves the highest accuracy (98.08%) with ResNet50, leveraging the model's capacity for effective multimodal integration, while late-fusion is better suited for lighter backbones like ResNet18. These results highlight the potential of multimodal architectures for sports action recognition and the critical role of skeleton pose information in capturing complex motion patterns. Visit the project page at https://edowhite.github.io/Gate-Shift-Pose", "date": "2025-03-06", "categories": []}
{"id": "2506.09650", "title": "HopaDIFF: Holistic-Partial Aware Fourier Conditioned Diffusion for Referring Human Action Segmentation in Multi-Person Scenarios", "abstract": "Action segmentation is a core challenge in high-level video understanding, aiming to partition untrimmed videos into segments and assign each a label from a predefined action set. Existing methods primarily address single-person activities with fixed action sequences, overlooking multi-person scenarios. In this work, we pioneer textual reference-guided human action segmentation in multi-person settings, where a textual description specifies the target person for segmentation. We introduce the first dataset for Referring Human Action Segmentation, i.e., RHAS133, built from 133 movies and annotated with 137 fine-grained actions with 33h video data, together with textual descriptions for this new task. Benchmarking existing action segmentation methods on RHAS133 using VLM-based feature extractors reveals limited performance and poor aggregation of visual cues for the target person. To address this, we propose a holistic-partial aware Fourier-conditioned diffusion framework, i.e., HopaDIFF, leveraging a novel cross-input gate attentional xLSTM to enhance holistic-partial long-range reasoning and a novel Fourier condition to introduce more fine-grained control to improve the action segmentation generation. HopaDIFF achieves state-of-the-art results on RHAS133 in diverse evaluation settings. The dataset and code are available at https://github.com/KPeng9510/HopaDIFF.", "date": "2025-06-11", "categories": []}
{"id": "2509.23009", "title": "Disentangling Static and Dynamic Information for Reducing Static Bias in Action Recognition", "abstract": "Action recognition models rely excessively on static cues rather than dynamic human motion, which is known as static bias. This bias leads to poor performance in real-world applications and zero-shot action recognition. In this paper, we propose a method to reduce static bias by separating temporal dynamic information from static scene information. Our approach uses a statistical independence loss between biased and unbiased streams, combined with a scene prediction loss. Our experiments demonstrate that this method effectively reduces static bias and confirm the importance of scene prediction loss.", "date": "2025-09-27", "categories": []}
{"id": "2509.23044", "title": "MMeViT: Multi-Modal ensemble ViT for Post-Stroke Rehabilitation Action Recognition", "abstract": "Rehabilitation therapy for stroke patients faces a supply shortage despite the increasing demand. To address this issue, remote monitoring systems that reduce the burden on medical staff are emerging as a viable alternative. A key component of these remote monitoring systems is Human Action Recognition (HAR) technology, which classifies actions. However, existing HAR studies have primarily focused on non-disable individuals, making them unsuitable for recognizing the actions of stroke patients. HAR research for stroke has largely concentrated on classifying relatively simple actions using machine learning rather than deep learning. In this study, we designed a system to monitor the actions of stroke patients, focusing on domiciliary upper limb Activities of Daily Living (ADL). Our system utilizes IMU (Inertial Measurement Unit) sensors and an RGB-D camera, which are the most common modalities in HAR. We directly collected a dataset through this system, investigated an appropriate preprocess and proposed a deep learning model suitable for processing multimodal data. We analyzed the collected dataset and found that the action data of stroke patients is less clustering than that of non-disabled individuals. Simultaneously, we found that the proposed model learns similar tendencies for each label in data with features that are difficult to clustering. This study suggests the possibility of expanding the deep learning model, which has learned the action features of stroke patients, to not only simple action recognition but also feedback such as assessment contributing to domiciliary rehabilitation in future research. The code presented in this study is available at https://github.com/ye-Kim/MMeViT.", "date": "2025-09-27", "categories": []}
{"id": "2509.23255", "title": "LiDAR-based Human Activity Recognition through Laplacian Spectral Analysis", "abstract": "Human Activity Recognition supports applications in healthcare, manufacturing, and human-machine interaction. LiDAR point clouds offer a privacy-preserving alternative to cameras and are robust to illumination. We propose a HAR method based on graph spectral analysis. Each LiDAR frame is mapped to a proximity graph (epsilon-graph) and the Laplacian spectrum is computed. Eigenvalues and statistics of eigenvectors form pose descriptors, and temporal statistics over sliding windows yield fixed vectors for classification with support vector machines and random forests. On the MM-Fi dataset with 40 subjects and 27 activities, under a strict subject-independent protocol, the method reaches 94.4% accuracy on a 13-class rehabilitation set and 90.3% on all 27 activities. It also surpasses the skeleton-based baselines reported for MM-Fi. The contribution is a compact and interpretable feature set derived directly from point cloud geometry that provides an accurate and efficient alternative to end-to-end deep learning.", "date": "2025-09-27", "categories": []}
{"id": "2509.23517", "title": "Evaluating point-light biological motion in multimodal large language models", "abstract": "Humans can extract rich semantic information from minimal visual cues, as demonstrated by point-light displays (PLDs), which consist of sparse sets of dots localized to key joints of the human body. This ability emerges early in development and is largely attributed to human embodied experience. Since PLDs isolate body motion as the sole source of meaning, they represent key stimuli for testing the constraints of action understanding in these systems. Here we introduce ActPLD, the first benchmark to evaluate action processing in MLLMs from human PLDs. Tested models include state-of-the-art proprietary and open-source systems on single-actor and socially interacting PLDs. Our results reveal consistently low performance across models, introducing fundamental gaps in action and spatiotemporal understanding.", "date": "2025-09-27", "categories": []}
{"id": "2509.23888", "title": "AssemblyHands-X: Modeling 3D Hand-Body Coordination for Understanding Bimanual Human Activities", "abstract": "Bimanual human activities inherently involve coordinated movements of both hands and body. However, the impact of this coordination in activity understanding has not been systematically evaluated due to the lack of suitable datasets. Such evaluation demands kinematic-level annotations (e.g., 3D pose) for the hands and body, yet existing 3D activity datasets typically annotate either hand or body pose. Another line of work employs marker-based motion capture to provide full-body pose, but the physical markers introduce visual artifacts, thereby limiting models' generalization to natural, markerless videos. To address these limitations, we present AssemblyHands-X, the first markerless 3D hand-body benchmark for bimanual activities, designed to study the effect of hand-body coordination for action recognition. We begin by constructing a pipeline for 3D pose annotation from synchronized multi-view videos. Our approach combines multi-view triangulation with SMPL-X mesh fitting, yielding reliable 3D registration of hands and upper body. We then validate different input representations (e.g., video, hand pose, body pose, or hand-body pose) across recent action recognition models based on graph convolution or spatio-temporal attention. Our extensive experiments show that pose-based action inference is more efficient and accurate than video baselines. Moreover, joint modeling of hand and body cues improves action recognition over using hands or upper body alone, highlighting the importance of modeling interdependent hand-body dynamics for a holistic understanding of bimanual activities.", "date": "2025-09-28", "categories": []}
{"id": "2509.24303", "title": "Experience Paper: Adopting Activity Recognition in On-demand Food Delivery Business", "abstract": "This paper presents the first nationwide deployment of human activity recognition (HAR) technology in the on-demand food delivery industry. We successfully adapted the state-of-the-art LIMU-BERT foundation model to the delivery platform. Spanning three phases over two years, the deployment progresses from a feasibility study in Yangzhou City to nationwide adoption involving 500,000 couriers across 367 cities in China. The adoption enables a series of downstream applications, and large-scale tests demonstrate its significant operational and economic benefits, showcasing the transformative potential of HAR technology in real-world applications. Additionally, we share lessons learned from this deployment and open-source our LIMU-BERT pretrained with millions of hours of sensor data.", "date": "2025-09-29", "categories": []}
{"id": "2509.25042", "title": "Fast Real-Time Pipeline for Robust Arm Gesture Recognition", "abstract": "This paper presents a real-time pipeline for dynamic arm gesture recognition based on OpenPose keypoint estimation, keypoint normalization, and a recurrent neural network classifier. The 1 x 1 normalization scheme and two feature representations (coordinate- and angle-based) are presented for the pipeline. In addition, an efficient method to improve robustness against camera angle variations is also introduced by using artificially rotated training data. Experiments on a custom traffic-control gesture dataset demonstrate high accuracy across varying viewing angles and speeds. Finally, an approach to calculate the speed of the arm signal (if necessary) is also presented.", "date": "2025-09-29", "categories": []}
{"id": "2509.21595", "title": "Temporal vs. Spatial: Comparing DINOv3 and V-JEPA2 Feature Representations for Video Action Analysis", "abstract": "This study presents a comprehensive comparative analysis of two prominent self-supervised learning architectures for video action recognition: DINOv3, which processes frames independently through spatial feature extraction, and V-JEPA2, which employs joint temporal modeling across video sequences. We evaluate both approaches on the UCF Sports dataset, examining feature quality through multiple dimensions including classification accuracy, clustering performance, intra-class consistency, and inter-class discrimination. Our analysis reveals fundamental architectural trade-offs: DINOv3 achieves superior clustering performance (Silhouette score: 0.31 vs 0.21) and demonstrates exceptional discrimination capability (6.16x separation ratio) particularly for pose-identifiable actions, while V-JEPA2 exhibits consistent reliability across all action types with significantly lower performance variance (0.094 vs 0.288). Through action-specific evaluation, we identify that DINOv3's spatial processing architecture excels at static pose recognition but shows degraded performance on motion-dependent actions, whereas V-JEPA2's temporal modeling provides balanced representation quality across diverse action categories. These findings contribute to the understanding of architectural design choices in video analysis systems and provide empirical guidance for selecting appropriate feature extraction methods based on task requirements and reliability constraints.", "date": "2025-09-25", "categories": []}
{"id": "2509.17084", "title": "MoCLIP-Lite: Efficient Video Recognition by Fusing CLIP with Motion Vectors", "abstract": "Video action recognition is a fundamental task in computer vision, but state-of-the-art models are often computationally expensive and rely on extensive video pre-training. In parallel, large-scale vision-language models like Contrastive Language-Image Pre-training (CLIP) offer powerful zero-shot capabilities on static images, while motion vectors (MV) provide highly efficient temporal information directly from compressed video streams. To synergize the strengths of these paradigms, we propose MoCLIP-Lite, a simple yet powerful two-stream late fusion framework for efficient video recognition. Our approach combines features from a frozen CLIP image encoder with features from a lightweight, supervised network trained on raw MV. During fusion, both backbones are frozen, and only a tiny Multi-Layer Perceptron (MLP) head is trained, ensuring extreme efficiency. Through comprehensive experiments on the UCF101 dataset, our method achieves a remarkable 89.2% Top-1 accuracy, significantly outperforming strong zero-shot (65.0%) and MV-only (66.5%) baselines. Our work provides a new, highly efficient baseline for video understanding that effectively bridges the gap between large static models and dynamic, low-cost motion cues. Our code and models are available at https://github.com/microa/MoCLIP-Lite.", "date": "2025-09-21", "categories": []}
{"id": "2509.08908", "title": "Diffusion-Based Action Recognition Generalizes to Untrained Domains", "abstract": "Humans can recognize the same actions despite large context and viewpoint variations, such as differences between species (walking in spiders vs. horses), viewpoints (egocentric vs. third-person), and contexts (real life vs movies). Current deep learning models struggle with such generalization. We propose using features generated by a Vision Diffusion Model (VDM), aggregated via a transformer, to achieve human-like action recognition across these challenging conditions. We find that generalization is enhanced by the use of a model conditioned on earlier timesteps of the diffusion process to highlight semantic information over pixel level details in the extracted features. We experimentally explore the generalization properties of our approach in classifying actions across animal species, across different viewing angles, and different recording contexts. Our model sets a new state-of-the-art across all three generalization benchmarks, bringing machine action recognition closer to human-like robustness. Project page: https://www.vision.caltech.edu/actiondiff. Code: https://github.com/frankyaoxiao/ActionDiff", "date": "2025-09-10", "categories": []}
{"id": "2509.16452", "title": "KRAST: Knowledge-Augmented Robotic Action Recognition with Structured Text for Vision-Language Models", "abstract": "Accurate vision-based action recognition is crucial for developing autonomous robots that can operate safely and reliably in complex, real-world environments. In this work, we advance video-based recognition of indoor daily actions for robotic perception by leveraging vision-language models (VLMs) enriched with domain-specific knowledge. We adapt a prompt-learning framework in which class-level textual descriptions of each action are embedded as learnable prompts into a frozen pre-trained VLM backbone. Several strategies for structuring and encoding these textual descriptions are designed and evaluated. Experiments on the ETRI-Activity3D dataset demonstrate that our method, using only RGB video inputs at test time, achieves over 95\\% accuracy and outperforms state-of-the-art approaches. These results highlight the effectiveness of knowledge-augmented prompts in enabling robust action recognition with minimal supervision.", "date": "2025-09-19", "categories": []}
{"id": "2509.17638", "title": "A$^2$M$^2$-Net: Adaptively Aligned Multi-Scale Moment for Few-Shot Action Recognition", "abstract": "Thanks to capability to alleviate the cost of large-scale annotation, few-shot action recognition (FSAR) has attracted increased attention of researchers in recent years. Existing FSAR approaches typically neglect the role of individual motion pattern in comparison, and under-explore the feature statistics for video dynamics. Thereby, they struggle to handle the challenging temporal misalignment in video dynamics, particularly by using 2D backbones. To overcome these limitations, this work proposes an adaptively aligned multi-scale second-order moment network, namely A$^2$M$^2$-Net, to describe the latent video dynamics with a collection of powerful representation candidates and adaptively align them in an instance-guided manner. To this end, our A$^2$M$^2$-Net involves two core components, namely, adaptive alignment (A$^2$ module) for matching, and multi-scale second-order moment (M$^2$ block) for strong representation. Specifically, M$^2$ block develops a collection of semantic second-order descriptors at multiple spatio-temporal scales. Furthermore, A$^2$ module aims to adaptively select informative candidate descriptors while considering the individual motion pattern. By such means, our A$^2$M$^2$-Net is able to handle the challenging temporal misalignment problem by establishing an adaptive alignment protocol for strong representation. Notably, our proposed method generalizes well to various few-shot settings and diverse metrics. The experiments are conducted on five widely used FSAR benchmarks, and the results show our A$^2$M$^2$-Net achieves very competitive performance compared to state-of-the-arts, demonstrating its effectiveness and generalization.", "date": "2025-09-22", "categories": []}
{"id": "2509.17888", "title": "Trainee Action Recognition through Interaction Analysis in CCATT Mixed-Reality Training", "abstract": "This study examines how Critical Care Air Transport Team (CCATT) members are trained using mixed-reality simulations that replicate the high-pressure conditions of aeromedical evacuation. Each team - a physician, nurse, and respiratory therapist - must stabilize severely injured soldiers by managing ventilators, IV pumps, and suction devices during flight. Proficient performance requires clinical expertise and cognitive skills, such as situational awareness, rapid decision-making, effective communication, and coordinated task management, all of which must be maintained under stress. Recent advances in simulation and multimodal data analytics enable more objective and comprehensive performance evaluation. In contrast, traditional instructor-led assessments are subjective and may overlook critical events, thereby limiting generalizability and consistency. However, AI-based automated and more objective evaluation metrics still demand human input to train the AI algorithms to assess complex team dynamics in the presence of environmental noise and the need for accurate re-identification in multi-person tracking. To address these challenges, we introduce a systematic, data-driven assessment framework that combines Cognitive Task Analysis (CTA) with Multimodal Learning Analytics (MMLA). We have developed a domain-specific CTA model for CCATT training and a vision-based action recognition pipeline using a fine-tuned Human-Object Interaction model, the Cascade Disentangling Network (CDN), to detect and track trainee-equipment interactions over time. These interactions automatically yield performance indicators (e.g., reaction time, task duration), which are mapped onto a hierarchical CTA model tailored to CCATT operations, enabling interpretable, domain-relevant performance evaluations.", "date": "2025-09-22", "categories": []}
{"id": "2509.14619", "title": "LSTC-MDA: A Unified Framework for Long-Short Term Temporal Convolution and Mixed Data Augmentation in Skeleton-Based Action Recognition", "abstract": "Skeleton-based action recognition faces two longstanding challenges: the scarcity of labeled training samples and difficulty modeling short- and long-range temporal dependencies. To address these issues, we propose a unified framework, LSTC-MDA, which simultaneously improves temporal modeling and data diversity. We introduce a novel Long-Short Term Temporal Convolution (LSTC) module with parallel short- and long-term branches, these two feature branches are then aligned and fused adaptively using learned similarity weights to preserve critical long-range cues lost by conventional stride-2 temporal convolutions. We also extend Joint Mixing Data Augmentation (JMDA) with an Additive Mixup at the input level, diversifying training samples and restricting mixup operations to the same camera view to avoid distribution shifts. Ablation studies confirm each component contributes. LSTC-MDA achieves state-of-the-art results: 94.1% and 97.5% on NTU 60 (X-Sub and X-View), 90.4% and 92.0% on NTU 120 (X-Sub and X-Set),97.2% on NW-UCLA. Code: https://github.com/xiaobaoxia/LSTC-MDA.", "date": "2025-09-18", "categories": []}
{"id": "2509.09067", "title": "Improvement of Human-Object Interaction Action Recognition Using Scene Information and Multi-Task Learning Approach", "abstract": "Recent graph convolutional neural networks (GCNs) have shown high performance in the field of human action recognition by using human skeleton poses. However, it fails to detect human-object interaction cases successfully due to the lack of effective representation of the scene information and appropriate learning architectures. In this context, we propose a methodology to utilize human action recognition performance by considering fixed object information in the environment and following a multi-task learning approach. In order to evaluate the proposed method, we collected real data from public environments and prepared our data set, which includes interaction classes of hands-on fixed objects (e.g., ATM ticketing machines, check-in/out machines, etc.) and non-interaction classes of walking and standing. The multi-task learning approach, along with interaction area information, succeeds in recognizing the studied interaction and non-interaction actions with an accuracy of 99.25%, outperforming the accuracy of the base model using only human skeleton poses by 2.75%.", "date": "2025-09-11", "categories": []}
{"id": "2509.12145", "title": "Open-ended Hierarchical Streaming Video Understanding with Vision Language Models", "abstract": "We introduce Hierarchical Streaming Video Understanding, a task that combines online temporal action localization with free-form description generation. Given the scarcity of datasets with hierarchical and fine-grained temporal annotations, we demonstrate that LLMs can effectively group atomic actions into higher-level events, enriching existing datasets. We then propose OpenHOUSE (Open-ended Hierarchical Online Understanding System for Events), which extends streaming action perception beyond action classification. OpenHOUSE features a specialized streaming module that accurately detects boundaries between closely adjacent actions, nearly doubling the performance of direct extensions of existing methods. We envision the future of streaming action perception in the integration of powerful generative models, with OpenHOUSE representing a key step in that direction.", "date": "2025-09-15", "categories": []}
{"id": "2509.12193", "title": "Domain-Adaptive Pretraining Improves Primate Behavior Recognition", "abstract": "Computer vision for animal behavior offers promising tools to aid research in ecology, cognition, and to support conservation efforts. Video camera traps allow for large-scale data collection, but high labeling costs remain a bottleneck to creating large-scale datasets. We thus need data-efficient learning approaches. In this work, we show that we can utilize self-supervised learning to considerably improve action recognition on primate behavior. On two datasets of great ape behavior (PanAf and ChimpACT), we outperform published state-of-the-art action recognition models by 6.1 %pt. accuracy and 6.3 %pt. mAP, respectively. We achieve this by utilizing a pretrained V-JEPA model and applying domain-adaptive pretraining (DAP), i.e. continuing the pretraining with in-domain data. We show that most of the performance gain stems from the DAP. Our method promises great potential for improving the recognition of animal behavior, as DAP does not require labeled samples. Code is available at https://github.com/ecker-lab/dap-behavior", "date": "2025-09-15", "categories": []}
{"id": "2304.11631", "title": "TSGCNeXt: Dynamic-Static Multi-Graph Convolution for Efficient Skeleton-Based Action Recognition with Long-term Learning Potential", "abstract": "Skeleton-based action recognition has achieved remarkable results in human action recognition with the development of graph convolutional networks (GCNs). However, the recent works tend to construct complex learning mechanisms with redundant training and exist a bottleneck for long time-series. To solve these problems, we propose the Temporal-Spatio Graph ConvNeXt (TSGCNeXt) to explore efficient learning mechanism of long temporal skeleton sequences. Firstly, a new graph learning mechanism with simple structure, Dynamic-Static Separate Multi-graph Convolution (DS-SMG) is proposed to aggregate features of multiple independent topological graphs and avoid the node information being ignored during dynamic convolution. Next, we construct a graph convolution training acceleration mechanism to optimize the back-propagation computing of dynamic graph learning with 55.08\\% speed-up. Finally, the TSGCNeXt restructure the overall structure of GCN with three Spatio-temporal learning modules,efficiently modeling long temporal features. In comparison with existing previous methods on large-scale datasets NTU RGB+D 60 and 120, TSGCNeXt outperforms on single-stream networks. In addition, with the ema model introduced into the multi-stream fusion, TSGCNeXt achieves SOTA levels. On the cross-subject and cross-set of the NTU 120, accuracies reach 90.22% and 91.74%.", "date": "2023-04-23", "categories": []}
{"id": "2509.08225", "title": "Ensemble Distribution Distillation for Self-Supervised Human Activity Recognition", "abstract": "Human Activity Recognition (HAR) has seen significant advancements with the adoption of deep learning techniques, yet challenges remain in terms of data requirements, reliability and robustness. This paper explores a novel application of Ensemble Distribution Distillation (EDD) within a self-supervised learning framework for HAR aimed at overcoming these challenges. By leveraging unlabeled data and a partially supervised training strategy, our approach yields an increase in predictive accuracy, robust estimates of uncertainty, and substantial increases in robustness against adversarial perturbation; thereby significantly improving reliability in real-world scenarios without increasing computational complexity at inference. We demonstrate this with an evaluation on several publicly available datasets. The contributions of this work include the development of a self-supervised EDD framework, an innovative data augmentation technique designed for HAR, and empirical validation of the proposed method's effectiveness in increasing robustness and reliability.", "date": "2025-09-10", "categories": []}
{"id": "2509.05695", "title": "Leveraging Vision-Language Large Models for Interpretable Video Action Recognition with Semantic Tokenization", "abstract": "Human action recognition often struggles with deep semantic understanding, complex contextual information, and fine-grained distinction, limitations that traditional methods frequently encounter when dealing with diverse video data. Inspired by the remarkable capabilities of large language models, this paper introduces LVLM-VAR, a novel framework that pioneers the application of pre-trained Vision-Language Large Models (LVLMs) to video action recognition, emphasizing enhanced accuracy and interpretability. Our method features a Video-to-Semantic-Tokens (VST) Module, which innovatively transforms raw video sequences into discrete, semantically and temporally consistent \"semantic action tokens,\" effectively crafting an \"action narrative\" that is comprehensible to an LVLM. These tokens, combined with natural language instructions, are then processed by a LoRA-fine-tuned LVLM (e.g., LLaVA-13B) for robust action classification and semantic reasoning. LVLM-VAR not only achieves state-of-the-art or highly competitive performance on challenging benchmarks such as NTU RGB+D and NTU RGB+D 120, demonstrating significant improvements (e.g., 94.1% on NTU RGB+D X-Sub and 90.0% on NTU RGB+D 120 X-Set), but also substantially boosts model interpretability by generating natural language explanations for its predictions.", "date": "2025-09-06", "categories": []}
{"id": "2509.04736", "title": "WatchHAR: Real-time On-device Human Activity Recognition System for Smartwatches", "abstract": "Despite advances in practical and multimodal fine-grained Human Activity Recognition (HAR), a system that runs entirely on smartwatches in unconstrained environments remains elusive. We present WatchHAR, an audio and inertial-based HAR system that operates fully on smartwatches, addressing privacy and latency issues associated with external data processing. By optimizing each component of the pipeline, WatchHAR achieves compounding performance gains. We introduce a novel architecture that unifies sensor data preprocessing and inference into an end-to-end trainable module, achieving 5x faster processing while maintaining over 90% accuracy across more than 25 activity classes. WatchHAR outperforms state-of-the-art models for event detection and activity classification while running directly on the smartwatch, achieving 9.3 ms processing time for activity event detection and 11.8 ms for multimodal activity classification. This research advances on-device activity recognition, realizing smartwatches' potential as standalone, privacy-aware, and minimally-invasive continuous activity tracking devices.", "date": "2025-09-05", "categories": []}
{"id": "2409.05662", "title": "Real-Time Human Action Recognition on Embedded Platforms", "abstract": "With advancements in computer vision and deep learning, video-based human action recognition (HAR) has become practical. However, due to the complexity of the computation pipeline, running HAR on live video streams incurs excessive delays on embedded platforms. This work tackles the real-time performance challenges of HAR with four contributions: 1) an experimental study identifying a standard Optical Flow (OF) extraction technique as the latency bottleneck in a state-of-the-art HAR pipeline, 2) an exploration of the latency-accuracy tradeoff between the standard and deep learning approaches to OF extraction, which highlights the need for a novel, efficient motion feature extractor, 3) the design of Integrated Motion Feature Extractor (IMFE), a novel single-shot neural network architecture for motion feature extraction with drastic improvement in latency, 4) the development of RT-HARE, a real-time HAR system tailored for embedded platforms. Experimental results on an Nvidia Jetson Xavier NX platform demonstrated that RT-HARE realizes real-time HAR at a video frame rate of 30 frames per second while delivering high levels of recognition accuracy.", "date": "2024-09-09", "categories": []}
{"id": "2509.00692", "title": "CascadeFormer: A Family of Two-stage Cascading Transformers for Skeleton-based Human Action Recognition", "abstract": "Skeleton-based human action recognition leverages sequences of human joint coordinates to identify actions performed in videos. Owing to the intrinsic spatiotemporal structure of skeleton data, Graph Convolutional Networks (GCNs) have been the dominant architecture in this field. However, recent advances in transformer models and masked pretraining frameworks open new avenues for representation learning. In this work, we propose CascadeFormer, a family of two-stage cascading transformers for skeleton-based human action recognition. Our framework consists of a masked pretraining stage to learn generalizable skeleton representations, followed by a cascading fine-tuning stage tailored for discriminative action classification. We evaluate CascadeFormer across three benchmark datasets (Penn Action N-UCLA, and NTU RGB+D 60), achieving competitive performance on all tasks. To promote reproducibility, we release our code and model checkpoints.", "date": "2025-08-31", "categories": []}
{"id": "2403.08420", "title": "ALow-Cost Real-Time Framework for Industrial Action Recognition Using Foundation Models", "abstract": "Action recognition (AR) in industrial environments -- particularly for identifying actions and operational gestures -- faces persistent challenges due to high deployment costs, poor cross-scenario generalization, and limited real-time performance. To address these issues, we propose a low-cost real-time framework for industrial action recognition using foundation models, denoted as LRIAR, to enhance recognition accuracy and transferability while minimizing human annotation and computational overhead. The proposed framework constructs an automatically labeled dataset by coupling Grounding DINO with the pretrained BLIP-2 image encoder, enabling efficient and scalable action labeling. Leveraging the constructed dataset, we train YOLOv5 for real-time action detection, and a Vision Transformer (ViT) classifier is deceloped via LoRA-based fine-tuning for action classification. Extensive experiments conducted in real-world industrial settings validate the effectiveness of LRIAR, demonstrating consistent improvements over state-of-the-art methods in recognition accuracy, scenario generalization, and deployment efficiency.", "date": "2024-03-13", "categories": []}
{"id": "2503.06316", "title": "End-to-End Action Segmentation Transformer", "abstract": "Most recent work on action segmentation relies on pre-computed frame features from models trained on other tasks and typically focuses on framewise encoding and labeling without explicitly modeling action segments. To overcome these limitations, we introduce the End-to-End Action Segmentation Transformer (EAST), which processes raw video frames directly -- eliminating the need for pre-extracted features and enabling true end-to-end training. Our contributions are as follows: (1) a lightweight adapter design for effective fine-tuning of large backbones; (2) an efficient segmentation-by-detection framework for leveraging action proposals predicted over a coarsely downsampled video; and (3) a novel action-proposal-based data augmentation strategy. EAST achieves SOTA performance on standard benchmarks, including GTEA, 50Salads, Breakfast, and Assembly-101.", "date": "2025-03-08", "categories": []}
{"id": "2508.19647", "title": "UTAL-GNN: Unsupervised Temporal Action Localization using Graph Neural Networks", "abstract": "Fine-grained action localization in untrimmed sports videos presents a significant challenge due to rapid and subtle motion transitions over short durations. Existing supervised and weakly supervised solutions often rely on extensive annotated datasets and high-capacity models, making them computationally intensive and less adaptable to real-world scenarios. In this work, we introduce a lightweight and unsupervised skeleton-based action localization pipeline that leverages spatio-temporal graph neural representations. Our approach pre-trains an Attention-based Spatio-Temporal Graph Convolutional Network (ASTGCN) on a pose-sequence denoising task with blockwise partitions, enabling it to learn intrinsic motion dynamics without any manual labeling. At inference, we define a novel Action Dynamics Metric (ADM), computed directly from low-dimensional ASTGCN embeddings, which detects motion boundaries by identifying inflection points in its curvature profile. Our method achieves a mean Average Precision (mAP) of 82.66% and average localization latency of 29.09 ms on the DSV Diving dataset, matching state-of-the-art supervised performance while maintaining computational efficiency. Furthermore, it generalizes robustly to unseen, in-the-wild diving footage without retraining, demonstrating its practical applicability for lightweight, real-time action analysis systems in embedded or dynamic environments.", "date": "2025-08-27", "categories": []}
{"id": "2503.08933", "title": "PromptGAR: Flexible Promptive Group Activity Recognition", "abstract": "We present PromptGAR, a novel framework for Group Activity Recognition (GAR) that offering both input flexibility and high recognition accuracy. The existing approaches suffer from limited real-world applicability due to their reliance on full prompt annotations, fixed number of frames and instances, and the lack of actor consistency. To bridge the gap, we proposed PromptGAR, which is the first GAR model to provide input flexibility across prompts, frames, and instances without the need for retraining. We leverage diverse visual prompts, like bounding boxes, skeletal keypoints, and instance identities, by unifying them as point prompts. A recognition decoder then cross-updates class and prompt tokens for enhanced performance. To ensure actor consistency for extended activity durations, we also introduce a relative instance attention mechanism that directly encodes instance identities. Comprehensive evaluations demonstrate that PromptGAR achieves competitive performances both on full prompts and partial prompt inputs, establishing its effectiveness on input flexibility and generalization ability for real-world applications.", "date": "2025-03-11", "categories": []}
{"id": "2507.02826", "title": "Confidence-driven Gradient Modulation for Multimodal Human Activity Recognition: A Dynamic Contrastive Dual-Path Learning Approach", "abstract": "Sensor-based Human Activity Recognition (HAR) is a core technology that enables intelligent systems to perceive and interact with their environment. However, multimodal HAR systems still encounter key challenges, such as difficulties in cross-modal feature alignment and imbalanced modality contributions. To address these issues, we propose a novel framework called the Dynamic Contrastive Dual-Path Network (DCDP-HAR). The framework comprises three key components. First, a dual-path feature extraction architecture is employed, where ResNet and DenseNet branches collaboratively process multimodal sensor data. Second, a multi-stage contrastive learning mechanism is introduced to achieve progressive alignment from local perception to semantic abstraction. Third, we present a confidence-driven gradient modulation strategy that dynamically monitors and adjusts the learning intensity of each modality branch during backpropagation, effectively alleviating modality competition. In addition, a momentum-based gradient accumulation strategy is adopted to enhance training stability. We conduct ablation studies to validate the effectiveness of each component and perform extensive comparative experiments on four public benchmark datasets.", "date": "2025-07-03", "categories": []}
{"id": "2507.09512", "title": "Online Micro-gesture Recognition Using Data Augmentation and Spatial-Temporal Attention", "abstract": "In this paper, we introduce the latest solution developed by our team, HFUT-VUT, for the Micro-gesture Online Recognition track of the IJCAI 2025 MiGA Challenge. The Micro-gesture Online Recognition task is a highly challenging problem that aims to locate the temporal positions and recognize the categories of multiple micro-gesture instances in untrimmed videos. Compared to traditional temporal action detection, this task places greater emphasis on distinguishing between micro-gesture categories and precisely identifying the start and end times of each instance. Moreover, micro-gestures are typically spontaneous human actions, with greater differences than those found in other human actions. To address these challenges, we propose hand-crafted data augmentation and spatial-temporal attention to enhance the model's ability to classify and localize micro-gestures more accurately. Our solution achieved an F1 score of 38.03, outperforming the previous state-of-the-art by 37.9%. As a result, our method ranked first in the Micro-gesture Online Recognition track.", "date": "2025-07-13", "categories": []}
{"id": "2209.11750", "title": "Transformer-based Models to Deal with Heterogeneous Environments in Human Activity Recognition", "abstract": "Human Activity Recognition (HAR) on mobile devices has been demonstrated to be possible using neural models trained on data collected from the device's inertial measurement units. These models have used Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTMs), Transformers or a combination of these to achieve state-of-the-art results with real-time performance. However, these approaches have not been extensively evaluated in real-world situations where the input data may be different from the training data. This paper highlights the issue of data heterogeneity in machine learning applications and how it can hinder their deployment in pervasive settings. To address this problem, we propose and publicly release the code of two sensor-wise Transformer architectures called HART and MobileHART for Human Activity Recognition Transformer. Our experiments on several publicly available datasets show that these HART architectures outperform previous architectures with fewer floating point operations and parameters than conventional Transformers. The results also show they are more robust to changes in mobile position or device brand and hence better suited for the heterogeneous environments encountered in real-life settings. Finally, the source code has been made publicly available.", "date": "2022-09-22", "categories": []}
{"id": "2410.10624", "title": "SensorLLM: Aligning Large Language Models with Motion Sensors for Human Activity Recognition", "abstract": "We introduce SensorLLM, a two-stage framework that enables Large Language Models (LLMs) to perform human activity recognition (HAR) from sensor time-series data. Despite their strong reasoning and generalization capabilities, LLMs remain underutilized for motion sensor data due to the lack of semantic context in time-series, computational constraints, and challenges in processing numerical inputs. SensorLLM addresses these limitations through a Sensor-Language Alignment stage, where the model aligns sensor inputs with trend descriptions. Special tokens are introduced to mark channel boundaries. This alignment enables LLMs to capture numerical variations, channel-specific features, and data of varying durations, without requiring human annotations. In the subsequent Task-Aware Tuning stage, we refine the model for HAR classification, achieving performance that matches or surpasses state-of-the-art methods. Our results demonstrate that SensorLLM evolves into an effective sensor learner, reasoner, and classifier through human-intuitive Sensor-Language Alignment, generalizing across diverse HAR datasets. We believe this work establishes a foundation for future research on time-series and text alignment, paving the way for foundation models in sensor data analysis. Our codes are available at https://github.com/zechenli03/SensorLLM.", "date": "2024-10-14", "categories": []}
{"id": "2506.01608", "title": "EPFL-Smart-Kitchen-30: Densely annotated cooking dataset with 3D kinematics to challenge video and language models", "abstract": "Understanding behavior requires datasets that capture humans while carrying out complex tasks. The kitchen is an excellent environment for assessing human motor and cognitive function, as many complex actions are naturally exhibited in kitchens from chopping to cleaning. Here, we introduce the EPFL-Smart-Kitchen-30 dataset, collected in a noninvasive motion capture platform inside a kitchen environment. Nine static RGB-D cameras, inertial measurement units (IMUs) and one head-mounted HoloLens~2 headset were used to capture 3D hand, body, and eye movements. The EPFL-Smart-Kitchen-30 dataset is a multi-view action dataset with synchronized exocentric, egocentric, depth, IMUs, eye gaze, body and hand kinematics spanning 29.7 hours of 16 subjects cooking four different recipes. Action sequences were densely annotated with 33.78 action segments per minute. Leveraging this multi-modal dataset, we propose four benchmarks to advance behavior understanding and modeling through 1) a vision-language benchmark, 2) a semantic text-to-motion generation benchmark, 3) a multi-modal action recognition benchmark, 4) a pose-based action segmentation benchmark. We expect the EPFL-Smart-Kitchen-30 dataset to pave the way for better methods as well as insights to understand the nature of ecologically-valid human behavior. Code and data are available at https://github.com/amathislab/EPFL-Smart-Kitchen", "date": "2025-06-02", "categories": []}
{"id": "2507.16287", "title": "Beyond Label Semantics: Language-Guided Action Anatomy for Few-shot Action Recognition", "abstract": "Few-shot action recognition (FSAR) aims to classify human actions in videos with only a small number of labeled samples per category. The scarcity of training data has driven recent efforts to incorporate additional modalities, particularly text. However, the subtle variations in human posture, motion dynamics, and the object interactions that occur during different phases, are critical inherent knowledge of actions that cannot be fully exploited by action labels alone. In this work, we propose Language-Guided Action Anatomy (LGA), a novel framework that goes beyond label semantics by leveraging Large Language Models (LLMs) to dissect the essential representational characteristics hidden beneath action labels. Guided by the prior knowledge encoded in LLM, LGA effectively captures rich spatiotemporal cues in few-shot scenarios. Specifically, for text, we prompt an off-the-shelf LLM to anatomize labels into sequences of atomic action descriptions, focusing on the three core elements of action (subject, motion, object). For videos, a Visual Anatomy Module segments actions into atomic video phases to capture the sequential structure of actions. A fine-grained fusion strategy then integrates textual and visual features at the atomic level, resulting in more generalizable prototypes. Finally, we introduce a Multimodal Matching mechanism, comprising both video-video and video-text matching, to ensure robust few-shot classification. Experimental results demonstrate that LGA achieves state-of-the-art performance across multipe FSAR benchmarks.", "date": "2025-07-22", "categories": []}
{"id": "2508.17442", "title": "Multi-Level LVLM Guidance for Untrimmed Video Action Recognition", "abstract": "Action recognition and localization in complex, untrimmed videos remain a formidable challenge in computer vision, largely due to the limitations of existing methods in capturing fine-grained actions, long-term temporal dependencies, and high-level semantic information from low-level visual features. This paper introduces the Event-Contextualized Video Transformer (ECVT), a novel architecture that leverages the advanced semantic understanding capabilities of Large Vision-Language Models (LVLMs) to bridge this gap. ECVT employs a dual-branch design, comprising a Video Encoding Branch for spatio-temporal feature extraction and a Cross-Modal Guidance Branch. The latter utilizes an LVLM to generate multi-granularity semantic descriptions, including Global Event Prompting for macro-level narrative and Temporal Sub-event Prompting for fine-grained action details. These multi-level textual cues are integrated into the video encoder's learning process through sophisticated mechanisms such as adaptive gating for high-level semantic fusion, cross-modal attention for fine-grained feature refinement, and an event graph module for temporal context calibration. Trained end-to-end with a comprehensive loss function incorporating semantic consistency and temporal calibration terms, ECVT significantly enhances the model's ability to understand video temporal structures and event logic. Extensive experiments on ActivityNet v1.3 and THUMOS14 datasets demonstrate that ECVT achieves state-of-the-art performance, with an average mAP of 40.5% on ActivityNet v1.3 and mAP@0.5 of 67.1% on THUMOS14, outperforming leading baselines.", "date": "2025-08-24", "categories": []}
{"id": "2409.14336", "title": "Zero-Shot Skeleton-based Action Recognition with Dual Visual-Text Alignment", "abstract": "Zero-shot action recognition, which addresses the issue of scalability and generalization in action recognition and allows the models to adapt to new and unseen actions dynamically, is an important research topic in computer vision communities. The key to zero-shot action recognition lies in aligning visual features with semantic vectors representing action categories. Most existing methods either directly project visual features onto the semantic space of text category or learn a shared embedding space between the two modalities. However, a direct projection cannot accurately align the two modalities, and learning robust and discriminative embedding space between visual and text representations is often difficult. To address these issues, we introduce Dual Visual-Text Alignment (DVTA) for skeleton-based zero-shot action recognition. The DVTA consists of two alignment modules--Direct Alignment (DA) and Augmented Alignment (AA)--along with a designed Semantic Description Enhancement (SDE). The DA module maps the skeleton features to the semantic space through a specially designed visual projector, followed by the SDE, which is based on cross-attention to enhance the connection between skeleton and text, thereby reducing the gap between modalities. The AA module further strengthens the learning of the embedding space by utilizing deep metric learning to learn the similarity between skeleton and text. Our approach achieves state-of-the-art performances on several popular zero-shot skeleton-based action recognition benchmarks. The code is available at: https://github.com/jidongkuang/DVTA.", "date": "2024-09-22", "categories": []}
{"id": "2508.15903", "title": "VT-LVLM-AR: A Video-Temporal Large Vision-Language Model Adapter for Fine-Grained Action Recognition in Long-Term Videos", "abstract": "Human action recognition in long-term videos, characterized by complex backgrounds and subtle action differences, poses significant challenges for traditional deep learning models due to computational overhead, difficulty in capturing long-range temporal dependencies, and limited semantic understanding. While Large Language Models (LLMs) and Large Vision-Language Models (LVLMs) have shown remarkable capabilities in multi-modal understanding and reasoning, their direct application to continuous video streams for fine-grained action recognition remains an open problem. This paper introduces VT-LVLM-AR (Video-Temporal Large Vision-Language Model Adapter for Action Recognition), a novel framework designed to bridge this gap. VT-LVLM-AR comprises a Video-to-Event Mapper (VTEM) that efficiently transforms raw video into compact, semantically rich, and temporally coherent \"visual event sequences\" through lightweight spatio-temporal feature extraction, adaptive temporal pooling, and conceptual quantization with an event coherence bias. These visual event sequences are then fed into an LVLM-based Action Reasoning module, specifically a frozen LLaVA-1.5 model, adapted using parameter-efficient Prompt Tuning (P-Tuning v2) for action classification. Comprehensive evaluations on the NTU RGB+D and NTU RGB+D 120 datasets demonstrate that VT-LVLM-AR consistently achieves state-of-the-art performance, surpassing existing methods (e.g., 94.1% accuracy on NTU RGB+D X-Sub). Ablation studies confirm the critical contributions of VTEM's components and the efficacy of Prompt Tuning, while human evaluations underscore the interpretability of our visual event representations. This work highlights the immense potential of leveraging LVLMs for robust and interpretable video action understanding through effective video-to-language translation and efficient model adaptation.", "date": "2025-08-21", "categories": []}
{"id": "2508.16291", "title": "Learning Long-Range Action Representation by Two-Stream Mamba Pyramid Network for Figure Skating Assessment", "abstract": "Technical Element Score (TES) and Program Component Score (PCS) evaluations in figure skating demand precise assessment of athletic actions and artistic interpretation, respectively. Existing methods face three major challenges. Firstly, video and audio cues are regarded as common features for both TES and PCS predictions in previous works without considering the prior evaluation criterion of figure skating. Secondly, action elements in competitions are separated in time, TES should be derived from each element's score, but existing methods try to give an overall TES prediction without evaluating each action element. Thirdly, lengthy competition videos make it difficult and inefficient to handle long-range contexts. To address these challenges, we propose a two-stream Mamba pyramid network that aligns with actual judging criteria to predict TES and PCS by separating visual-feature based TES evaluation stream from audio-visual-feature based PCS evaluation stream. In the PCS evaluation stream, we introduce a multi-level fusion mechanism to guarantee that video-based features remain unaffected when assessing TES, and enhance PCS estimation by fusing visual and auditory cues across each contextual level of the pyramid. In the TES evaluation stream, the multi-scale Mamba pyramid and TES head we proposed effectively address the challenges of localizing and evaluating action elements with various temporal scales and give score predictions. With Mamba's superior ability to capture long-range dependencies and its linear computational complexity, our method is ideal for handling lengthy figure skating videos. Comprehensive experimentation demonstrates that our framework attains state-of-the-art performance on the FineFS benchmark. Our source code is available at https://github.com/ycwfs/Figure-Skating-Action-Quality-Assessment.", "date": "2025-08-22", "categories": []}
{"id": "2407.05311", "title": "MMAD: Multi-label Micro-Action Detection in Videos", "abstract": "Human body actions are an important form of non-verbal communication in social interactions. This paper specifically focuses on a subset of body actions known as micro-actions, which are subtle, low-intensity body movements with promising applications in human emotion analysis. In real-world scenarios, human micro-actions often temporally co-occur, with multiple micro-actions overlapping in time, such as concurrent head and hand movements. However, current research primarily focuses on recognizing individual micro-actions while overlooking their co-occurring nature. To address this gap, we propose a new task named Multi-label Micro-Action Detection (MMAD), which involves identifying all micro-actions in a given short video, determining their start and end times, and categorizing them. Accomplishing this requires a model capable of accurately capturing both long-term and short-term action relationships to detect multiple overlapping micro-actions. To facilitate the MMAD task, we introduce a new dataset named Multi-label Micro-Action-52 (MMA-52) and propose a baseline method equipped with a dual-path spatial-temporal adapter to address the challenges of subtle visual change in MMAD. We hope that MMA-52 can stimulate research on micro-action analysis in videos and prompt the development of spatio-temporal modeling in human-centric video understanding. The proposed MMA-52 dataset is available at: https://github.com/VUT-HFUT/Micro-Action.", "date": "2024-07-07", "categories": []}
{"id": "2508.14604", "title": "UST-SSM: Unified Spatio-Temporal State Space Models for Point Cloud Video Modeling", "abstract": "Point cloud videos capture dynamic 3D motion while reducing the effects of lighting and viewpoint variations, making them highly effective for recognizing subtle and continuous human actions. Although Selective State Space Models (SSMs) have shown good performance in sequence modeling with linear complexity, the spatio-temporal disorder of point cloud videos hinders their unidirectional modeling when directly unfolding the point cloud video into a 1D sequence through temporally sequential scanning. To address this challenge, we propose the Unified Spatio-Temporal State Space Model (UST-SSM), which extends the latest advancements in SSMs to point cloud videos. Specifically, we introduce Spatial-Temporal Selection Scanning (STSS), which reorganizes unordered points into semantic-aware sequences through prompt-guided clustering, thereby enabling the effective utilization of points that are spatially and temporally distant yet similar within the sequence. For missing 4D geometric and motion details, Spatio-Temporal Structure Aggregation (STSA) aggregates spatio-temporal features and compensates. To improve temporal interaction within the sampled sequence, Temporal Interaction Sampling (TIS) enhances fine-grained temporal dependencies through non-anchor frame utilization and expanded receptive fields. Experimental results on the MSR-Action3D, NTU RGB+D, and Synthia 4D datasets validate the effectiveness of our method. Our code is available at https://github.com/wangzy01/UST-SSM.", "date": "2025-08-20", "categories": []}
{"id": "2504.20530", "title": "Beyond the Horizon: Decoupling Multi-View UAV Action Recognition via Partial Order Transfer", "abstract": "Action recognition in unmanned aerial vehicles (UAVs) poses unique challenges due to significant view variations along the vertical spatial axis. Unlike traditional ground-based settings, UAVs capture actions at a wide range of altitudes, resulting in considerable appearance discrepancies. We introduce a multi-view formulation tailored to varying UAV altitudes and empirically observe a partial order among views, where recognition accuracy consistently decreases as altitude increases. This observation motivates a novel approach that explicitly models the hierarchical structure of UAV views to improve recognition performance across altitudes. To this end, we propose the Partial Order Guided Multi-View Network (POG-MVNet), designed to address drastic view variations by effectively leveraging view-dependent information across different altitude levels. The framework comprises three key components: a View Partition (VP) module, which uses the head-to-body ratio to group views by altitude; an Order-aware Feature Decoupling (OFD) module, which disentangles action-relevant and view-specific features under partial order guidance; and an Action Partial Order Guide (APOG), which uses the partial order to transfer informative knowledge from easier views to more challenging ones. We conduct experiments on Drone-Action, MOD20, and UAV, demonstrating that POG-MVNet significantly outperforms competing methods. For example, POG-MVNet achieves a 4.7% improvement on Drone-Action and a 3.5% improvement on UAV compared to state-of-the-art methods ASAT and FAR. Code will be released soon.", "date": "2025-04-29", "categories": []}
{"id": "2508.13507", "title": "Bridging the Gap: Doubles Badminton Analysis with Singles-Trained Models", "abstract": "Badminton is known as one of the fastest racket sports in the world. Despite doubles matches being more prevalent in international tournaments than singles, previous research has mainly focused on singles due to the challenges in data availability and multi-person tracking. To address this gap, we designed an approach that transfers singles-trained models to doubles analysis. We extracted keypoints from the ShuttleSet single matches dataset using ViT-Pose and embedded them through a contrastive learning framework based on ST-GCN. To improve tracking stability, we incorporated a custom multi-object tracking algorithm that resolves ID switching issues from fast and overlapping player movements. A Transformer-based classifier then determines shot occurrences based on the learned embeddings. Our findings demonstrate the feasibility of extending pose-based shot recognition to doubles badminton, broadening analytics capabilities. This work establishes a foundation for doubles-specific datasets to enhance understanding of this predominant yet understudied format of the fast racket sport.", "date": "2025-08-19", "categories": []}
{"id": "2508.13565", "title": "Generative Model-Based Feature Attention Module for Video Action Analysis", "abstract": "Video action analysis is a foundational technology within the realm of intelligent video comprehension, particularly concerning its application in Internet of Things(IoT). However, existing methodologies overlook feature semantics in feature extraction and focus on optimizing action proposals, thus these solutions are unsuitable for widespread adoption in high-performance IoT applications due to the limitations in precision, such as autonomous driving, which necessitate robust and scalable intelligent video analytics analysis. To address this issue, we propose a novel generative attention-based model to learn the relation of feature semantics. Specifically, by leveraging the differences of actions' foreground and background, our model simultaneously learns the frame- and segment-dependencies of temporal action feature semantics, which takes advantage of feature semantics in the feature extraction effectively. To evaluate the effectiveness of our model, we conduct extensive experiments on two benchmark video task, action recognition and action detection. In the context of action detection tasks, we substantiate the superiority of our approach through comprehensive validation on widely recognized datasets. Moreover, we extend the validation of the effectiveness of our proposed method to a broader task, video action recognition. Our code is available at https://github.com/Generative-Feature-Model/GAF.", "date": "2025-08-19", "categories": []}
{"id": "2508.13983", "title": "OmViD: Omni-supervised active learning for video action detection", "abstract": "Video action detection requires dense spatio-temporal annotations, which are both challenging and expensive to obtain. However, real-world videos often vary in difficulty and may not require the same level of annotation. This paper analyzes the appropriate annotation types for each sample and their impact on spatio-temporal video action detection. It focuses on two key aspects: 1) how to obtain varying levels of annotation for videos, and 2) how to learn action detection from different annotation types. The study explores video-level tags, points, scribbles, bounding boxes, and pixel-level masks. First, a simple active learning strategy is proposed to estimate the necessary annotation type for each video. Then, a novel spatio-temporal 3D-superpixel approach is introduced to generate pseudo-labels from these annotations, enabling effective training. The approach is validated on UCF101-24 and JHMDB-21 datasets, significantly cutting annotation costs with minimal performance loss.", "date": "2025-08-19", "categories": []}
{"id": "2505.10312", "title": "SOS: A Shuffle Order Strategy for Data Augmentation in Industrial Human Activity Recognition", "abstract": "In the realm of Human Activity Recognition (HAR), obtaining high quality and variance data is still a persistent challenge due to high costs and the inherent variability of real-world activities. This study introduces a generation dataset by deep learning approaches (Attention Autoencoder and conditional Generative Adversarial Networks). Another problem that data heterogeneity is a critical challenge, one of the solutions is to shuffle the data to homogenize the distribution. Experimental results demonstrate that the random sequence strategy significantly improves classification performance, achieving an accuracy of up to 0.70 $\\pm$ 0.03 and a macro F1 score of 0.64 $\\pm$ 0.01. For that, disrupting temporal dependencies through random sequence reordering compels the model to focus on instantaneous recognition, thereby improving robustness against activity transitions. This approach not only broadens the effective training dataset but also offers promising avenues for enhancing HAR systems in complex, real-world scenarios.", "date": "2025-05-15", "categories": []}
{"id": "2508.12586", "title": "Foundation Model for Skeleton-Based Human Action Understanding", "abstract": "Human action understanding serves as a foundational pillar in the field of intelligent motion perception. Skeletons serve as a modality- and device-agnostic representation for human modeling, and skeleton-based action understanding has potential applications in humanoid robot control and interaction. \\RED{However, existing works often lack the scalability and generalization required to handle diverse action understanding tasks. There is no skeleton foundation model that can be adapted to a wide range of action understanding tasks}. This paper presents a Unified Skeleton-based Dense Representation Learning (USDRL) framework, which serves as a foundational model for skeleton-based human action understanding. USDRL consists of a Transformer-based Dense Spatio-Temporal Encoder (DSTE), Multi-Grained Feature Decorrelation (MG-FD), and Multi-Perspective Consistency Training (MPCT). The DSTE module adopts two parallel streams to learn temporal dynamic and spatial structure features. The MG-FD module collaboratively performs feature decorrelation across temporal, spatial, and instance domains to reduce dimensional redundancy and enhance information extraction. The MPCT module employs both multi-view and multi-modal self-supervised consistency training. The former enhances the learning of high-level semantics and mitigates the impact of low-level discrepancies, while the latter effectively facilitates the learning of informative multimodal features. We perform extensive experiments on 25 benchmarks across across 9 skeleton-based action understanding tasks, covering coarse prediction, dense prediction, and transferred prediction. Our approach significantly outperforms the current state-of-the-art methods. We hope that this work would broaden the scope of research in skeleton-based action understanding and encourage more attention to dense prediction tasks.", "date": "2025-08-18", "categories": []}
{"id": "2508.12948", "title": "MaskSem: Semantic-Guided Masking for Learning 3D Hybrid High-Order Motion Representation", "abstract": "Human action recognition is a crucial task for intelligent robotics, particularly within the context of human-robot collaboration research. In self-supervised skeleton-based action recognition, the mask-based reconstruction paradigm learns the spatial structure and motion patterns of the skeleton by masking joints and reconstructing the target from unlabeled data. However, existing methods focus on a limited set of joints and low-order motion patterns, limiting the model's ability to understand complex motion patterns. To address this issue, we introduce MaskSem, a novel semantic-guided masking method for learning 3D hybrid high-order motion representations. This novel framework leverages Grad-CAM based on relative motion to guide the masking of joints, which can be represented as the most semantically rich temporal orgions. The semantic-guided masking process can encourage the model to explore more discriminative features. Furthermore, we propose using hybrid high-order motion as the reconstruction target, enabling the model to learn multi-order motion patterns. Specifically, low-order motion velocity and high-order motion acceleration are used together as the reconstruction target. This approach offers a more comprehensive description of the dynamic motion process, enhancing the model's understanding of motion patterns. Experiments on the NTU60, NTU120, and PKU-MMD datasets show that MaskSem, combined with a vanilla transformer, improves skeleton-based action recognition, making it more suitable for applications in human-robot interaction.", "date": "2025-08-18", "categories": []}
{"id": "2504.03096", "title": "Scaling Open-Vocabulary Action Detection", "abstract": "In this work, we focus on scaling open-vocabulary action detection. Existing approaches for action detection are predominantly limited to closed-set scenarios and rely on complex, parameter-heavy architectures. Extending these models to the open-vocabulary setting poses two key challenges: (1) the lack of large-scale datasets with many action classes for robust training, and (2) parameter-heavy adaptations to a pretrained vision-language contrastive model to convert it for detection, risking overfitting the additional non-pretrained parameters to base action classes. Firstly, we introduce an encoder-only multimodal model for video action detection, reducing the reliance on parameter-heavy additions for video action detection. Secondly, we introduce a simple weakly supervised training strategy to exploit an existing closed-set action detection dataset for pretraining. Finally, we depart from the ill-posed base-to-novel benchmark used by prior works in open-vocabulary action detection and devise a new benchmark to evaluate on existing closed-set action detection datasets without ever using them for training, showing novel results to serve as baselines for future work. Our code is available at https://siatheindochinese.github.io/sia_act_page/ .", "date": "2025-04-04", "categories": []}
{"id": "2508.10281", "title": "VIFSS: View-Invariant and Figure Skating-Specific Pose Representation Learning for Temporal Action Segmentation", "abstract": "Understanding human actions from videos plays a critical role across various domains, including sports analytics. In figure skating, accurately recognizing the type and timing of jumps a skater performs is essential for objective performance evaluation. However, this task typically requires expert-level knowledge due to the fine-grained and complex nature of jump procedures. While recent approaches have attempted to automate this task using Temporal Action Segmentation (TAS), there are two major limitations to TAS for figure skating: the annotated data is insufficient, and existing methods do not account for the inherent three-dimensional aspects and procedural structure of jump actions. In this work, we propose a new TAS framework for figure skating jumps that explicitly incorporates both the three-dimensional nature and the semantic procedure of jump movements. First, we propose a novel View-Invariant, Figure Skating-Specific pose representation learning approach (VIFSS) that combines contrastive learning as pre-training and action classification as fine-tuning. For view-invariant contrastive pre-training, we construct FS-Jump3D, the first publicly available 3D pose dataset specialized for figure skating jumps. Second, we introduce a fine-grained annotation scheme that marks the ``entry (preparation)'' and ``landing'' phases, enabling TAS models to learn the procedural structure of jumps. Extensive experiments demonstrate the effectiveness of our framework. Our method achieves over 92% F1@50 on element-level TAS, which requires recognizing both jump types and rotation levels. Furthermore, we show that view-invariant contrastive pre-training is particularly effective when fine-tuning data is limited, highlighting the practicality of our approach in real-world scenarios.", "date": "2025-08-14", "categories": []}
{"id": "2508.10469", "title": "Enhanced Sparse Point Cloud Data Processing for Privacy-aware Human Action Recognition", "abstract": "Human Action Recognition (HAR) plays a crucial role in healthcare, fitness tracking, and ambient assisted living technologies. While traditional vision based HAR systems are effective, they pose privacy concerns. mmWave radar sensors offer a privacy preserving alternative but present challenges due to the sparse and noisy nature of their point cloud data. In the literature, three primary data processing methods: Density-Based Spatial Clustering of Applications with Noise (DBSCAN), the Hungarian Algorithm, and Kalman Filtering have been widely used to improve the quality and continuity of radar data. However, a comprehensive evaluation of these methods, both individually and in combination, remains lacking. This paper addresses that gap by conducting a detailed performance analysis of the three methods using the MiliPoint dataset. We evaluate each method individually, all possible pairwise combinations, and the combination of all three, assessing both recognition accuracy and computational cost. Furthermore, we propose targeted enhancements to the individual methods aimed at improving accuracy. Our results provide crucial insights into the strengths and trade-offs of each method and their integrations, guiding future work on mmWave based HAR systems", "date": "2025-08-14", "categories": []}
{"id": "2507.03936", "title": "Learning Adaptive Node Selection with External Attention for Human Interaction Recognition", "abstract": "Most GCN-based methods model interacting individuals as independent graphs, neglecting their inherent inter-dependencies. Although recent approaches utilize predefined interaction adjacency matrices to integrate participants, these matrices fail to adaptively capture the dynamic and context-specific joint interactions across different actions. In this paper, we propose the Active Node Selection with External Attention Network (ASEA), an innovative approach that dynamically captures interaction relationships without predefined assumptions. Our method models each participant individually using a GCN to capture intra-personal relationships, facilitating a detailed representation of their actions. To identify the most relevant nodes for interaction modeling, we introduce the Adaptive Temporal Node Amplitude Calculation (AT-NAC) module, which estimates global node activity by combining spatial motion magnitude with adaptive temporal weighting, thereby highlighting salient motion patterns while reducing irrelevant or redundant information. A learnable threshold, regularized to prevent extreme variations, is defined to selectively identify the most informative nodes for interaction modeling. To capture interactions, we design the External Attention (EA) module to operate on active nodes, effectively modeling the interaction dynamics and semantic relationships between individuals. Extensive evaluations show that our method captures interaction relationships more effectively and flexibly, achieving state-of-the-art performance.", "date": "2025-07-05", "categories": []}
{"id": "2508.08944", "title": "UniSTFormer: Unified Spatio-Temporal Lightweight Transformer for Efficient Skeleton-Based Action Recognition", "abstract": "Skeleton-based action recognition (SAR) has achieved impressive progress with transformer architectures. However, existing methods often rely on complex module compositions and heavy designs, leading to increased parameter counts, high computational costs, and limited scalability. In this paper, we propose a unified spatio-temporal lightweight transformer framework that integrates spatial and temporal modeling within a single attention module, eliminating the need for separate temporal modeling blocks. This approach reduces redundant computations while preserving temporal awareness within the spatial modeling process. Furthermore, we introduce a simplified multi-scale pooling fusion module that combines local and global pooling pathways to enhance the model's ability to capture fine-grained local movements and overarching global motion patterns. Extensive experiments on benchmark datasets demonstrate that our lightweight model achieves a superior balance between accuracy and efficiency, reducing parameter complexity by over 58% and lowering computational cost by over 60% compared to state-of-the-art transformer-based baselines, while maintaining competitive recognition performance.", "date": "2025-08-12", "categories": []}
{"id": "2504.11966", "title": "Exploring Video-Based Driver Activity Recognition under Noisy Labels", "abstract": "As an open research topic in the field of deep learning, learning with noisy labels has attracted much attention and grown rapidly over the past ten years. Learning with label noise is crucial for driver distraction behavior recognition, as real-world video data often contains mislabeled samples, impacting model reliability and performance. However, label noise learning is barely explored in the driver activity recognition field. In this paper, we propose the first label noise learning approach for the driver activity recognition task. Based on the cluster assumption, we initially enable the model to learn clustering-friendly low-dimensional representations from given videos and assign the resultant embeddings into clusters. We subsequently perform co-refinement within each cluster to smooth the classifier outputs. Furthermore, we propose a flexible sample selection strategy that combines two selection criteria without relying on any hyperparameters to filter clean samples from the training dataset. We also incorporate a self-adaptive parameter into the sample selection process to enforce balancing across classes. A comprehensive variety of experiments on the public Drive&Act dataset for all granularity levels demonstrates the superior performance of our method in comparison with other label-denoising methods derived from the image classification field. The source code is available at https://github.com/ilonafan/DAR-noisy-labels.", "date": "2025-04-16", "categories": []}
{"id": "2508.02320", "title": "Zero-shot Compositional Action Recognition with Neural Logic Constraints", "abstract": "Zero-shot compositional action recognition (ZS-CAR) aims to identify unseen verb-object compositions in the videos by exploiting the learned knowledge of verb and object primitives during training. Despite compositional learning's progress in ZS-CAR, two critical challenges persist: 1) Missing compositional structure constraint, leading to spurious correlations between primitives; 2) Neglecting semantic hierarchy constraint, leading to semantic ambiguity and impairing the training process. In this paper, we argue that human-like symbolic reasoning offers a principled solution to these challenges by explicitly modeling compositional and hierarchical structured abstraction. To this end, we propose a logic-driven ZS-CAR framework LogicCAR that integrates dual symbolic constraints: Explicit Compositional Logic and Hierarchical Primitive Logic. Specifically, the former models the restrictions within the compositions, enhancing the compositional reasoning ability of our model. The latter investigates the semantical dependencies among different primitives, empowering the models with fine-to-coarse reasoning capacity. By formalizing these constraints in first-order logic and embedding them into neural network architectures, LogicCAR systematically bridges the gap between symbolic abstraction and existing models. Extensive experiments on the Sth-com dataset demonstrate that our LogicCAR outperforms existing baseline methods, proving the effectiveness of our logic-driven constraints.", "date": "2025-08-04", "categories": []}
{"id": "2508.06528", "title": "A Framework Combining 3D CNN and Transformer for Video-Based Behavior Recognition", "abstract": "Video-based behavior recognition is essential in fields such as public safety, intelligent surveillance, and human-computer interaction. Traditional 3D Convolutional Neural Network (3D CNN) effectively capture local spatiotemporal features but struggle with modeling long-range dependencies. Conversely, Transformers excel at learning global contextual information but face challenges with high computational costs. To address these limitations, we propose a hybrid framework combining 3D CNN and Transformer architectures. The 3D CNN module extracts low-level spatiotemporal features, while the Transformer module captures long-range temporal dependencies, with a fusion mechanism integrating both representations. Evaluated on benchmark datasets, the proposed model outperforms traditional 3D CNN and standalone Transformers, achieving higher recognition accuracy with manageable complexity. Ablation studies further validate the complementary strengths of the two modules. This hybrid framework offers an effective and scalable solution for video-based behavior recognition.", "date": "2025-08-02", "categories": []}
{"id": "2506.21586", "title": "Can Vision Language Models Understand Mimed Actions?", "abstract": "Nonverbal communication (NVC) plays an integral role in human language, but studying NVC in general is challenging because of its broad scope and high variance in interpretation among individuals and cultures. However, mime -- the theatrical technique of suggesting intent using only gesture, expression, and movement -- is a subset of NVC that consists of explicit and embodied actions with much lower human interpretation variance. We argue that a solid understanding of mimed actions is a crucial prerequisite for vision-language models capable of interpreting and commanding more subtle aspects of NVC. Hence, we propose Mime Identification Multimodal Evaluation (MIME), a novel video-based question answering benchmark comprising of 86 mimed actions. Constructed with motion capture data, MIME consists of variations of each action with perturbations applied to the character, background, and viewpoint for evaluating recognition robustness. We find that both open-weight and API-based vision-language models perform significantly worse than humans on MIME, motivating the need for increased research for instilling more robust understanding of human gestures.", "date": "2025-06-17", "categories": []}
{"id": "2507.03539", "title": "CLOT: Closed Loop Optimal Transport for Unsupervised Action Segmentation", "abstract": "Unsupervised action segmentation has recently pushed its limits with ASOT, an optimal transport (OT)-based method that simultaneously learns action representations and performs clustering using pseudo-labels. Unlike other OT-based approaches, ASOT makes no assumptions about action ordering and can decode a temporally consistent segmentation from a noisy cost matrix between video frames and action labels. However, the resulting segmentation lacks segment-level supervision, limiting the effectiveness of feedback between frames and action representations. To address this limitation, we propose Closed Loop Optimal Transport (CLOT), a novel OT-based framework with a multi-level cyclic feature learning mechanism. Leveraging its encoder-decoder architecture, CLOT learns pseudo-labels alongside frame and segment embeddings by solving two separate OT problems. It then refines both frame embeddings and pseudo-labels through cross-attention between the learned frame and segment embeddings, by integrating a third OT problem. Experimental results on four benchmark datasets demonstrate the benefits of cyclical learning for unsupervised action segmentation.", "date": "2025-07-04", "categories": []}
{"id": "2508.03266", "title": "EgoPrompt: Prompt Learning for Egocentric Action Recognition", "abstract": "Driven by the increasing demand for applications in augmented and virtual reality, egocentric action recognition has emerged as a prominent research area. It is typically divided into two subtasks: recognizing the performed behavior (i.e., verb component) and identifying the objects being acted upon (i.e., noun component) from the first-person perspective. However, most existing approaches treat these two components as independent classification tasks, focusing on extracting component-specific knowledge while overlooking their inherent semantic and contextual relationships, leading to fragmented representations and sub-optimal generalization capability. To address these challenges, we propose a prompt learning-based framework, EgoPrompt, to conduct the egocentric action recognition task. Building on the existing prompting strategy to capture the component-specific knowledge, we construct a Unified Prompt Pool space to establish interaction between the two types of component representations. Specifically, the component representations (from verbs and nouns) are first decomposed into fine-grained patterns with the prompt pair form. Then, these pattern-level representations are fused through an attention-based mechanism to facilitate cross-component interaction. To ensure the prompt pool is informative, we further introduce a novel training objective, Diverse Pool Criteria. This objective realizes our goals from two perspectives: Prompt Selection Frequency Regularization and Prompt Knowledge Orthogonalization. Extensive experiments are conducted on the Ego4D, EPIC-Kitchens, and EGTEA datasets. The results consistently show that EgoPrompt achieves state-of-the-art performance across within-dataset, cross-dataset, and base-to-novel generalization benchmarks.", "date": "2025-08-05", "categories": []}
{"id": "2508.05529", "title": "Looking into the Unknown: Exploring Action Discovery for Segmentation of Known and Unknown Actions", "abstract": "We introduce Action Discovery, a novel setup within Temporal Action Segmentation that addresses the challenge of defining and annotating ambiguous actions and incomplete annotations in partially labeled datasets. In this setup, only a subset of actions - referred to as known actions - is annotated in the training data, while other unknown actions remain unlabeled. This scenario is particularly relevant in domains like neuroscience, where well-defined behaviors (e.g., walking, eating) coexist with subtle or infrequent actions that are often overlooked, as well as in applications where datasets are inherently partially annotated due to ambiguous or missing labels. To address this problem, we propose a two-step approach that leverages the known annotations to guide both the temporal and semantic granularity of unknown action segments. First, we introduce the Granularity-Guided Segmentation Module (GGSM), which identifies temporal intervals for both known and unknown actions by mimicking the granularity of annotated actions. Second, we propose the Unknown Action Segment Assignment (UASA), which identifies semantically meaningful classes within the unknown actions, based on learned embedding similarities. We systematically explore the proposed setting of Action Discovery on three challenging datasets - Breakfast, 50Salads, and Desktop Assembly - demonstrating that our method considerably improves upon existing baselines.", "date": "2025-08-07", "categories": []}
{"id": "2508.04513", "title": "Skeleton Motion Words for Unsupervised Skeleton-Based Temporal Action Segmentation", "abstract": "Current state-of-the-art methods for skeleton-based temporal action segmentation are predominantly supervised and require annotated data, which is expensive to collect. In contrast, existing unsupervised temporal action segmentation methods have focused primarily on video data, while skeleton sequences remain underexplored, despite their relevance to real-world applications, robustness, and privacy-preserving nature. In this paper, we propose a novel approach for unsupervised skeleton-based temporal action segmentation. Our method utilizes a sequence-to-sequence temporal autoencoder that keeps the information of the different joints disentangled in the embedding space. Latent skeleton sequences are then divided into non-overlapping patches and quantized to obtain distinctive skeleton motion words, driving the discovery of semantically meaningful action clusters. We thoroughly evaluate the proposed approach on three widely used skeleton-based datasets, namely HuGaDB, LARa, and BABEL. The results demonstrate that our model outperforms the current state-of-the-art unsupervised temporal action segmentation methods. Code is available at https://github.com/bachlab/SMQ .", "date": "2025-08-06", "categories": []}
{"id": "2507.08344", "title": "MM-Gesture: Towards Precise Micro-Gesture Recognition through Multimodal Fusion", "abstract": "In this paper, we present MM-Gesture, the solution developed by our team HFUT-VUT, which ranked 1st in the micro-gesture classification track of the 3rd MiGA Challenge at IJCAI 2025, achieving superior performance compared to previous state-of-the-art methods. MM-Gesture is a multimodal fusion framework designed specifically for recognizing subtle and short-duration micro-gestures (MGs), integrating complementary cues from joint, limb, RGB video, Taylor-series video, optical-flow video, and depth video modalities. Utilizing PoseConv3D and Video Swin Transformer architectures with a novel modality-weighted ensemble strategy, our method further enhances RGB modality performance through transfer learning pre-trained on the larger MA-52 dataset. Extensive experiments on the iMiGUE benchmark, including ablation studies across different modalities, validate the effectiveness of our proposed approach, achieving a top-1 accuracy of 73.213%. Code is available at: https://github.com/momiji-bit/MM-Gesture.", "date": "2025-07-11", "categories": []}
{"id": "2508.02981", "title": "MoExDA: Domain Adaptation for Edge-based Action Recognition", "abstract": "Modern action recognition models suffer from static bias, leading to reduced generalization performance. In this paper, we propose MoExDA, a lightweight domain adaptation between RGB and edge information using edge frames in addition to RGB frames to counter the static bias issue. Experiments demonstrate that the proposed method effectively suppresses static bias with a lower computational cost, allowing for more robust action recognition than previous approaches.", "date": "2025-08-05", "categories": []}
{"id": "2508.03695", "title": "Trokens: Semantic-Aware Relational Trajectory Tokens for Few-Shot Action Recognition", "abstract": "Video understanding requires effective modeling of both motion and appearance information, particularly for few-shot action recognition. While recent advances in point tracking have been shown to improve few-shot action recognition, two fundamental challenges persist: selecting informative points to track and effectively modeling their motion patterns. We present Trokens, a novel approach that transforms trajectory points into semantic-aware relational tokens for action recognition. First, we introduce a semantic-aware sampling strategy to adaptively distribute tracking points based on object scale and semantic relevance. Second, we develop a motion modeling framework that captures both intra-trajectory dynamics through the Histogram of Oriented Displacements (HoD) and inter-trajectory relationships to model complex action patterns. Our approach effectively combines these trajectory tokens with semantic features to enhance appearance features with motion information, achieving state-of-the-art performance across six diverse few-shot action recognition benchmarks: Something-Something-V2 (both full and small splits), Kinetics, UCF101, HMDB51, and FineGym. For project page see https://trokens-iccv25.github.io", "date": "2025-08-05", "categories": []}
{"id": "2411.14796", "title": "Adaptive Hyper-Graph Convolution Network for Skeleton-based Human Action Recognition with Virtual Connections", "abstract": "The shared topology of human skeletons motivated the recent investigation of graph convolutional network (GCN) solutions for action recognition. However, most of the existing GCNs rely on the binary connection of two neighboring vertices (joints) formed by an edge (bone), overlooking the potential of constructing multi-vertex convolution structures. Although some studies have attempted to utilize hyper-graphs to represent the topology, they rely on a fixed construction strategy, which limits their adaptivity in uncovering the intricate latent relationships within the action. In this paper, we address this oversight and explore the merits of an adaptive hyper-graph convolutional network (Hyper-GCN) to achieve the aggregation of rich semantic information conveyed by skeleton vertices. In particular, our Hyper-GCN adaptively optimises the hyper-graphs during training, revealing the action-driven multi-vertex relations. Besides, virtual connections are often designed to support efficient feature aggregation, implicitly extending the spectrum of dependencies within the skeleton. By injecting virtual connections into hyper-graphs, the semantic clues of diverse action categories can be highlighted. The results of experiments conducted on the NTU-60, NTU-120, and NW-UCLA datasets demonstrate the merits of our Hyper-GCN, compared to the state-of-the-art methods. The code is available at https://github.com/6UOOON9/Hyper-GCN.", "date": "2024-11-22", "categories": []}
{"id": "2508.01701", "title": "MHARFedLLM: Multimodal Human Activity Recognition Using Federated Large Language Model", "abstract": "Human Activity Recognition (HAR) plays a vital role in applications such as fitness tracking, smart homes, and healthcare monitoring. Traditional HAR systems often rely on single modalities, such as motion sensors or cameras, limiting robustness and accuracy in real-world environments. This work presents FedTime-MAGNET, a novel multimodal federated learning framework that advances HAR by combining heterogeneous data sources: depth cameras, pressure mats, and accelerometers. At its core is the Multimodal Adaptive Graph Neural Expert Transformer (MAGNET), a fusion architecture that uses graph attention and a Mixture of Experts to generate unified, discriminative embeddings across modalities. To capture complex temporal dependencies, a lightweight T5 encoder only architecture is customized and adapted within this framework. Extensive experiments show that FedTime-MAGNET significantly improves HAR performance, achieving a centralized F1 Score of 0.934 and a strong federated F1 Score of 0.881. These results demonstrate the effectiveness of combining multimodal fusion, time series LLMs, and federated learning for building accurate and robust HAR systems.", "date": "2025-08-03", "categories": []}
{"id": "2507.06603", "title": "Cross-Modal Dual-Causal Learning for Long-Term Action Recognition", "abstract": "Long-term action recognition (LTAR) is challenging due to extended temporal spans with complex atomic action correlations and visual confounders. Although vision-language models (VLMs) have shown promise, they often rely on statistical correlations instead of causal mechanisms. Moreover, existing causality-based methods address modal-specific biases but lack cross-modal causal modeling, limiting their utility in VLM-based LTAR. This paper proposes \\textbf{C}ross-\\textbf{M}odal \\textbf{D}ual-\\textbf{C}ausal \\textbf{L}earning (CMDCL), which introduces a structural causal model to uncover causal relationships between videos and label texts.\n  CMDCL addresses cross-modal biases in text embeddings via textual causal intervention and removes confounders inherent in the visual modality through visual causal intervention guided by the debiased text.\n  These dual-causal interventions enable robust action representations to address LTAR challenges. Experimental results on three benchmarks including Charades, Breakfast and COIN, demonstrate the effectiveness of the proposed model. Our code is available at https://github.com/xushaowu/CMDCL.", "date": "2025-07-09", "categories": []}
{"id": "2508.00085", "title": "Punching Bag vs. Punching Person: Motion Transferability in Videos", "abstract": "Action recognition models demonstrate strong generalization, but can they effectively transfer high-level motion concepts across diverse contexts, even within similar distributions? For example, can a model recognize the broad action \"punching\" when presented with an unseen variation such as \"punching person\"? To explore this, we introduce a motion transferability framework with three datasets: (1) Syn-TA, a synthetic dataset with 3D object motions; (2) Kinetics400-TA; and (3) Something-Something-v2-TA, both adapted from natural video datasets. We evaluate 13 state-of-the-art models on these benchmarks and observe a significant drop in performance when recognizing high-level actions in novel contexts. Our analysis reveals: 1) Multimodal models struggle more with fine-grained unknown actions than with coarse ones; 2) The bias-free Syn-TA proves as challenging as real-world datasets, with models showing greater performance drops in controlled settings; 3) Larger models improve transferability when spatial cues dominate but struggle with intensive temporal reasoning, while reliance on object and background cues hinders generalization. We further explore how disentangling coarse and fine motions can improve recognition in temporally challenging datasets. We believe this study establishes a crucial benchmark for assessing motion transferability in action recognition. Datasets and relevant code: https://github.com/raiyaan-abdullah/Motion-Transfer.", "date": "2025-07-31", "categories": []}
{"id": "2507.18675", "title": "Advancing Vision-based Human Action Recognition: Exploring Vision-Language CLIP Model for Generalisation in Domain-Independent Tasks", "abstract": "Human action recognition plays a critical role in healthcare and medicine, supporting applications such as patient behavior monitoring, fall detection, surgical robot supervision, and procedural skill assessment. While traditional models like CNNs and RNNs have achieved moderate success, they often struggle to generalize across diverse and complex actions. Recent advancements in vision-language models, especially the transformer-based CLIP model, offer promising capabilities for generalizing action recognition from video data. In this work, we evaluate CLIP on the UCF-101 dataset and systematically analyze its performance under three masking strategies: (1) percentage-based and shape-based black masking at 10%, 30%, and 50%, (2) feature-specific masking to suppress bias-inducing elements, and (3) isolation masking that retains only class-specific regions. Our results reveal that CLIP exhibits inconsistent behavior and frequent misclassifications, particularly when essential visual cues are obscured. To overcome these limitations, we propose incorporating class-specific noise, learned via a custom loss function, to reinforce attention to class-defining features. This enhancement improves classification accuracy and model confidence while reducing bias. We conclude with a discussion on the challenges of applying such models in clinical domains and outline directions for future work to improve generalizability across domain-independent healthcare scenarios.", "date": "2025-07-24", "categories": []}
{"id": "2507.22421", "title": "Efficient Spatial-Temporal Modeling for Real-Time Video Analysis: A Unified Framework for Action Recognition and Object Tracking", "abstract": "Real-time video analysis remains a challenging problem in computer vision, requiring efficient processing of both spatial and temporal information while maintaining computational efficiency. Existing approaches often struggle to balance accuracy and speed, particularly in resource-constrained environments. In this work, we present a unified framework that leverages advanced spatial-temporal modeling techniques for simultaneous action recognition and object tracking. Our approach builds upon recent advances in parallel sequence modeling and introduces a novel hierarchical attention mechanism that adaptively focuses on relevant spatial regions across temporal sequences. We demonstrate that our method achieves state-of-the-art performance on standard benchmarks while maintaining real-time inference speeds. Extensive experiments on UCF-101, HMDB-51, and MOT17 datasets show improvements of 3.2% in action recognition accuracy and 2.8% in tracking precision compared to existing methods, with 40% faster inference time.", "date": "2025-07-30", "categories": []}
{"id": "2507.22522", "title": "Recognizing Actions from Robotic View for Natural Human-Robot Interaction", "abstract": "Natural Human-Robot Interaction (N-HRI) requires robots to recognize human actions at varying distances and states, regardless of whether the robot itself is in motion or stationary. This setup is more flexible and practical than conventional human action recognition tasks. However, existing benchmarks designed for traditional action recognition fail to address the unique complexities in N-HRI due to limited data, modalities, task categories, and diversity of subjects and environments. To address these challenges, we introduce ACTIVE (Action from Robotic View), a large-scale dataset tailored specifically for perception-centric robotic views prevalent in mobile service robots. ACTIVE comprises 30 composite action categories, 80 participants, and 46,868 annotated video instances, covering both RGB and point cloud modalities. Participants performed various human actions in diverse environments at distances ranging from 3m to 50m, while the camera platform was also mobile, simulating real-world scenarios of robot perception with varying camera heights due to uneven ground. This comprehensive and challenging benchmark aims to advance action and attribute recognition research in N-HRI. Furthermore, we propose ACTIVE-PC, a method that accurately perceives human actions at long distances using Multilevel Neighborhood Sampling, Layered Recognizers, Elastic Ellipse Query, and precise decoupling of kinematic interference from human actions. Experimental results demonstrate the effectiveness of ACTIVE-PC. Our code is available at: https://github.com/wangzy01/ACTIVE-Action-from-Robotic-View.", "date": "2025-07-30", "categories": []}
{"id": "2507.21247", "title": "Dual Guidance Semi-Supervised Action Detection", "abstract": "Semi-Supervised Learning (SSL) has shown tremendous potential to improve the predictive performance of deep learning models when annotations are hard to obtain. However, the application of SSL has so far been mainly studied in the context of image classification. In this work, we present a semi-supervised approach for spatial-temporal action localization. We introduce a dual guidance network to select better pseudo-bounding boxes. It combines a frame-level classification with a bounding-box prediction to enforce action class consistency across frames and boxes. Our evaluation across well-known spatial-temporal action localization datasets, namely UCF101-24 , J-HMDB-21 and AVA shows that the proposed module considerably enhances the model's performance in limited labeled data settings. Our framework achieves superior results compared to extended image-based semi-supervised baselines.", "date": "2025-07-28", "categories": []}
{"id": "2507.21353", "title": "Group Relative Augmentation for Data Efficient Action Detection", "abstract": "Adapting large Video-Language Models (VLMs) for action detection using only a few examples poses challenges like overfitting and the granularity mismatch between scene-level pre-training and required person-centric understanding. We propose an efficient adaptation strategy combining parameter-efficient tuning (LoRA) with a novel learnable internal feature augmentation. Applied within the frozen VLM backbone using FiLM, these augmentations generate diverse feature variations directly relevant to the task. Additionally, we introduce a group-weighted loss function that dynamically modulates the training contribution of each augmented sample based on its prediction divergence relative to the group average. This promotes robust learning by prioritizing informative yet reasonable augmentations. We demonstrate our method's effectiveness on complex multi-label, multi-person action detection datasets (AVA, MOMA), achieving strong mAP performance and showcasing significant data efficiency for adapting VLMs from limited examples.", "date": "2025-07-28", "categories": []}
{"id": "2507.21964", "title": "Thou Shalt Not Prompt: Zero-Shot Human Activity Recognition in Smart Homes via Language Modeling of Sensor Data & Activities", "abstract": "Developing zero-shot human activity recognition (HAR) methods is a critical direction in smart home research -- considering its impact on making HAR systems work across smart homes having diverse sensing modalities, layouts, and activities of interest. The state-of-the-art solutions along this direction are based on generating natural language descriptions of the sensor data and feeding it via a carefully crafted prompt to the LLM to perform classification. Despite their performance guarantees, such ``prompt-the-LLM'' approaches carry several risks, including privacy invasion, reliance on an external service, and inconsistent predictions due to version changes, making a case for alternative zero-shot HAR methods that do not require prompting the LLMs. In this paper, we propose one such solution that models sensor data and activities using natural language, leveraging its embeddings to perform zero-shot classification and thereby bypassing the need to prompt the LLMs for activity predictions. The impact of our work lies in presenting a detailed case study on six datasets, highlighting how language modeling can bolster HAR systems in zero-shot recognition.", "date": "2025-07-29", "categories": []}
{"id": "2403.18915", "title": "PLOT-TAL: Prompt Learning with Optimal Transport for Few-Shot Temporal Action Localization", "abstract": "Few-shot temporal action localization (TAL) methods that adapt large models via single-prompt tuning often fail to produce precise temporal boundaries. This stems from the model learning a non-discriminative mean representation of an action from sparse data, which compromises generalization. We address this by proposing a new paradigm based on multi-prompt ensembles, where a set of diverse, learnable prompts for each action is encouraged to specialize on compositional sub-events. To enforce this specialization, we introduce PLOT-TAL, a framework that leverages Optimal Transport (OT) to find a globally optimal alignment between the prompt ensemble and the video's temporal features. Our method establishes a new state-of-the-art on the challenging few-shot benchmarks of THUMOS'14 and EPIC-Kitchens, without requiring complex meta-learning. The significant performance gains, particularly at high IoU thresholds, validate our hypothesis and demonstrate the superiority of learning distributed, compositional representations for precise temporal localization.", "date": "2024-03-27", "categories": []}
{"id": "2507.00566", "title": "Zero-Shot Skeleton-Based Action Recognition With Prototype-Guided Feature Alignment", "abstract": "Zero-shot skeleton-based action recognition aims to classify unseen skeleton-based human actions without prior exposure to such categories during training. This task is extremely challenging due to the difficulty in generalizing from known to unknown actions. Previous studies typically use two-stage training: pre-training skeleton encoders on seen action categories using cross-entropy loss and then aligning pre-extracted skeleton and text features, enabling knowledge transfer to unseen classes through skeleton-text alignment and language models' generalization. However, their efficacy is hindered by 1) insufficient discrimination for skeleton features, as the fixed skeleton encoder fails to capture necessary alignment information for effective skeleton-text alignment; 2) the neglect of alignment bias between skeleton and unseen text features during testing. To this end, we propose a prototype-guided feature alignment paradigm for zero-shot skeleton-based action recognition, termed PGFA. Specifically, we develop an end-to-end cross-modal contrastive training framework to improve skeleton-text alignment, ensuring sufficient discrimination for skeleton features. Additionally, we introduce a prototype-guided text feature alignment strategy to mitigate the adverse impact of the distribution discrepancy during testing. We provide a theoretical analysis to support our prototype-guided text feature alignment strategy and empirically evaluate our overall PGFA on three well-known datasets. Compared with the top competitor SMIE method, our PGFA achieves absolute accuracy improvements of 22.96%, 12.53%, and 18.54% on the NTU-60, NTU-120, and PKU-MMD datasets, respectively.", "date": "2025-07-01", "categories": []}
{"id": "2502.06631", "title": "Conformal Predictions for Human Action Recognition with Vision-Language Models", "abstract": "Human-in-the-Loop (HITL) systems are essential in high-stakes, real-world applications where AI must collaborate with human decision-makers. This work investigates how Conformal Prediction (CP) techniques, which provide rigorous coverage guarantees, can enhance the reliability of state-of-the-art human action recognition (HAR) systems built upon Vision-Language Models (VLMs). We demonstrate that CP can significantly reduce the average number of candidate classes without modifying the underlying VLM. However, these reductions often result in distributions with long tails which can hinder their practical utility. To mitigate this, we propose tuning the temperature of the softmax prediction, without using additional calibration data. This work contributes to ongoing efforts for multi-modal human-AI interaction in dynamic real-world environments.", "date": "2025-02-10", "categories": []}
{"id": "2507.14657", "title": "AI-Enhanced Precision in Sport Taekwondo: Increasing Fairness, Speed, and Trust in Competition (FST.ai)", "abstract": "The integration of Artificial Intelligence (AI) into sports officiating represents a paradigm shift in how decisions are made in competitive environments. Traditional manual systems, even when supported by Instant Video Replay (IVR), often suffer from latency, subjectivity, and inconsistent enforcement, undermining fairness and athlete trust. This paper introduces 'FST.ai' -- which is developed under the 'R3AL.ai' project, which serves as its Principal Investigator: r3al.ai -- a novel AI-powered framework designed to enhance officiating in Sport Taekwondo, particularly focusing on the complex task of real-time head kick detection and scoring. Leveraging computer vision, deep learning, and edge inference, the system automates the identification and classification of key actions, significantly reducing decision time from minutes to seconds while improving consistency and transparency. Importantly, the methodology is not limited to Taekwondo. The underlying framework -- based on pose estimation, motion classification, and impact analysis -- can be adapted to a wide range of sports requiring action detection, such as judo, karate, fencing, or even team sports like football and basketball, where foul recognition or performance tracking is critical. By addressing one of Taekwondo's most challenging scenarios -- head kick scoring -- we demonstrate the robustness, scalability, and sport-agnostic potential of 'FST.ai' to transform officiating standards across multiple disciplines.", "date": "2025-07-19", "categories": []}
{"id": "2507.16342", "title": "Mamba-OTR: a Mamba-based Solution for Online Take and Release Detection from Untrimmed Egocentric Video", "abstract": "This work tackles the problem of Online detection of Take and Release (OTR) of an object in untrimmed egocentric videos. This task is challenging due to severe label imbalance, with temporally sparse positive annotations, and the need for precise temporal predictions. Furthermore, methods need to be computationally efficient in order to be deployed in real-world online settings. To address these challenges, we propose Mamba-OTR, a model based on the Mamba architecture. Mamba-OTR is designed to exploit temporal recurrence during inference while being trained on short video clips. To address label imbalance, our training pipeline incorporates the focal loss and a novel regularization scheme that aligns model predictions with the evaluation metric. Extensive experiments on EPIC-KITCHENS-100, the comparisons with transformer-based approach, and the evaluation of different training and test schemes demonstrate the superiority of Mamba-OTR in both accuracy and efficiency. These finding are particularly evident when evaluating full-length videos or high frame-rate sequences, even when trained on short video snippets for computational convenience. The proposed Mamba-OTR achieves a noteworthy mp-mAP of 45.48 when operating in a sliding-window fashion, and 43.35 in streaming mode, versus the 20.32 of a vanilla transformer and 25.16 of a vanilla Mamba, thus providing a strong baseline for OTR. We will publicly release the source code of Mamba-OTR to support future research.", "date": "2025-07-22", "categories": []}
{"id": "2406.06703", "title": "Video-based Exercise Classification and Activated Muscle Group Prediction with Hybrid X3D-SlowFast Network", "abstract": "This paper introduces a simple yet effective strategy for exercise classification and muscle group activation prediction (MGAP). These tasks have significant implications for personal fitness, facilitating more affordable, accessible, safer, and simpler exercise routines. This is particularly relevant for novices and individuals with disabilities. Previous research in the field is mostly dominated by the reliance on mounted sensors and a limited scope of exercises, reducing practicality for everyday use. Furthermore, existing MGAP methodologies suffer from a similar dependency on sensors and a restricted range of muscle groups, often excluding strength training exercises, which are pivotal for a comprehensive fitness regimen. Addressing these limitations, our research employs a video-based deep learning framework that encompasses a broad spectrum of exercises and muscle groups, including those vital for strength training. Utilizing the \"Workout/Exercises Video\" dataset, our approach integrates the X3D and SlowFast video activity recognition models in an effective way to enhance exercise classification and MGAP performance. Our findings demonstrate that this hybrid method, obtained via weighted ensemble, outperforms existing baseline models in accuracy. Pretrained models play a crucial role in enhancing overall performance, with optimal channel reduction values for the SlowFast model identified near 10. Through an ablation study that explores fine-tuning, we further elucidate the interrelation between the two tasks. Our composite model, a weighted-average ensemble of X3D and SlowFast, sets a new benchmark in both exercise classification and MGAP across all evaluated categories, offering a robust solution to the limitations of previous approaches.", "date": "2024-06-10", "categories": []}
{"id": "2507.06411", "title": "Hierarchical Multi-Stage Transformer Architecture for Context-Aware Temporal Action Localization", "abstract": "Inspired by the recent success of transformers and multi-stage architectures in video recognition and object detection domains. We thoroughly explore the rich spatio-temporal properties of transformers within a multi-stage architecture paradigm for the temporal action localization (TAL) task. This exploration led to the development of a hierarchical multi-stage transformer architecture called PCL-Former, where each subtask is handled by a dedicated transformer module with a specialized loss function. Specifically, the Proposal-Former identifies candidate segments in an untrimmed video that may contain actions, the Classification-Former classifies the action categories within those segments, and the Localization-Former precisely predicts the temporal boundaries (i.e., start and end) of the action instances. To evaluate the performance of our method, we have conducted extensive experiments on three challenging benchmark datasets: THUMOS-14, ActivityNet-1.3, and HACS Segments. We also conducted detailed ablation experiments to assess the impact of each individual module of our PCL-Former. The obtained quantitative results validate the effectiveness of the proposed PCL-Former, outperforming state-of-the-art TAL approaches by 2.8%, 1.2%, and 4.8% on THUMOS14, ActivityNet-1.3, and HACS datasets, respectively.", "date": "2025-07-08", "categories": []}
{"id": "2507.12426", "title": "DVFL-Net: A Lightweight Distilled Video Focal Modulation Network for Spatio-Temporal Action Recognition", "abstract": "The landscape of video recognition has evolved significantly, shifting from traditional Convolutional Neural Networks (CNNs) to Transformer-based architectures for improved accuracy. While 3D CNNs have been effective at capturing spatiotemporal dynamics, recent Transformer models leverage self-attention to model long-range spatial and temporal dependencies. Despite achieving state-of-the-art performance on major benchmarks, Transformers remain computationally expensive, particularly with dense video data. To address this, we propose a lightweight Video Focal Modulation Network, DVFL-Net, which distills spatiotemporal knowledge from a large pre-trained teacher into a compact nano student model, enabling efficient on-device deployment. DVFL-Net utilizes knowledge distillation and spatial-temporal feature modulation to significantly reduce computation while preserving high recognition performance. We employ forward Kullback-Leibler (KL) divergence alongside spatio-temporal focal modulation to effectively transfer both local and global context from the Video-FocalNet Base (teacher) to the proposed VFL-Net (student). We evaluate DVFL-Net on UCF50, UCF101, HMDB51, SSV2, and Kinetics-400, benchmarking it against recent state-of-the-art methods in Human Action Recognition (HAR). Additionally, we conduct a detailed ablation study analyzing the impact of forward KL divergence. The results confirm the superiority of DVFL-Net in achieving an optimal balance between performance and efficiency, demonstrating lower memory usage, reduced GFLOPs, and strong accuracy, making it a practical solution for real-time HAR applications.", "date": "2025-07-16", "categories": []}
{"id": "2507.13482", "title": "Improving Out-of-distribution Human Activity Recognition via IMU-Video Cross-modal Representation Learning", "abstract": "Human Activity Recognition (HAR) based on wearable inertial sensors plays a critical role in remote health monitoring. In patients with movement disorders, the ability to detect abnormal patient movements in their home environments can enable continuous optimization of treatments and help alert caretakers as needed. Machine learning approaches have been proposed for HAR tasks using Inertial Measurement Unit (IMU) data; however, most rely on application-specific labels and lack generalizability to data collected in different environments or populations. To address this limitation, we propose a new cross-modal self-supervised pretraining approach to learn representations from large-sale unlabeled IMU-video data and demonstrate improved generalizability in HAR tasks on out of distribution (OOD) IMU datasets, including a dataset collected from patients with Parkinson's disease. Specifically, our results indicate that the proposed cross-modal pretraining approach outperforms the current state-of-the-art IMU-video pretraining approach and IMU-only pretraining under zero-shot and few-shot evaluations. Broadly, our study provides evidence that in highly dynamic data modalities, such as IMU signals, cross-modal pretraining may be a useful tool to learn generalizable data representations. Our software is available at https://github.com/scheshmi/IMU-Video-OOD-HAR.", "date": "2025-07-17", "categories": []}
{"id": "2501.19034", "title": "XRF V2: A Dataset for Action Summarization with Wi-Fi Signals, and IMUs in Phones, Watches, Earbuds, and Glasses", "abstract": "Human Action Recognition (HAR) plays a crucial role in applications such as health monitoring, smart home automation, and human-computer interaction. While HAR has been extensively studied, action summarization using Wi-Fi and IMU signals in smart-home environments , which involves identifying and summarizing continuous actions, remains an emerging task. This paper introduces the novel XRF V2 dataset, designed for indoor daily activity Temporal Action Localization (TAL) and action summarization. XRF V2 integrates multimodal data from Wi-Fi signals, IMU sensors (smartphones, smartwatches, headphones, and smart glasses), and synchronized video recordings, offering a diverse collection of indoor activities from 16 volunteers across three distinct environments. To tackle TAL and action summarization, we propose the XRFMamba neural network, which excels at capturing long-term dependencies in untrimmed sensory sequences and achieves the best performance with an average mAP of 78.74, outperforming the recent WiFiTAD by 5.49 points in mAP@avg while using 35% fewer parameters. In action summarization, we introduce a new metric, Response Meaning Consistency (RMC), to evaluate action summarization performance. And it achieves an average Response Meaning Consistency (mRMC) of 0.802. We envision XRF V2 as a valuable resource for advancing research in human action localization, action forecasting, pose estimation, multimodal foundation models pre-training, synthetic data generation, and more. The data and code are available at https://github.com/aiotgroup/XRFV2.", "date": "2025-01-31", "categories": []}
{"id": "2507.12617", "title": "Predicting Soccer Penalty Kick Direction Using Human Action Recognition", "abstract": "Action anticipation has become a prominent topic in Human Action Recognition (HAR). However, its application to real-world sports scenarios remains limited by the availability of suitable annotated datasets. This work presents a novel dataset of manually annotated soccer penalty kicks to predict shot direction based on pre-kick player movements. We propose a deep learning classifier to benchmark this dataset that integrates HAR-based feature embeddings with contextual metadata. We evaluate twenty-two backbone models across seven architecture families (MViTv2, MViTv1, SlowFast, Slow, X3D, I3D, C2D), achieving up to 63.9% accuracy in predicting shot direction (left or right), outperforming the real goalkeepers' decisions. These results demonstrate the dataset's value for anticipatory action recognition and validate our model's potential as a generalizable approach for sports-based predictive tasks.", "date": "2025-07-16", "categories": []}
{"id": "2302.00646", "title": "Epic-Sounds: A Large-scale Dataset of Actions That Sound", "abstract": "We introduce EPIC-SOUNDS, a large-scale dataset of audio annotations capturing temporal extents and class labels within the audio stream of the egocentric videos. We propose an annotation pipeline where annotators temporally label distinguishable audio segments and describe the action that could have caused this sound. We identify actions that can be discriminated purely from audio, through grouping these free-form descriptions of audio into classes. For actions that involve objects colliding, we collect human annotations of the materials of these objects (e.g. a glass object being placed on a wooden surface), which we verify from video, discarding ambiguities. Overall, EPIC-SOUNDS includes 78.4k categorised segments of audible events and actions, distributed across 44 classes as well as 39.2k non-categorised segments. We train and evaluate state-of-the-art audio recognition and detection models on our dataset, for both audio-only and audio-visual methods. We also conduct analysis on: the temporal overlap between audio events, the temporal and label correlations between audio and visual modalities, the ambiguities in annotating materials from audio-only input, the importance of audio-only labels and the limitations of current models to understand actions that sound.", "date": "2023-02-01", "categories": []}
{"id": "2411.10745", "title": "Bridging the Skeleton-Text Modality Gap: Diffusion-Powered Modality Alignment for Zero-shot Skeleton-based Action Recognition", "abstract": "In zero-shot skeleton-based action recognition (ZSAR), aligning skeleton features with the text features of action labels is essential for accurately predicting unseen actions. ZSAR faces a fundamental challenge in bridging the modality gap between the two-kind features, which severely limits generalization to unseen actions. Previous methods focus on direct alignment between skeleton and text latent spaces, but the modality gaps between these spaces hinder robust generalization learning. Motivated by the success of diffusion models in multi-modal alignment (e.g., text-to-image, text-to-video), we firstly present a diffusion-based skeleton-text alignment framework for ZSAR. Our approach, Triplet Diffusion for Skeleton-Text Matching (TDSM), focuses on cross-alignment power of diffusion models rather than their generative capability. Specifically, TDSM aligns skeleton features with text prompts by incorporating text features into the reverse diffusion process, where skeleton features are denoised under text guidance, forming a unified skeleton-text latent space for robust matching. To enhance discriminative power, we introduce a triplet diffusion (TD) loss that encourages our TDSM to correct skeleton-text matches while pushing them apart for different action classes. Our TDSM significantly outperforms very recent state-of-the-art methods with significantly large margins of 2.36%-point to 13.05%-point, demonstrating superior accuracy and scalability in zero-shot settings through effective skeleton-text matching.", "date": "2024-11-16", "categories": []}
{"id": "2507.12132", "title": "DoRF: Doppler Radiance Fields for Robust Human Activity Recognition Using Wi-Fi", "abstract": "Wi-Fi Channel State Information (CSI) has gained increasing interest for remote sensing applications. Recent studies show that Doppler velocity projections extracted from CSI can enable human activity recognition (HAR) that is robust to environmental changes and generalizes to new users. However, despite these advances, generalizability still remains insufficient for practical deployment. Inspired by neural radiance fields (NeRF), which learn a volumetric representation of a 3D scene from 2D images, this work proposes a novel approach to reconstruct an informative 3D latent motion representation from one-dimensional Doppler velocity projections extracted from Wi-Fi CSI. The resulting latent representation is then used to construct a uniform Doppler radiance field (DoRF) of the motion, providing a comprehensive view of the performed activity and improving the robustness to environmental variability. The results show that the proposed approach noticeably enhances the generalization accuracy of Wi-Fi-based HAR, highlighting the strong potential of DoRFs for practical sensing applications.", "date": "2025-07-16", "categories": []}
{"id": "2507.12245", "title": "Calisthenics Skills Temporal Video Segmentation", "abstract": "Calisthenics is a fast-growing bodyweight discipline that consists of different categories, one of which is focused on skills. Skills in calisthenics encompass both static and dynamic elements performed by athletes. The evaluation of static skills is based on their difficulty level and the duration of the hold. Automated tools able to recognize isometric skills from a video by segmenting them to estimate their duration would be desirable to assist athletes in their training and judges during competitions. Although the video understanding literature on action recognition through body pose analysis is rich, no previous work has specifically addressed the problem of calisthenics skill temporal video segmentation. This study aims to provide an initial step towards the implementation of automated tools within the field of Calisthenics. To advance knowledge in this context, we propose a dataset of video footage of static calisthenics skills performed by athletes. Each video is annotated with a temporal segmentation which determines the extent of each skill. We hence report the results of a baseline approach to address the problem of skill temporal segmentation on the proposed dataset. The results highlight the feasibility of the proposed problem, while there is still room for improvement.", "date": "2025-07-16", "categories": []}
{"id": "2507.12292", "title": "Efficient Calisthenics Skills Classification through Foreground Instance Selection and Depth Estimation", "abstract": "Calisthenics skill classification is the computer vision task of inferring the skill performed by an athlete from images, enabling automatic performance assessment and personalized analytics. Traditional methods for calisthenics skill recognition are based on pose estimation methods to determine the position of skeletal data from images, which is later fed to a classification algorithm to infer the performed skill. Despite the progress in human pose estimation algorithms, they still involve high computational costs, long inference times, and complex setups, which limit the applicability of such approaches in real-time applications or mobile devices. This work proposes a direct approach to calisthenics skill recognition, which leverages depth estimation and athlete patch retrieval to avoid the computationally expensive human pose estimation module. Using Depth Anything V2 for depth estimation and YOLOv10 for athlete localization, we segment the subject from the background rather than relying on traditional pose estimation techniques. This strategy increases efficiency, reduces inference time, and improves classification accuracy. Our approach significantly outperforms skeleton-based methods, achieving 38.3x faster inference with RGB image patches and improved classification accuracy with depth patches (0.837 vs. 0.815). Beyond these performance gains, the modular design of our pipeline allows for flexible replacement of components, enabling future enhancements and adaptation to real-world applications.", "date": "2025-07-16", "categories": []}
{"id": "2507.10969", "title": "Women Sport Actions Dataset for Visual Classification Using Small Scale Training Data", "abstract": "Sports action classification representing complex body postures and player-object interactions is an emerging area in image-based sports analysis. Some works have contributed to automated sports action recognition using machine learning techniques over the past decades. However, sufficient image datasets representing women sports actions with enough intra- and inter-class variations are not available to the researchers. To overcome this limitation, this work presents a new dataset named WomenSports for women sports classification using small-scale training data. This dataset includes a variety of sports activities, covering wide variations in movements, environments, and interactions among players. In addition, this study proposes a convolutional neural network (CNN) for deep feature extraction. A channel attention scheme upon local contextual regions is applied to refine and enhance feature representation. The experiments are carried out on three different sports datasets and one dance dataset for generalizing the proposed algorithm, and the performances on these datasets are noteworthy. The deep learning method achieves 89.15% top-1 classification accuracy using ResNet-50 on the proposed WomenSports dataset, which is publicly available for research at Mendeley Data.", "date": "2025-07-15", "categories": []}
{"id": "2401.11406", "title": "Adversarial Augmentation Training Makes Action Recognition Models More Robust to Realistic Video Distribution Shifts", "abstract": "Despite recent advances in video action recognition achieving strong performance on existing benchmarks, these models often lack robustness when faced with natural distribution shifts between training and test data. We propose two novel evaluation methods to assess model resilience to such distribution disparity. One method uses two different datasets collected from different sources and uses one for training and validation, and the other for testing. More precisely, we created dataset splits of HMDB-51 or UCF-101 for training, and Kinetics-400 for testing, using the subset of the classes that are overlapping in both train and test datasets. The other proposed method extracts the feature mean of each class from the target evaluation dataset's training data (i.e. class prototype) and estimates test video prediction as a cosine similarity score between each sample to the class prototypes of each target class. This procedure does not alter model weights using the target dataset and it does not require aligning overlapping classes of two different datasets, thus is a very efficient method to test the model robustness to distribution shifts without prior knowledge of the target distribution. We address the robustness problem by adversarial augmentation training - generating augmented views of videos that are \"hard\" for the classification model by applying gradient ascent on the augmentation parameters - as well as \"curriculum\" scheduling the strength of the video augmentations. We experimentally demonstrate the superior performance of the proposed adversarial augmentation approach over baselines across three state-of-the-art action recognition models - TSM, Video Swin Transformer, and Uniformer. The presented work provides critical insight into model robustness to distribution shifts and presents effective techniques to enhance video action recognition performance in a real-world deployment.", "date": "2024-01-21", "categories": []}
{"id": "2507.09323", "title": "Dynamic Inter-Class Confusion-Aware Encoder for Audio-Visual Fusion in Human Activity Recognition", "abstract": "Humans do not understand individual events in isolation; rather, they generalize concepts within classes and compare them to others. Existing audio-video pre-training paradigms only focus on the alignment of the overall audio-video modalities, without considering the reinforcement of distinguishing easily confused classes through cognitive induction and contrast during training. This paper proposes the Dynamic Inter-Class Confusion-Aware Encoder (DICCAE), an encoder that aligns audio-video representations at a fine-grained, category-level. DICCAE addresses category confusion by dynamically adjusting the confusion loss based on inter-class confusion degrees, thereby enhancing the model's ability to distinguish between similar activities. To further extend the application of DICCAE, we also introduce a novel training framework that incorporates both audio and video modalities, as well as their fusion. To mitigate the scarcity of audio-video data in the human activity recognition task, we propose a cluster-guided audio-video self-supervised pre-training strategy for DICCAE. DICCAE achieves near state-of-the-art performance on the VGGSound dataset, with a top-1 accuracy of 65.5%. We further evaluate its feature representation quality through extensive ablation studies, validating the necessity of each module.", "date": "2025-07-12", "categories": []}
{"id": "2507.02827", "title": "USAD: End-to-End Human Activity Recognition via Diffusion Model with Spatiotemporal Attention", "abstract": "The primary objective of human activity recognition (HAR) is to infer ongoing human actions from sensor data, a task that finds broad applications in health monitoring, safety protection, and sports analysis. Despite proliferating research, HAR still faces key challenges, including the scarcity of labeled samples for rare activities, insufficient extraction of high-level features, and suboptimal model performance on lightweight devices. To address these issues, this paper proposes a comprehensive optimization approach centered on multi-attention interaction mechanisms. First, an unsupervised, statistics-guided diffusion model is employed to perform data augmentation, thereby alleviating the problems of labeled data scarcity and severe class imbalance. Second, a multi-branch spatio-temporal interaction network is designed, which captures multi-scale features of sequential data through parallel residual branches with 3*3, 5*5, and 7*7 convolutional kernels. Simultaneously, temporal attention mechanisms are incorporated to identify critical time points, while spatial attention enhances inter-sensor interactions. A cross-branch feature fusion unit is further introduced to improve the overall feature representation capability. Finally, an adaptive multi-loss function fusion strategy is integrated, allowing for dynamic adjustment of loss weights and overall model optimization. Experimental results on three public datasets, WISDM, PAMAP2, and OPPORTUNITY, demonstrate that the proposed unsupervised data augmentation spatio-temporal attention diffusion network (USAD) achieves accuracies of 98.84%, 93.81%, and 80.92% respectively, significantly outperforming existing approaches. Furthermore, practical deployment on embedded devices verifies the efficiency and feasibility of the proposed method.", "date": "2025-07-03", "categories": []}
{"id": "2407.16803", "title": "C3T: Cross-modal Transfer Through Time for Sensor-based Human Activity Recognition", "abstract": "In order to unlock the potential of diverse sensors, we investigate a method to transfer knowledge between time-series modalities using a multimodal \\textit{temporal} representation space for Human Activity Recognition (HAR). Specifically, we explore the setting where the modality used in testing has no labeled data during training, which we refer to as Unsupervised Modality Adaptation (UMA). We categorize existing UMA approaches as Student-Teacher or Contrastive Alignment methods. These methods typically compress continuous-time data samples into single latent vectors during alignment, inhibiting their ability to transfer temporal information through real-world temporal distortions. To address this, we introduce Cross-modal Transfer Through Time (C3T), which preserves temporal information during alignment to handle dynamic sensor data better. C3T achieves this by aligning a set of temporal latent vectors across sensing modalities. Our extensive experiments on various camera+IMU datasets demonstrate that C3T outperforms existing methods in UMA by at least 8% in accuracy and shows superior robustness to temporal distortions such as time-shift, misalignment, and dilation. Our findings suggest that C3T has significant potential for developing generalizable models for time-series sensor data, opening new avenues for various multimodal applications.", "date": "2024-07-23", "categories": []}
{"id": "2507.07734", "title": "EEvAct: Early Event-Based Action Recognition with High-Rate Two-Stream Spiking Neural Networks", "abstract": "Recognizing human activities early is crucial for the safety and responsiveness of human-robot and human-machine interfaces. Due to their high temporal resolution and low latency, event-based vision sensors are a perfect match for this early recognition demand. However, most existing processing approaches accumulate events to low-rate frames or space-time voxels which limits the early prediction capabilities. In contrast, spiking neural networks (SNNs) can process the events at a high-rate for early predictions, but most works still fall short on final accuracy. In this work, we introduce a high-rate two-stream SNN which closes this gap by outperforming previous work by 2% in final accuracy on the large-scale THU EACT-50 dataset. We benchmark the SNNs within a novel early event-based recognition framework by reporting Top-1 and Top-5 recognition scores for growing observation time. Finally, we exemplify the impact of these methods on a real-world task of early action triggering for human motion capture in sports.", "date": "2025-07-10", "categories": []}
{"id": "2507.07949", "title": "TinierHAR: Towards Ultra-Lightweight Deep Learning Models for Efficient Human Activity Recognition on Edge Devices", "abstract": "Human Activity Recognition (HAR) on resource-constrained wearable devices demands inference models that harmonize accuracy with computational efficiency. This paper introduces TinierHAR, an ultra-lightweight deep learning architecture that synergizes residual depthwise separable convolutions, gated recurrent units (GRUs), and temporal aggregation to achieve SOTA efficiency without compromising performance. Evaluated across 14 public HAR datasets, TinierHAR reduces Parameters by 2.7x (vs. TinyHAR) and 43.3x (vs. DeepConvLSTM), and MACs by 6.4x and 58.6x, respectively, while maintaining the averaged F1-scores. Beyond quantitative gains, this work provides the first systematic ablation study dissecting the contributions of spatial-temporal components across proposed TinierHAR, prior SOTA TinyHAR, and the classical DeepConvLSTM, offering actionable insights for designing efficient HAR systems. We finally discussed the findings and suggested principled design guidelines for future efficient HAR. To catalyze edge-HAR research, we open-source all materials in this work for future benchmarking\\footnote{https://github.com/zhaxidele/TinierHAR}", "date": "2025-07-10", "categories": []}
{"id": "2507.06442", "title": "THOR: Thermal-guided Hand-Object Reasoning via Adaptive Vision Sampling", "abstract": "Wearable cameras are increasingly used as an observational and interventional tool for human behaviors by providing detailed visual data of hand-related activities. This data can be leveraged to facilitate memory recall for logging of behavior or timely interventions aimed at improving health. However, continuous processing of RGB images from these cameras consumes significant power impacting battery lifetime, generates a large volume of unnecessary video data for post-processing, raises privacy concerns, and requires substantial computational resources for real-time analysis. We introduce THOR, a real-time adaptive spatio-temporal RGB frame sampling method that leverages thermal sensing to capture hand-object patches and classify them in real-time. We use low-resolution thermal camera data to identify moments when a person switches from one hand-related activity to another, and adjust the RGB frame sampling rate by increasing it during activity transitions and reducing it during periods of sustained activity. Additionally, we use the thermal cues from the hand to localize the region of interest (i.e., the hand-object interaction) in each RGB frame, allowing the system to crop and process only the necessary part of the image for activity recognition. We develop a wearable device to validate our method through an in-the-wild study with 14 participants and over 30 activities, and further evaluate it on Ego4D (923 participants across 9 countries, totaling 3,670 hours of video). Our results show that using only 3% of the original RGB video data, our method captures all the activity segments, and achieves hand-related activity recognition F1-score (95%) comparable to using the entire RGB video (94%). Our work provides a more practical path for the longitudinal use of wearable cameras to monitor hand-related activities and health-risk behaviors in real time.", "date": "2025-07-08", "categories": []}
{"id": "2507.00050", "title": "SEZ-HARN: Self-Explainable Zero-shot Human Activity Recognition Network", "abstract": "Human Activity Recognition (HAR), which uses data from Inertial Measurement Unit (IMU) sensors, has many practical applications in healthcare and assisted living environments. However, its use in real-world scenarios has been limited by the lack of comprehensive IMU-based HAR datasets that cover a wide range of activities and the lack of transparency in existing HAR models. Zero-shot HAR (ZS-HAR) overcomes the data limitations, but current models struggle to explain their decisions, making them less transparent. This paper introduces a novel IMU-based ZS-HAR model called the Self-Explainable Zero-shot Human Activity Recognition Network (SEZ-HARN). It can recognize activities not encountered during training and provide skeleton videos to explain its decision-making process. We evaluate the effectiveness of the proposed SEZ-HARN on four benchmark datasets PAMAP2, DaLiAc, HTD-MHAD and MHealth and compare its performance against three state-of-the-art black-box ZS-HAR models. The experiment results demonstrate that SEZ-HARN produces realistic and understandable explanations while achieving competitive Zero-shot recognition accuracy. SEZ-HARN achieves a Zero-shot prediction accuracy within 3\\% of the best-performing black-box model on PAMAP2 while maintaining comparable performance on the other three datasets.", "date": "2025-06-25", "categories": []}
{"id": "2507.03250", "title": "Subject Invariant Contrastive Learning for Human Activity Recognition", "abstract": "The high cost of annotating data makes self-supervised approaches, such as contrastive learning methods, appealing for Human Activity Recognition (HAR). Effective contrastive learning relies on selecting informative positive and negative samples. However, HAR sensor signals are subject to significant domain shifts caused by subject variability. These domain shifts hinder model generalization to unseen subjects by embedding subject-specific variations rather than activity-specific features. As a result, human activity recognition models trained with contrastive learning often struggle to generalize to new subjects. We introduce Subject-Invariant Contrastive Learning (SICL), a simple yet effective loss function to improve generalization in human activity recognition. SICL re-weights negative pairs drawn from the same subject to suppress subject-specific cues and emphasize activity-specific information. We evaluate our loss function on three public benchmarks: UTD-MHAD, MMAct, and DARai. We show that SICL improves performance by up to 11% over traditional contrastive learning methods. Additionally, we demonstrate the adaptability of our loss function across various settings, including multiple self-supervised methods, multimodal scenarios, and supervised learning frameworks.", "date": "2025-07-04", "categories": []}
{"id": "2505.06002", "title": "Task-Adapter++: Task-specific Adaptation with Order-aware Alignment for Few-shot Action Recognition", "abstract": "Large-scale pre-trained models have achieved remarkable success in language and image tasks, leading an increasing number of studies to explore the application of pre-trained image models, such as CLIP, in the domain of few-shot action recognition (FSAR). However, current methods generally suffer from several problems: 1) Direct fine-tuning often undermines the generalization capability of the pre-trained model; 2) The exploration of task-specific information is insufficient in the visual tasks; 3) The semantic order information is typically overlooked during text modeling; 4) Existing cross-modal alignment techniques ignore the temporal coupling of multimodal information. To address these, we propose Task-Adapter++, a parameter-efficient dual adaptation method for both image and text encoders. Specifically, to make full use of the variations across different few-shot learning tasks, we design a task-specific adaptation for the image encoder so that the most discriminative information can be well noticed during feature extraction. Furthermore, we leverage large language models (LLMs) to generate detailed sequential sub-action descriptions for each action class, and introduce semantic order adapters into the text encoder to effectively model the sequential relationships between these sub-actions. Finally, we develop an innovative fine-grained cross-modal alignment strategy that actively maps visual features to reside in the same temporal stage as semantic descriptions. Extensive experiments fully demonstrate the effectiveness and superiority of the proposed method, which achieves state-of-the-art performance on 5 benchmarks consistently. The code is open-sourced at https://github.com/Jaulin-Bage/Task-Adapter-pp.", "date": "2025-05-09", "categories": []}
{"id": "2312.01431", "title": "D$^2$ST-Adapter: Disentangled-and-Deformable Spatio-Temporal Adapter for Few-shot Action Recognition", "abstract": "Adapting pre-trained image models to video modality has proven to be an effective strategy for robust few-shot action recognition. In this work, we explore the potential of adapter tuning in image-to-video model adaptation and propose a novel video adapter tuning framework, called Disentangled-and-Deformable Spatio-Temporal Adapter (D$^2$ST-Adapter). It features a lightweight design, low adaptation overhead and powerful spatio-temporal feature adaptation capabilities. D$^2$ST-Adapter is structured with an internal dual-pathway architecture that enables built-in disentangled encoding of spatial and temporal features within the adapter, seamlessly integrating into the single-stream feature learning framework of pre-trained image models. In particular, we develop an efficient yet effective implementation of the D$^2$ST-Adapter, incorporating the specially devised anisotropic Deformable Spatio-Temporal Attention as its pivotal operation. This mechanism can be individually tailored for two pathways with anisotropic sampling densities along the spatial and temporal domains in 3D spatio-temporal space, enabling disentangled encoding of spatial and temporal features while maintaining a lightweight design. Extensive experiments by instantiating our method on both pre-trained ResNet and ViT demonstrate the superiority of our method over state-of-the-art methods. Our method is particularly well-suited to challenging scenarios where temporal dynamics are critical for action recognition. Code is available at https://github.com/qizhongtan/D2ST-Adapter.", "date": "2023-12-03", "categories": []}
{"id": "2506.23196", "title": "DEL: Dense Event Localization for Multi-modal Audio-Visual Understanding", "abstract": "Real-world videos often contain overlapping events and complex temporal dependencies, making multimodal interaction modeling particularly challenging. We introduce DEL, a framework for dense semantic action localization, aiming to accurately detect and classify multiple actions at fine-grained temporal resolutions in long untrimmed videos. DEL consists of two key modules: the alignment of audio and visual features that leverage masked self-attention to enhance intra-mode consistency and a multimodal interaction refinement module that models cross-modal dependencies across multiple scales, enabling high-level semantics and fine-grained details. Our method achieves state-of-the-art performance on multiple real-world Temporal Action Localization (TAL) datasets, UnAV-100, THUMOS14, ActivityNet 1.3, and EPIC-Kitchens-100, surpassing previous approaches with notable average mAP gains of +3.3%, +2.6%, +1.2%, +1.7% (verb), and +1.4% (noun), respectively.", "date": "2025-06-29", "categories": []}
{"id": "2506.23573", "title": "Online Human Action Detection during Escorting", "abstract": "The deployment of robot assistants in large indoor spaces has seen significant growth, with escorting tasks becoming a key application. However, most current escorting robots primarily rely on navigation-focused strategies, assuming that the person being escorted will follow without issue. In crowded environments, this assumption often falls short, as individuals may struggle to keep pace, become obstructed, get distracted, or need to stop unexpectedly. As a result, conventional robotic systems are often unable to provide effective escorting services due to their limited understanding of human movement dynamics. To address these challenges, an effective escorting robot must continuously detect and interpret human actions during the escorting process and adjust its movement accordingly. However, there is currently no existing dataset designed specifically for human action detection in the context of escorting. Given that escorting often occurs in crowded environments, where other individuals may enter the robot's camera view, the robot also needs to identify the specific human it is escorting (the subject) before predicting their actions. Since no existing model performs both person re-identification and action prediction in real-time, we propose a novel neural network architecture that can accomplish both tasks. This enables the robot to adjust its speed dynamically based on the escortee's movements and seamlessly resume escorting after any disruption. In comparative evaluations against strong baselines, our system demonstrates superior efficiency and effectiveness, showcasing its potential to significantly improve robotic escorting services in complex, real-world scenarios.", "date": "2025-06-30", "categories": []}
{"id": "2506.22179", "title": "Frequency-Semantic Enhanced Variational Autoencoder for Zero-Shot Skeleton-based Action Recognition", "abstract": "Zero-shot skeleton-based action recognition aims to develop models capable of identifying actions beyond the categories encountered during training. Previous approaches have primarily focused on aligning visual and semantic representations but often overlooked the importance of fine-grained action patterns in the semantic space (e.g., the hand movements in drinking water and brushing teeth). To address these limitations, we propose a Frequency-Semantic Enhanced Variational Autoencoder (FS-VAE) to explore the skeleton semantic representation learning with frequency decomposition. FS-VAE consists of three key components: 1) a frequency-based enhancement module with high- and low-frequency adjustments to enrich the skeletal semantics learning and improve the robustness of zero-shot action recognition; 2) a semantic-based action description with multilevel alignment to capture both local details and global correspondence, effectively bridging the semantic gap and compensating for the inherent loss of information in skeleton sequences; 3) a calibrated cross-alignment loss that enables valid skeleton-text pairs to counterbalance ambiguous ones, mitigating discrepancies and ambiguities in skeleton and text features, thereby ensuring robust alignment. Evaluations on the benchmarks demonstrate the effectiveness of our approach, validating that frequency-enhanced semantic features enable robust differentiation of visually and semantically similar action clusters, improving zero-shot action recognition.", "date": "2025-06-27", "categories": []}
{"id": "2506.20342", "title": "Feature Hallucination for Self-supervised Action Recognition", "abstract": "Understanding human actions in videos requires more than raw pixel analysis; it relies on high-level semantic reasoning and effective integration of multimodal features. We propose a deep translational action recognition framework that enhances recognition accuracy by jointly predicting action concepts and auxiliary features from RGB video frames. At test time, hallucination streams infer missing cues, enriching feature representations without increasing computational overhead. To focus on action-relevant regions beyond raw pixels, we introduce two novel domain-specific descriptors. Object Detection Features (ODF) aggregate outputs from multiple object detectors to capture contextual cues, while Saliency Detection Features (SDF) highlight spatial and intensity patterns crucial for action recognition. Our framework seamlessly integrates these descriptors with auxiliary modalities such as optical flow, Improved Dense Trajectories, skeleton data, and audio cues. It remains compatible with state-of-the-art architectures, including I3D, AssembleNet, Video Transformer Network, FASTER, and recent models like VideoMAE V2 and InternVideo2. To handle uncertainty in auxiliary features, we incorporate aleatoric uncertainty modeling in the hallucination step and introduce a robust loss function to mitigate feature noise. Our multimodal self-supervised action recognition framework achieves state-of-the-art performance on multiple benchmarks, including Kinetics-400, Kinetics-600, and Something-Something V2, demonstrating its effectiveness in capturing fine-grained action dynamics.", "date": "2025-06-25", "categories": []}
{"id": "2312.03799", "title": "Low-power, Continuous Remote Behavioral Localization with Event Cameras", "abstract": "Researchers in natural science need reliable methods for quantifying animal behavior. Recently, numerous computer vision methods emerged to automate the process. However, observing wild species at remote locations remains a challenging task due to difficult lighting conditions and constraints on power supply and data storage. Event cameras offer unique advantages for battery-dependent remote monitoring due to their low power consumption and high dynamic range capabilities. We use this novel sensor to quantify a behavior in Chinstrap penguins called ecstatic display. We formulate the problem as a temporal action detection task, determining the start and end times of the behavior. For this purpose, we recorded a colony of breeding penguins in Antarctica for several weeks and labeled event data on 16 nests. The developed method consists of a generator of candidate time intervals (proposals) and a classifier of the actions within them. The experiments show that the event cameras' natural response to motion is effective for continuous behavior monitoring and detection, reaching a mean average precision (mAP) of 58% (which increases to 63% in good weather conditions). The results also demonstrate the robustness against various lighting conditions contained in the challenging dataset. The low-power capabilities of the event camera allow it to record significantly longer than with a conventional camera. This work pioneers the use of event cameras for remote wildlife observation, opening new interdisciplinary opportunities. https://tub-rip.github.io/eventpenguins/", "date": "2023-12-06", "categories": []}
{"id": "2506.17685", "title": "Domain Generalization using Action Sequences for Egocentric Action Recognition", "abstract": "Recognizing human activities from visual inputs, particularly through a first-person viewpoint, is essential for enabling robots to replicate human behavior. Egocentric vision, characterized by cameras worn by observers, captures diverse changes in illumination, viewpoint, and environment. This variability leads to a notable drop in the performance of Egocentric Action Recognition models when tested in environments not seen during training. In this paper, we tackle these challenges by proposing a domain generalization approach for Egocentric Action Recognition. Our insight is that action sequences often reflect consistent user intent across visual domains. By leveraging action sequences, we aim to enhance the model's generalization ability across unseen environments. Our proposed method, named SeqDG, introduces a visual-text sequence reconstruction objective (SeqRec) that uses contextual cues from both text and visual inputs to reconstruct the central action of the sequence. Additionally, we enhance the model's robustness by training it on mixed sequences of actions from different domains (SeqMix). We validate SeqDG on the EGTEA and EPIC-KITCHENS-100 datasets. Results on EPIC-KITCHENS-100, show that SeqDG leads to +2.4% relative average improvement in cross-domain action recognition in unseen environments, and on EGTEA the model achieved +0.6% Top-1 accuracy over SOTA in intra-domain action recognition.", "date": "2025-06-21", "categories": []}
{"id": "2506.18261", "title": "Improving Weakly Supervised Temporal Action Localization by Exploiting Multi-resolution Information in Temporal Domain", "abstract": "Weakly supervised temporal action localization is a challenging task as only the video-level annotation is available during the training process. To address this problem, we propose a two-stage approach to fully exploit multi-resolution information in the temporal domain and generate high quality frame-level pseudo labels based on both appearance and motion streams. Specifically, in the first stage, we generate reliable initial frame-level pseudo labels, and in the second stage, we iteratively refine the pseudo labels and use a set of selected frames with highly confident pseudo labels to train neural networks and better predict action class scores at each frame. We fully exploit temporal information at multiple scales to improve temporal action localization performance. Specifically, in order to obtain reliable initial frame-level pseudo labels, in the first stage, we propose an Initial Label Generation (ILG) module, which leverages temporal multi-resolution consistency to generate high quality class activation sequences (CASs), which consist of a number of sequences with each sequence measuring how likely each video frame belongs to one specific action class. In the second stage, we propose a Progressive Temporal Label Refinement (PTLR) framework. In our PTLR framework, two networks called Network-OTS and Network-RTS, which are respectively used to generate CASs for the original temporal scale and the reduced temporal scales, are used as two streams (i.e., the OTS stream and the RTS stream) to refine the pseudo labels in turn. By this way, the multi-resolution information in the temporal domain is exchanged at the pseudo label level, and our work can help improve each stream (i.e., the OTS/RTS stream) by exploiting the refined pseudo labels from another stream (i.e., the RTS/OTS stream).", "date": "2025-06-23", "categories": []}
{"id": "2506.18721", "title": "Including Semantic Information via Word Embeddings for Skeleton-based Action Recognition", "abstract": "Effective human action recognition is widely used for cobots in Industry 4.0 to assist in assembly tasks. However, conventional skeleton-based methods often lose keypoint semantics, limiting their effectiveness in complex interactions. In this work, we introduce a novel approach to skeleton-based action recognition that enriches input representations by leveraging word embeddings to encode semantic information. Our method replaces one-hot encodings with semantic volumes, enabling the model to capture meaningful relationships between joints and objects. Through extensive experiments on multiple assembly datasets, we demonstrate that our approach significantly improves classification performance, and enhances generalization capabilities by simultaneously supporting different skeleton types and object classes. Our findings highlight the potential of incorporating semantic information to enhance skeleton-based action recognition in dynamic and diverse environments.", "date": "2025-06-23", "categories": []}
{"id": "2506.16701", "title": "Language-driven Description Generation and Common Sense Reasoning for Video Action Recognition", "abstract": "Recent video action recognition methods have shown excellent performance by adapting large-scale pre-trained language-image models to the video domain. However, language models contain rich common sense priors - the scene contexts that humans use to constitute an understanding of objects, human-object interactions, and activities - that have not been fully exploited. In this paper, we introduce a framework incorporating language-driven common sense priors to identify cluttered video action sequences from monocular views that are often heavily occluded. We propose: (1) A video context summary component that generates candidate objects, activities, and the interactions between objects and activities; (2) A description generation module that describes the current scene given the context and infers subsequent activities, through auxiliary prompts and common sense reasoning; (3) A multi-modal activity recognition head that combines visual and textual cues to recognize video actions. We demonstrate the effectiveness of our approach on the challenging Action Genome and Charades datasets.", "date": "2025-06-20", "categories": []}
{"id": "2506.13910", "title": "Intelligent Image Sensing for Crime Analysis: A ML Approach towards Enhanced Violence Detection and Investigation", "abstract": "The increasing global crime rate, coupled with substantial human and property losses, highlights the limitations of traditional surveillance methods in promptly detecting diverse and unexpected acts of violence. Addressing this pressing need for automatic violence detection, we leverage Machine Learning to detect and categorize violent events in video streams. This paper introduces a comprehensive framework for violence detection and classification, employing Supervised Learning for both binary and multi-class violence classification. The detection model relies on 3D Convolutional Neural Networks, while the classification model utilizes the separable convolutional 3D model for feature extraction and bidirectional LSTM for temporal processing. Training is conducted on a diverse customized datasets with frame-level annotations, incorporating videos from surveillance cameras, human recordings, hockey fight, sohas and wvd dataset across various platforms. Additionally, a camera module integrated with raspberry pi is used to capture live video feed, which is sent to the ML model for processing. Thus, demonstrating improved performance in terms of computational resource efficiency and accuracy.", "date": "2025-06-16", "categories": []}
{"id": "2506.12585", "title": "DejaVid: Encoder-Agnostic Learned Temporal Matching for Video Classification", "abstract": "In recent years, large transformer-based video encoder models have greatly advanced state-of-the-art performance on video classification tasks. However, these large models typically process videos by averaging embedding outputs from multiple clips over time to produce fixed-length representations. This approach fails to account for a variety of time-related features, such as variable video durations, chronological order of events, and temporal variance in feature significance. While methods for temporal modeling do exist, they often require significant architectural changes and expensive retraining, making them impractical for off-the-shelf, fine-tuned large encoders. To overcome these limitations, we propose DejaVid, an encoder-agnostic method that enhances model performance without the need for retraining or altering the architecture. Our framework converts a video into a variable-length temporal sequence of embeddings, which we call a multivariate time series (MTS). An MTS naturally preserves temporal order and accommodates variable video durations. We then learn per-timestep, per-feature weights over the encoded MTS frames, allowing us to account for variations in feature importance over time. We introduce a new neural network architecture inspired by traditional time series alignment algorithms for this learning task. Our evaluation demonstrates that DejaVid substantially improves the performance of a state-of-the-art large encoder, achieving leading Top-1 accuracy of 77.2% on Something-Something V2, 89.1% on Kinetics-400, and 88.6% on HMDB51, while adding fewer than 1.8% additional learnable parameters and requiring less than 3 hours of training time. Our code is available at https://github.com/darrylho/DejaVid.", "date": "2025-06-14", "categories": []}
{"id": "2506.12848", "title": "Towards Fine-Grained Emotion Understanding via Skeleton-Based Micro-Gesture Recognition", "abstract": "We present our solution to the MiGA Challenge at IJCAI 2025, which aims to recognize micro-gestures (MGs) from skeleton sequences for the purpose of hidden emotion understanding. MGs are characterized by their subtlety, short duration, and low motion amplitude, making them particularly challenging to model and classify. We adopt PoseC3D as the baseline framework and introduce three key enhancements: (1) a topology-aware skeleton representation specifically designed for the iMiGUE dataset to better capture fine-grained motion patterns; (2) an improved temporal processing strategy that facilitates smoother and more temporally consistent motion modeling; and (3) the incorporation of semantic label embeddings as auxiliary supervision to improve the model generalization. Our method achieves a Top-1 accuracy of 67.01\\% on the iMiGUE test set. As a result of these contributions, our approach ranks third on the official MiGA Challenge leaderboard. The source code is available at \\href{https://github.com/EGO-False-Sleep/Miga25_track1}{https://github.com/EGO-False-Sleep/Miga25\\_track1}.", "date": "2025-06-15", "categories": []}
{"id": "2506.13322", "title": "Active Multimodal Distillation for Few-shot Action Recognition", "abstract": "Owing to its rapid progress and broad application prospects, few-shot action recognition has attracted considerable interest. However, current methods are predominantly based on limited single-modal data, which does not fully exploit the potential of multimodal information. This paper presents a novel framework that actively identifies reliable modalities for each sample using task-specific contextual cues, thus significantly improving recognition performance. Our framework integrates an Active Sample Inference (ASI) module, which utilizes active inference to predict reliable modalities based on posterior distributions and subsequently organizes them accordingly. Unlike reinforcement learning, active inference replaces rewards with evidence-based preferences, making more stable predictions. Additionally, we introduce an active mutual distillation module that enhances the representation learning of less reliable modalities by transferring knowledge from more reliable ones. Adaptive multimodal inference is employed during the meta-test to assign higher weights to reliable modalities. Extensive experiments across multiple benchmarks demonstrate that our method significantly outperforms existing approaches.", "date": "2025-06-16", "categories": []}
{"id": "2506.13458", "title": "Leveraging Vision-Language Pre-training for Human Activity Recognition in Still Images", "abstract": "Recognising human activity in a single photo enables indexing, safety and assistive applications, yet lacks motion cues. Using 285 MSCOCO images labelled as walking, running, sitting, and standing, scratch CNNs scored 41% accuracy. Fine-tuning multimodal CLIP raised this to 76%, demonstrating that contrastive vision-language pre-training decisively improves still-image action recognition in real-world deployments.", "date": "2025-06-16", "categories": []}
{"id": "2506.07612", "title": "Scaling Human Activity Recognition: A Comparative Evaluation of Synthetic Data Generation and Augmentation Techniques", "abstract": "Human activity recognition (HAR) is often limited by the scarcity of labeled datasets due to the high cost and complexity of real-world data collection. To mitigate this, recent work has explored generating virtual inertial measurement unit (IMU) data via cross-modality transfer. While video-based and language-based pipelines have each shown promise, they differ in assumptions and computational cost. Moreover, their effectiveness relative to traditional sensor-level data augmentation remains unclear. In this paper, we present a direct comparison between these two virtual IMU generation approaches against classical data augmentation techniques. We construct a large-scale virtual IMU dataset spanning 100 diverse activities from Kinetics-400 and simulate sensor signals at 22 body locations. The three data generation strategies are evaluated on benchmark HAR datasets (UTD-MHAD, PAMAP2, HAD-AW) using four popular models. Results show that virtual IMU data significantly improves performance over real or augmented data alone, particularly under limited-data conditions. We offer practical guidance on choosing data generation strategies and highlight the distinct advantages and disadvantages of each approach.", "date": "2025-06-09", "categories": []}
{"id": "2506.11165", "title": "Evaluating BiLSTM and CNN+GRU Approaches for Human Activity Recognition Using WiFi CSI Data", "abstract": "This paper compares the performance of BiLSTM and CNN+GRU deep learning models for Human Activity Recognition (HAR) on two WiFi-based Channel State Information (CSI) datasets: UT-HAR and NTU-Fi HAR. The findings indicate that the CNN+GRU model has a higher accuracy on the UT-HAR dataset (95.20%) thanks to its ability to extract spatial features. In contrast, the BiLSTM model performs better on the high-resolution NTU-Fi HAR dataset (92.05%) by extracting long-term temporal dependencies more effectively. The findings strongly emphasize the critical role of dataset characteristics and preprocessing techniques in model performance improvement. We also show the real-world applicability of such models in applications like healthcare and intelligent home systems, highlighting their potential for unobtrusive activity recognition.", "date": "2025-06-11", "categories": []}
{"id": "2506.11678", "title": "Pose Matters: Evaluating Vision Transformers and CNNs for Human Action Recognition on Small COCO Subsets", "abstract": "This study explores human action recognition using a three-class subset of the COCO image corpus, benchmarking models from simple fully connected networks to transformer architectures. The binary Vision Transformer (ViT) achieved 90% mean test accuracy, significantly exceeding multiclass classifiers such as convolutional networks (approximately 35%) and CLIP-based models (approximately 62-64%). A one-way ANOVA (F = 61.37, p < 0.001) confirmed these differences are statistically significant. Qualitative analysis with SHAP explainer and LeGrad heatmaps indicated that the ViT localizes pose-specific regions (e.g., lower limbs for walking or running), while simpler feed-forward models often focus on background textures, explaining their errors. These findings emphasize the data efficiency of transformer representations and the importance of explainability techniques in diagnosing class-specific failures.", "date": "2025-06-13", "categories": []}
{"id": "2408.10688", "title": "TDS-CLIP: Temporal Difference Side Network for Efficient VideoAction Recognition", "abstract": "Recently, large-scale pre-trained vision-language models (e.g., CLIP), have garnered significant attention thanks to their powerful representative capabilities. This inspires researchers in transferring the knowledge from these large pre-trained models to other task-specific models, e.g., Video Action Recognition (VAR) models, via particularly leveraging side networks to enhance the efficiency of parameter-efficient fine-tuning (PEFT). However, current transferring approaches in VAR tend to directly transfer the frozen knowledge from large pre-trained models to action recognition networks with minimal cost, instead of exploiting the temporal modeling capabilities of the action recognition models themselves. Therefore, in this paper, we propose a novel memory-efficient Temporal Difference Side Network (TDS-CLIP) to balance knowledge transferring and temporal modeling, avoiding backpropagation in frozen parameter models. Specifically, we introduce a Temporal Difference Adapter (TD-Adapter), which can effectively capture local temporal differences in motion features to strengthen the model's global temporal modeling capabilities. Furthermore, we designed a Side Motion Enhancement Adapter (SME-Adapter) to guide the proposed side network in efficiently learning the rich motion information in videos, thereby improving the side network's ability to capture and learn motion information. Extensive experiments are conducted on three benchmark datasets, including Something-Something V1&V2, and Kinetics-400. Experimental results show that our method achieves competitive performance in video action recognition tasks.", "date": "2024-08-20", "categories": []}
{"id": "2404.12258", "title": "DeepLocalization: Using change point detection for Temporal Action Localization", "abstract": "In this study, we introduce DeepLocalization, an innovative framework devised for the real-time localization of actions tailored explicitly for monitoring driver behavior. Utilizing the power of advanced deep learning methodologies, our objective is to tackle the critical issue of distracted driving-a significant factor contributing to road accidents. Our strategy employs a dual approach: leveraging Graph-Based Change-Point Detection for pinpointing actions in time alongside a Video Large Language Model (Video-LLM) for precisely categorizing activities. Through careful prompt engineering, we customize the Video-LLM to adeptly handle driving activities' nuances, ensuring its classification efficacy even with sparse data. Engineered to be lightweight, our framework is optimized for consumer-grade GPUs, making it vastly applicable in practical scenarios. We subjected our method to rigorous testing on the SynDD2 dataset, a complex benchmark for distracted driving behaviors, where it demonstrated commendable performance-achieving 57.5% accuracy in event classification and 51% in event detection. These outcomes underscore the substantial promise of DeepLocalization in accurately identifying diverse driver behaviors and their temporal occurrences, all within the bounds of limited computational resources.", "date": "2024-04-18", "categories": []}
{"id": "2506.02472", "title": "HRTR: A Single-stage Transformer for Fine-grained Sub-second Action Segmentation in Stroke Rehabilitation", "abstract": "Stroke rehabilitation often demands precise tracking of patient movements to monitor progress, with complexities of rehabilitation exercises presenting two critical challenges: fine-grained and sub-second (under one-second) action detection. In this work, we propose the High Resolution Temporal Transformer (HRTR), to time-localize and classify high-resolution (fine-grained), sub-second actions in a single-stage transformer, eliminating the need for multi-stage methods and post-processing. Without any refinements, HRTR outperforms state-of-the-art systems on both stroke related and general datasets, achieving Edit Score (ES) of 70.1 on StrokeRehab Video, 69.4 on StrokeRehab IMU, and 88.4 on 50Salads.", "date": "2025-06-03", "categories": []}
{"id": "2506.02550", "title": "Technical Report for Ego4D Long-Term Action Anticipation Challenge 2025", "abstract": "In this report, we present a novel three-stage framework developed for the Ego4D Long-Term Action Anticipation (LTA) task. Inspired by recent advances in foundation models, our method consists of three stages: feature extraction, action recognition, and long-term action anticipation. First, visual features are extracted using a high-performance visual encoder. The features are then fed into a Transformer to predict verbs and nouns, with a verb-noun co-occurrence matrix incorporated to enhance recognition accuracy. Finally, the predicted verb-noun pairs are formatted as textual prompts and input into a fine-tuned large language model (LLM) to anticipate future action sequences. Our framework achieves first place in this challenge at CVPR 2025, establishing a new state-of-the-art in long-term action prediction. Our code will be released at https://github.com/CorrineQiu/Ego4D-LTA-Challenge-2025.", "date": "2025-06-03", "categories": []}
{"id": "2506.09345", "title": "An Effective End-to-End Solution for Multimodal Action Recognition", "abstract": "Recently, multimodal tasks have strongly advanced the field of action recognition with their rich multimodal information. However, due to the scarcity of tri-modal data, research on tri-modal action recognition tasks faces many challenges. To this end, we have proposed a comprehensive multimodal action recognition solution that effectively utilizes multimodal information. First, the existing data are transformed and expanded by optimizing data enhancement techniques to enlarge the training scale. At the same time, more RGB datasets are used to pre-train the backbone network, which is better adapted to the new task by means of transfer learning. Secondly, multimodal spatial features are extracted with the help of 2D CNNs and combined with the Temporal Shift Module (TSM) to achieve multimodal spatial-temporal feature extraction comparable to 3D CNNs and improve the computational efficiency. In addition, common prediction enhancement methods, such as Stochastic Weight Averaging (SWA), Ensemble and Test-Time augmentation (TTA), are used to integrate the knowledge of models from different training periods of the same architecture and different architectures, so as to predict the actions from different perspectives and fully exploit the target information. Ultimately, we achieved the Top-1 accuracy of 99% and the Top-5 accuracy of 100% on the competition leaderboard, demonstrating the superiority of our solution.", "date": "2025-06-11", "categories": []}
{"id": "2411.08466", "title": "Weakly Supervised Temporal Action Localization via Dual-Prior Collaborative Learning Guided by Multimodal Large Language Models", "abstract": "Recent breakthroughs in Multimodal Large Language Models (MLLMs) have gained significant recognition within the deep learning community, where the fusion of the Video Foundation Models (VFMs) and Large Language Models(LLMs) has proven instrumental in constructing robust video understanding systems, effectively surmounting constraints associated with predefined visual tasks. These sophisticated MLLMs exhibit remarkable proficiency in comprehending videos, swiftly attaining unprecedented performance levels across diverse benchmarks. However, their operation demands substantial memory and computational resources, underscoring the continued importance of traditional models in video comprehension tasks. In this paper, we introduce a novel learning paradigm termed MLLM4WTAL. This paradigm harnesses the potential of MLLM to offer temporal action key semantics and complete semantic priors for conventional Weakly-supervised Temporal Action Localization (WTAL) methods. MLLM4WTAL facilitates the enhancement of WTAL by leveraging MLLM guidance. It achieves this by integrating two distinct modules: Key Semantic Matching (KSM) and Complete Semantic Reconstruction (CSR). These modules work in tandem to effectively address prevalent issues like incomplete and over-complete outcomes common in WTAL methods. Rigorous experiments are conducted to validate the efficacy of our proposed approach in augmenting the performance of various heterogeneous WTAL models.", "date": "2024-11-13", "categories": []}
{"id": "2411.15729", "title": "OccludeNet: A Causal Journey into Mixed-View Actor-Centric Video Action Recognition under Occlusions", "abstract": "The lack of occlusion data in common action recognition video datasets limits model robustness and hinders consistent performance gains. We build OccludeNet, a large-scale occluded video dataset including both real and synthetic occlusion scenes in different natural settings. OccludeNet includes dynamic occlusion, static occlusion, and multi-view interactive occlusion, addressing gaps in current datasets. Our analysis shows occlusion affects action classes differently: actions with low scene relevance and partial body visibility see larger drops in accuracy. To overcome the limits of existing occlusion-aware methods, we propose a structural causal model for occluded scenes and introduce the Causal Action Recognition (CAR) method, which uses backdoor adjustment and counterfactual reasoning. This approach strengthens key actor information and improves model robustness to occlusion. We hope the challenges of OccludeNet will encourage more study of causal links in occluded scenes and lead to a fresh look at class relations, ultimately leading to lasting performance improvements. Our code and data is availibale at: https://github.com/The-Martyr/OccludeNet-Dataset", "date": "2024-11-24", "categories": []}
{"id": "2506.06480", "title": "(LiFT) Lightweight Fitness Transformer: A language-vision model for Remote Monitoring of Physical Training", "abstract": "We introduce a fitness tracking system that enables remote monitoring for exercises using only a RGB smartphone camera, making fitness tracking more private, scalable, and cost effective. Although prior work explored automated exercise supervision, existing models are either too limited in exercise variety or too complex for real-world deployment. Prior approaches typically focus on a small set of exercises and fail to generalize across diverse movements. In contrast, we develop a robust, multitask motion analysis model capable of performing exercise detection and repetition counting across hundreds of exercises, a scale far beyond previous methods. We overcome previous data limitations by assembling a large-scale fitness dataset, Olympia covering more than 1,900 exercises. To our knowledge, our vision-language model is the first that can perform multiple tasks on skeletal fitness data. On Olympia, our model can detect exercises with 76.5% accuracy and count repetitions with 85.3% off-by-one accuracy, using only RGB video. By presenting a single vision-language transformer model for both exercise identification and rep counting, we take a significant step toward democratizing AI-powered fitness tracking.", "date": "2025-06-06", "categories": []}
{"id": "2505.23524", "title": "CLIP-AE: CLIP-assisted Cross-view Audio-Visual Enhancement for Unsupervised Temporal Action Localization", "abstract": "Temporal Action Localization (TAL) has garnered significant attention in information retrieval. Existing supervised or weakly supervised methods heavily rely on labeled temporal boundaries and action categories, which are labor-intensive and time-consuming. Consequently, unsupervised temporal action localization (UTAL) has gained popularity. However, current methods face two main challenges: 1) Classification pre-trained features overly focus on highly discriminative regions; 2) Solely relying on visual modality information makes it difficult to determine contextual boundaries. To address these issues, we propose a CLIP-assisted cross-view audiovisual enhanced UTAL method. Specifically, we introduce visual language pre-training (VLP) and classification pre-training-based collaborative enhancement to avoid excessive focus on highly discriminative regions; we also incorporate audio perception to provide richer contextual boundary information. Finally, we introduce a self-supervised cross-view learning paradigm to achieve multi-view perceptual enhancement without additional annotations. Extensive experiments on two public datasets demonstrate our model's superiority over several state-of-the-art competitors.", "date": "2025-05-29", "categories": []}
{"id": "2409.09444", "title": "KAN-HyperpointNet for Point Cloud Sequence-Based 3D Human Action Recognition", "abstract": "Point cloud sequence-based 3D action recognition has achieved impressive performance and efficiency. However, existing point cloud sequence modeling methods cannot adequately balance the precision of limb micro-movements with the integrity of posture macro-structure, leading to the loss of crucial information cues in action inference. To overcome this limitation, we introduce D-Hyperpoint, a novel data type generated through a D-Hyperpoint Embedding module. D-Hyperpoint encapsulates both regional-momentary motion and global-static posture, effectively summarizing the unit human action at each moment. In addition, we present a D-Hyperpoint KANsMixer module, which is recursively applied to nested groupings of D-Hyperpoints to learn the action discrimination information and creatively integrates Kolmogorov-Arnold Networks (KAN) to enhance spatio-temporal interaction within D-Hyperpoints. Finally, we propose KAN-HyperpointNet, a spatio-temporal decoupled network architecture for 3D action recognition. Extensive experiments on two public datasets: MSR Action3D and NTU-RGB+D 60, demonstrate the state-of-the-art performance of our method.", "date": "2024-09-14", "categories": []}
{"id": "2506.03174", "title": "Multimodal Foundation Model for Cross-Modal Retrieval and Activity Recognition Tasks", "abstract": "In recent years, the widespread adoption of wearable devices has highlighted the growing importance of behavior analysis using IMU. While applications span diverse fields such as healthcare and robotics, recent studies have increasingly focused on multimodal analysis, in addition to unimodal analysis. Several studies have proposed multimodal foundation models that incorporate first-person video and text data; however, these models still fall short in providing a detailed analysis of full-body human activity. To address this limitation, we propose Activity Understanding and Representations Alignment - Multimodal Foundation Model (AURA-MFM), a foundational model integrating four modalities: third-person video, motion capture, IMU, and text. By incorporating third-person video and motion capture data, the model enables a detailed and multidimensional understanding of human activity, which first-person perspectives alone fail to capture. Additionally, a Transformer-based IMU encoder is employed to enhance the model's overall performance. Experimental evaluations on retrieval and activity recognition tasks demonstrate that our model surpasses existing methods. Notably, in the zero-shot classification for action recognition, our method achieved significantly higher performance, with an F1-score of 0.6226 and an accuracy of 0.7320, whereas the existing method recorded an F1-score of 0.0747 and an accuracy of 0.1961.", "date": "2025-05-29", "categories": []}
{"id": "2506.03481", "title": "Heterogeneous Skeleton-Based Action Representation Learning", "abstract": "Skeleton-based human action recognition has received widespread attention in recent years due to its diverse range of application scenarios. Due to the different sources of human skeletons, skeleton data naturally exhibit heterogeneity. The previous works, however, overlook the heterogeneity of human skeletons and solely construct models tailored for homogeneous skeletons. This work addresses the challenge of heterogeneous skeleton-based action representation learning, specifically focusing on processing skeleton data that varies in joint dimensions and topological structures. The proposed framework comprises two primary components: heterogeneous skeleton processing and unified representation learning. The former first converts two-dimensional skeleton data into three-dimensional skeleton via an auxiliary network, and then constructs a prompted unified skeleton using skeleton-specific prompts. We also design an additional modality named semantic motion encoding to harness the semantic information within skeletons. The latter module learns a unified action representation using a shared backbone network that processes different heterogeneous skeletons. Extensive experiments on the NTU-60, NTU-120, and PKU-MMD II datasets demonstrate the effectiveness of our method in various tasks of action understanding. Our approach can be applied to action recognition in robots with different humanoid structures.", "date": "2025-06-04", "categories": []}
{"id": "2506.02382", "title": "Multi-level and Multi-modal Action Anticipation", "abstract": "Action anticipation, the task of predicting future actions from partially observed videos, is crucial for advancing intelligent systems. Unlike action recognition, which operates on fully observed videos, action anticipation must handle incomplete information. Hence, it requires temporal reasoning, and inherent uncertainty handling. While recent advances have been made, traditional methods often focus solely on visual modalities, neglecting the potential of integrating multiple sources of information. Drawing inspiration from human behavior, we introduce \\textit{Multi-level and Multi-modal Action Anticipation (m\\&m-Ant)}, a novel multi-modal action anticipation approach that combines both visual and textual cues, while explicitly modeling hierarchical semantic information for more accurate predictions. To address the challenge of inaccurate coarse action labels, we propose a fine-grained label generator paired with a specialized temporal consistency loss function to optimize performance. Extensive experiments on widely used datasets, including Breakfast, 50 Salads, and DARai, demonstrate the effectiveness of our approach, achieving state-of-the-art results with an average anticipation accuracy improvement of 3.08\\% over existing methods. This work underscores the potential of multi-modal and hierarchical modeling in advancing action anticipation and establishes a new benchmark for future research in the field. Our code is available at: https://github.com/olivesgatech/mM-ant.", "date": "2025-06-03", "categories": []}
{"id": "2506.00827", "title": "Improving Keystep Recognition in Ego-Video via Dexterous Focus", "abstract": "In this paper, we address the challenge of understanding human activities from an egocentric perspective. Traditional activity recognition techniques face unique challenges in egocentric videos due to the highly dynamic nature of the head during many activities. We propose a framework that seeks to address these challenges in a way that is independent of network architecture by restricting the ego-video input to a stabilized, hand-focused video. We demonstrate that this straightforward video transformation alone outperforms existing egocentric video baselines on the Ego-Exo4D Fine-Grained Keystep Recognition benchmark without requiring any alteration of the underlying model infrastructure.", "date": "2025-06-01", "categories": []}
{"id": "2506.00915", "title": "3D Skeleton-Based Action Recognition: A Review", "abstract": "With the inherent advantages of skeleton representation, 3D skeleton-based action recognition has become a prominent topic in the field of computer vision. However, previous reviews have predominantly adopted a model-oriented perspective, often neglecting the fundamental steps involved in skeleton-based action recognition. This oversight tends to ignore key components of skeleton-based action recognition beyond model design and has hindered deeper, more intrinsic understanding of the task. To bridge this gap, our review aims to address these limitations by presenting a comprehensive, task-oriented framework for understanding skeleton-based action recognition. We begin by decomposing the task into a series of sub-tasks, placing particular emphasis on preprocessing steps such as modality derivation and data augmentation. The subsequent discussion delves into critical sub-tasks, including feature extraction and spatio-temporal modeling techniques. Beyond foundational action recognition networks, recently advanced frameworks such as hybrid architectures, Mamba models, large language models (LLMs), and generative models have also been highlighted. Finally, a comprehensive overview of public 3D skeleton datasets is presented, accompanied by an analysis of state-of-the-art algorithms evaluated on these benchmarks. By integrating task-oriented discussions, comprehensive examinations of sub-tasks, and an emphasis on the latest advancements, our review provides a fundamental and accessible structured roadmap for understanding and advancing the field of 3D skeleton-based action recognition.", "date": "2025-06-01", "categories": []}
{"id": "2506.01214", "title": "A Review on Coarse to Fine-Grained Animal Action Recognition", "abstract": "This review provides an in-depth exploration of the field of animal action recognition, focusing on coarse-grained (CG) and fine-grained (FG) techniques. The primary aim is to examine the current state of research in animal behaviour recognition and to elucidate the unique challenges associated with recognising subtle animal actions in outdoor environments. These challenges differ significantly from those encountered in human action recognition due to factors such as non-rigid body structures, frequent occlusions, and the lack of large-scale, annotated datasets. The review begins by discussing the evolution of human action recognition, a more established field, highlighting how it progressed from broad, coarse actions in controlled settings to the demand for fine-grained recognition in dynamic environments. This shift is particularly relevant for animal action recognition, where behavioural variability and environmental complexity present unique challenges that human-centric models cannot fully address. The review then underscores the critical differences between human and animal action recognition, with an emphasis on high intra-species variability, unstructured datasets, and the natural complexity of animal habitats. Techniques like spatio-temporal deep learning frameworks (e.g., SlowFast) are evaluated for their effectiveness in animal behaviour analysis, along with the limitations of existing datasets. By assessing the strengths and weaknesses of current methodologies and introducing a recently-published dataset, the review outlines future directions for advancing fine-grained action recognition, aiming to improve accuracy and generalisability in behaviour analysis across species.", "date": "2025-06-01", "categories": []}
{"id": "2506.01757", "title": "Efficient Egocentric Action Recognition with Multimodal Data", "abstract": "The increasing availability of wearable XR devices opens new perspectives for Egocentric Action Recognition (EAR) systems, which can provide deeper human understanding and situation awareness. However, deploying real-time algorithms on these devices can be challenging due to the inherent trade-offs between portability, battery life, and computational resources. In this work, we systematically analyze the impact of sampling frequency across different input modalities - RGB video and 3D hand pose - on egocentric action recognition performance and CPU usage. By exploring a range of configurations, we provide a comprehensive characterization of the trade-offs between accuracy and computational efficiency. Our findings reveal that reducing the sampling rate of RGB frames, when complemented with higher-frequency 3D hand pose input, can preserve high accuracy while significantly lowering CPU demands. Notably, we observe up to a 3x reduction in CPU usage with minimal to no loss in recognition performance. This highlights the potential of multimodal input strategies as a viable approach to achieving efficient, real-time EAR on XR devices.", "date": "2025-06-02", "categories": []}
{"id": "2505.24375", "title": "Spatiotemporal Analysis of Forest Machine Operations Using 3D Video Classification", "abstract": "This paper presents a deep learning-based framework for classifying forestry operations from dashcam video footage. Focusing on four key work elements - crane-out, cutting-and-to-processing, driving, and processing - the approach employs a 3D ResNet-50 architecture implemented with PyTorchVideo. Trained on a manually annotated dataset of field recordings, the model achieves strong performance, with a validation F1 score of 0.88 and precision of 0.90. These results underscore the effectiveness of spatiotemporal convolutional networks for capturing both motion patterns and appearance in real-world forestry environments.\n  The system integrates standard preprocessing and augmentation techniques to improve generalization, but overfitting is evident, highlighting the need for more training data and better class balance. Despite these challenges, the method demonstrates clear potential for reducing the manual workload associated with traditional time studies, offering a scalable solution for operational monitoring and efficiency analysis in forestry.\n  This work contributes to the growing application of AI in natural resource management and sets the foundation for future systems capable of real-time activity recognition in forest machinery. Planned improvements include dataset expansion, enhanced regularization, and deployment trials on embedded systems for in-field use.", "date": "2025-05-30", "categories": []}
{"id": "2505.22858", "title": "A Probabilistic Jump-Diffusion Framework for Open-World Egocentric Activity Recognition", "abstract": "Open-world egocentric activity recognition poses a fundamental challenge due to its unconstrained nature, requiring models to infer unseen activities from an expansive, partially observed search space. We introduce ProbRes, a Probabilistic Residual search framework based on jump-diffusion that efficiently navigates this space by balancing prior-guided exploration with likelihood-driven exploitation. Our approach integrates structured commonsense priors to construct a semantically coherent search space, adaptively refines predictions using Vision-Language Models (VLMs) and employs a stochastic search mechanism to locate high-likelihood activity labels while minimizing exhaustive enumeration efficiently. We systematically evaluate ProbRes across multiple openness levels (L0--L3), demonstrating its adaptability to increasing search space complexity. In addition to achieving state-of-the-art performance on benchmark datasets (GTEA Gaze, GTEA Gaze+, EPIC-Kitchens, and Charades-Ego), we establish a clear taxonomy for open-world recognition, delineating the challenges and methodological advancements necessary for egocentric activity understanding.", "date": "2025-05-28", "categories": []}
{"id": "2505.23012", "title": "Spatio-Temporal Joint Density Driven Learning for Skeleton-Based Action Recognition", "abstract": "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.", "date": "2025-05-29", "categories": []}
{"id": "2505.21918", "title": "Self-supervised Learning Method Using Transformer for Multi-dimensional Sensor Data Processing", "abstract": "We developed a deep learning algorithm for human activity recognition using sensor signals as input. In this study, we built a pretrained language model based on the Transformer architecture, which is widely used in natural language processing. By leveraging this pretrained model, we aimed to improve performance on the downstream task of human activity recognition. While this task can be addressed using a vanilla Transformer, we propose an enhanced n-dimensional numerical processing Transformer that incorporates three key features: embedding n-dimensional numerical data through a linear layer, binning-based pre-processing, and a linear transformation in the output layer. We evaluated the effectiveness of our proposed model across five different datasets. Compared to the vanilla Transformer, our model demonstrated 10%-15% improvements in accuracy.", "date": "2025-05-28", "categories": []}
{"id": "2505.18048", "title": "SHARDeg: A Benchmark for Skeletal Human Action Recognition in Degraded Scenarios", "abstract": "Computer vision (CV) models for detection, prediction or classification tasks operate on video data-streams that are often degraded in the real world, due to deployment in real-time or on resource-constrained hardware. It is therefore critical that these models are robust to degraded data, but state of the art (SoTA) models are often insufficiently assessed with these real-world constraints in mind. This is exemplified by Skeletal Human Action Recognition (SHAR), which is critical in many CV pipelines operating in real-time and at the edge, but robustness to degraded data has previously only been shallowly and inconsistently assessed. Here we address this issue for SHAR by providing an important first data degradation benchmark on the most detailed and largest 3D open dataset, NTU-RGB+D-120, and assess the robustness of five leading SHAR models to three forms of degradation that represent real-world issues. We demonstrate the need for this benchmark by showing that the form of degradation, which has not previously been considered, has a large impact on model accuracy; at the same effective frame rate, model accuracy can vary by >40% depending on degradation type. We also identify that temporal regularity of frames in degraded SHAR data is likely a major driver of differences in model performance, and harness this to improve performance of existing models by up to >40%, through employing a simple mitigation approach based on interpolation. Finally, we highlight how our benchmark has helped identify an important degradation-resistant SHAR model based in Rough Path Theory; the LogSigRNN SHAR model outperforms the SoTA DeGCN model in five out of six cases at low frame rates by an average accuracy of 6%, despite trailing the SoTA model by 11-12% on un-degraded data at high frame rates (30 FPS).", "date": "2025-05-23", "categories": []}
{"id": "2505.20739", "title": "Detecting Informative Channels: ActionFormer", "abstract": "Human Activity Recognition (HAR) has recently witnessed advancements with Transformer-based models. Especially, ActionFormer shows us a new perspectives for HAR in the sense that this approach gives us additional outputs which detect the border of the activities as well as the activity labels. ActionFormer was originally proposed with its input as image/video. However, this was converted to with its input as sensor signals as well. We analyze this extensively in terms of deep learning architectures. Based on the report of high temporal dynamics which limits the model's ability to capture subtle changes effectively and of the interdependencies between the spatial and temporal features. We propose the modified ActionFormer which will decrease these defects for sensor signals. The key to our approach lies in accordance with the Sequence-and-Excitation strategy to minimize the increase in additional parameters and opt for the swish activation function to retain the information about direction in the negative range. Experiments on the WEAR dataset show that our method achieves substantial improvement of a 16.01\\% in terms of average mAP for inertial data.", "date": "2025-05-27", "categories": []}
{"id": "2505.20920", "title": "HuMoCon: Concept Discovery for Human Motion Understanding", "abstract": "We present HuMoCon, a novel motion-video understanding framework designed for advanced human behavior analysis. The core of our method is a human motion concept discovery framework that efficiently trains multi-modal encoders to extract semantically meaningful and generalizable features. HuMoCon addresses key challenges in motion concept discovery for understanding and reasoning, including the lack of explicit multi-modality feature alignment and the loss of high-frequency information in masked autoencoding frameworks. Our approach integrates a feature alignment strategy that leverages video for contextual understanding and motion for fine-grained interaction modeling, further with a velocity reconstruction mechanism to enhance high-frequency feature expression and mitigate temporal over-smoothing. Comprehensive experiments on standard benchmarks demonstrate that HuMoCon enables effective motion concept discovery and significantly outperforms state-of-the-art methods in training large models for human motion understanding. We will open-source the associated code with our paper.", "date": "2025-05-27", "categories": []}
{"id": "2503.01353", "title": "Dendron: Enhancing Human Activity Recognition with On-Device TinyML Learning", "abstract": "Human activity recognition (HAR) is a research field that employs Machine Learning (ML) techniques to identify user activities. Recent studies have prioritized the development of HAR solutions directly executed on wearable devices, enabling the on-device activity recognition. This approach is supported by the Tiny Machine Learning (TinyML) paradigm, which integrates ML within embedded devices with limited resources. However, existing approaches in the field lack in the capability for on-device learning of new HAR tasks, particularly when supervised data are scarce. To address this limitation, our paper introduces Dendron, a novel TinyML methodology designed to facilitate the on-device learning of new tasks for HAR, even in conditions of limited supervised data. Experimental results on two public-available datasets and an off-the-shelf device (STM32-NUCLEO-F401RE) show the effectiveness and efficiency of the proposed solution.", "date": "2025-03-03", "categories": []}
{"id": "2505.19928", "title": "CA3D: Convolutional-Attentional 3D Nets for Efficient Video Activity Recognition on the Edge", "abstract": "In this paper, we introduce a deep learning solution for video activity recognition that leverages an innovative combination of convolutional layers with a linear-complexity attention mechanism. Moreover, we introduce a novel quantization mechanism to further improve the efficiency of our model during both training and inference. Our model maintains a reduced computational cost, while preserving robust learning and generalization capabilities. Our approach addresses the issues related to the high computing requirements of current models, with the goal of achieving competitive accuracy on consumer and edge devices, enabling smart home and smart healthcare applications where efficiency and privacy issues are of concern. We experimentally validate our model on different established and publicly available video activity recognition benchmarks, improving accuracy over alternative models at a competitive computing cost.", "date": "2025-05-26", "categories": []}
{"id": "2505.17555", "title": "ProTAL: A Drag-and-Link Video Programming Framework for Temporal Action Localization", "abstract": "Temporal Action Localization (TAL) aims to detect the start and end timestamps of actions in a video. However, the training of TAL models requires a substantial amount of manually annotated data. Data programming is an efficient method to create training labels with a series of human-defined labeling functions. However, its application in TAL faces difficulties of defining complex actions in the context of temporal video frames. In this paper, we propose ProTAL, a drag-and-link video programming framework for TAL. ProTAL enables users to define \\textbf{key events} by dragging nodes representing body parts and objects and linking them to constrain the relations (direction, distance, etc.). These definitions are used to generate action labels for large-scale unlabelled videos. A semi-supervised method is then employed to train TAL models with such labels. We demonstrate the effectiveness of ProTAL through a usage scenario and a user study, providing insights into designing video programming framework.", "date": "2025-05-23", "categories": []}
{"id": "2505.17867", "title": "Multi-task Learning For Joint Action and Gesture Recognition", "abstract": "In practical applications, computer vision tasks often need to be addressed simultaneously. Multitask learning typically achieves this by jointly training a single deep neural network to learn shared representations, providing efficiency and improving generalization. Although action and gesture recognition are closely related tasks, since they focus on body and hand movements, current state-of-the-art methods handle them separately. In this paper, we show that employing a multi-task learning paradigm for action and gesture recognition results in more efficient, robust and generalizable visual representations, by leveraging the synergies between these tasks. Extensive experiments on multiple action and gesture datasets demonstrate that handling actions and gestures in a single architecture can achieve better performance for both tasks in comparison to their single-task learning variants.", "date": "2025-05-23", "categories": []}
{"id": "2502.03364", "title": "Scaling laws in wearable human activity recognition", "abstract": "Many deep architectures and self-supervised pre-training techniques have been proposed for human activity recognition (HAR) from wearable multimodal sensors. Scaling laws have the potential to help move towards more principled design by linking model capacity with pre-training data volume. Yet, scaling laws have not been established for HAR to the same extent as in language and vision. By conducting an exhaustive grid search on both amount of pre-training data and Transformer architectures, we establish the first known scaling laws for HAR. We show that pre-training loss scales with a power law relationship to amount of data and parameter count and that increasing the number of users in a dataset results in a steeper improvement in performance than increasing data per user, indicating that diversity of pre-training data is important, which contrasts to some previously reported findings in self-supervised HAR. We show that these scaling laws translate to downstream performance improvements on three HAR benchmark datasets of postures, modes of locomotion and activities of daily living: UCI HAR and WISDM Phone and WISDM Watch. Finally, we suggest some previously published works should be revisited in light of these scaling laws with more adequate model capacities.", "date": "2025-02-05", "categories": []}
{"id": "2505.05711", "title": "DiGIT: Multi-Dilated Gated Encoder and Central-Adjacent Region Integrated Decoder for Temporal Action Detection Transformer", "abstract": "In this paper, we examine a key limitation in query-based detectors for temporal action detection (TAD), which arises from their direct adaptation of originally designed architectures for object detection. Despite the effectiveness of the existing models, they struggle to fully address the unique challenges of TAD, such as the redundancy in multi-scale features and the limited ability to capture sufficient temporal context. To address these issues, we propose a multi-dilated gated encoder and central-adjacent region integrated decoder for temporal action detection transformer (DiGIT). Our approach replaces the existing encoder that consists of multi-scale deformable attention and feedforward network with our multi-dilated gated encoder. Our proposed encoder reduces the redundant information caused by multi-level features while maintaining the ability to capture fine-grained and long-range temporal information. Furthermore, we introduce a central-adjacent region integrated decoder that leverages a more comprehensive sampling strategy for deformable cross-attention to capture the essential information. Extensive experiments demonstrate that DiGIT achieves state-of-the-art performance on THUMOS14, ActivityNet v1.3, and HACS-Segment. Code is available at: https://github.com/Dotori-HJ/DiGIT", "date": "2025-05-09", "categories": []}
{"id": "2504.02287", "title": "MultiSensor-Home: A Wide-area Multi-modal Multi-view Dataset for Action\n  Recognition and Transformer-based Sensor Fusion", "abstract": "  Multi-modal multi-view action recognition is a rapidly growing field in\ncomputer vision, offering significant potential for applications in\nsurveillance. However, current datasets often fail to address real-world\nchallenges such as wide-area distributed settings, asynchronous data streams,\nand the lack of frame-level annotations. Furthermore, existing methods face\ndifficulties in effectively modeling inter-view relationships and enhancing\nspatial feature learning. In this paper, we introduce the MultiSensor-Home\ndataset, a novel benchmark designed for comprehensive action recognition in\nhome environments, and also propose the Multi-modal Multi-view\nTransformer-based Sensor Fusion (MultiTSF) method. The proposed\nMultiSensor-Home dataset features untrimmed videos captured by distributed\nsensors, providing high-resolution RGB and audio data along with detailed\nmulti-view frame-level action labels. The proposed MultiTSF method leverages a\nTransformer-based fusion mechanism to dynamically model inter-view\nrelationships. Furthermore, the proposed method integrates a human detection\nmodule to enhance spatial feature learning, guiding the model to prioritize\nframes with human activity to enhance action the recognition accuracy.\nExperiments on the proposed MultiSensor-Home and the existing MM-Office\ndatasets demonstrate the superiority of MultiTSF over the state-of-the-art\nmethods. Quantitative and qualitative results highlight the effectiveness of\nthe proposed method in advancing real-world multi-modal multi-view action\nrecognition. The source code is available at\nhttps://github.com/thanhhff/MultiTSF.\n", "date": "2025-04-03", "categories": []}
{"id": "2411.15106", "title": "About Time: Advances, Challenges, and Outlooks of Action Understanding", "abstract": "  We have witnessed impressive advances in video action understanding.\nIncreased dataset sizes, variability, and computation availability have enabled\nleaps in performance and task diversification. Current systems can provide\ncoarse- and fine-grained descriptions of video scenes, extract segments\ncorresponding to queries, synthesize unobserved parts of videos, and predict\ncontext across multiple modalities. This survey comprehensively reviews\nadvances in uni- and multi-modal action understanding across a range of tasks.\nWe focus on prevalent challenges, overview widely adopted datasets, and survey\nseminal works with an emphasis on recent advances. We broadly distinguish\nbetween three temporal scopes: (1) recognition tasks of actions observed in\nfull, (2) prediction tasks for ongoing partially observed actions, and (3)\nforecasting tasks for subsequent unobserved action(s). This division allows us\nto identify specific action modeling and video representation challenges.\nFinally, we outline future directions to address current shortcomings.\n", "date": "2024-11-22", "categories": []}
{"id": "2505.01713", "title": "Vision and Intention Boost Large Language Model in Long-Term Action\n  Anticipation", "abstract": "  Long-term action anticipation (LTA) aims to predict future actions over an\nextended period. Previous approaches primarily focus on learning exclusively\nfrom video data but lack prior knowledge. Recent researches leverage large\nlanguage models (LLMs) by utilizing text-based inputs which suffer severe\ninformation loss. To tackle these limitations single-modality methods face, we\npropose a novel Intention-Conditioned Vision-Language (ICVL) model in this\nstudy that fully leverages the rich semantic information of visual data and the\npowerful reasoning capabilities of LLMs. Considering intention as a high-level\nconcept guiding the evolution of actions, we first propose to employ a\nvision-language model (VLM) to infer behavioral intentions as comprehensive\ntextual features directly from video inputs. The inferred intentions are then\nfused with visual features through a multi-modality fusion strategy, resulting\nin intention-enhanced visual representations. These enhanced visual\nrepresentations, along with textual prompts, are fed into LLM for future action\nanticipation. Furthermore, we propose an effective example selection strategy\njointly considers visual and textual similarities, providing more relevant and\ninformative examples for in-context learning. Extensive experiments with\nstate-of-the-art performance on Ego4D, EPIC-Kitchens-55, and EGTEA GAZE+\ndatasets fully demonstrate the effectiveness and superiority of the proposed\nmethod.\n", "date": "2025-05-03", "categories": []}
{"id": "2505.02052", "title": "TxP: Reciprocal Generation of Ground Pressure Dynamics and Activity\n  Descriptions for Improving Human Activity Recognition", "abstract": "  Sensor-based human activity recognition (HAR) has predominantly focused on\nInertial Measurement Units and vision data, often overlooking the capabilities\nunique to pressure sensors, which capture subtle body dynamics and shifts in\nthe center of mass. Despite their potential for postural and balance-based\nactivities, pressure sensors remain underutilized in the HAR domain due to\nlimited datasets. To bridge this gap, we propose to exploit generative\nfoundation models with pressure-specific HAR techniques. Specifically, we\npresent a bidirectional Text$\\times$Pressure model that uses generative\nfoundation models to interpret pressure data as natural language. TxP\naccomplishes two tasks: (1) Text2Pressure, converting activity text\ndescriptions into pressure sequences, and (2) Pressure2Text, generating\nactivity descriptions and classifications from dynamic pressure maps.\nLeveraging pre-trained models like CLIP and LLaMA 2 13B Chat, TxP is trained on\nour synthetic PressLang dataset, containing over 81,100 text-pressure pairs.\nValidated on real-world data for activities such as yoga and daily tasks, TxP\nprovides novel approaches to data augmentation and classification grounded in\natomic actions. This consequently improved HAR performance by up to 12.4\\% in\nmacro F1 score compared to the state-of-the-art, advancing pressure-based HAR\nwith broader applications and deeper insights into human movement.\n", "date": "2025-05-04", "categories": []}
{"id": "2501.11124", "title": "Rethinking Pseudo-Label Guided Learning for Weakly Supervised Temporal\n  Action Localization from the Perspective of Noise Correction", "abstract": "  Pseudo-label learning methods have been widely applied in weakly-supervised\ntemporal action localization. Existing works directly utilize weakly-supervised\nbase model to generate instance-level pseudo-labels for training the\nfully-supervised detection head. We argue that the noise in pseudo-labels would\ninterfere with the learning of fully-supervised detection head, leading to\nsignificant performance leakage. Issues with noisy labels include:(1)\ninaccurate boundary localization; (2) undetected short action clips; (3)\nmultiple adjacent segments incorrectly detected as one segment. To target these\nissues, we introduce a two-stage noisy label learning strategy to harness every\npotential useful signal in noisy labels. First, we propose a frame-level\npseudo-label generation model with a context-aware denoising algorithm to\nrefine the boundaries. Second, we introduce an online-revised teacher-student\nframework with a missing instance compensation module and an ambiguous instance\ncorrection module to solve the short-action-missing and many-to-one problems.\nBesides, we apply a high-quality pseudo-label mining loss in our online-revised\nteacher-student framework to add different weights to the noisy labels to train\nmore effectively. Our model outperforms the previous state-of-the-art method in\ndetection accuracy and inference speed greatly upon the THUMOS14 and\nActivityNet v1.2 benchmarks.\n", "date": "2025-01-19", "categories": []}
{"id": "2504.21266", "title": "CoCoDiff: Diversifying Skeleton Action Features via Coarse-Fine\n  Text-Co-Guided Latent Diffusion", "abstract": "  In action recognition tasks, feature diversity is essential for enhancing\nmodel generalization and performance. Existing methods typically promote\nfeature diversity by expanding the training data in the sample space, which\noften leads to inefficiencies and semantic inconsistencies. To overcome these\nproblems, we propose a novel Coarse-fine text co-guidance Diffusion model\n(CoCoDiff). CoCoDiff generates diverse yet semantically consistent features in\nthe latent space by leveraging diffusion and multi-granularity textual\nguidance. Specifically, our approach feeds spatio-temporal features extracted\nfrom skeleton sequences into a latent diffusion model to generate diverse\naction representations. Meanwhile, we introduce a coarse-fine text co-guided\nstrategy that leverages textual information from large language models (LLMs)\nto ensure semantic consistency between the generated features and the original\ninputs. It is noted that CoCoDiff operates as a plug-and-play auxiliary module\nduring training, incurring no additional inference cost. Extensive experiments\ndemonstrate that CoCoDiff achieves SOTA performance on skeleton-based action\nrecognition benchmarks, including NTU RGB+D, NTU RGB+D 120 and\nKinetics-Skeleton.\n", "date": "2025-04-30", "categories": []}
{"id": "2504.18977", "title": "3DPyranet Features Fusion for Spatio-temporal Feature Learning", "abstract": "  Convolutional neural network (CNN) slides a kernel over the whole image to\nproduce an output map. This kernel scheme reduces the number of parameters with\nrespect to a fully connected neural network (NN). While CNN has proven to be an\neffective model in recognition of handwritten characters and traffic signal\nsign boards, etc. recently, its deep variants have proven to be effective in\nsimilar as well as more challenging applications like object, scene and action\nrecognition. Deep CNN add more layers and kernels to the classical CNN,\nincreasing the number of parameters, and partly reducing the main advantage of\nCNN which is less parameters. In this paper, a 3D pyramidal neural network\ncalled 3DPyraNet and a discriminative approach for spatio-temporal feature\nlearning based on it, called 3DPyraNet-F, are proposed. 3DPyraNet introduces a\nnew weighting scheme which learns features from both spatial and temporal\ndimensions analyzing multiple adjacent frames and keeping a biological\nplausible structure. It keeps the spatial topology of the input image and\npresents fewer parameters and lower computational and memory costs compared to\nboth fully connected NNs and recent deep CNNs. 3DPyraNet-F extract the features\nmaps of the highest layer of the learned network, fuse them in a single vector,\nand provide it as input in such a way to a linear-SVM classifier that enhances\nthe recognition of human actions and dynamic scenes from the videos.\nEncouraging results are reported with 3DPyraNet in real-world environments,\nespecially in the presence of camera induced motion. Further, 3DPyraNet-F\nclearly outperforms the state-of-the-art on three benchmark datasets and shows\ncomparable result for the fourth.\n", "date": "2025-04-26", "categories": []}
{"id": "2504.18059", "title": "POET: Prompt Offset Tuning for Continual Human Action Adaptation", "abstract": "  As extended reality (XR) is redefining how users interact with computing\ndevices, research in human action recognition is gaining prominence. Typically,\nmodels deployed on immersive computing devices are static and limited to their\ndefault set of classes. The goal of our research is to provide users and\ndevelopers with the capability to personalize their experience by adding new\naction classes to their device models continually. Importantly, a user should\nbe able to add new classes in a low-shot and efficient manner, while this\nprocess should not require storing or replaying any of user's sensitive\ntraining data. We formalize this problem as privacy-aware few-shot continual\naction recognition. Towards this end, we propose POET: Prompt-Offset Tuning.\nWhile existing prompt tuning approaches have shown great promise for continual\nlearning of image, text, and video modalities; they demand access to\nextensively pretrained transformers. Breaking away from this assumption, POET\ndemonstrates the efficacy of prompt tuning a significantly lightweight\nbackbone, pretrained exclusively on the base class data. We propose a novel\nspatio-temporal learnable prompt offset tuning approach, and are the first to\napply such prompt tuning to Graph Neural Networks. We contribute two new\nbenchmarks for our new problem setting in human action recognition: (i) NTU\nRGB+D dataset for activity recognition, and (ii) SHREC-2017 dataset for hand\ngesture recognition. We find that POET consistently outperforms comprehensive\nbenchmarks. Source code at\nhttps://github.com/humansensinglab/POET-continual-action-recognition.\n", "date": "2025-04-25", "categories": []}
{"id": "2504.16655", "title": "WiFi based Human Fall and Activity Recognition using Transformer based\n  Encoder Decoder and Graph Neural Networks", "abstract": "  Human pose estimation and action recognition have received attention due to\ntheir critical roles in healthcare monitoring, rehabilitation, and assistive\ntechnologies. In this study, we proposed a novel architecture named Transformer\nbased Encoder Decoder Network (TED Net) designed for estimating human skeleton\nposes from WiFi Channel State Information (CSI). TED Net integrates\nconvolutional encoders with transformer based attention mechanisms to capture\nspatiotemporal features from CSI signals. The estimated skeleton poses were\nused as input to a customized Directed Graph Neural Network (DGNN) for action\nrecognition. We validated our model on two datasets: a publicly available multi\nmodal dataset for assessing general pose estimation, and a newly collected\ndataset focused on fall related scenarios involving 20 participants.\nExperimental results demonstrated that TED Net outperformed existing approaches\nin pose estimation, and that the DGNN achieves reliable action classification\nusing CSI based skeletons, with performance comparable to RGB based systems.\nNotably, TED Net maintains robust performance across both fall and non fall\ncases. These findings highlight the potential of CSI driven human skeleton\nestimation for effective action recognition, particularly in home environments\nsuch as elderly fall detection. In such settings, WiFi signals are often\nreadily available, offering a privacy preserving alternative to vision based\nmethods, which may raise concerns about continuous camera monitoring.\n", "date": "2025-04-23", "categories": []}
{"id": "2304.12304", "title": "A Survey on Multi-Resident Activity Recognition in Smart Environments", "abstract": "  Human activity recognition (HAR) is a rapidly growing field that utilizes\nsmart devices, sensors, and algorithms to automatically classify and identify\nthe actions of individuals within a given environment. These systems have a\nwide range of applications, including assisting with caring tasks, increasing\nsecurity, and improving energy efficiency. However, there are several\nchallenges that must be addressed in order to effectively utilize HAR systems\nin multi-resident environments. One of the key challenges is accurately\nassociating sensor observations with the identities of the individuals\ninvolved, which can be particularly difficult when residents are engaging in\ncomplex and collaborative activities. This paper provides a brief overview of\nthe design and implementation of HAR systems, including a summary of the\nvarious data collection devices and approaches used for human activity\nidentification. It also reviews previous research on the use of these systems\nin multi-resident environments and offers conclusions on the current state of\nthe art in the field.\n", "date": "2023-04-24", "categories": []}
{"id": "2504.14860", "title": "Bridge the Gap: From Weak to Full Supervision for Temporal Action\n  Localization with PseudoFormer", "abstract": "  Weakly-supervised Temporal Action Localization (WTAL) has achieved notable\nsuccess but still suffers from a lack of temporal annotations, leading to a\nperformance and framework gap compared with fully-supervised methods. While\nrecent approaches employ pseudo labels for training, three key challenges:\ngenerating high-quality pseudo labels, making full use of different priors, and\noptimizing training methods with noisy labels remain unresolved. Due to these\nperspectives, we propose PseudoFormer, a novel two-branch framework that\nbridges the gap between weakly and fully-supervised Temporal Action\nLocalization (TAL). We first introduce RickerFusion, which maps all predicted\naction proposals to a global shared space to generate pseudo labels with better\nquality. Subsequently, we leverage both snippet-level and proposal-level labels\nwith different priors from the weak branch to train the regression-based model\nin the full branch. Finally, the uncertainty mask and iterative refinement\nmechanism are applied for training with noisy pseudo labels. PseudoFormer\nachieves state-of-the-art WTAL results on the two commonly used benchmarks,\nTHUMOS14 and ActivityNet1.3. Besides, extensive ablation studies demonstrate\nthe contribution of each component of our method.\n", "date": "2025-04-21", "categories": []}
{"id": "2412.18780", "title": "Skeleton-based Action Recognition with Non-linear Dependency Modeling\n  and Hilbert-Schmidt Independence Criterion", "abstract": "  Human skeleton-based action recognition has long been an indispensable aspect\nof artificial intelligence. Current state-of-the-art methods tend to consider\nonly the dependencies between connected skeletal joints, limiting their ability\nto capture non-linear dependencies between physically distant joints. Moreover,\nmost existing approaches distinguish action classes by estimating the\nprobability density of motion representations, yet the high-dimensional nature\nof human motions invokes inherent difficulties in accomplishing such\nmeasurements. In this paper, we seek to tackle these challenges from two\ndirections: (1) We propose a novel dependency refinement approach that\nexplicitly models dependencies between any pair of joints, effectively\ntranscending the limitations imposed by joint distance. (2) We further propose\na framework that utilizes the Hilbert-Schmidt Independence Criterion to\ndifferentiate action classes without being affected by data dimensionality, and\nmathematically derive learning objectives guaranteeing precise recognition.\nEmpirically, our approach sets the state-of-the-art performance on NTU RGB+D,\nNTU RGB+D 120, and Northwestern-UCLA datasets.\n", "date": "2024-12-25", "categories": []}
{"id": "2504.13140", "title": "PCBEAR: Pose Concept Bottleneck for Explainable Action Recognition", "abstract": "  Human action recognition (HAR) has achieved impressive results with deep\nlearning models, but their decision-making process remains opaque due to their\nblack-box nature. Ensuring interpretability is crucial, especially for\nreal-world applications requiring transparency and accountability. Existing\nvideo XAI methods primarily rely on feature attribution or static textual\nconcepts, both of which struggle to capture motion dynamics and temporal\ndependencies essential for action understanding. To address these challenges,\nwe propose Pose Concept Bottleneck for Explainable Action Recognition (PCBEAR),\na novel concept bottleneck framework that introduces human pose sequences as\nmotion-aware, structured concepts for video action recognition. Unlike methods\nbased on pixel-level features or static textual descriptions, PCBEAR leverages\nhuman skeleton poses, which focus solely on body movements, providing robust\nand interpretable explanations of motion dynamics. We define two types of\npose-based concepts: static pose concepts for spatial configurations at\nindividual frames, and dynamic pose concepts for motion patterns across\nmultiple frames. To construct these concepts, PCBEAR applies clustering to\nvideo pose sequences, allowing for automatic discovery of meaningful concepts\nwithout manual annotation. We validate PCBEAR on KTH, Penn-Action, and HAA500,\nshowing that it achieves high classification performance while offering\ninterpretable, motion-driven explanations. Our method provides both strong\npredictive performance and human-understandable insights into the model's\nreasoning process, enabling test-time interventions for debugging and improving\nmodel behavior.\n", "date": "2025-04-17", "categories": []}
{"id": "2309.12029", "title": "Exploring Self-supervised Skeleton-based Action Recognition in Occluded\n  Environments", "abstract": "  To integrate action recognition into autonomous robotic systems, it is\nessential to address challenges such as person occlusions-a common yet often\noverlooked scenario in existing self-supervised skeleton-based action\nrecognition methods. In this work, we propose IosPSTL, a simple and effective\nself-supervised learning framework designed to handle occlusions. IosPSTL\ncombines a cluster-agnostic KNN imputer with an Occluded Partial\nSpatio-Temporal Learning (OPSTL) strategy. First, we pre-train the model on\noccluded skeleton sequences. Then, we introduce a cluster-agnostic KNN imputer\nthat performs semantic grouping using k-means clustering on sequence\nembeddings. It imputes missing skeleton data by applying K-Nearest Neighbors in\nthe latent space, leveraging nearby sample representations to restore occluded\njoints. This imputation generates more complete skeleton sequences, which\nsignificantly benefits downstream self-supervised models. To further enhance\nlearning, the OPSTL module incorporates Adaptive Spatial Masking (ASM) to make\nbetter use of intact, high-quality skeleton sequences during training. Our\nmethod achieves state-of-the-art performance on the occluded versions of the\nNTU-60 and NTU-120 datasets, demonstrating its robustness and effectiveness\nunder challenging conditions. Code is available at\nhttps://github.com/cyfml/OPSTL.\n", "date": "2023-09-21", "categories": []}
{"id": "2504.12021", "title": "Action Anticipation from SoccerNet Football Video Broadcasts", "abstract": "  Artificial intelligence has revolutionized the way we analyze sports videos,\nwhether to understand the actions of games in long untrimmed videos or to\nanticipate the player's motion in future frames. Despite these efforts, little\nattention has been given to anticipating game actions before they occur. In\nthis work, we introduce the task of action anticipation for football broadcast\nvideos, which consists in predicting future actions in unobserved future\nframes, within a five- or ten-second anticipation window. To benchmark this\ntask, we release a new dataset, namely the SoccerNet Ball Action Anticipation\ndataset, based on SoccerNet Ball Action Spotting. Additionally, we propose a\nFootball Action ANticipation TRAnsformer (FAANTRA), a baseline method that\nadapts FUTR, a state-of-the-art action anticipation model, to predict\nball-related actions. To evaluate action anticipation, we introduce new\nmetrics, including mAP@$\\delta$, which evaluates the temporal precision of\npredicted future actions, as well as mAP@$\\infty$, which evaluates their\noccurrence within the anticipation window. We also conduct extensive ablation\nstudies to examine the impact of various task settings, input configurations,\nand model architectures. Experimental results highlight both the feasibility\nand challenges of action anticipation in football videos, providing valuable\ninsights into the design of predictive models for sports analytics. By\nforecasting actions before they unfold, our work will enable applications in\nautomated broadcasting, tactical analysis, and player decision-making. Our\ndataset and code are publicly available at\nhttps://github.com/MohamadDalal/FAANTRA.\n", "date": "2025-04-16", "categories": []}
{"id": "2412.14719", "title": "Prototypical Calibrating Ambiguous Samples for Micro-Action Recognition", "abstract": "  Micro-Action Recognition (MAR) has gained increasing attention due to its\ncrucial role as a form of non-verbal communication in social interactions, with\npromising potential for applications in human communication and emotion\nanalysis. However, current approaches often overlook the inherent ambiguity in\nmicro-actions, which arises from the wide category range and subtle visual\ndifferences between categories. This oversight hampers the accuracy of\nmicro-action recognition. In this paper, we propose a novel Prototypical\nCalibrating Ambiguous Network (PCAN) to unleash and mitigate the ambiguity of\nMAR. Firstly, we employ a hierarchical action-tree to identify the ambiguous\nsample, categorizing them into distinct sets of ambiguous samples of false\nnegatives and false positives, considering both body- and action-level\ncategories. Secondly, we implement an ambiguous contrastive refinement module\nto calibrate these ambiguous samples by regulating the distance between\nambiguous samples and their corresponding prototypes. This calibration process\naims to pull false negative (FN) samples closer to their respective prototypes\nand push false positive (FP) samples apart from their affiliated prototypes. In\naddition, we propose a new prototypical diversity amplification loss to\nstrengthen the model's capacity by amplifying the differences between different\nprototypes. Finally, we propose a prototype-guided rectification to rectify\nprediction by incorporating the representability of prototypes. Extensive\nexperiments conducted on the benchmark dataset demonstrate the superior\nperformance of our method compared to existing approaches. The code is\navailable at https://github.com/kunli-cs/PCAN.\n", "date": "2024-12-19", "categories": []}
{"id": "2504.08222", "title": "F$^3$Set: Towards Analyzing Fast, Frequent, and Fine-grained Events from\n  Videos", "abstract": "  Analyzing Fast, Frequent, and Fine-grained (F$^3$) events presents a\nsignificant challenge in video analytics and multi-modal LLMs. Current methods\nstruggle to identify events that satisfy all the F$^3$ criteria with high\naccuracy due to challenges such as motion blur and subtle visual discrepancies.\nTo advance research in video understanding, we introduce F$^3$Set, a benchmark\nthat consists of video datasets for precise F$^3$ event detection. Datasets in\nF$^3$Set are characterized by their extensive scale and comprehensive detail,\nusually encompassing over 1,000 event types with precise timestamps and\nsupporting multi-level granularity. Currently, F$^3$Set contains several sports\ndatasets, and this framework may be extended to other applications as well. We\nevaluated popular temporal action understanding methods on F$^3$Set, revealing\nsubstantial challenges for existing techniques. Additionally, we propose a new\nmethod, F$^3$ED, for F$^3$ event detections, achieving superior performance.\nThe dataset, model, and benchmark code are available at\nhttps://github.com/F3Set/F3Set.\n", "date": "2025-04-11", "categories": []}
{"id": "2504.07785", "title": "Towards Micro-Action Recognition with Limited Annotations: An\n  Asynchronous Pseudo Labeling and Training Approach", "abstract": "  Micro-Action Recognition (MAR) aims to classify subtle human actions in\nvideo. However, annotating MAR datasets is particularly challenging due to the\nsubtlety of actions. To this end, we introduce the setting of Semi-Supervised\nMAR (SSMAR), where only a part of samples are labeled. We first evaluate\ntraditional Semi-Supervised Learning (SSL) methods to SSMAR and find that these\nmethods tend to overfit on inaccurate pseudo-labels, leading to error\naccumulation and degraded performance. This issue primarily arises from the\ncommon practice of directly using the predictions of classifier as\npseudo-labels to train the model. To solve this issue, we propose a novel\nframework, called Asynchronous Pseudo Labeling and Training (APLT), which\nexplicitly separates the pseudo-labeling process from model training.\nSpecifically, we introduce a semi-supervised clustering method during the\noffline pseudo-labeling phase to generate more accurate pseudo-labels.\nMoreover, a self-adaptive thresholding strategy is proposed to dynamically\nfilter noisy labels of different classes. We then build a memory-based\nprototype classifier based on the filtered pseudo-labels, which is fixed and\nused to guide the subsequent model training phase. By alternating the two\npseudo-labeling and model training phases in an asynchronous manner, the model\ncan not only be learned with more accurate pseudo-labels but also avoid the\noverfitting issue. Experiments on three MAR datasets show that our APLT largely\noutperforms state-of-the-art SSL methods. For instance, APLT improves accuracy\nby 14.5\\% over FixMatch on the MA-12 dataset when using only 50\\% labeled data.\nCode will be publicly available.\n", "date": "2025-04-10", "categories": []}
{"id": "2504.07916", "title": "Semantically Encoding Activity Labels for Context-Aware Human Activity\n  Recognition", "abstract": "  Prior work has primarily formulated CA-HAR as a multi-label classification\nproblem, where model inputs are time-series sensor data and target labels are\nbinary encodings representing whether a given activity or context occurs. These\nCA-HAR methods either predicted each label independently or manually imposed\nrelationships using graphs. However, both strategies often neglect an essential\naspect: activity labels have rich semantic relationships. For instance,\nwalking, jogging, and running activities share similar movement patterns but\ndiffer in pace and intensity, indicating that they are semantically related.\nConsequently, prior CA-HAR methods often struggled to accurately capture these\ninherent and nuanced relationships, particularly on datasets with noisy labels\ntypically used for CA-HAR or situations where the ideal sensor type is\nunavailable (e.g., recognizing speech without audio sensors). To address this\nlimitation, we propose SEAL, which leverage LMs to encode CA-HAR activity\nlabels to capture semantic relationships. LMs generate vector embeddings that\npreserve rich semantic information from natural language. Our SEAL approach\nencodes input-time series sensor data from smart devices and their associated\nactivity and context labels (text) as vector embeddings. During training, SEAL\naligns the sensor data representations with their corresponding\nactivity/context label embeddings in a shared embedding space. At inference\ntime, SEAL performs a similarity search, returning the CA-HAR label with the\nembedding representation closest to the input data. Although LMs have been\nwidely explored in other domains, surprisingly, their potential in CA-HAR has\nbeen underexplored, making our approach a novel contribution to the field. Our\nresearch opens up new possibilities for integrating more advanced LMs into\nCA-HAR tasks.\n", "date": "2025-04-10", "categories": []}
{"id": "2504.05700", "title": "Pose-Aware Weakly-Supervised Action Segmentation", "abstract": "  Understanding human behavior is an important problem in the pursuit of visual\nintelligence. A challenge in this endeavor is the extensive and costly effort\nrequired to accurately label action segments. To address this issue, we\nconsider learning methods that demand minimal supervision for segmentation of\nhuman actions in long instructional videos. Specifically, we introduce a\nweakly-supervised framework that uniquely incorporates pose knowledge during\ntraining while omitting its use during inference, thereby distilling pose\nknowledge pertinent to each action component. We propose a pose-inspired\ncontrastive loss as a part of the whole weakly-supervised framework which is\ntrained to distinguish action boundaries more effectively. Our approach,\nvalidated through extensive experiments on representative datasets, outperforms\nprevious state-of-the-art (SOTA) in segmenting long instructional videos under\nboth online and offline settings. Additionally, we demonstrate the framework's\nadaptability to various segmentation backbones and pose extractors across\ndifferent datasets.\n", "date": "2025-04-08", "categories": []}
{"id": "2504.05830", "title": "Human Activity Recognition using RGB-Event based Sensors: A Multi-modal\n  Heat Conduction Model and A Benchmark Dataset", "abstract": "  Human Activity Recognition (HAR) primarily relied on traditional RGB cameras\nto achieve high-performance activity recognition. However, the challenging\nfactors in real-world scenarios, such as insufficient lighting and rapid\nmovements, inevitably degrade the performance of RGB cameras. To address these\nchallenges, biologically inspired event cameras offer a promising solution to\novercome the limitations of traditional RGB cameras. In this work, we rethink\nhuman activity recognition by combining the RGB and event cameras. The first\ncontribution is the proposed large-scale multi-modal RGB-Event human activity\nrecognition benchmark dataset, termed HARDVS 2.0, which bridges the dataset\ngaps. It contains 300 categories of everyday real-world actions with a total of\n107,646 paired videos covering various challenging scenarios. Inspired by the\nphysics-informed heat conduction model, we propose a novel multi-modal heat\nconduction operation framework for effective activity recognition, termed\nMMHCO-HAR. More in detail, given the RGB frames and event streams, we first\nextract the feature embeddings using a stem network. Then, multi-modal Heat\nConduction blocks are designed to fuse the dual features, the key module of\nwhich is the multi-modal Heat Conduction Operation layer. We integrate RGB and\nevent embeddings through a multi-modal DCT-IDCT layer while adaptively\nincorporating the thermal conductivity coefficient via FVEs into this module.\nAfter that, we propose an adaptive fusion module based on a policy routing\nstrategy for high-performance classification. Comprehensive experiments\ndemonstrate that our method consistently performs well, validating its\neffectiveness and robustness. The source code and benchmark dataset will be\nreleased on https://github.com/Event-AHU/HARDVS/tree/HARDVSv2\n", "date": "2025-04-08", "categories": []}
{"id": "2504.05956", "title": "Temporal Alignment-Free Video Matching for Few-shot Action Recognition", "abstract": "  Few-Shot Action Recognition (FSAR) aims to train a model with only a few\nlabeled video instances. A key challenge in FSAR is handling divergent\nnarrative trajectories for precise video matching. While the frame- and\ntuple-level alignment approaches have been promising, their methods heavily\nrely on pre-defined and length-dependent alignment units (e.g., frames or\ntuples), which limits flexibility for actions of varying lengths and speeds. In\nthis work, we introduce a novel TEmporal Alignment-free Matching (TEAM)\napproach, which eliminates the need for temporal units in action representation\nand brute-force alignment during matching. Specifically, TEAM represents each\nvideo with a fixed set of pattern tokens that capture globally discriminative\nclues within the video instance regardless of action length or speed, ensuring\nits flexibility. Furthermore, TEAM is inherently efficient, using token-wise\ncomparisons to measure similarity between videos, unlike existing methods that\nrely on pairwise comparisons for temporal alignment. Additionally, we propose\nan adaptation process that identifies and removes common information across\nclasses, establishing clear boundaries even between novel categories. Extensive\nexperiments demonstrate the effectiveness of TEAM. Codes are available at\ngithub.com/leesb7426/TEAM.\n", "date": "2025-04-08", "categories": []}
{"id": "2408.17422", "title": "Open-Vocabulary Action Localization with Iterative Visual Prompting", "abstract": "  Video action localization aims to find the timings of specific actions from a\nlong video. Although existing learning-based approaches have been successful,\nthey require annotating videos, which comes with a considerable labor cost.\nThis paper proposes a training-free, open-vocabulary approach based on emerging\noff-the-shelf vision-language models (VLMs). The challenge stems from the fact\nthat VLMs are neither designed to process long videos nor tailored for finding\nactions. We overcome these problems by extending an iterative visual prompting\ntechnique. Specifically, we sample video frames and create a concatenated image\nwith frame index labels, allowing a VLM to identify the frames that most likely\ncorrespond to the start and end of the action. By iteratively narrowing the\nsampling window around the selected frames, the estimation gradually converges\nto more precise temporal boundaries. We demonstrate that this technique yields\nreasonable performance, achieving results comparable to state-of-the-art\nzero-shot action localization. These results support the use of VLMs as a\npractical tool for understanding videos. Sample code is available at\nhttps://microsoft.github.io/VLM-Video-Action-Localization/\n", "date": "2024-08-30", "categories": []}
{"id": "2504.01890", "title": "Is Temporal Prompting All We Need For Limited Labeled Action\n  Recognition?", "abstract": "  Video understanding has shown remarkable improvements in recent years,\nlargely dependent on the availability of large scaled labeled datasets. Recent\nadvancements in visual-language models, especially based on contrastive\npretraining, have shown remarkable generalization in zero-shot tasks, helping\nto overcome this dependence on labeled datasets. Adaptations of such models for\nvideos, typically involve modifying the architecture of vision-language models\nto cater to video data. However, this is not trivial, since such adaptations\nare mostly computationally intensive and struggle with temporal modeling. We\npresent TP-CLIP, an adaptation of CLIP that leverages temporal visual prompting\nfor temporal adaptation without modifying the core CLIP architecture. This\npreserves its generalization abilities. TP-CLIP efficiently integrates into the\nCLIP architecture, leveraging its pre-trained capabilities for video data.\nExtensive experiments across various datasets demonstrate its efficacy in\nzero-shot and few-shot learning, outperforming existing approaches with fewer\nparameters and computational efficiency. In particular, we use just 1/3 the\nGFLOPs and 1/28 the number of tuneable parameters in comparison to recent\nstate-of-the-art and still outperform it by up to 15.8% depending on the task\nand dataset.\n", "date": "2025-04-02", "categories": []}
{"id": "2504.02279", "title": "MultiTSF: Transformer-based Sensor Fusion for Human-Centric Multi-view\n  and Multi-modal Action Recognition", "abstract": "  Action recognition from multi-modal and multi-view observations holds\nsignificant potential for applications in surveillance, robotics, and smart\nenvironments. However, existing methods often fall short of addressing\nreal-world challenges such as diverse environmental conditions, strict sensor\nsynchronization, and the need for fine-grained annotations. In this study, we\npropose the Multi-modal Multi-view Transformer-based Sensor Fusion (MultiTSF).\nThe proposed method leverages a Transformer-based to dynamically model\ninter-view relationships and capture temporal dependencies across multiple\nviews. Additionally, we introduce a Human Detection Module to generate\npseudo-ground-truth labels, enabling the model to prioritize frames containing\nhuman activity and enhance spatial feature learning. Comprehensive experiments\nconducted on our in-house MultiSensor-Home dataset and the existing MM-Office\ndataset demonstrate that MultiTSF outperforms state-of-the-art methods in both\nvideo sequence-level and frame-level action recognition settings.\n", "date": "2025-04-03", "categories": []}
{"id": "2504.03687", "title": "Process Optimization and Deployment for Sensor-Based Human Activity\n  Recognition Based on Deep Learning", "abstract": "  Sensor-based human activity recognition is a key technology for many\nhuman-centered intelligent applications. However, this research is still in its\ninfancy and faces many unresolved challenges. To address these, we propose a\ncomprehensive optimization process approach centered on multi-attention\ninteraction. We first utilize unsupervised statistical feature-guided diffusion\nmodels for highly adaptive data enhancement, and introduce a novel network\narchitecture-Multi-branch Spatiotemporal Interaction Network, which uses\nmulti-branch features at different levels to effectively Sequential ), which\nuses multi-branch features at different levels to effectively Sequential\nspatio-temporal interaction to enhance the ability to mine advanced latent\nfeatures. In addition, we adopt a multi-loss function fusion strategy in the\ntraining phase to dynamically adjust the fusion weights between batches to\noptimize the training results. Finally, we also conducted actual deployment on\nembedded devices to extensively test the practical feasibility of the proposed\nmethod in existing work. We conduct extensive testing on three public datasets,\nincluding ablation studies, comparisons of related work, and embedded\ndeployments.\n", "date": "2025-03-22", "categories": []}
{"id": "2411.19041", "title": "TAMT: Temporal-Aware Model Tuning for Cross-Domain Few-Shot Action\n  Recognition", "abstract": "  Going beyond few-shot action recognition (FSAR), cross-domain FSAR (CDFSAR)\nhas attracted recent research interests by solving the domain gap lying in\nsource-to-target transfer learning. Existing CDFSAR methods mainly focus on\njoint training of source and target data to mitigate the side effect of domain\ngap. However, such kind of methods suffer from two limitations: First,\npair-wise joint training requires retraining deep models in case of one source\ndata and multiple target ones, which incurs heavy computation cost, especially\nfor large source and small target data. Second, pre-trained models after joint\ntraining are adopted to target domain in a straightforward manner, hardly\ntaking full potential of pre-trained models and then limiting recognition\nperformance. To overcome above limitations, this paper proposes a simple yet\neffective baseline, namely Temporal-Aware Model Tuning (TAMT) for CDFSAR.\nSpecifically, our TAMT involves a decoupled paradigm by performing pre-training\non source data and fine-tuning target data, which avoids retraining for\nmultiple target data with single source. To effectively and efficiently explore\nthe potential of pre-trained models in transferring to target domain, our TAMT\nproposes a Hierarchical Temporal Tuning Network (HTTN), whose core involves\nlocal temporal-aware adapters (TAA) and a global temporal-aware moment tuning\n(GTMT). Particularly, TAA learns few parameters to recalibrate the intermediate\nfeatures of frozen pre-trained models, enabling efficient adaptation to target\ndomains. Furthermore, GTMT helps to generate powerful video representations,\nimproving match performance on the target domain. Experiments on several widely\nused video benchmarks show our TAMT outperforms the recently proposed\ncounterparts by 13%$\\sim$31%, achieving new state-of-the-art CDFSAR results.\n", "date": "2024-11-28", "categories": []}
{"id": "2504.02512", "title": "Towards Generalizing Temporal Action Segmentation to Unseen Views", "abstract": "  While there has been substantial progress in temporal action segmentation,\nthe challenge to generalize to unseen views remains unaddressed. Hence, we\ndefine a protocol for unseen view action segmentation where camera views for\nevaluating the model are unavailable during training. This includes changing\nfrom top-frontal views to a side view or even more challenging from exocentric\nto egocentric views. Furthermore, we present an approach for temporal action\nsegmentation that tackles this challenge. Our approach leverages a shared\nrepresentation at both the sequence and segment levels to reduce the impact of\nview differences during training. We achieve this by introducing a sequence\nloss and an action loss, which together facilitate consistent video and action\nrepresentations across different views. The evaluation on the Assembly101,\nIkeaASM, and EgoExoLearn datasets demonstrate significant improvements, with a\n12.8% increase in F1@50 for unseen exocentric views and a substantial 54%\nimprovement for unseen egocentric views.\n", "date": "2025-04-03", "categories": []}
{"id": "2504.02778", "title": "Multi-Head Adaptive Graph Convolution Network for Sparse Point\n  Cloud-Based Human Activity Recognition", "abstract": "  Human activity recognition is increasingly vital for supporting independent\nliving, particularly for the elderly and those in need of assistance. Domestic\nservice robots with monitoring capabilities can enhance safety and provide\nessential support. Although image-based methods have advanced considerably in\nthe past decade, their adoption remains limited by concerns over privacy and\nsensitivity to low-light or dark conditions. As an alternative, millimetre-wave\n(mmWave) radar can produce point cloud data which is privacy-preserving.\nHowever, processing the sparse and noisy point clouds remains a long-standing\nchallenge. While graph-based methods and attention mechanisms show promise,\nthey predominantly rely on \"fixed\" kernels; kernels that are applied uniformly\nacross all neighbourhoods, highlighting the need for adaptive approaches that\ncan dynamically adjust their kernels to the specific geometry of each local\nneighbourhood in point cloud data. To overcome this limitation, we introduce an\nadaptive approach within the graph convolutional framework. Instead of a single\nshared weight function, our Multi-Head Adaptive Kernel (MAK) module generates\nmultiple dynamic kernels, each capturing different aspects of the local feature\nspace. By progressively refining local features while maintaining global\nspatial context, our method enables convolution kernels to adapt to varying\nlocal features. Experimental results on benchmark datasets confirm the\neffectiveness of our approach, achieving state-of-the-art performance in human\nactivity recognition. Our source code is made publicly available at:\nhttps://github.com/Gbouna/MAK-GCN\n", "date": "2025-04-03", "categories": []}
{"id": "2504.00149", "title": "Towards Precise Action Spotting: Addressing Temporal Misalignment in\n  Labels with Dynamic Label Assignment", "abstract": "  Precise action spotting has attracted considerable attention due to its\npromising applications. While existing methods achieve substantial performance\nby employing well-designed model architecture, they overlook a significant\nchallenge: the temporal misalignment inherent in ground-truth labels. This\nmisalignment arises when frames labeled as containing events do not align\naccurately with the actual event times, often as a result of human annotation\nerrors or the inherent difficulties in precisely identifying event boundaries\nacross neighboring frames. To tackle this issue, we propose a novel dynamic\nlabel assignment strategy that allows predictions to have temporal offsets from\nground-truth action times during training, ensuring consistent event spotting.\nOur method extends the concept of minimum-cost matching, which is utilized in\nthe spatial domain for object detection, to the temporal domain. By calculating\nmatching costs based on predicted action class scores and temporal offsets, our\nmethod dynamically assigns labels to the most likely predictions, even when the\npredicted times of these predictions deviate from ground-truth times,\nalleviating the negative effects of temporal misalignment in labels. We conduct\nextensive experiments and demonstrate that our method achieves state-of-the-art\nperformance, particularly in conditions where events are visually distinct and\ntemporal misalignment in labels is common.\n", "date": "2025-03-31", "categories": []}
{"id": "2504.00647", "title": "FDDet: Frequency-Decoupling for Boundary Refinement in Temporal Action\n  Detection", "abstract": "  Temporal action detection aims to locate and classify actions in untrimmed\nvideos. While recent works focus on designing powerful feature processors for\npre-trained representations, they often overlook the inherent noise and\nredundancy within these features. Large-scale pre-trained video encoders tend\nto introduce background clutter and irrelevant semantics, leading to context\nconfusion and imprecise boundaries. To address this, we propose a\nfrequency-aware decoupling network that improves action discriminability by\nfiltering out noisy semantics captured by pre-trained models. Specifically, we\nintroduce an adaptive temporal decoupling scheme that suppresses irrelevant\ninformation while preserving fine-grained atomic action details, yielding more\ntask-specific representations. In addition, we enhance inter-frame modeling by\ncapturing temporal variations to better distinguish actions from background\nredundancy. Furthermore, we present a long-short-term category-aware relation\nnetwork that jointly models local transitions and long-range dependencies,\nimproving localization precision. The refined atomic features and\nfrequency-guided dynamics are fed into a standard detection head to produce\naccurate action predictions. Extensive experiments on THUMOS14, HACS, and\nActivityNet-1.3 show that our method, powered by InternVideo2-6B features,\nachieves state-of-the-art performance on temporal action detection benchmarks.\n", "date": "2025-04-01", "categories": []}
{"id": "2503.23214", "title": "Action Recognition in Real-World Ambient Assisted Living Environment", "abstract": "  The growing ageing population and their preference to maintain independence\nby living in their own homes require proactive strategies to ensure safety and\nsupport. Ambient Assisted Living (AAL) technologies have emerged to facilitate\nageing in place by offering continuous monitoring and assistance within the\nhome. Within AAL technologies, action recognition plays a crucial role in\ninterpreting human activities and detecting incidents like falls, mobility\ndecline, or unusual behaviours that may signal worsening health conditions.\nHowever, action recognition in practical AAL applications presents challenges,\nincluding occlusions, noisy data, and the need for real-time performance. While\nadvancements have been made in accuracy, robustness to noise, and computation\nefficiency, achieving a balance among them all remains a challenge. To address\nthis challenge, this paper introduces the Robust and Efficient Temporal\nConvolution network (RE-TCN), which comprises three main elements: Adaptive\nTemporal Weighting (ATW), Depthwise Separable Convolutions (DSC), and data\naugmentation techniques. These elements aim to enhance the model's accuracy,\nrobustness against noise and occlusion, and computational efficiency within\nreal-world AAL contexts. RE-TCN outperforms existing models in terms of\naccuracy, noise and occlusion robustness, and has been validated on four\nbenchmark datasets: NTU RGB+D 60, Northwestern-UCLA, SHREC'17, and DHG-14/28.\nThe code is publicly available at: https://github.com/Gbouna/RE-TCN\n", "date": "2025-03-29", "categories": []}
{"id": "2503.23266", "title": "OwlSight: A Robust Illumination Adaptation Framework for Dark Video\n  Human Action Recognition", "abstract": "  Human action recognition in low-light environments is crucial for various\nreal-world applications. However, the existing approaches overlook the full\nutilization of brightness information throughout the training phase, leading to\nsuboptimal performance. To address this limitation, we propose OwlSight, a\nbiomimetic-inspired framework with whole-stage illumination enhancement to\ninteract with action classification for accurate dark video human action\nrecognition. Specifically, OwlSight incorporates a Time-Consistency Module\n(TCM) to capture shallow spatiotemporal features meanwhile maintaining temporal\ncoherence, which are then processed by a Luminance Adaptation Module (LAM) to\ndynamically adjust the brightness based on the input luminance distribution.\nFurthermore, a Reflect Augmentation Module (RAM) is presented to maximize\nillumination utilization and simultaneously enhance action recognition via two\ninteractive paths. Additionally, we build Dark-101, a large-scale dataset\ncomprising 18,310 dark videos across 101 action categories, significantly\nsurpassing existing datasets (e.g., ARID1.5 and Dark-48) in scale and\ndiversity. Extensive experiments demonstrate that the proposed OwlSight\nachieves state-of-the-art performance across four low-light action recognition\nbenchmarks. Notably, it outperforms previous best approaches by 5.36% on\nARID1.5 and 1.72% on Dark-101, highlighting its effectiveness in challenging\ndark environments.\n", "date": "2025-03-30", "categories": []}
{"id": "2407.14744", "title": "A Comprehensive Review of Few-shot Action Recognition", "abstract": "  Few-shot action recognition aims to address the high cost and impracticality\nof manually labeling complex and variable video data in action recognition. It\nrequires accurately classifying human actions in videos using only a few\nlabeled examples per class. Compared to few-shot learning in image scenarios,\nfew-shot action recognition is more challenging due to the intrinsic complexity\nof video data. Numerous approaches have driven significant advancements in\nfew-shot action recognition, which underscores the need for a comprehensive\nsurvey. Unlike early surveys that focus on few-shot image or text\nclassification, we deeply consider the unique challenges of few-shot action\nrecognition. In this survey, we provide a comprehensive review of recent\nmethods and introduce a novel and systematic taxonomy of existing approaches,\naccompanied by a detailed analysis. We categorize the methods into\ngenerative-based and meta-learning frameworks, and further elaborate on the\nmethods within the meta-learning framework, covering aspects: video instance\nrepresentation, category prototype learning, and generalized video alignment.\nAdditionally, the survey presents the commonly used benchmarks and discusses\nrelevant advanced topics and promising future directions. We hope this survey\ncan serve as a valuable resource for researchers, offering essential guidance\nto newcomers and stimulating seasoned researchers with fresh insights.\n", "date": "2024-07-20", "categories": []}
{"id": "2501.12086", "title": "DSTSA-GCN: Advancing Skeleton-Based Gesture Recognition with\n  Semantic-Aware Spatio-Temporal Topology Modeling", "abstract": "  Graph convolutional networks (GCNs) have emerged as a powerful tool for\nskeleton-based action and gesture recognition, thanks to their ability to model\nspatial and temporal dependencies in skeleton data. However, existing GCN-based\nmethods face critical limitations: (1) they lack effective spatio-temporal\ntopology modeling that captures dynamic variations in skeletal motion, and (2)\nthey struggle to model multiscale structural relationships beyond local joint\nconnectivity. To address these issues, we propose a novel framework called\nDynamic Spatial-Temporal Semantic Awareness Graph Convolutional Network\n(DSTSA-GCN). DSTSA-GCN introduces three key modules: Group Channel-wise Graph\nConvolution (GC-GC), Group Temporal-wise Graph Convolution (GT-GC), and\nMulti-Scale Temporal Convolution (MS-TCN). GC-GC and GT-GC operate in parallel\nto independently model channel-specific and frame-specific correlations,\nenabling robust topology learning that accounts for temporal variations.\nAdditionally, both modules employ a grouping strategy to adaptively capture\nmultiscale structural relationships. Complementing this, MS-TCN enhances\ntemporal modeling through group-wise temporal convolutions with diverse\nreceptive fields. Extensive experiments demonstrate that DSTSA-GCN\nsignificantly improves the topology modeling capabilities of GCNs, achieving\nstate-of-the-art performance on benchmark datasets for gesture and action\nrecognition, including SHREC17 Track, DHG-14\\/28, NTU-RGB+D, and NTU-RGB+D-120.\n", "date": "2025-01-21", "categories": []}
{"id": "2503.20209", "title": "BEAR: A Video Dataset For Fine-grained Behaviors Recognition Oriented\n  with Action and Environment Factors", "abstract": "  Behavior recognition is an important task in video representation learning.\nAn essential aspect pertains to effective feature learning conducive to\nbehavior recognition. Recently, researchers have started to study fine-grained\nbehavior recognition, which provides similar behaviors and encourages the model\nto concern with more details of behaviors with effective features for\ndistinction. However, previous fine-grained behaviors limited themselves to\ncontrolling partial information to be similar, leading to an unfair and not\ncomprehensive evaluation of existing works. In this work, we develop a new\nvideo fine-grained behavior dataset, named BEAR, which provides fine-grained\n(i.e. similar) behaviors that uniquely focus on two primary factors defining\nbehavior: Environment and Action. It includes two fine-grained behavior\nprotocols including Fine-grained Behavior with Similar Environments and\nFine-grained Behavior with Similar Actions as well as multiple sub-protocols as\ndifferent scenarios. Furthermore, with this new dataset, we conduct multiple\nexperiments with different behavior recognition models. Our research primarily\nexplores the impact of input modality, a critical element in studying the\nenvironmental and action-based aspects of behavior recognition. Our\nexperimental results yield intriguing insights that have substantial\nimplications for further research endeavors.\n", "date": "2025-03-26", "categories": []}
{"id": "2503.20348", "title": "VideoGEM: Training-free Action Grounding in Videos", "abstract": "  Vision-language foundation models have shown impressive capabilities across\nvarious zero-shot tasks, including training-free localization and grounding,\nprimarily focusing on localizing objects in images. However, leveraging those\ncapabilities to localize actions and events in videos is challenging, as\nactions have less physical outline and are usually described by higher-level\nconcepts. In this work, we propose VideoGEM, the first training-free spatial\naction grounding method based on pretrained image- and video-language\nbackbones. Namely, we adapt the self-self attention formulation of GEM to\nspatial activity grounding. We observe that high-level semantic concepts, such\nas actions, usually emerge in the higher layers of the image- and\nvideo-language models. We, therefore, propose a layer weighting in the\nself-attention path to prioritize higher layers. Additionally, we introduce a\ndynamic weighting method to automatically tune layer weights to capture each\nlayer`s relevance to a specific prompt. Finally, we introduce a prompt\ndecomposition, processing action, verb, and object prompts separately,\nresulting in a better spatial localization of actions. We evaluate the proposed\napproach on three image- and video-language backbones, CLIP, OpenCLIP, and\nViCLIP, and on four video grounding datasets, V-HICO, DALY,\nYouCook-Interactions, and GroundingYouTube, showing that the proposed\ntraining-free approach is able to outperform current trained state-of-the-art\napproaches for spatial video grounding.\n", "date": "2025-03-26", "categories": []}
{"id": "2503.20781", "title": "BASKET: A Large-Scale Video Dataset for Fine-Grained Skill Estimation", "abstract": "  We present BASKET, a large-scale basketball video dataset for fine-grained\nskill estimation. BASKET contains 4,477 hours of video capturing 32,232\nbasketball players from all over the world. Compared to prior skill estimation\ndatasets, our dataset includes a massive number of skilled participants with\nunprecedented diversity in terms of gender, age, skill level, geographical\nlocation, etc. BASKET includes 20 fine-grained basketball skills, challenging\nmodern video recognition models to capture the intricate nuances of player\nskill through in-depth video analysis. Given a long highlight video (8-10\nminutes) of a particular player, the model needs to predict the skill level\n(e.g., excellent, good, average, fair, poor) for each of the 20 basketball\nskills. Our empirical analysis reveals that the current state-of-the-art video\nmodels struggle with this task, significantly lagging behind the human\nbaseline. We believe that BASKET could be a useful resource for developing new\nvideo models with advanced long-range, fine-grained recognition capabilities.\nIn addition, we hope that our dataset will be useful for domain-specific\napplications such as fair basketball scouting, personalized player development,\nand many others. Dataset and code are available at\nhttps://github.com/yulupan00/BASKET.\n", "date": "2025-03-26", "categories": []}
{"id": "2410.14340", "title": "Zero-shot Action Localization via the Confidence of Large\n  Vision-Language Models", "abstract": "  Precise action localization in untrimmed video is vital for fields such as\nprofessional sports and minimally invasive surgery, where the delineation of\nparticular motions in recordings can dramatically enhance analysis. But in many\ncases, large scale datasets with video-label pairs for localization are\nunavailable, limiting the opportunity to fine-tune video-understanding models.\nRecent developments in large vision-language models (LVLM) address this need\nwith impressive zero-shot capabilities in a variety of video understanding\ntasks. However, the adaptation of LVLMs, with their powerful visual question\nanswering capabilities, to zero-shot localization in long-form video is still\nrelatively unexplored. To this end, we introduce a true Zero-shot Action\nLocalization method (ZEAL). Specifically, we leverage the built-in action\nknowledge of a large language model (LLM) to inflate actions into detailed\ndescriptions of the archetypal start and end of the action. These descriptions\nserve as queries to LVLM for generating frame-level confidence scores which can\nbe aggregated to produce localization outputs. The simplicity and flexibility\nof our method lends it amenable to more capable LVLMs as they are developed,\nand we demonstrate remarkable results in zero-shot action localization on a\nchallenging benchmark, without any training. Our code is publicly available at\n$\\href{https://github.com/josaklil-ai/zeal}{github.com/josaklil-ai/zeal}$.\n", "date": "2024-10-18", "categories": []}
{"id": "2503.18957", "title": "A Real-Time Human Action Recognition Model for Assisted Living", "abstract": "  Ensuring the safety and well-being of elderly and vulnerable populations in\nassisted living environments is a critical concern. Computer vision presents an\ninnovative and powerful approach to predicting health risks through video\nmonitoring, employing human action recognition (HAR) technology. However,\nreal-time prediction of human actions with high performance and efficiency is a\nchallenge. This research proposes a real-time human action recognition model\nthat combines a deep learning model and a live video prediction and alert\nsystem, in order to predict falls, staggering and chest pain for residents in\nassisted living. Six thousand RGB video samples from the NTU RGB+D 60 dataset\nwere selected to create a dataset with four classes: Falling, Staggering, Chest\nPain, and Normal, with the Normal class comprising 40 daily activities.\nTransfer learning technique was applied to train four state-of-the-art HAR\nmodels on a GPU server, namely, UniFormerV2, TimeSformer, I3D, and SlowFast.\nResults of the four models are presented in this paper based on class-wise and\nmacro performance metrics, inference efficiency, model complexity and\ncomputational costs. TimeSformer is proposed for developing the real-time human\naction recognition model, leveraging its leading macro F1 score (95.33%),\nrecall (95.49%), and precision (95.19%) along with significantly higher\ninference throughput compared to the others. This research provides insights to\nenhance safety and health of the elderly and people with chronic illnesses in\nassisted living environments, fostering sustainable care, smarter communities\nand industry innovation.\n", "date": "2025-03-18", "categories": []}
{"id": "2503.17978", "title": "PIM: Physics-Informed Multi-task Pre-training for Improving Inertial\n  Sensor-Based Human Activity Recognition", "abstract": "  Human activity recognition (HAR) with deep learning models relies on large\namounts of labeled data, often challenging to obtain due to associated cost,\ntime, and labor. Self-supervised learning (SSL) has emerged as an effective\napproach to leverage unlabeled data through pretext tasks, such as masked\nreconstruction and multitask learning with signal processing-based data\naugmentations, to pre-train encoder models. However, such methods are often\nderived from computer vision approaches that disregard physical mechanisms and\nconstraints that govern wearable sensor data and the phenomena they reflect. In\nthis paper, we propose a physics-informed multi-task pre-training (PIM)\nframework for IMU-based HAR. PIM generates pre-text tasks based on the\nunderstanding of basic physical aspects of human motion: including movement\nspeed, angles of movement, and symmetry between sensor placements. Given a\nsensor signal, we calculate corresponding features using physics-based\nequations and use them as pretext tasks for SSL. This enables the model to\ncapture fundamental physical characteristics of human activities, which is\nespecially relevant for multi-sensor systems. Experimental evaluations on four\nHAR benchmark datasets demonstrate that the proposed method outperforms\nexisting state-of-the-art methods, including data augmentation and masked\nreconstruction, in terms of accuracy and F1 score. We have observed gains of\nalmost 10\\% in macro f1 score and accuracy with only 2 to 8 labeled examples\nper class and up to 3% when there is no reduction in the amount of training\ndata.\n", "date": "2025-03-23", "categories": []}
{"id": "2503.18358", "title": "Cost-Sensitive Learning for Long-Tailed Temporal Action Segmentation", "abstract": "  Temporal action segmentation in untrimmed procedural videos aims to densely\nlabel frames into action classes. These videos inherently exhibit long-tailed\ndistributions, where actions vary widely in frequency and duration. In temporal\naction segmentation approaches, we identified a bi-level learning bias. This\nbias encompasses (1) a class-level bias, stemming from class imbalance favoring\nhead classes, and (2) a transition-level bias arising from variations in\ntransitions, prioritizing commonly observed transitions. As a remedy, we\nintroduce a constrained optimization problem to alleviate both biases. We\ndefine learning states for action classes and their associated transitions and\nintegrate them into the optimization process. We propose a novel cost-sensitive\nloss function formulated as a weighted cross-entropy loss, with weights\nadaptively adjusted based on the learning state of actions and their\ntransitions. Experiments on three challenging temporal segmentation benchmarks\nand various frameworks demonstrate the effectiveness of our approach, resulting\nin significant improvements in both per-class frame-wise and segment-wise\nperformance.\n", "date": "2025-03-24", "categories": []}
{"id": "2503.18359", "title": "Context-Enhanced Memory-Refined Transformer for Online Action Detection", "abstract": "  Online Action Detection (OAD) detects actions in streaming videos using past\nobservations. State-of-the-art OAD approaches model past observations and their\ninteractions with an anticipated future. The past is encoded using short- and\nlong-term memories to capture immediate and long-range dependencies, while\nanticipation compensates for missing future context. We identify a\ntraining-inference discrepancy in existing OAD methods that hinders learning\neffectiveness. The training uses varying lengths of short-term memory, while\ninference relies on a full-length short-term memory. As a remedy, we propose a\nContext-enhanced Memory-Refined Transformer (CMeRT). CMeRT introduces a\ncontext-enhanced encoder to improve frame representations using additional\nnear-past context. It also features a memory-refined decoder to leverage\nnear-future generation to enhance performance. CMeRT achieves state-of-the-art\nin online detection and anticipation on THUMOS'14, CrossTask, and\nEPIC-Kitchens-100.\n", "date": "2025-03-24", "categories": []}
{"id": "2503.18553", "title": "ATARS: An Aerial Traffic Atomic Activity Recognition and Temporal\n  Segmentation Dataset", "abstract": "  Traffic Atomic Activity which describes traffic patterns for topological\nintersection dynamics is a crucial topic for the advancement of intelligent\ndriving systems. However, existing atomic activity datasets are collected from\nan egocentric view, which cannot support the scenarios where traffic activities\nin an entire intersection are required. Moreover, existing datasets only\nprovide video-level atomic activity annotations, which require exhausting\nefforts to manually trim the videos for recognition and limit their\napplications to untrimmed videos. To bridge this gap, we introduce the Aerial\nTraffic Atomic Activity Recognition and Segmentation (ATARS) dataset, the first\naerial dataset designed for multi-label atomic activity analysis. We offer\natomic activity labels for each frame, which accurately record the intervals\nfor traffic activities. Moreover, we propose a novel task, Multi-label Temporal\nAtomic Activity Recognition, enabling the study of accurate temporal\nlocalization for atomic activity and easing the burden of manual video trimming\nfor recognition. We conduct extensive experiments to evaluate existing\nstate-of-the-art models on both atomic activity recognition and temporal atomic\nactivity segmentation. The results highlight the unique challenges of our ATARS\ndataset, such as recognizing extremely small objects' activities. We further\nprovide comprehensive discussion analyzing these challenges and offer valuable\ninsights for future direction to improve recognizing atomic activity in aerial\nview. Our source code and dataset are available at\nhttps://github.com/magecliff96/ATARS/\n", "date": "2025-03-24", "categories": []}
{"id": "2408.05421", "title": "EPAM-Net: An Efficient Pose-driven Attention-guided Multimodal Network\n  for Video Action Recognition", "abstract": "  Existing multimodal-based human action recognition approaches are\ncomputationally intensive, limiting their deployment in real-time applications.\nIn this work, we present a novel and efficient pose-driven attention-guided\nmultimodal network (EPAM-Net) for action recognition in videos. Specifically,\nwe propose eXpand temporal Shift (X-ShiftNet) convolutional architectures for\nRGB and pose streams to capture spatio-temporal features from RGB videos and\ntheir skeleton sequences. The X-ShiftNet tackles the high computational cost of\nthe 3D CNNs by integrating the Temporal Shift Module (TSM) into an efficient 2D\nCNN, enabling efficient spatiotemporal learning. Then skeleton features are\nutilized to guide the visual network stream, focusing on keyframes and their\nsalient spatial regions using the proposed spatial-temporal attention block.\nFinally, the predictions of the two streams are fused for final classification.\nThe experimental results show that our method, with a significant reduction in\nfloating-point operations (FLOPs), outperforms and competes with the\nstate-of-the-art methods on NTU RGB-D 60, NTU RGB-D 120, PKU-MMD, and Toyota\nSmartHome datasets. The proposed EPAM-Net provides up to a 72.8x reduction in\nFLOPs and up to a 48.6x reduction in the number of network parameters. The code\nwill be available at\nhttps://github.com/ahmed-nady/Multimodal-Action-Recognition.\n", "date": "2024-08-10", "categories": []}
{"id": "2411.18941", "title": "Revealing Key Details to See Differences: A Novel Prototypical\n  Perspective for Skeleton-based Action Recognition", "abstract": "  In skeleton-based action recognition, a key challenge is distinguishing\nbetween actions with similar trajectories of joints due to the lack of\nimage-level details in skeletal representations. Recognizing that the\ndifferentiation of similar actions relies on subtle motion details in specific\nbody parts, we direct our approach to focus on the fine-grained motion of local\nskeleton components. To this end, we introduce ProtoGCN, a Graph Convolutional\nNetwork (GCN)-based model that breaks down the dynamics of entire skeleton\nsequences into a combination of learnable prototypes representing core motion\npatterns of action units. By contrasting the reconstruction of prototypes,\nProtoGCN can effectively identify and enhance the discriminative representation\nof similar actions. Without bells and whistles, ProtoGCN achieves\nstate-of-the-art performance on multiple benchmark datasets, including NTU\nRGB+D, NTU RGB+D 120, Kinetics-Skeleton, and FineGYM, which demonstrates the\neffectiveness of the proposed method. The code is available at\nhttps://github.com/firework8/ProtoGCN.\n", "date": "2024-11-28", "categories": []}
{"id": "2503.14547", "title": "Matching Skeleton-based Activity Representations with Heterogeneous\n  Signals for HAR", "abstract": "  In human activity recognition (HAR), activity labels have typically been\nencoded in one-hot format, which has a recent shift towards using textual\nrepresentations to provide contextual knowledge. Here, we argue that HAR should\nbe anchored to physical motion data, as motion forms the basis of activity and\napplies effectively across sensing systems, whereas text is inherently limited.\nWe propose SKELAR, a novel HAR framework that pretrains activity\nrepresentations from skeleton data and matches them with heterogeneous HAR\nsignals. Our method addresses two major challenges: (1) capturing core motion\nknowledge without context-specific details. We achieve this through a\nself-supervised coarse angle reconstruction task that recovers joint rotation\nangles, invariant to both users and deployments; (2) adapting the\nrepresentations to downstream tasks with varying modalities and focuses. To\naddress this, we introduce a self-attention matching module that dynamically\nprioritizes relevant body parts in a data-driven manner. Given the lack of\ncorresponding labels in existing skeleton data, we establish MASD, a new HAR\ndataset with IMU, WiFi, and skeleton, collected from 20 subjects performing 27\nactivities. This is the first broadly applicable HAR dataset with\ntime-synchronized data across three modalities. Experiments show that SKELAR\nachieves the state-of-the-art performance in both full-shot and few-shot\nsettings. We also demonstrate that SKELAR can effectively leverage synthetic\nskeleton data to extend its use in scenarios without skeleton collections.\n", "date": "2025-03-17", "categories": []}
{"id": "2410.12819", "title": "Deep Adversarial Learning with Activity-Based User Discrimination Task\n  for Human Activity Recognition", "abstract": "  We present a new adversarial deep learning framework for the problem of human\nactivity recognition (HAR) using inertial sensors worn by people. Our framework\nincorporates a novel adversarial activity-based discrimination task that\naddresses inter-person variability-i.e., the fact that different people perform\nthe same activity in different ways. Overall, our proposed framework\noutperforms previous approaches on three HAR datasets using a\nleave-one-(person)-out cross-validation (LOOCV) benchmark. Additional results\ndemonstrate that our discrimination task yields better classification results\ncompared to previous tasks within the same adversarial framework.\n", "date": "2024-10-01", "categories": []}
{"id": "2410.21113", "title": "Zero-Shot Action Recognition in Surveillance Videos", "abstract": "  The growing demand for surveillance in public spaces presents significant\nchallenges due to the shortage of human resources. Current AI-based video\nsurveillance systems heavily rely on core computer vision models that require\nextensive finetuning, which is particularly difficult in surveillance settings\ndue to limited datasets and difficult setting (viewpoint, low quality, etc.).\nIn this work, we propose leveraging Large Vision-Language Models (LVLMs), known\nfor their strong zero and few-shot generalization, to tackle video\nunderstanding tasks in surveillance. Specifically, we explore VideoLLaMA2, a\nstate-of-the-art LVLM, and an improved token-level sampling method,\nSelf-Reflective Sampling (Self-ReS). Our experiments on the UCF-Crime dataset\nshow that VideoLLaMA2 represents a significant leap in zero-shot performance,\nwith 20% boost over the baseline. Self-ReS additionally increases zero-shot\naction recognition performance to 44.6%. These results highlight the potential\nof LVLMs, paired with improved sampling techniques, for advancing surveillance\nvideo analysis in diverse scenarios.\n", "date": "2024-10-28", "categories": []}
{"id": "2411.11223", "title": "Efficient Transfer Learning for Video-language Foundation Models", "abstract": "  Pre-trained vision-language models provide a robust foundation for efficient\ntransfer learning across various downstream tasks. In the field of video action\nrecognition, mainstream approaches often introduce additional modules to\ncapture temporal information. Although the additional modules increase the\ncapacity of model, enabling it to better capture video-specific inductive\nbiases, existing methods typically introduce a substantial number of new\nparameters and are prone to catastrophic forgetting of previously acquired\ngeneralizable knowledge. In this paper, we propose a parameter-efficient\nMulti-modal Spatio-Temporal Adapter (MSTA) to enhance the alignment between\ntextual and visual representations, achieving a balance between generalizable\nknowledge and task-specific adaptation. Furthermore, to mitigate over-fitting\nand enhance generalizability, we introduce a spatio-temporal description-guided\nconsistency constraint.This constraint involves providing template inputs\n(e.g., \"a video of \\{\\textbf{cls}\\}\") to the trainable language branch and\nLLM-generated spatio-temporal descriptions to the pre-trained language branch,\nenforcing output consistency between the branches. This approach reduces\noverfitting to downstream tasks and enhances the distinguishability of the\ntrainable branch within the spatio-temporal semantic space. We evaluate the\neffectiveness of our approach across four tasks: zero-shot transfer, few-shot\nlearning, base-to-novel generalization, and fully-supervised learning. Compared\nto many state-of-the-art methods, our MSTA achieves outstanding performance\nacross all evaluations, while using only 2-7\\% of the trainable parameters in\nthe original model.\n", "date": "2024-11-18", "categories": []}
{"id": "2503.13542", "title": "HAR-DoReMi: Optimizing Data Mixture for Self-Supervised Human Activity\n  Recognition Across Heterogeneous IMU Datasets", "abstract": "  Cross-dataset Human Activity Recognition (HAR) suffers from limited model\ngeneralization, hindering its practical deployment. To address this critical\nchallenge, inspired by the success of DoReMi in Large Language Models (LLMs),\nwe introduce a data mixture optimization strategy for pre-training HAR models,\naiming to improve the recognition performance across heterogeneous datasets.\nHowever, directly applying DoReMi to the HAR field encounters new challenges\ndue to the continuous, multi-channel and intrinsic heterogeneous\ncharacteristics of IMU sensor data. To overcome these limitations, we propose a\nnovel framework HAR-DoReMi, which introduces a masked reconstruction task based\non Mean Squared Error (MSE) loss. By raplacing the discrete language sequence\nprediction task, which relies on the Negative Log-Likelihood (NLL) loss, in the\noriginal DoReMi framework, the proposed framework is inherently more\nappropriate for handling the continuous and multi-channel characteristics of\nIMU data. In addition, HAR-DoReMi integrates the Mahony fusion algorithm into\nthe self-supervised HAR pre-training, aiming to mitigate the heterogeneity of\nvarying sensor orientation. This is achieved by estimating the sensor\norientation within each dataset and facilitating alignment with a unified\ncoordinate system, thereby improving the cross-dataset generalization ability\nof the HAR model. Experimental evaluation on multiple cross-dataset HAR\ntransfer tasks demonstrates that HAR-DoReMi improves the accuracy by an average\nof 6.51%, compared to the current state-of-the-art method with only\napproximately 30% to 50% of the data usage. These results confirm the\neffectiveness of HAR-DoReMi in improving the generalization and data efficiency\nof pre-training HAR models, underscoring its significant potential to\nfacilitate the practical deployment of HAR technology.\n", "date": "2025-03-16", "categories": []}
{"id": "2503.13724", "title": "Towards Scalable Modeling of Compressed Videos for Efficient Action\n  Recognition", "abstract": "  Training robust deep video representations has proven to be computationally\nchallenging due to substantial decoding overheads, the enormous size of raw\nvideo streams, and their inherent high temporal redundancy. Different from\nexisting schemes, operating exclusively in the compressed video domain and\nexploiting all freely available modalities, i.e., I-frames, and P-frames\n(motion vectors and residuals) offers a compute-efficient alternative. Existing\nmethods approach this task as a naive multi-modality problem, ignoring the\ntemporal correlation and implicit sparsity across P-frames for modeling\nstronger shared representations for videos of the same action, making training\nand generalization easier. By revisiting the high-level design of dominant\nvideo understanding backbones, we increase inference speed by a factor of $56$\nwhile retaining similar performance. For this, we propose a hybrid end-to-end\nframework that factorizes learning across three key concepts to reduce\ninference cost by $330\\times$ versus prior art: First, a specially designed\ndual-encoder scheme with efficient Spiking Temporal Modulators to minimize\nlatency while retaining cross-domain feature aggregation. Second, a unified\ntransformer model to capture inter-modal dependencies using global\nself-attention to enhance I-frame -- P-frame contextual interactions. Third, a\nMulti-Modal Mixer Block to model rich representations from the joint\nspatiotemporal token embeddings. Experiments show that our method results in a\nlightweight architecture achieving state-of-the-art video recognition\nperformance on UCF-101, HMDB-51, K-400, K-600 and SS-v2 datasets with favorable\ncosts ($0.73$J/V) and fast inference ($16$V/s). Our observations bring new\ninsights into practical design choices for efficient next-generation\nspatiotemporal learners. Code is available.\n", "date": "2025-03-17", "categories": []}
{"id": "2503.14430", "title": "Joint Image-Instance Spatial-Temporal Attention for Few-shot Action\n  Recognition", "abstract": "  Few-shot Action Recognition (FSAR) constitutes a crucial challenge in\ncomputer vision, entailing the recognition of actions from a limited set of\nexamples. Recent approaches mainly focus on employing image-level features to\nconstruct temporal dependencies and generate prototypes for each action\ncategory. However, a considerable number of these methods utilize mainly\nimage-level features that incorporate background noise and focus insufficiently\non real foreground (action-related instances), thereby compromising the\nrecognition capability, particularly in the few-shot scenario. To tackle this\nissue, we propose a novel joint Image-Instance level Spatial-temporal attention\napproach (I2ST) for Few-shot Action Recognition. The core concept of I2ST is to\nperceive the action-related instances and integrate them with image features\nvia spatial-temporal attention. Specifically, I2ST consists of two key\ncomponents: Action-related Instance Perception and Joint Image-Instance\nSpatial-temporal Attention. Given the basic representations from the feature\nextractor, the Action-related Instance Perception is introduced to perceive\naction-related instances under the guidance of a text-guided segmentation\nmodel. Subsequently, the Joint Image-Instance Spatial-temporal Attention is\nused to construct the feature dependency between instances and images...\n", "date": "2025-03-11", "categories": []}
{"id": "2312.15144", "title": "A Generically Contrastive Spatiotemporal Representation Enhancement for\n  3D Skeleton Action Recognition", "abstract": "  Skeleton-based action recognition is a central task in computer vision and\nhuman-robot interaction. However, most previous methods suffer from overlooking\nthe explicit exploitation of the latent data distributions (i.e., the\nintra-class variations and inter-class relations), thereby leading to confusion\nabout ambiguous samples and sub-optimum solutions of the skeleton encoders. To\nmitigate this, we propose a Contrastive Spatiotemporal Representation\nEnhancement (CSRE) framework to obtain more discriminative representations from\nthe sequences, which can be incorporated into various previous skeleton\nencoders and can be removed when testing. Specifically, we decompose the\nrepresentation into spatial- and temporal-specific features to explore\nfine-grained motion patterns along the corresponding dimensions. Furthermore,\nto explicitly exploit the latent data distributions, we employ the attentive\nfeatures to contrastive learning, which models the cross-sequence semantic\nrelations by pulling together the features from the positive pairs and pushing\naway the negative pairs. Extensive experiments show that CSRE with five various\nskeleton encoders (HCN, 2S-AGCN, CTR-GCN, Hyperformer, and BlockGCN) achieves\nsolid improvements on five benchmarks. The code will be released at\nhttps://github.com/zhshj0110/CSRE.\n", "date": "2023-12-23", "categories": []}
{"id": "2503.12034", "title": "Real-Time Manipulation Action Recognition with a Factorized Graph\n  Sequence Encoder", "abstract": "  Recognition of human manipulation actions in real-time is essential for safe\nand effective human-robot interaction and collaboration. The challenge lies in\ndeveloping a model that is both lightweight enough for real-time execution and\ncapable of generalization. While some existing methods in the literature can\nrun in real-time, they struggle with temporal scalability, i.e., they fail to\nadapt to long-duration manipulations effectively. To address this, leveraging\nthe generalizable scene graph representations, we propose a new Factorized\nGraph Sequence Encoder network that not only runs in real-time but also scales\neffectively in the temporal dimension, thanks to its factorized encoder\narchitecture. Additionally, we introduce Hand Pooling operation, a simple\npooling operation for more focused extraction of the graph-level embeddings.\nOur model outperforms the previous state-of-the-art real-time approach,\nachieving a 14.3\\% and 5.6\\% improvement in F1-macro score on the KIT Bimanual\nAction (Bimacs) Dataset and Collaborative Action (CoAx) Dataset, respectively.\nMoreover, we conduct an extensive ablation study to validate our network design\nchoices. Finally, we compare our model with its architecturally similar\nRGB-based model on the Bimacs dataset and show the limitations of this model in\ncontrast to ours on such an object-centric manipulation dataset.\n", "date": "2025-03-15", "categories": []}
{"id": "2503.12852", "title": "ACT360: An Efficient 360-Degree Action Detection and Summarization\n  Framework for Mission-Critical Training and Debriefing", "abstract": "  Effective training and debriefing are critical in high-stakes,\nmission-critical environments such as disaster response, military simulations,\nand industrial safety, where precision and minimizing errors are paramount. The\ntraditional post-training analysis relies on manually reviewing 2D videos, a\ntime-consuming process that lacks comprehensive situational awareness. To\naddress these limitations, we introduce ACT360, a system that leverages\n360-degree videos and machine learning for automated action detection and\nstructured debriefing. ACT360 integrates 360YOWO, an enhanced You Only Watch\nOnce (YOWO) model with spatial attention and equirectangular-aware convolution\n(EAC) to mitigate panoramic video distortions. To enable deployment in\nresource-constrained environments, we apply quantization and model pruning,\nreducing the model size by 74% while maintaining robust accuracy (mAP drop of\nonly 1.5%, from 0.865 to 0.850) and improving inference speed. We validate our\napproach on a publicly available dataset of 55 labeled 360-degree videos\ncovering seven key operational actions, recorded across various real-world\ntraining sessions and environmental conditions. Additionally, ACT360 integrates\n360AIE (Action Insight Explorer), a web-based interface for automatic action\ndetection, retrieval, and textual summarization using large language models\n(LLMs), significantly enhancing post-incident analysis efficiency. ACT360\nserves as a generalized framework for mission-critical debriefing,\nincorporating EAC, spatial attention, summarization, and model optimization.\nThese innovations apply to any training environment requiring lightweight\naction detection and structured post-exercise analysis.\n", "date": "2025-03-17", "categories": []}
{"id": "2503.12969", "title": "Action tube generation by person query matching for spatio-temporal\n  action detection", "abstract": "  This paper proposes a method for spatio-temporal action detection (STAD) that\ndirectly generates action tubes from the original video without relying on\npost-processing steps such as IoU-based linking and clip splitting. Our\napproach applies query-based detection (DETR) to each frame and matches DETR\nqueries to link the same person across frames. We introduce the Query Matching\nModule (QMM), which uses metric learning to bring queries for the same person\ncloser together across frames compared to queries for different people. Action\nclasses are predicted using the sequence of queries obtained from QMM matching,\nallowing for variable-length inputs from videos longer than a single clip.\nExperimental results on JHMDB, UCF101-24, and AVA datasets demonstrate that our\nmethod performs well for large position changes of people while offering\nsuperior computational efficiency and lower resource requirements.\n", "date": "2025-03-17", "categories": []}
{"id": "2503.11004", "title": "VA-AR: Learning Velocity-Aware Action Representations with Mixture of\n  Window Attention", "abstract": "  Action recognition is a crucial task in artificial intelligence, with\nsignificant implications across various domains. We initially perform a\ncomprehensive analysis of seven prominent action recognition methods across\nfive widely-used datasets. This analysis reveals a critical, yet previously\noverlooked, observation: as the velocity of actions increases, the performance\nof these methods variably declines, undermining their robustness. This decline\nin performance poses significant challenges for their application in real-world\nscenarios. Building on these findings, we introduce the Velocity-Aware Action\nRecognition (VA-AR) framework to obtain robust action representations across\ndifferent velocities. Our principal insight is that rapid actions (e.g., the\ngiant circle backward in uneven bars or a smash in badminton) occur within\nshort time intervals, necessitating smaller temporal attention windows to\naccurately capture intricate changes. Conversely, slower actions (e.g.,\ndrinking water or wiping face) require larger windows to effectively encompass\nthe broader context. VA-AR employs a Mixture of Window Attention (MoWA)\nstrategy, dynamically adjusting its attention window size based on the action's\nvelocity. This adjustment enables VA-AR to obtain a velocity-aware\nrepresentation, thereby enhancing the accuracy of action recognition. Extensive\nexperiments confirm that VA-AR achieves state-of-the-art performance on the\nsame five datasets, demonstrating VA-AR's effectiveness across a broad spectrum\nof action recognition scenarios.\n", "date": "2025-03-14", "categories": []}
{"id": "2410.14993", "title": "Making Every Frame Matter: Continuous Activity Recognition in Streaming\n  Video via Adaptive Video Context Modeling", "abstract": "  Video activity recognition has become increasingly important in robots and\nembodied AI. Recognizing continuous video activities poses considerable\nchallenges due to the fast expansion of streaming video, which contains\nmulti-scale and untrimmed activities. We introduce a novel system, CARS, to\novercome these issues through adaptive video context modeling. Adaptive video\ncontext modeling refers to selectively maintaining activity-related features in\ntemporal and spatial dimensions. CARS has two key designs. The first is an\nactivity spatial feature extraction by eliminating irrelevant visual features\nwhile maintaining recognition accuracy. The second is an activity-aware state\nupdate introducing dynamic adaptability to better preserve the video context\nfor multi-scale activity recognition. Our CARS runs at speeds $>$30 FPS on\ntypical edge devices and outperforms all baselines by 1.2\\% to 79.7\\% in\naccuracy. Moreover, we explore applying CARS to a large video model as a video\nencoder. Experimental results show that our CARS can result in a 0.46-point\nenhancement (on a 5-point scale) on the in-distribution video activity dataset,\nand an improvement ranging from 1.19\\% to 4\\% on zero-shot video activity\ndatasets.\n", "date": "2024-10-19", "categories": []}
{"id": "2501.18509", "title": "Reframing Dense Action Detection (RefDense): A Paradigm Shift in Problem\n  Solving & a Novel Optimization Strategy", "abstract": "  Dense action detection involves detecting multiple co-occurring actions while\naction classes are often ambiguous and represent overlapping concepts. We argue\nthat handling the dual challenge of temporal and class overlaps is too complex\nto effectively be tackled by a single network. To address this, we propose to\ndecompose the task of detecting dense ambiguous actions into detecting dense,\nunambiguous sub-concepts that form the action classes (i.e., action entities\nand action motions), and assigning these sub-tasks to distinct sub-networks. By\nisolating these unambiguous concepts, the sub-networks can focus exclusively on\nresolving a single challenge, dense temporal overlaps. Furthermore,\nsimultaneous actions in a video often exhibit interrelationships, and\nexploiting these relationships can improve the method performance. However,\ncurrent dense action detection networks fail to effectively learn these\nrelationships due to their reliance on binary cross-entropy optimization, which\ntreats each class independently. To address this limitation, we propose\nproviding explicit supervision on co-occurring concepts during network\noptimization through a novel language-guided contrastive learning loss. Our\nextensive experiments demonstrate the superiority of our approach over\nstate-of-the-art methods, achieving substantial improvements of 3.8% and 1.7%\non average across all metrics on the challenging benchmark datasets, Charades\nand MultiTHUMOS.\n", "date": "2025-01-30", "categories": []}
{"id": "2503.07821", "title": "Elderly Activity Recognition in the Wild: Results from the EAR Challenge", "abstract": "  This paper presents our solution for the Elderly Action Recognition (EAR)\nChallenge, part of the Computer Vision for Smalls Workshop at WACV 2025. The\ncompetition focuses on recognizing Activities of Daily Living (ADLs) performed\nby the elderly, covering six action categories with a diverse dataset. Our\napproach builds upon a state-of-the-art action recognition model, fine-tuned\nthrough transfer learning on elderly-specific datasets to enhance adaptability.\nTo improve generalization and mitigate dataset bias, we carefully curated\ntraining data from multiple publicly available sources and applied targeted\npre-processing techniques. Our solution currently achieves 0.81455 accuracy on\nthe public leaderboard, highlighting its effectiveness in classifying elderly\nactivities. Source codes are publicly available at\nhttps://github.com/ffyyytt/EAR-WACV25-DAKiet-TSM.\n", "date": "2025-03-10", "categories": []}
{"id": "2503.07860", "title": "Video Action Differencing", "abstract": "  How do two individuals differ when performing the same action? In this work,\nwe introduce Video Action Differencing (VidDiff), the novel task of identifying\nsubtle differences between videos of the same action, which has many\napplications, such as coaching and skill learning. To enable development on\nthis new task, we first create VidDiffBench, a benchmark dataset containing 549\nvideo pairs, with human annotations of 4,469 fine-grained action differences\nand 2,075 localization timestamps indicating where these differences occur. Our\nexperiments demonstrate that VidDiffBench poses a significant challenge for\nstate-of-the-art large multimodal models (LMMs), such as GPT-4o and Qwen2-VL.\nBy analyzing failure cases of LMMs on VidDiffBench, we highlight two key\nchallenges for this task: localizing relevant sub-actions over two videos and\nfine-grained frame comparison. To overcome these, we propose the VidDiff\nmethod, an agentic workflow that breaks the task into three stages: action\ndifference proposal, keyframe localization, and frame differencing, each stage\nutilizing specialized foundation models. To encourage future research in this\nnew task, we release the benchmark at\nhttps://huggingface.co/datasets/jmhb/VidDiffBench and code at\nhttp://jmhb0.github.io/viddiff.\n", "date": "2025-03-10", "categories": []}
{"id": "2412.14833", "title": "Synchronized and Fine-Grained Head for Skeleton-Based Ambiguous Action\n  Recognition", "abstract": "  Skeleton-based action recognition using GCNs has achieved remarkable\nperformance, but recognizing ambiguous actions, such as \"waving\" and\n\"saluting\", remains a significant challenge. Existing methods typically rely on\na serial combination of GCNs and TCNs, where spatial and temporal features are\nextracted independently, leading to an unbalanced spatial-temporal information,\nwhich hinders accurate action recognition. Moreover, existing methods for\nambiguous actions often overemphasize local details, resulting in the loss of\ncrucial global context, which further complicates the task of differentiating\nambiguous actions. To address these challenges, we propose a lightweight\nplug-and-play module called SF-Head, inserted between GCN and TCN layers.\nSF-Head first conducts SSTE with a Feature Redundancy Loss (F-RL), ensuring a\nbalanced interaction. It then performs AC-FA, with a Feature Consistency Loss\n(F-CL), which aligns the aggregated feature with their original\nspatial-temporal feature. Experimental results on NTU RGB+D 60, NTU RGB+D 120,\nNW-UCLA and PKU-MMD I datasets demonstrate significant improvements in\ndistinguishing ambiguous actions.\n", "date": "2024-12-19", "categories": []}
{"id": "2503.06522", "title": "SGA-INTERACT: A 3D Skeleton-based Benchmark for Group Activity\n  Understanding in Modern Basketball Tactic", "abstract": "  Group Activity Understanding is predominantly studied as Group Activity\nRecognition (GAR) task. However, existing GAR benchmarks suffer from\ncoarse-grained activity vocabularies and the only data form in single-view,\nwhich hinder the evaluation of state-of-the-art algorithms. To address these\nlimitations, we introduce SGA-INTERACT, the first 3D skeleton-based benchmark\nfor group activity understanding. It features complex activities inspired by\nbasketball tactics, emphasizing rich spatial interactions and long-term\ndependencies. SGA-INTERACT introduces Temporal Group Activity Localization\n(TGAL) task, extending group activity understanding to untrimmed sequences,\nfilling the gap left by GAR as a standalone task. In addition to the benchmark,\nwe propose One2Many, a novel framework that employs a pretrained 3D skeleton\nbackbone for unified individual feature extraction. This framework aligns with\nthe feature extraction paradigm in RGB-based methods, enabling direct\nevaluation of RGB-based models on skeleton-based benchmarks. We conduct\nextensive evaluations on SGA-INTERACT using two skeleton-based methods, three\nRGB-based methods, and a proposed baseline within the One2Many framework. The\ngeneral low performance of baselines highlights the benchmark's challenges,\nmotivating advancements in group activity understanding.\n", "date": "2025-03-09", "categories": []}
{"id": "2503.07008", "title": "SDFA: Structure Aware Discriminative Feature Aggregation for Efficient\n  Human Fall Detection in Video", "abstract": "  Older people are susceptible to fall due to instability in posture and\ndeteriorating health. Immediate access to medical support can greatly reduce\nrepercussions. Hence, there is an increasing interest in automated fall\ndetection, often incorporated into a smart healthcare system to provide better\nmonitoring. Existing systems focus on wearable devices which are inconvenient\nor video monitoring which has privacy concerns. Moreover, these systems provide\na limited perspective of their generalization ability as they are tested on\ndatasets containing few activities that have wide disparity in the action space\nand are easy to differentiate. Complex daily life scenarios pose much greater\nchallenges with activities that overlap in action spaces due to similar posture\nor motion. To overcome these limitations, we propose a fall detection model,\ncoined SDFA, based on human skeletons extracted from low-resolution videos. The\nuse of skeleton data ensures privacy and low-resolution videos ensures low\nhardware and computational cost. Our model captures discriminative structural\ndisplacements and motion trends using unified joint and motion features\nprojected onto a shared high dimensional space. Particularly, the use of\nseparable convolution combined with a powerful GCN architecture provides\nimproved performance. Extensive experiments on five large-scale datasets with a\nwide range of evaluation settings show that our model achieves competitive\nperformance with extremely low computational complexity and runs faster than\nexisting models.\n", "date": "2025-03-10", "categories": []}
{"id": "2409.00926", "title": "Towards Student Actions in Classroom Scenes: New Dataset and Baseline", "abstract": "  Analyzing student actions is an important and challenging task in educational\nresearch. Existing efforts have been hampered by the lack of accessible\ndatasets to capture the nuanced action dynamics in classrooms. In this paper,\nwe present a new multi-label Student Action Video (SAV) dataset, specifically\ndesigned for action detection in classroom settings. The SAV dataset consists\nof 4,324 carefully trimmed video clips from 758 different classrooms, annotated\nwith 15 distinct student actions. Compared to existing action detection\ndatasets, the SAV dataset stands out by providing a wide range of real\nclassroom scenarios, high-quality video data, and unique challenges, including\nsubtle movement differences, dense object engagement, significant scale\ndifferences, varied shooting angles, and visual occlusion. These complexities\nintroduce new opportunities and challenges to advance action detection methods.\nTo benchmark this, we propose a novel baseline method based on a visual\ntransformer, designed to enhance attention to key local details within small\nand dense object regions. Our method demonstrates excellent performance with a\nmean Average Precision (mAP) of 67.9% and 27.4% on the SAV and AVA datasets,\nrespectively. This paper not only provides the dataset but also calls for\nfurther research into AI-driven educational tools that may transform teaching\nmethodologies and learning outcomes. The code and dataset are released at\nhttps://github.com/Ritatanz/SAV.\n", "date": "2024-09-02", "categories": []}
{"id": "2503.04078", "title": "Spatial-Temporal Perception with Causal Inference for Naturalistic\n  Driving Action Recognition", "abstract": "  Naturalistic driving action recognition is essential for vehicle cabin\nmonitoring systems. However, the complexity of real-world backgrounds presents\nsignificant challenges for this task, and previous approaches have struggled\nwith practical implementation due to their limited ability to observe subtle\nbehavioral differences and effectively learn inter-frame features from video.\nIn this paper, we propose a novel Spatial-Temporal Perception (STP)\narchitecture that emphasizes both temporal information and spatial\nrelationships between key objects, incorporating a causal decoder to perform\nbehavior recognition and temporal action localization. Without requiring\nmultimodal input, STP directly extracts temporal and spatial distance features\nfrom RGB video clips. Subsequently, these dual features are jointly encoded by\nmaximizing the expected likelihood across all possible permutations of the\nfactorization order. By integrating temporal and spatial features at different\nscales, STP can perceive subtle behavioral changes in challenging scenarios.\nAdditionally, we introduce a causal-aware module to explore relationships\nbetween video frame features, significantly enhancing detection efficiency and\nperformance. We validate the effectiveness of our approach using two publicly\navailable driver distraction detection benchmarks. The results demonstrate that\nour framework achieves state-of-the-art performance.\n", "date": "2025-03-06", "categories": []}
{"id": "2405.02077", "title": "MVP-Shot: Multi-Velocity Progressive-Alignment Framework for Few-Shot\n  Action Recognition", "abstract": "  Recent few-shot action recognition (FSAR) methods typically perform semantic\nmatching on learned discriminative features to achieve promising performance.\nHowever, most FSAR methods focus on single-scale (e.g., frame-level,\nsegment-level, etc) feature alignment, which ignores that human actions with\nthe same semantic may appear at different velocities. To this end, we develop a\nnovel Multi-Velocity Progressive-alignment (MVP-Shot) framework to\nprogressively learn and align semantic-related action features at\nmulti-velocity levels. Concretely, a Multi-Velocity Feature Alignment (MVFA)\nmodule is designed to measure the similarity between features from support and\nquery videos with different velocity scales and then merge all similarity\nscores in a residual fashion. To avoid the multiple velocity features deviating\nfrom the underlying motion semantic, our proposed Progressive Semantic-Tailored\nInteraction (PSTI) module injects velocity-tailored text information into the\nvideo feature via feature interaction on channel and temporal domains at\ndifferent velocities. The above two modules compensate for each other to make\nmore accurate query sample predictions under the few-shot settings.\nExperimental results show our method outperforms current state-of-the-art\nmethods on multiple standard few-shot benchmarks (i.e., HMDB51, UCF101,\nKinetics, and SSv2-small).\n", "date": "2024-05-03", "categories": []}
{"id": "2410.01962", "title": "LS-HAR: Language Supervised Human Action Recognition with Salient\n  Fusion, Construction Sites as a Use-Case", "abstract": "  Detecting human actions is a crucial task for autonomous robots and vehicles,\noften requiring the integration of various data modalities for improved\naccuracy. In this study, we introduce a novel approach to Human Action\nRecognition (HAR) using language supervision named LS-HAR based on skeleton and\nvisual cues. Our method leverages a language model to guide the feature\nextraction process in the skeleton encoder. Specifically, we employ learnable\nprompts for the language model conditioned on the skeleton modality to optimize\nfeature representation. Furthermore, we propose a fusion mechanism that\ncombines dual-modality features using a salient fusion module, incorporating\nattention and transformer mechanisms to address the modalities' high\ndimensionality. This fusion process prioritizes informative video frames and\nbody joints, enhancing the recognition accuracy of human actions. Additionally,\nwe introduce a new dataset tailored for real-world robotic applications in\nconstruction sites, featuring visual, skeleton, and depth data modalities,\nnamed VolvoConstAct. This dataset serves to facilitate the training and\nevaluation of machine learning models to instruct autonomous construction\nmachines for performing necessary tasks in real-world construction sites. To\nevaluate our approach, we conduct experiments on our dataset as well as three\nwidely used public datasets: NTU-RGB+D, NTU-RGB+D 120, and NW-UCLA. Results\nreveal that our proposed method achieves promising performance across all\ndatasets, demonstrating its robustness and potential for various applications.\nThe code, dataset, and demonstration of real-machine experiments are available\nat: https://mmahdavian.github.io/ls_har/\n", "date": "2024-10-02", "categories": []}
{"id": "2403.15423", "title": "Cross-user activity recognition via temporal relation optimal transport", "abstract": "  Current research on human activity recognition (HAR) mainly assumes that\ntraining and testing data are drawn from the same distribution to achieve a\ngeneralised model, which means all the data are considered to be independent\nand identically distributed $\\displaystyle (i.i.d.) $. In many real-world\napplications, this assumption does not hold, and collected training and target\ntesting datasets have non-uniform distribution, such as in the case of\ncross-user HAR. Domain adaptation is a promising approach for cross-user HAR\ntasks. Existing domain adaptation works based on the assumption that samples in\neach domain are $\\displaystyle i.i.d. $ and do not consider the knowledge of\ntemporal relation hidden in time series data for aligning data distribution.\nThis strong assumption of $\\displaystyle i.i.d. $ may not be suitable for time\nseries-related domain adaptation methods because the samples formed by time\nseries segmentation and feature extraction techniques are only coarse\napproximations to $\\displaystyle i.i.d. $ assumption in each domain. In this\npaper, we propose the temporal relation optimal transport (TROT) method to\nutilise temporal relation and relax the $\\displaystyle i.i.d. $ assumption for\nthe samples in each domain for accurate and efficient knowledge transfer. We\nobtain the temporal relation representation and implement temporal relation\nalignment of activities via the Hidden Markov model (HMM) and optimal transport\n(OT) techniques. Besides, a new regularisation term that preserves temporal\nrelation order information for an improved optimal transport mapping is\nproposed to enhance the domain adaptation performance. Comprehensive\nexperiments are conducted on three public activity recognition datasets (i.e.\nOPPT, PAMAP2 and DSADS), demonstrating that TROT outperforms other\nstate-of-the-art methods.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.17958", "title": "Deep Generative Domain Adaptation with Temporal Attention for Cross-User\n  Activity Recognition", "abstract": "  In Human Activity Recognition (HAR), a predominant assumption is that the\ndata utilized for training and evaluation purposes are drawn from the same\ndistribution. It is also assumed that all data samples are independent and\nidentically distributed ($\\displaystyle i.i.d.$). Contrarily, practical\nimplementations often challenge this notion, manifesting data distribution\ndiscrepancies, especially in scenarios such as cross-user HAR. Domain\nadaptation is the promising approach to address these challenges inherent in\ncross-user HAR tasks. However, a clear gap in domain adaptation techniques is\nthe neglect of the temporal relation embedded within time series data during\nthe phase of aligning data distributions. Addressing this oversight, our\nresearch presents the Deep Generative Domain Adaptation with Temporal Attention\n(DGDATA) method. This novel method uniquely recognises and integrates temporal\nrelations during the domain adaptation process. By synergizing the capabilities\nof generative models with the Temporal Relation Attention mechanism, our method\nimproves the classification performance in cross-user HAR. A comprehensive\nevaluation has been conducted on three public sensor-based HAR datasets\ntargeting different scenarios and applications to demonstrate the efficacy of\nthe proposed DGDATA method.\n", "date": "2024-03-12", "categories": []}
{"id": "2503.00796", "title": "An Efficient 3D Convolutional Neural Network with Channel-wise,\n  Spatial-grouped, and Temporal Convolutions", "abstract": "  There has been huge progress on video action recognition in recent years.\nHowever, many works focus on tweaking existing 2D backbones due to the reliance\nof ImageNet pretraining, which restrains the models from achieving higher\nefficiency for video recognition. In this work we introduce a simple and very\nefficient 3D convolutional neural network for video action recognition. The\ndesign of the building block consists of a channel-wise convolution, followed\nby a spatial group convolution, and finally a temporal convolution. We evaluate\nthe performance and efficiency of our proposed network on several video action\nrecognition datasets by directly training on the target dataset without relying\non pertaining. On Something-Something-V1&V2, Kinetics-400 and Multi-Moments in\nTime, our network can match or even surpass the performance of other models\nwhich are several times larger. On the fine-grained action recognition dataset\nFineGym, we beat the previous state-of-the-art accuracy achieved with 2-stream\nmethods by more than 5% using only RGB input.\n", "date": "2025-03-02", "categories": []}
{"id": "2503.02284", "title": "Semi-Supervised Audio-Visual Video Action Recognition with Audio Source\n  Localization Guided Mixup", "abstract": "  Video action recognition is a challenging but important task for\nunderstanding and discovering what the video does. However, acquiring\nannotations for a video is costly, and semi-supervised learning (SSL) has been\nstudied to improve performance even with a small number of labeled data in the\ntask. Prior studies for semi-supervised video action recognition have mostly\nfocused on using single modality - visuals - but the video is multi-modal, so\nutilizing both visuals and audio would be desirable and improve performance\nfurther, which has not been explored well. Therefore, we propose audio-visual\nSSL for video action recognition, which uses both visual and audio together,\neven with quite a few labeled data, which is challenging. In addition, to\nmaximize the information of audio and video, we propose a novel audio source\nlocalization-guided mixup method that considers inter-modal relations between\nvideo and audio modalities. In experiments on UCF-51, Kinetics-400, and\nVGGSound datasets, our model shows the superior performance of the proposed\nsemi-supervised audio-visual action recognition framework and audio source\nlocalization-guided mixup.\n", "date": "2025-03-04", "categories": []}
{"id": "2502.00156", "title": "ALBAR: Adversarial Learning approach to mitigate Biases in Action\n  Recognition", "abstract": "  Bias in machine learning models can lead to unfair decision making, and while\nit has been well-studied in the image and text domains, it remains\nunderexplored in action recognition. Action recognition models often suffer\nfrom background bias (i.e., inferring actions based on background cues) and\nforeground bias (i.e., relying on subject appearance), which can be detrimental\nto real-life applications such as autonomous vehicles or assisted living\nmonitoring. While prior approaches have mainly focused on mitigating background\nbias using specialized augmentations, we thoroughly study both foreground and\nbackground bias. We propose ALBAR, a novel adversarial training method that\nmitigates foreground and background biases without requiring specialized\nknowledge of the bias attributes. Our framework applies an adversarial\ncross-entropy loss to the sampled static clip (where all the frames are the\nsame) and aims to make its class probabilities uniform using a proposed entropy\nmaximization loss. Additionally, we introduce a gradient penalty loss for\nregularization against the debiasing process. We evaluate our method on\nestablished background and foreground bias protocols, setting a new\nstate-of-the-art and strongly improving combined debiasing performance by over\n12% absolute on HMDB51. Furthermore, we identify an issue of background leakage\nin the existing UCF101 protocol for bias evaluation which provides a shortcut\nto predict actions and does not provide an accurate measure of the debiasing\ncapability of a model. We address this issue by proposing more fine-grained\nsegmentation boundaries for the actor, where our method also outperforms\nexisting approaches. Project Page:\nhttps://joefioresi718.github.io/ALBAR_webpage/\n", "date": "2025-01-31", "categories": []}
{"id": "2503.01436", "title": "Fall Detection from Indoor Videos using MediaPipe and Handcrafted\n  Feature", "abstract": "  Falls are a common cause of fatal injuries and hospitalization. However,\nhaving fall detection on person, in particular for senior citizens can prove to\nbe critical. Presently,there are handheld, ambient detector and vision-based\ndetection techniques being utilized for fall detection. However, the approaches\nhave issues with accuracy and cost. In this regard, in this research, an\napproach is proposed to detect falls in indoor environments utilizing the\nhandcrafted features extracted from human body skeleton. The human body\nskeleton is formed using MediaPipe framework. Results on UR Fall detection show\nthe superiority of our model, capable of detecting falls correctly in a wide\nnumber of settings involving people belonging to different ages and genders.\nThis proposed model using MediaPipe for fall classification in daily activities\nachieves significant accuracy compare to the present existing approaches.\n", "date": "2025-03-03", "categories": []}
{"id": "2410.12337", "title": "ARIC: An Activity Recognition Dataset in Classroom Surveillance Images", "abstract": "  The application of activity recognition in the ``AI + Education\" field is\ngaining increasing attention. However, current work mainly focuses on the\nrecognition of activities in manually captured videos and a limited number of\nactivity types, with little attention given to recognizing activities in\nsurveillance images from real classrooms. Activity recognition in classroom\nsurveillance images faces multiple challenges, such as class imbalance and high\nactivity similarity. To address this gap, we constructed a novel multimodal\ndataset focused on classroom surveillance image activity recognition called\nARIC (Activity Recognition In Classroom). The ARIC dataset has advantages of\nmultiple perspectives, 32 activity categories, three modalities, and real-world\nclassroom scenarios. In addition to the general activity recognition tasks, we\nalso provide settings for continual learning and few-shot continual learning.\nWe hope that the ARIC dataset can act as a facilitator for future analysis and\nresearch for open teaching scenarios. You can download preliminary data from\nhttps://ivipclab.github.io/publication_ARIC/ARIC.\n", "date": "2024-10-16", "categories": []}
{"id": "2411.05692", "title": "Autoregressive Adaptive Hypergraph Transformer for Skeleton-based\n  Activity Recognition", "abstract": "  Extracting multiscale contextual information and higher-order correlations\namong skeleton sequences using Graph Convolutional Networks (GCNs) alone is\ninadequate for effective action classification. Hypergraph convolution\naddresses the above issues but cannot harness the long-range dependencies. The\ntransformer proves to be effective in capturing these dependencies and making\ncomplex contextual features accessible. We propose an Autoregressive Adaptive\nHyperGraph Transformer (AutoregAd-HGformer) model for in-phase (autoregressive\nand discrete) and out-phase (adaptive) hypergraph generation. The vector\nquantized in-phase hypergraph equipped with powerful autoregressive learned\npriors produces a more robust and informative representation suitable for\nhyperedge formation. The out-phase hypergraph generator provides a\nmodel-agnostic hyperedge learning technique to align the attributes with input\nskeleton embedding. The hybrid (supervised and unsupervised) learning in\nAutoregAd-HGformer explores the action-dependent feature along spatial,\ntemporal, and channel dimensions. The extensive experimental results and\nablation study indicate the superiority of our model over state-of-the-art\nhypergraph architectures on the NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets.\n", "date": "2024-11-08", "categories": []}
{"id": "2502.20361", "title": "OpenTAD: A Unified Framework and Comprehensive Study of Temporal Action\n  Detection", "abstract": "  Temporal action detection (TAD) is a fundamental video understanding task\nthat aims to identify human actions and localize their temporal boundaries in\nvideos. Although this field has achieved remarkable progress in recent years,\nfurther progress and real-world applications are impeded by the absence of a\nstandardized framework. Currently, different methods are compared under\ndifferent implementation settings, evaluation protocols, etc., making it\ndifficult to assess the real effectiveness of a specific technique. To address\nthis issue, we propose \\textbf{OpenTAD}, a unified TAD framework consolidating\n16 different TAD methods and 9 standard datasets into a modular codebase. In\nOpenTAD, minimal effort is required to replace one module with a different\ndesign, train a feature-based TAD model in end-to-end mode, or switch between\nthe two. OpenTAD also facilitates straightforward benchmarking across various\ndatasets and enables fair and in-depth comparisons among different methods.\nWith OpenTAD, we comprehensively study how innovations in different network\ncomponents affect detection performance and identify the most effective design\nchoices through extensive experiments. This study has led to a new\nstate-of-the-art TAD method built upon existing techniques for each component.\nWe have made our code and models available at\nhttps://github.com/sming256/OpenTAD.\n", "date": "2025-02-27", "categories": []}
{"id": "2502.18555", "title": "Application of Attention Mechanism with Bidirectional Long Short-Term\n  Memory (BiLSTM) and CNN for Human Conflict Detection using Computer Vision", "abstract": "  The automatic detection of human conflicts through videos is a crucial area\nin computer vision, with significant applications in monitoring and public\nsafety policies. However, the scarcity of public datasets and the complexity of\nhuman interactions make this task challenging. This study investigates the\nintegration of advanced deep learning techniques, including Attention\nMechanism, Convolutional Neural Networks (CNNs), and Bidirectional Long\nShortTerm Memory (BiLSTM), to improve the detection of violent behaviors in\nvideos. The research explores how the use of the attention mechanism can help\nfocus on the most relevant parts of the video, enhancing the accuracy and\nrobustness of the model. The experiments indicate that the combination of CNNs\nwith BiLSTM and the attention mechanism provides a promising solution for\nconflict monitoring, offering insights into the effectiveness of different\nstrategies. This work opens new possibilities for the development of automated\nsurveillance systems that can operate more efficiently in real-time detection\nof violent events.\n", "date": "2025-02-25", "categories": []}
{"id": "2312.13509", "title": "MR-STGN: Multi-Residual Spatio Temporal Graph Network Using Attention\n  Fusion for Patient Action Assessment", "abstract": "  Accurate assessment of patient actions plays a crucial role in healthcare as\nit contributes significantly to disease progression monitoring and treatment\neffectiveness. However, traditional approaches to assess patient actions often\nrely on manual observation and scoring, which are subjective and\ntime-consuming. In this paper, we propose an automated approach for patient\naction assessment using a Multi-Residual Spatio Temporal Graph Network\n(MR-STGN) that incorporates both angular and positional 3D skeletons. The\nMR-STGN is specifically designed to capture the spatio-temporal dynamics of\npatient actions. It achieves this by integrating information from multiple\nresidual layers, with each layer extracting features at distinct levels of\nabstraction. Furthermore, we integrate an attention fusion mechanism into the\nnetwork, which facilitates the adaptive weighting of various features. This\nempowers the model to concentrate on the most pertinent aspects of the\npatient's movements, offering precise instructions regarding specific body\nparts or movements that require attention. Ablation studies are conducted to\nanalyze the impact of individual components within the proposed model. We\nevaluate our model on the UI-PRMD dataset demonstrating its performance in\naccurately predicting real-time patient action scores, surpassing\nstate-of-the-art methods.\n", "date": "2023-12-21", "categories": []}
{"id": "2502.05432", "title": "MoFM: A Large-Scale Human Motion Foundation Model", "abstract": "  Foundation Models (FM) have increasingly drawn the attention of researchers\ndue to their scalability and generalization across diverse tasks. Inspired by\nthe success of FMs and the principles that have driven advancements in Large\nLanguage Models (LLMs), we introduce MoFM as a novel Motion Foundation Model.\nMoFM is designed for the semantic understanding of complex human motions in\nboth time and space. To facilitate large-scale training, MotionBook, a\ncomprehensive human motion dictionary of discretized motions is designed and\nemployed. MotionBook utilizes Thermal Cubes to capture spatio-temporal motion\nheatmaps, applying principles from discrete variational models to encode human\nmovements into discrete units for a more efficient and scalable representation.\nMoFM, trained on a large corpus of motion data, provides a foundational\nbackbone adaptable to diverse downstream tasks, supporting paradigms such as\none-shot, unsupervised, and supervised tasks. This versatility makes MoFM\nwell-suited for a wide range of motion-based applications.\n", "date": "2025-02-08", "categories": []}
{"id": "2502.17483", "title": "ConSense: Continually Sensing Human Activity with WiFi via Growing and\n  Picking", "abstract": "  WiFi-based human activity recognition (HAR) holds significant application\npotential across various fields. To handle dynamic environments where new\nactivities are continuously introduced, WiFi-based HAR systems must adapt by\nlearning new concepts without forgetting previously learned ones. Furthermore,\nretaining knowledge from old activities by storing historical exemplar is\nimpractical for WiFi-based HAR due to privacy concerns and limited storage\ncapacity of edge devices. In this work, we propose ConSense, a lightweight and\nfast-adapted exemplar-free class incremental learning framework for WiFi-based\nHAR. The framework leverages the transformer architecture and involves dynamic\nmodel expansion and selective retraining to preserve previously learned\nknowledge while integrating new information. Specifically, during incremental\nsessions, small-scale trainable parameters that are trained specifically on the\ndata of each task are added in the multi-head self-attention layer. In\naddition, a selective retraining strategy that dynamically adjusts the weights\nin multilayer perceptron based on the performance stability of neurons across\ntasks is used. Rather than training the entire model, the proposed strategies\nof dynamic model expansion and selective retraining reduce the overall\ncomputational load while balancing stability on previous tasks and plasticity\non new tasks. Evaluation results on three public WiFi datasets demonstrate that\nConSense not only outperforms several competitive approaches but also requires\nfewer parameters, highlighting its practical utility in class-incremental\nscenarios for HAR.\n", "date": "2025-02-18", "categories": []}
{"id": "2501.02593", "title": "Evolving Skeletons: Motion Dynamics in Action Recognition", "abstract": "  Skeleton-based action recognition has gained significant attention for its\nability to efficiently represent spatiotemporal information in a lightweight\nformat. Most existing approaches use graph-based models to process skeleton\nsequences, where each pose is represented as a skeletal graph structured around\nhuman physical connectivity. Among these, the Spatiotemporal Graph\nConvolutional Network (ST-GCN) has become a widely used framework.\nAlternatively, hypergraph-based models, such as the Hyperformer, capture\nhigher-order correlations, offering a more expressive representation of complex\njoint interactions. A recent advancement, termed Taylor Videos, introduces\nmotion-enhanced skeleton sequences by embedding motion concepts, providing a\nfresh perspective on interpreting human actions in skeleton-based action\nrecognition. In this paper, we conduct a comprehensive evaluation of both\ntraditional skeleton sequences and Taylor-transformed skeletons using ST-GCN\nand Hyperformer models on the NTU-60 and NTU-120 datasets. We compare skeletal\ngraph and hypergraph representations, analyzing static poses against\nmotion-injected poses. Our findings highlight the strengths and limitations of\nTaylor-transformed skeletons, demonstrating their potential to enhance motion\ndynamics while exposing current challenges in fully using their benefits. This\nstudy underscores the need for innovative skeletal modelling techniques to\neffectively handle motion-rich data and advance the field of action\nrecognition.\n", "date": "2025-01-05", "categories": []}
{"id": "2502.16493", "title": "Trunk-branch Contrastive Network with Multi-view Deformable Aggregation\n  for Multi-view Action Recognition", "abstract": "  Multi-view action recognition aims to identify actions in a given multi-view\nscene. Traditional studies initially extracted refined features from each view,\nfollowed by implemented paired interaction and integration, but they\npotentially overlooked the critical local features in each view. When observing\nobjects from multiple perspectives, individuals typically form a comprehensive\nimpression and subsequently fill in specific details. Drawing inspiration from\nthis cognitive process, we propose a novel trunk-branch contrastive network\n(TBCNet) for RGB-based multi-view action recognition. Distinctively, TBCNet\nfirst obtains fused features in the trunk block and then implicitly supplements\nvital details provided by the branch block via contrastive learning, generating\na more informative and comprehensive action representation. Within this\nframework, we construct two core components: the multi-view deformable\naggregation and the trunk-branch contrastive learning. MVDA employed in the\ntrunk block effectively facilitates multi-view feature fusion and adaptive\ncross-view spatio-temporal correlation, where a global aggregation module is\nutilized to emphasize significant spatial information and a composite relative\nposition bias is designed to capture the intra- and cross-view relative\npositions. Moreover, a trunk-branch contrastive loss is constructed between\naggregated features and refined details from each view. By incorporating two\ndistinct weights for positive and negative samples, a weighted trunk-branch\ncontrastive loss is proposed to extract valuable information and emphasize\nsubtle inter-class differences. The effectiveness of TBCNet is verified by\nextensive experiments on four datasets including NTU-RGB+D 60, NTU-RGB+D 120,\nPKU-MMD, and N-UCLA dataset. Compared to other RGB-based methods, our approach\nachieves state-of-the-art performance in cross-subject and cross-setting\nprotocols.\n", "date": "2025-02-23", "categories": []}
{"id": "2308.14024", "title": "Balanced Representation Learning for Long-tailed Skeleton-based Action\n  Recognition", "abstract": "  Skeleton-based action recognition has recently made significant progress.\nHowever, data imbalance is still a great challenge in real-world scenarios. The\nperformance of current action recognition algorithms declines sharply when\ntraining data suffers from heavy class imbalance. The imbalanced data actually\ndegrades the representations learned by these methods and becomes the\nbottleneck for action recognition. How to learn unbiased representations from\nimbalanced action data is the key to long-tailed action recognition. In this\npaper, we propose a novel balanced representation learning method to address\nthe long-tailed problem in action recognition. Firstly, a spatial-temporal\naction exploration strategy is presented to expand the sample space\neffectively, generating more valuable samples in a rebalanced manner. Secondly,\nwe design a detached action-aware learning schedule to further mitigate the\nbias in the representation space. The schedule detaches the representation\nlearning of tail classes from training and proposes an action-aware loss to\nimpose more effective constraints. Additionally, a skip-modal representation is\nproposed to provide complementary structural information. The proposed method\nis validated on four skeleton datasets, NTU RGB+D 60, NTU RGB+D 120, NW-UCLA,\nand Kinetics. It not only achieves consistently large improvement compared to\nthe state-of-the-art (SOTA) methods, but also demonstrates a superior\ngeneralization capacity through extensive experiments. Our code is available at\nhttps://github.com/firework8/BRL.\n", "date": "2023-08-27", "categories": []}
{"id": "2502.14939", "title": "Online hand gesture recognition using Continual Graph Transformers", "abstract": "  Online continuous action recognition has emerged as a critical research area\ndue to its practical implications in real-world applications, such as\nhuman-computer interaction, healthcare, and robotics. Among various modalities,\nskeleton-based approaches have gained significant popularity, demonstrating\ntheir effectiveness in capturing 3D temporal data while ensuring robustness to\nenvironmental variations. However, most existing works focus on segment-based\nrecognition, making them unsuitable for real-time, continuous recognition\nscenarios. In this paper, we propose a novel online recognition system designed\nfor real-time skeleton sequence streaming. Our approach leverages a hybrid\narchitecture combining Spatial Graph Convolutional Networks (S-GCN) for spatial\nfeature extraction and a Transformer-based Graph Encoder (TGE) for capturing\ntemporal dependencies across frames. Additionally, we introduce a continual\nlearning mechanism to enhance model adaptability to evolving data\ndistributions, ensuring robust recognition in dynamic environments. We evaluate\nour method on the SHREC'21 benchmark dataset, demonstrating its superior\nperformance in online hand gesture recognition. Our approach not only achieves\nstate-of-the-art accuracy but also significantly reduces false positive rates,\nmaking it a compelling solution for real-time applications. The proposed system\ncan be seamlessly integrated into various domains, including human-robot\ncollaboration and assistive technologies, where natural and intuitive\ninteraction is crucial.\n", "date": "2025-02-20", "categories": []}
{"id": "2502.13883", "title": "Multi-view Video-Pose Pretraining for Operating Room Surgical Activity\n  Recognition", "abstract": "  Understanding the workflow of surgical procedures in complex operating rooms\nrequires a deep understanding of the interactions between clinicians and their\nenvironment. Surgical activity recognition (SAR) is a key computer vision task\nthat detects activities or phases from multi-view camera recordings. Existing\nSAR models often fail to account for fine-grained clinician movements and\nmulti-view knowledge, or they require calibrated multi-view camera setups and\nadvanced point-cloud processing to obtain better results. In this work, we\npropose a novel calibration-free multi-view multi-modal pretraining framework\ncalled Multiview Pretraining for Video-Pose Surgical Activity Recognition\nPreViPS, which aligns 2D pose and vision embeddings across camera views. Our\nmodel follows CLIP-style dual-encoder architecture: one encoder processes\nvisual features, while the other encodes human pose embeddings. To handle the\ncontinuous 2D human pose coordinates, we introduce a tokenized discrete\nrepresentation to convert the continuous 2D pose coordinates into discrete pose\nembeddings, thereby enabling efficient integration within the dual-encoder\nframework. To bridge the gap between these two modalities, we propose several\npretraining objectives using cross- and in-modality geometric constraints\nwithin the embedding space and incorporating masked pose token prediction\nstrategy to enhance representation learning. Extensive experiments and ablation\nstudies demonstrate improvements over the strong baselines, while\ndata-efficiency experiments on two distinct operating room datasets further\nhighlight the effectiveness of our approach. We highlight the benefits of our\napproach for surgical activity recognition in both multi-view and single-view\nsettings, showcasing its practical applicability in complex surgical\nenvironments. Code will be made available at:\nhttps://github.com/CAMMA-public/PreViPS.\n", "date": "2025-02-19", "categories": []}
{"id": "2502.12582", "title": "Adaptive Prototype Model for Attribute-based Multi-label Few-shot Action\n  Recognition", "abstract": "  In real-world action recognition systems, incorporating more attributes helps\nachieve a more comprehensive understanding of human behavior. However, using a\nsingle model to simultaneously recognize multiple attributes can lead to a\ndecrease in accuracy. In this work, we propose a novel method i.e. Adaptive\nAttribute Prototype Model (AAPM) for human action recognition, which captures\nrich action-relevant attribute information and strikes a balance between\naccuracy and robustness. Firstly, we introduce the Text-Constrain Module (TCM)\nto incorporate textual information from potential labels, and constrain the\nconstruction of different attributes prototype representations. In addition, we\nexplore the Attribute Assignment Method (AAM) to address the issue of training\nbias and increase robustness during the training process.Furthermore, we\nconstruct a new video dataset with attribute-based multi-label called\nMulti-Kinetics for evaluation, which contains various attribute labels (e.g.\naction, scene, object, etc.) related to human behavior. Extensive experiments\ndemonstrate that our AAPM achieves the state-of-the-art performance in both\nattribute-based multi-label few-shot action recognition and single-label\nfew-shot action recognition. The project and dataset are available at an\nanonymous account https://github.com/theAAPM/AAPM\n", "date": "2025-02-18", "categories": []}
{"id": "2304.04220", "title": "Towards Active Learning for Action Spotting in Association Football\n  Videos", "abstract": "  Association football is a complex and dynamic sport, with numerous actions\noccurring simultaneously in each game. Analyzing football videos is challenging\nand requires identifying subtle and diverse spatio-temporal patterns. Despite\nrecent advances in computer vision, current algorithms still face significant\nchallenges when learning from limited annotated data, lowering their\nperformance in detecting these patterns. In this paper, we propose an active\nlearning framework that selects the most informative video samples to be\nannotated next, thus drastically reducing the annotation effort and\naccelerating the training of action spotting models to reach the highest\naccuracy at a faster pace. Our approach leverages the notion of uncertainty\nsampling to select the most challenging video clips to train on next, hastening\nthe learning process of the algorithm. We demonstrate that our proposed active\nlearning framework effectively reduces the required training data for accurate\naction spotting in football videos. We achieve similar performances for action\nspotting with NetVLAD++ on SoccerNet-v2, using only one-third of the dataset,\nindicating significant capabilities for reducing annotation time and improving\ndata efficiency. We further validate our approach on two new datasets that\nfocus on temporally localizing actions of headers and passes, proving its\neffectiveness across different action semantics in football. We believe our\nactive learning framework for action spotting would support further\napplications of action spotting algorithms and accelerate annotation campaigns\nin the sports domain.\n", "date": "2023-04-09", "categories": []}
{"id": "2412.18204", "title": "BoxMAC -- A Boxing Dataset for Multi-label Action Classification", "abstract": "  In competitive combat sports like boxing, analyzing a boxers's performance\nstatics is crucial for evaluating the quantity and variety of punches delivered\nduring bouts. These statistics provide valuable data and feedback, which are\nroutinely used for coaching and performance enhancement. We introduce BoxMAC, a\nreal-world boxing dataset featuring 15 professional boxers and encompassing 13\ndistinct action labels. Comprising over 60,000 frames, our dataset has been\nmeticulously annotated for multiple actions per frame with inputs from a boxing\ncoach. Since two boxers can execute different punches within a single\ntimestamp, this problem falls under the domain of multi-label action\nclassification. We propose a novel architecture for jointly recognizing\nmultiple actions in both individual images and videos. We investigate baselines\nusing deep neural network architectures to address both tasks. We believe that\nBoxMAC will enable researchers and practitioners to develop and evaluate more\nefficient models for performance analysis. With its realistic and diverse\nnature, BoxMAC can serve as a valuable resource for the advancement of boxing\nas a sport\n", "date": "2024-12-24", "categories": []}
{"id": "2502.10713", "title": "Improving action segmentation via explicit similarity measurement", "abstract": "  Existing supervised action segmentation methods depend on the quality of\nframe-wise classification using attention mechanisms or temporal convolutions\nto capture temporal dependencies. Even boundary detection-based methods\nprimarily depend on the accuracy of an initial frame-wise classification, which\ncan overlook precise identification of segments and boundaries in case of\nlow-quality prediction. To address this problem, this paper proposes ASESM\n(Action Segmentation via Explicit Similarity Measurement) to enhance the\nsegmentation accuracy by incorporating explicit similarity evaluation across\nframes and predictions. Our supervised learning architecture uses frame-level\nmulti-resolution features as input to multiple Transformer encoders. The\nresulting multiple frame-wise predictions are used for similarity voting to\nobtain high quality initial prediction. We apply a newly proposed boundary\ncorrection algorithm that operates based on feature similarity between\nconsecutive frames to adjust the boundary locations iteratively through the\nlearning process. The corrected prediction is then further refined through\nmultiple stages of temporal convolutions. As post-processing, we optionally\napply boundary correction again followed by a segment smoothing method that\nremoves outlier classes within segments using similarity measurement between\nconsecutive predictions. Additionally, we propose a fully unsupervised boundary\ndetection-correction algorithm that identifies segment boundaries based solely\non feature similarity without any training. Experiments on 50Salads, GTEA, and\nBreakfast datasets show the effectiveness of both the supervised and\nunsupervised algorithms. Code and models are made available on Github.\n", "date": "2025-02-15", "categories": []}
{"id": "2502.09967", "title": "VicKAM: Visual Conceptual Knowledge Guided Action Map for Weakly\n  Supervised Group Activity Recognition", "abstract": "  Existing weakly supervised group activity recognition methods rely on object\ndetectors or attention mechanisms to capture key areas automatically. However,\nthey overlook the semantic information associated with captured areas, which\nmay adversely affect the recognition performance. In this paper, we propose a\nnovel framework named Visual Conceptual Knowledge Guided Action Map (VicKAM)\nwhich effectively captures the locations of individual actions and integrates\nthem with action semantics for weakly supervised group activity recognition.It\ngenerates individual action prototypes from training set as visual conceptual\nknowledge to bridge action semantics and visual representations. Guided by this\nknowledge, VicKAM produces action maps that indicate the likelihood of each\naction occurring at various locations, based on image correlation theorem. It\nfurther augments individual action maps using group activity related\nstatistical information, representing individual action distribution under\ndifferent group activities, to establish connections between action maps and\nspecific group activities. The augmented action map is incorporated with action\nsemantic representations for group activity recognition.Extensive experiments\non two public benchmarks, the Volleyball and the NBA datasets, demonstrate the\neffectiveness of our proposed method, even in cases of limited training data.\nThe code will be released later.\n", "date": "2025-02-14", "categories": []}
{"id": "2209.09194", "title": "Differentiable Frequency-based Disentanglement for Aerial Video Action\n  Recognition", "abstract": "  We present a learning algorithm for human activity recognition in videos. Our\napproach is designed for UAV videos, which are mainly acquired from obliquely\nplaced dynamic cameras that contain a human actor along with background motion.\nTypically, the human actors occupy less than one-tenth of the spatial\nresolution. Our approach simultaneously harnesses the benefits of frequency\ndomain representations, a classical analysis tool in signal processing, and\ndata driven neural networks. We build a differentiable static-dynamic frequency\nmask prior to model the salient static and dynamic pixels in the video, crucial\nfor the underlying task of action recognition. We use this differentiable mask\nprior to enable the neural network to intrinsically learn disentangled feature\nrepresentations via an identity loss function. Our formulation empowers the\nnetwork to inherently compute disentangled salient features within its layers.\nFurther, we propose a cost-function encapsulating temporal relevance and\nspatial content to sample the most important frame within uniformly spaced\nvideo segments. We conduct extensive experiments on the UAV Human dataset and\nthe NEC Drone dataset and demonstrate relative improvements of 5.72% - 13.00%\nover the state-of-the-art and 14.28% - 38.05% over the corresponding baseline\nmodel.\n", "date": "2022-09-15", "categories": []}
{"id": "2412.09895", "title": "Building a Multi-modal Spatiotemporal Expert for Zero-shot Action\n  Recognition with CLIP", "abstract": "  Zero-shot action recognition (ZSAR) requires collaborative multi-modal\nspatiotemporal understanding. However, finetuning CLIP directly for ZSAR yields\nsuboptimal performance, given its inherent constraints in capturing essential\ntemporal dynamics from both vision and text perspectives, especially when\nencountering novel actions with fine-grained spatiotemporal discrepancies. In\nthis work, we propose Spatiotemporal Dynamic Duo (STDD), a novel CLIP-based\nframework to comprehend multi-modal spatiotemporal dynamics synergistically.\nFor the vision side, we propose an efficient Space-time Cross Attention, which\ncaptures spatiotemporal dynamics flexibly with simple yet effective operations\napplied before and after spatial attention, without adding additional\nparameters or increasing computational complexity. For the semantic side, we\nconduct spatiotemporal text augmentation by comprehensively constructing an\nAction Semantic Knowledge Graph (ASKG) to derive nuanced text prompts. The ASKG\nelaborates on static and dynamic concepts and their interrelations, based on\nthe idea of decomposing actions into spatial appearances and temporal motions.\nDuring the training phase, the frame-level video representations are\nmeticulously aligned with prompt-level nuanced text representations, which are\nconcurrently regulated by the video representations from the frozen CLIP to\nenhance generalizability. Extensive experiments validate the effectiveness of\nour approach, which consistently surpasses state-of-the-art approaches on\npopular video benchmarks (i.e., Kinetics-600, UCF101, and HMDB51) under\nchallenging ZSAR settings.\n", "date": "2024-12-13", "categories": []}
{"id": "2501.17550", "title": "Action Recognition Using Temporal Shift Module and Ensemble Learning", "abstract": "  This paper presents the first-rank solution for the Multi-Modal Action\nRecognition Challenge, part of the Multi-Modal Visual Pattern Recognition\nWorkshop at the \\acl{ICPR} 2024. The competition aimed to recognize human\nactions using a diverse dataset of 20 action classes, collected from\nmulti-modal sources. The proposed approach is built upon the \\acl{TSM}, a\ntechnique aimed at efficiently capturing temporal dynamics in video data,\nincorporating multiple data input types. Our strategy included transfer\nlearning to leverage pre-trained models, followed by meticulous fine-tuning on\nthe challenge's specific dataset to optimize performance for the 20 action\nclasses. We carefully selected a backbone network to balance computational\nefficiency and recognition accuracy and further refined the model using an\nensemble technique that integrates outputs from different modalities. This\nensemble approach proved crucial in boosting the overall performance. Our\nsolution achieved a perfect top-1 accuracy on the test set, demonstrating the\neffectiveness of the proposed approach in recognizing human actions across 20\nclasses. Our code is available online https://github.com/ffyyytt/TSM-MMVPR.\n", "date": "2025-01-29", "categories": []}
{"id": "2502.03549", "title": "Kronecker Mask and Interpretive Prompts are Language-Action Video\n  Learners", "abstract": "  Contrastive language-image pretraining (CLIP) has significantly advanced\nimage-based vision learning. A pressing topic subsequently arises: how can we\neffectively adapt CLIP to the video domain? Recent studies have focused on\nadjusting either the textual or visual branch of CLIP for action recognition.\nHowever, we argue that adaptations of both branches are crucial. In this paper,\nwe propose \\textbf{CLAVER}: a \\textbf{C}ontrastive\n\\textbf{L}anguage-\\textbf{A}ction \\textbf{V}ideo Learn\\textbf{er}, designed to\nshift CLIP's focus from the alignment of static visual objects and concrete\nnouns to the alignment of dynamic action behaviors and abstract verbs.\nSpecifically, we introduce a novel Kronecker mask attention for temporal\nmodeling. Our tailored Kronecker mask offers three benefits 1) it expands the\ntemporal receptive field for each token, 2) it serves as an effective\nspatiotemporal heterogeneity inductive bias, mitigating the issue of\nspatiotemporal homogenization, and 3) it can be seamlessly plugged into\ntransformer-based models. Regarding the textual branch, we leverage large\nlanguage models to generate diverse, sentence-level and semantically rich\ninterpretive prompts of actions, which shift the model's focus towards the verb\ncomprehension. Extensive experiments on various benchmarks and learning\nscenarios demonstrate the superiority and generality of our approach.\n", "date": "2025-02-05", "categories": []}
{"id": "2406.03857", "title": "MuJo: Multimodal Joint Feature Space Learning for Human Activity\n  Recognition", "abstract": "  Human activity recognition (HAR) is a long-standing problem in artificial\nintelligence with applications in a broad range of areas, including healthcare,\nsports and fitness, security, and more. The performance of HAR in real-world\nsettings is strongly dependent on the type and quality of the input signal that\ncan be acquired. Given an unobstructed, high-quality camera view of a scene,\ncomputer vision systems, in particular in conjunction with foundation models,\ncan today fairly reliably distinguish complex activities. On the other hand,\nrecognition using modalities such as wearable sensors (which are often more\nbroadly available, e.g., in mobile phones and smartwatches) is a more difficult\nproblem, as the signals often contain less information and labeled training\ndata is more difficult to acquire. To alleviate the need for labeled data, we\nintroduce our comprehensive Fitness Multimodal Activity Dataset (FiMAD) in this\nwork, which can be used with the proposed pre-training method MuJo (Multimodal\nJoint Feature Space Learning) to enhance HAR performance across various\nmodalities. FiMAD was created using YouTube fitness videos and contains\nparallel video, language, pose, and simulated IMU sensor data. MuJo utilizes\nthis dataset to learn a joint feature space for these modalities. We show that\nclassifiers pre-trained on FiMAD can increase the performance on real HAR\ndatasets such as MM-Fit, MyoGym, MotionSense, and MHEALTH. For instance, on\nMM-Fit, we achieve a Macro F1-Score of up to 0.855 when fine-tuning on only 2%\nof the training data and 0.942 when utilizing the complete training set for\nclassification tasks. We compare our approach with other self-supervised ones\nand show that, unlike them, ours consistently improves compared to the baseline\nnetwork performance while also providing better data efficiency.\n", "date": "2024-06-06", "categories": []}
{"id": "2308.10557", "title": "Local Spherical Harmonics Improve Skeleton-Based Hand Action Recognition", "abstract": "  Hand action recognition is essential. Communication, human-robot\ninteractions, and gesture control are dependent on it. Skeleton-based action\nrecognition traditionally includes hands, which belong to the classes which\nremain challenging to correctly recognize to date. We propose a method\nspecifically designed for hand action recognition which uses relative angular\nembeddings and local Spherical Harmonics to create novel hand representations.\nThe use of Spherical Harmonics creates rotation-invariant representations which\nmake hand action recognition even more robust against inter-subject differences\nand viewpoint changes. We conduct extensive experiments on the hand joints in\nthe First-Person Hand Action Benchmark with RGB-D Videos and 3D Hand Pose\nAnnotations, and on the NTU RGB+D 120 dataset, demonstrating the benefit of\nusing Local Spherical Harmonics Representations. Our code is available at\nhttps://github.com/KathPra/LSHR_LSHT.\n", "date": "2023-08-21", "categories": []}
{"id": "2412.13708", "title": "JoVALE: Detecting Human Actions in Video Using Audiovisual and Language\n  Contexts", "abstract": "  Video Action Detection (VAD) entails localizing and categorizing action\ninstances within videos, which inherently consist of diverse information\nsources such as audio, visual cues, and surrounding scene contexts. Leveraging\nthis multi-modal information effectively for VAD poses a significant challenge,\nas the model must identify action-relevant cues with precision. In this study,\nwe introduce a novel multi-modal VAD architecture, referred to as the Joint\nActor-centric Visual, Audio, Language Encoder (JoVALE). JoVALE is the first VAD\nmethod to integrate audio and visual features with scene descriptive context\nsourced from large-capacity image captioning models. At the heart of JoVALE is\nthe actor-centric aggregation of audio, visual, and scene descriptive\ninformation, enabling adaptive integration of crucial features for recognizing\neach actor's actions. We have developed a Transformer-based architecture, the\nActor-centric Multi-modal Fusion Network, specifically designed to capture the\ndynamic interactions among actors and their multi-modal contexts. Our\nevaluation on three prominent VAD benchmarks, including AVA, UCF101-24, and\nJHMDB51-21, demonstrates that incorporating multi-modal information\nsignificantly enhances performance, setting new state-of-the-art performances\nin the field.\n", "date": "2024-12-18", "categories": []}
{"id": "2501.11007", "title": "HFGCN:Hypergraph Fusion Graph Convolutional Networks for Skeleton-Based\n  Action Recognition", "abstract": "  In recent years, action recognition has received much attention and wide\napplication due to its important role in video understanding. Most of the\nresearches on action recognition methods focused on improving the performance\nvia various deep learning methods rather than the classification of skeleton\npoints. The topological modeling between skeleton points and body parts was\nseldom considered. Although some studies have used a data-driven approach to\nclassify the topology of the skeleton point, the nature of the skeleton point\nin terms of kinematics has not been taken into consideration. Therefore, in\nthis paper, we draw on the theory of kinematics to adapt the topological\nrelations of the skeleton point and propose a topological relation\nclassification based on body parts and distance from core of body. To\nsynthesize these topological relations for action recognition, we propose a\nnovel Hypergraph Fusion Graph Convolutional Network (HFGCN). In particular, the\nproposed model is able to focus on the human skeleton points and the different\nbody parts simultaneously, and thus construct the topology, which improves the\nrecognition accuracy obviously. We use a hypergraph to represent the\ncategorical relationships of these skeleton points and incorporate the\nhypergraph into a graph convolution network to model the higher-order\nrelationships among the skeleton points and enhance the feature representation\nof the network. In addition, our proposed hypergraph attention module and\nhypergraph graph convolution module optimize topology modeling in temporal and\nchannel dimensions, respectively, to further enhance the feature representation\nof the network. We conducted extensive experiments on three widely used\ndatasets.The results validate that our proposed method can achieve the best\nperformance when compared with the state-of-the-art skeleton-based methods.\n", "date": "2025-01-19", "categories": []}
{"id": "2401.07584", "title": "Collaboratively Self-supervised Video Representation Learning for Action\n  Recognition", "abstract": "  Considering the close connection between action recognition and human pose\nestimation, we design a Collaboratively Self-supervised Video Representation\n(CSVR) learning framework specific to action recognition by jointly factoring\nin generative pose prediction and discriminative context matching as pretext\ntasks. Specifically, our CSVR consists of three branches: a generative pose\nprediction branch, a discriminative context matching branch, and a video\ngenerating branch. Among them, the first one encodes dynamic motion feature by\nutilizing Conditional-GAN to predict the human poses of future frames, and the\nsecond branch extracts static context features by contrasting positive and\nnegative video feature and I-frame feature pairs. The third branch is designed\nto generate both current and future video frames, for the purpose of\ncollaboratively improving dynamic motion features and static context features.\nExtensive experiments demonstrate that our method achieves state-of-the-art\nperformance on multiple popular video datasets.\n", "date": "2024-01-15", "categories": []}
{"id": "2412.17640", "title": "Hierarchical Vector Quantization for Unsupervised Action Segmentation", "abstract": "  In this work, we address unsupervised temporal action segmentation, which\nsegments a set of long, untrimmed videos into semantically meaningful segments\nthat are consistent across videos. While recent approaches combine\nrepresentation learning and clustering in a single step for this task, they do\nnot cope with large variations within temporal segments of the same class. To\naddress this limitation, we propose a novel method, termed Hierarchical Vector\nQuantization (HVQ), that consists of two subsequent vector quantization\nmodules. This results in a hierarchical clustering where the additional\nsubclusters cover the variations within a cluster. We demonstrate that our\napproach captures the distribution of segment lengths much better than the\nstate of the art. To this end, we introduce a new metric based on the\nJensen-Shannon Distance (JSD) for unsupervised temporal action segmentation. We\nevaluate our approach on three public datasets, namely Breakfast, YouTube\nInstructional and IKEA ASM. Our approach outperforms the state of the art in\nterms of F1 score, recall and JSD.\n", "date": "2024-12-23", "categories": []}
{"id": "2501.13829", "title": "MV-GMN: State Space Model for Multi-View Action Recognition", "abstract": "  Recent advancements in multi-view action recognition have largely relied on\nTransformer-based models. While effective and adaptable, these models often\nrequire substantial computational resources, especially in scenarios with\nmultiple views and multiple temporal sequences. Addressing this limitation,\nthis paper introduces the MV-GMN model, a state-space model specifically\ndesigned to efficiently aggregate multi-modal data (RGB and skeleton),\nmulti-view perspectives, and multi-temporal information for action recognition\nwith reduced computational complexity. The MV-GMN model employs an innovative\nMulti-View Graph Mamba network comprising a series of MV-GMN blocks. Each block\nincludes a proposed Bidirectional State Space Block and a GCN module. The\nBidirectional State Space Block introduces four scanning strategies, including\nview-prioritized and time-prioritized approaches. The GCN module leverages\nrule-based and KNN-based methods to construct the graph network, effectively\nintegrating features from different viewpoints and temporal instances.\nDemonstrating its efficacy, MV-GMN outperforms the state-of-the-arts on several\ndatasets, achieving notable accuracies of 97.3\\% and 96.7\\% on the NTU RGB+D\n120 dataset in cross-subject and cross-view scenarios, respectively. MV-GMN\nalso surpasses Transformer-based baselines while requiring only linear\ninference complexity, underscoring the model's ability to reduce computational\nload and enhance the scalability and applicability of multi-view action\nrecognition technologies.\n", "date": "2025-01-23", "categories": []}
{"id": "2501.12681", "title": "Can masking background and object reduce static bias for zero-shot\n  action recognition?", "abstract": "  In this paper, we address the issue of static bias in zero-shot action\nrecognition. Action recognition models need to represent the action itself, not\nthe appearance. However, some fully-supervised works show that models often\nrely on static appearances, such as the background and objects, rather than\nhuman actions. This issue, known as static bias, has not been investigated for\nzero-shot. Although CLIP-based zero-shot models are now common, it remains\nunclear if they sufficiently focus on human actions, as CLIP primarily captures\nappearance features related to languages. In this paper, we investigate the\ninfluence of static bias in zero-shot action recognition with CLIP-based\nmodels. Our approach involves masking backgrounds, objects, and people\ndifferently during training and validation. Experiments with masking background\nshow that models depend on background bias as their performance decreases for\nKinetics400. However, for Mimetics, which has a weak background bias, masking\nthe background leads to improved performance even if the background is masked\nduring validation. Furthermore, masking both the background and objects in\ndifferent colors improves performance for SSv2, which has a strong object bias.\nThese results suggest that masking the background or objects during training\nprevents models from overly depending on static bias and makes them focus more\non human action.\n", "date": "2025-01-22", "categories": []}
{"id": "2501.13066", "title": "SMART-Vision: Survey of Modern Action Recognition Techniques in Vision", "abstract": "  Human Action Recognition (HAR) is a challenging domain in computer vision,\ninvolving recognizing complex patterns by analyzing the spatiotemporal dynamics\nof individuals' movements in videos. These patterns arise in sequential data,\nsuch as video frames, which are often essential to accurately distinguish\nactions that would be ambiguous in a single image. HAR has garnered\nconsiderable interest due to its broad applicability, ranging from robotics and\nsurveillance systems to sports motion analysis, healthcare, and the burgeoning\nfield of autonomous vehicles. While several taxonomies have been proposed to\ncategorize HAR approaches in surveys, they often overlook hybrid methodologies\nand fail to demonstrate how different models incorporate various architectures\nand modalities. In this comprehensive survey, we present the novel SMART-Vision\ntaxonomy, which illustrates how innovations in deep learning for HAR complement\none another, leading to hybrid approaches beyond traditional categories. Our\nsurvey provides a clear roadmap from foundational HAR works to current\nstate-of-the-art systems, highlighting emerging research directions and\naddressing unresolved challenges in discussion sections for architectures\nwithin the HAR domain. We provide details of the research datasets that various\napproaches used to measure and compare goodness HAR approaches. We also explore\nthe rapidly emerging field of Open-HAR systems, which challenges HAR systems by\npresenting samples from unknown, novel classes during test time.\n", "date": "2025-01-22", "categories": []}
{"id": "2410.06698", "title": "Fourier-based Action Recognition for Wildlife Behavior Quantification\n  with Event Cameras", "abstract": "  Event cameras are novel bio-inspired vision sensors that measure pixel-wise\nbrightness changes asynchronously instead of images at a given frame rate. They\noffer promising advantages, namely a high dynamic range, low latency, and\nminimal motion blur. Modern computer vision algorithms often rely on artificial\nneural network approaches, which require image-like representations of the data\nand cannot fully exploit the characteristics of event data. We propose\napproaches to action recognition based on the Fourier Transform. The approaches\nare intended to recognize oscillating motion patterns commonly present in\nnature. In particular, we apply our approaches to a recent dataset of breeding\npenguins annotated for \"ecstatic display\", a behavior where the observed\npenguins flap their wings at a certain frequency. We find that our approaches\nare both simple and effective, producing slightly lower results than a deep\nneural network (DNN) while relying just on a tiny fraction of the parameters\ncompared to the DNN (five orders of magnitude fewer parameters). They work well\ndespite the uncontrolled, diverse data present in the dataset. We hope this\nwork opens a new perspective on event-based processing and action recognition.\n", "date": "2024-10-09", "categories": []}
{"id": "2212.12141", "title": "Human Activity Recognition in an Open World", "abstract": "  Managing novelty in perception-based human activity recognition (HAR) is\ncritical in realistic settings to improve task performance over time and ensure\nsolution generalization outside of prior seen samples. Novelty manifests in HAR\nas unseen samples, activities, objects, environments, and sensor changes, among\nother ways. Novelty may be task-relevant, such as a new class or new features,\nor task-irrelevant resulting in nuisance novelty, such as never before seen\nnoise, blur, or distorted video recordings. To perform HAR optimally,\nalgorithmic solutions must be tolerant to nuisance novelty, and learn over time\nin the face of novelty. This paper 1) formalizes the definition of novelty in\nHAR building upon the prior definition of novelty in classification tasks, 2)\nproposes an incremental open world learning (OWL) protocol and applies it to\nthe Kinetics datasets to generate a new benchmark KOWL-718, 3) analyzes the\nperformance of current state-of-the-art HAR models when novelty is introduced\nover time, 4) provides a containerized and packaged pipeline for reproducing\nthe OWL protocol and for modifying for any future updates to Kinetics. The\nexperimental analysis includes an ablation study of how the different models\nperform under various conditions as annotated by Kinetics-AVA. The protocol as\nan algorithm for reproducing experiments using the KOWL-718 benchmark will be\npublicly released with code and containers at\nhttps://github.com/prijatelj/human-activity-recognition-in-an-open-world. The\ncode may be used to analyze different annotations and subsets of the Kinetics\ndatasets in an incremental open world fashion, as well as be extended as\nfurther updates to Kinetics are released.\n", "date": "2022-12-23", "categories": []}
{"id": "2501.08931", "title": "Visual WetlandBirds Dataset: Bird Species Identification and Behavior\n  Recognition in Videos", "abstract": "  The current biodiversity loss crisis makes animal monitoring a relevant field\nof study. In light of this, data collected through monitoring can provide\nessential insights, and information for decision-making aimed at preserving\nglobal biodiversity. Despite the importance of such data, there is a notable\nscarcity of datasets featuring videos of birds, and none of the existing\ndatasets offer detailed annotations of bird behaviors in video format. In\nresponse to this gap, our study introduces the first fine-grained video dataset\nspecifically designed for bird behavior detection and species classification.\nThis dataset addresses the need for comprehensive bird video datasets and\nprovides detailed data on bird actions, facilitating the development of deep\nlearning models to recognize these, similar to the advancements made in human\naction recognition. The proposed dataset comprises 178 videos recorded in\nSpanish wetlands, capturing 13 different bird species performing 7 distinct\nbehavior classes. In addition, we also present baseline results using state of\nthe art models on two tasks: bird behavior recognition and species\nclassification.\n", "date": "2025-01-15", "categories": []}
{"id": "2501.07039", "title": "IoT-Based Real-Time Medical-Related Human Activity Recognition Using\n  Skeletons and Multi-Stage Deep Learning for Healthcare", "abstract": "  The Internet of Things (IoT) and mobile technology have significantly\ntransformed healthcare by enabling real-time monitoring and diagnosis of\npatients. Recognizing medical-related human activities (MRHA) is pivotal for\nhealthcare systems, particularly for identifying actions that are critical to\npatient well-being. However, challenges such as high computational demands, low\naccuracy, and limited adaptability persist in Human Motion Recognition (HMR).\nWhile some studies have integrated HMR with IoT for real-time healthcare\napplications, limited research has focused on recognizing MRHA as essential for\neffective patient monitoring. This study proposes a novel HMR method for MRHA\ndetection, leveraging multi-stage deep learning techniques integrated with IoT.\nThe approach employs EfficientNet to extract optimized spatial features from\nskeleton frame sequences using seven Mobile Inverted Bottleneck Convolutions\n(MBConv) blocks, followed by ConvLSTM to capture spatio-temporal patterns. A\nclassification module with global average pooling, a fully connected layer, and\na dropout layer generates the final predictions. The model is evaluated on the\nNTU RGB+D 120 and HMDB51 datasets, focusing on MRHA, such as sneezing, falling,\nwalking, sitting, etc. It achieves 94.85% accuracy for cross-subject\nevaluations and 96.45% for cross-view evaluations on NTU RGB+D 120, along with\n89.00% accuracy on HMDB51. Additionally, the system integrates IoT capabilities\nusing a Raspberry Pi and GSM module, delivering real-time alerts via Twilios\nSMS service to caregivers and patients. This scalable and efficient solution\nbridges the gap between HMR and IoT, advancing patient monitoring, improving\nhealthcare outcomes, and reducing costs.\n", "date": "2025-01-13", "categories": []}
{"id": "2501.07408", "title": "Initial Findings on Sensor based Open Vocabulary Activity Recognition\n  via Text Embedding Inversion", "abstract": "  Conventional human activity recognition (HAR) relies on classifiers trained\nto predict discrete activity classes, inherently limiting recognition to\nactivities explicitly present in the training set. Such classifiers would\ninvariably fail, putting zero likelihood, when encountering unseen activities.\nWe propose Open Vocabulary HAR (OV-HAR), a framework that overcomes this\nlimitation by first converting each activity into natural language and breaking\nit into a sequence of elementary motions. This descriptive text is then encoded\ninto a fixed-size embedding. The model is trained to regress this embedding,\nwhich is subsequently decoded back into natural language using a pre-trained\nembedding inversion model. Unlike other works that rely on auto-regressive\nlarge language models (LLMs) at their core, OV-HAR achieves open vocabulary\nrecognition without the computational overhead of such models. The generated\ntext can be transformed into a single activity class using LLM prompt\nengineering. We have evaluated our approach on different modalities, including\nvision (pose), IMU, and pressure sensors, demonstrating robust generalization\nacross unseen activities and modalities, offering a fundamentally different\nparadigm from contemporary classifiers.\n", "date": "2025-01-13", "categories": []}
{"id": "2501.05066", "title": "Improving Skeleton-based Action Recognition with Interactive Object\n  Information", "abstract": "  Human skeleton information is important in skeleton-based action recognition,\nwhich provides a simple and efficient way to describe human pose. However,\nexisting skeleton-based methods focus more on the skeleton, ignoring the\nobjects interacting with humans, resulting in poor performance in recognizing\nactions that involve object interactions. We propose a new action recognition\nframework introducing object nodes to supplement absent interactive object\ninformation. We also propose Spatial Temporal Variable Graph Convolutional\nNetworks (ST-VGCN) to effectively model the Variable Graph (VG) containing\nobject nodes. Specifically, in order to validate the role of interactive object\ninformation, by leveraging a simple self-training approach, we establish a new\ndataset, JXGC 24, and an extended dataset, NTU RGB+D+Object 60, including more\nthan 2 million additional object nodes. At the same time, we designe the\nVariable Graph construction method to accommodate a variable number of nodes\nfor graph structure. Additionally, we are the first to explore the overfitting\nissue introduced by incorporating additional object information, and we propose\na VG-based data augmentation method to address this issue, called Random Node\nAttack. Finally, regarding the network structure, we introduce two fusion\nmodules, CAF and WNPool, along with a novel Node Balance Loss, to enhance the\ncomprehensive performance by effectively fusing and balancing skeleton and\nobject node information. Our method surpasses the previous state-of-the-art on\nmultiple skeleton-based action recognition benchmarks. The accuracy of our\nmethod on NTU RGB+D 60 cross-subject split is 96.7\\%, and on cross-view split,\nit is 99.2\\%.\n", "date": "2025-01-09", "categories": []}
{"id": "2305.18710", "title": "High-Performance Inference Graph Convolutional Networks for\n  Skeleton-Based Action Recognition", "abstract": "  Recently, the significant achievements have been made in skeleton-based human\naction recognition with the emergence of graph convolutional networks (GCNs).\nHowever, the state-of-the-art (SOTA) models used for this task focus on\nconstructing more complex higher-order connections between joint nodes to\ndescribe skeleton information, which leads to complex inference processes and\nhigh computational costs. To address the slow inference speed caused by overly\ncomplex model structures, we introduce re-parameterization and\nover-parameterization techniques to GCNs and propose two novel high-performance\ninference GCNs, namely HPI-GCN-RP and HPI-GCN-OP. After the completion of model\ntraining, model parameters are fixed. HPI-GCN-RP adopts re-parameterization\ntechnique to transform high-performance training model into fast inference\nmodel through linear transformations, which achieves a higher inference speed\nwith competitive model performance. HPI-GCN-OP further utilizes\nover-parameterization technique to achieve higher performance improvement by\nintroducing additional inference parameters, albeit with slightly decreased\ninference speed. The experimental results on the two skeleton-based action\nrecognition datasets demonstrate the effectiveness of our approach. Our\nHPI-GCN-OP achieves performance comparable to the current SOTA models, with\ninference speeds five times faster. Specifically, our HPI-GCN-OP achieves an\naccuracy of 93\\% on the cross-subject split of the NTU-RGB+D 60 dataset, and\n90.1\\% on the cross-subject benchmark of the NTU-RGB+D 120 dataset. Code is\navailable at github.com/lizaowo/HPI-GCN.\n", "date": "2023-05-30", "categories": []}
{"id": "2501.00432", "title": "OV-HHIR: Open Vocabulary Human Interaction Recognition Using Cross-modal\n  Integration of Large Language Models", "abstract": "  Understanding human-to-human interactions, especially in contexts like public\nsecurity surveillance, is critical for monitoring and maintaining safety.\nTraditional activity recognition systems are limited by fixed vocabularies,\npredefined labels, and rigid interaction categories that often rely on\nchoreographed videos and overlook concurrent interactive groups. These\nlimitations make such systems less adaptable to real-world scenarios, where\ninteractions are diverse and unpredictable. In this paper, we propose an open\nvocabulary human-to-human interaction recognition (OV-HHIR) framework that\nleverages large language models to generate open-ended textual descriptions of\nboth seen and unseen human interactions in open-world settings without being\nconfined to a fixed vocabulary. Additionally, we create a comprehensive,\nlarge-scale human-to-human interaction dataset by standardizing and combining\nexisting public human interaction datasets into a unified benchmark. Extensive\nexperiments demonstrate that our method outperforms traditional\nfixed-vocabulary classification systems and existing cross-modal language\nmodels for video understanding, setting the stage for more intelligent and\nadaptable visual understanding systems in surveillance and beyond.\n", "date": "2024-12-31", "categories": []}
{"id": "2501.00795", "title": "Multimodal Large Models Are Effective Action Anticipators", "abstract": "  The task of long-term action anticipation demands solutions that can\neffectively model temporal dynamics over extended periods while deeply\nunderstanding the inherent semantics of actions. Traditional approaches, which\nprimarily rely on recurrent units or Transformer layers to capture long-term\ndependencies, often fall short in addressing these challenges. Large Language\nModels (LLMs), with their robust sequential modeling capabilities and extensive\ncommonsense knowledge, present new opportunities for long-term action\nanticipation. In this work, we introduce the ActionLLM framework, a novel\napproach that treats video sequences as successive tokens, leveraging LLMs to\nanticipate future actions. Our baseline model simplifies the LLM architecture\nby setting future tokens, incorporating an action tuning module, and reducing\nthe textual decoder layer to a linear layer, enabling straightforward action\nprediction without the need for complex instructions or redundant descriptions.\nTo further harness the commonsense reasoning of LLMs, we predict action\ncategories for observed frames and use sequential textual clues to guide\nsemantic understanding. In addition, we introduce a Cross-Modality Interaction\nBlock, designed to explore the specificity within each modality and capture\ninteractions between vision and textual modalities, thereby enhancing\nmultimodal tuning. Extensive experiments on benchmark datasets demonstrate the\nsuperiority of the proposed ActionLLM framework, encouraging a promising\ndirection to explore LLMs in the context of action anticipation. Code is\navailable at https://github.com/2tianyao1/ActionLLM.git.\n", "date": "2025-01-01", "categories": []}
{"id": "2501.00935", "title": "Multiscaled Multi-Head Attention-based Video Transformer Network for\n  Hand Gesture Recognition", "abstract": "  Dynamic gesture recognition is one of the challenging research areas due to\nvariations in pose, size, and shape of the signer's hand. In this letter,\nMultiscaled Multi-Head Attention Video Transformer Network (MsMHA-VTN) for\ndynamic hand gesture recognition is proposed. A pyramidal hierarchy of\nmultiscale features is extracted using the transformer multiscaled head\nattention model. The proposed model employs different attention dimensions for\neach head of the transformer which enables it to provide attention at the\nmultiscale level. Further, in addition to single modality, recognition\nperformance using multiple modalities is examined. Extensive experiments\ndemonstrate the superior performance of the proposed MsMHA-VTN with an overall\naccuracy of 88.22\\% and 99.10\\% on NVGesture and Briareo datasets,\nrespectively.\n", "date": "2025-01-01", "categories": []}
{"id": "2501.01040", "title": "Event Masked Autoencoder: Point-wise Action Recognition with Event-Based\n  Cameras", "abstract": "  Dynamic vision sensors (DVS) are bio-inspired devices that capture visual\ninformation in the form of asynchronous events, which encode changes in pixel\nintensity with high temporal resolution and low latency. These events provide\nrich motion cues that can be exploited for various computer vision tasks, such\nas action recognition. However, most existing DVS-based action recognition\nmethods lose temporal information during data transformation or suffer from\nnoise and outliers caused by sensor imperfections or environmental factors. To\naddress these challenges, we propose a novel framework that preserves and\nexploits the spatiotemporal structure of event data for action recognition. Our\nframework consists of two main components: 1) a point-wise event masked\nautoencoder (MAE) that learns a compact and discriminative representation of\nevent patches by reconstructing them from masked raw event camera points data;\n2) an improved event points patch generation algorithm that leverages an event\ndata inlier model and point-wise data augmentation techniques to enhance the\nquality and diversity of event points patches. To the best of our knowledge,\nour approach introduces the pre-train method into event camera raw points data\nfor the first time, and we propose a novel event points patch embedding to\nutilize transformer-based models on event cameras.\n", "date": "2025-01-02", "categories": []}
{"id": "2501.01245", "title": "SeFAR: Semi-supervised Fine-grained Action Recognition with Temporal\n  Perturbation and Learning Stabilization", "abstract": "  Human action understanding is crucial for the advancement of multimodal\nsystems. While recent developments, driven by powerful large language models\n(LLMs), aim to be general enough to cover a wide range of categories, they\noften overlook the need for more specific capabilities. In this work, we\naddress the more challenging task of Fine-grained Action Recognition (FAR),\nwhich focuses on detailed semantic labels within shorter temporal duration\n(e.g., \"salto backward tucked with 1 turn\"). Given the high costs of annotating\nfine-grained labels and the substantial data needed for fine-tuning LLMs, we\npropose to adopt semi-supervised learning (SSL). Our framework, SeFAR,\nincorporates several innovative designs to tackle these challenges.\nSpecifically, to capture sufficient visual details, we construct Dual-level\ntemporal elements as more effective representations, based on which we design a\nnew strong augmentation strategy for the Teacher-Student learning paradigm\nthrough involving moderate temporal perturbation. Furthermore, to handle the\nhigh uncertainty within the teacher model's predictions for FAR, we propose the\nAdaptive Regulation to stabilize the learning process. Experiments show that\nSeFAR achieves state-of-the-art performance on two FAR datasets, FineGym and\nFineDiving, across various data scopes. It also outperforms other\nsemi-supervised methods on two classical coarse-grained datasets, UCF101 and\nHMDB51. Further analysis and ablation studies validate the effectiveness of our\ndesigns. Additionally, we show that the features extracted by our SeFAR could\nlargely promote the ability of multimodal foundation models to understand\nfine-grained and domain-specific semantics.\n", "date": "2025-01-02", "categories": []}
{"id": "2410.07153", "title": "CHASE: Learning Convex Hull Adaptive Shift for Skeleton-based\n  Multi-Entity Action Recognition", "abstract": "  Skeleton-based multi-entity action recognition is a challenging task aiming\nto identify interactive actions or group activities involving multiple diverse\nentities. Existing models for individuals often fall short in this task due to\nthe inherent distribution discrepancies among entity skeletons, leading to\nsuboptimal backbone optimization. To this end, we introduce a Convex Hull\nAdaptive Shift based multi-Entity action recognition method (CHASE), which\nmitigates inter-entity distribution gaps and unbiases subsequent backbones.\nSpecifically, CHASE comprises a learnable parameterized network and an\nauxiliary objective. The parameterized network achieves plausible,\nsample-adaptive repositioning of skeleton sequences through two key components.\nFirst, the Implicit Convex Hull Constrained Adaptive Shift ensures that the new\norigin of the coordinate system is within the skeleton convex hull. Second, the\nCoefficient Learning Block provides a lightweight parameterization of the\nmapping from skeleton sequences to their specific coefficients in convex\ncombinations. Moreover, to guide the optimization of this network for\ndiscrepancy minimization, we propose the Mini-batch Pair-wise Maximum Mean\nDiscrepancy as the additional objective. CHASE operates as a sample-adaptive\nnormalization method to mitigate inter-entity distribution discrepancies,\nthereby reducing data bias and improving the subsequent classifier's\nmulti-entity action recognition performance. Extensive experiments on six\ndatasets, including NTU Mutual 11/26, H2O, Assembly101, Collective Activity and\nVolleyball, consistently verify our approach by seamlessly adapting to\nsingle-entity backbones and boosting their performance in multi-entity\nscenarios. Our code is publicly available at https://github.com/Necolizer/CHASE .\n", "date": "2024-10-09", "categories": []}
{"id": "2412.20290", "title": "Transformer-Based Contrastive Meta-Learning For Low-Resource\n  Generalizable Activity Recognition", "abstract": "  Deep learning has been widely adopted for human activity recognition (HAR)\nwhile generalizing a trained model across diverse users and scenarios remains\nchallenging due to distribution shifts. The inherent low-resource challenge in\nHAR, i.e., collecting and labeling adequate human-involved data can be\nprohibitively costly, further raising the difficulty of tackling DS. We propose\nTACO, a novel transformer-based contrastive meta-learning approach for\ngeneralizable HAR. TACO addresses DS by synthesizing virtual target domains in\ntraining with explicit consideration of model generalizability. Additionally,\nwe extract expressive feature with the attention mechanism of Transformer and\nincorporate the supervised contrastive loss function within our\nmeta-optimization to enhance representation learning. Our evaluation\ndemonstrates that TACO achieves notably better performance across various\nlow-resource DS scenarios.\n", "date": "2024-12-28", "categories": []}
{"id": "2412.20621", "title": "FreqMixFormerV2: Lightweight Frequency-aware Mixed Transformer for Human\n  Skeleton Action Recognition", "abstract": "  Transformer-based human skeleton action recognition has been developed for\nyears. However, the complexity and high parameter count demands of these models\nhinder their practical applications, especially in resource-constrained\nenvironments. In this work, we propose FreqMixForemrV2, which was built upon\nthe Frequency-aware Mixed Transformer (FreqMixFormer) for identifying subtle\nand discriminative actions with pioneered frequency-domain analysis. We design\na lightweight architecture that maintains robust performance while\nsignificantly reducing the model complexity. This is achieved through a\nredesigned frequency operator that optimizes high-frequency and low-frequency\nparameter adjustments, and a simplified frequency-aware attention module. These\nimprovements result in a substantial reduction in model parameters, enabling\nefficient deployment with only a minimal sacrifice in accuracy. Comprehensive\nevaluations of standard datasets (NTU RGB+D, NTU RGB+D 120, and NW-UCLA\ndatasets) demonstrate that the proposed model achieves a superior balance\nbetween efficiency and accuracy, outperforming state-of-the-art methods with\nonly 60% of the parameters.\n", "date": "2024-12-29", "categories": []}
{"id": "2412.21205", "title": "Action-Agnostic Point-Level Supervision for Temporal Action Detection", "abstract": "  We propose action-agnostic point-level (AAPL) supervision for temporal action\ndetection to achieve accurate action instance detection with a lightly\nannotated dataset. In the proposed scheme, a small portion of video frames is\nsampled in an unsupervised manner and presented to human annotators, who then\nlabel the frames with action categories. Unlike point-level supervision, which\nrequires annotators to search for every action instance in an untrimmed video,\nframes to annotate are selected without human intervention in AAPL supervision.\nWe also propose a detection model and learning method to effectively utilize\nthe AAPL labels. Extensive experiments on the variety of datasets (THUMOS '14,\nFineAction, GTEA, BEOID, and ActivityNet 1.3) demonstrate that the proposed\napproach is competitive with or outperforms prior methods for video-level and\npoint-level supervision in terms of the trade-off between the annotation cost\nand detection performance.\n", "date": "2024-12-30", "categories": []}
{"id": "2306.07576", "title": "Action Recognition with Multi-stream Motion Modeling and Mutual\n  Information Maximization", "abstract": "  Action recognition has long been a fundamental and intriguing problem in\nartificial intelligence. The task is challenging due to the high dimensionality\nnature of an action, as well as the subtle motion details to be considered.\nCurrent state-of-the-art approaches typically learn from articulated motion\nsequences in the straightforward 3D Euclidean space. However, the vanilla\nEuclidean space is not efficient for modeling important motion characteristics\nsuch as the joint-wise angular acceleration, which reveals the driving force\nbehind the motion. Moreover, current methods typically attend to each channel\nequally and lack theoretical constrains on extracting task-relevant features\nfrom the input.\n  In this paper, we seek to tackle these challenges from three aspects: (1) We\npropose to incorporate an acceleration representation, explicitly modeling the\nhigher-order variations in motion. (2) We introduce a novel Stream-GCN network\nequipped with multi-stream components and channel attention, where different\nrepresentations (i.e., streams) supplement each other towards a more precise\naction recognition while attention capitalizes on those important channels. (3)\nWe explore feature-level supervision for maximizing the extraction of\ntask-relevant information and formulate this into a mutual information loss.\nEmpirically, our approach sets the new state-of-the-art performance on three\nbenchmark datasets, NTU RGB+D, NTU RGB+D 120, and NW-UCLA. Our code is\nanonymously released at https://github.com/ActionR-Group/Stream-GCN, hoping to\ninspire the community.\n", "date": "2023-06-13", "categories": []}
{"id": "2412.19418", "title": "Generalized Uncertainty-Based Evidential Fusion with Hybrid Multi-Head\n  Attention for Weak-Supervised Temporal Action Localization", "abstract": "  Weakly supervised temporal action localization (WS-TAL) is a task of\ntargeting at localizing complete action instances and categorizing them with\nvideo-level labels. Action-background ambiguity, primarily caused by background\nnoise resulting from aggregation and intra-action variation, is a significant\nchallenge for existing WS-TAL methods. In this paper, we introduce a hybrid\nmulti-head attention (HMHA) module and generalized uncertainty-based evidential\nfusion (GUEF) module to address the problem. The proposed HMHA effectively\nenhances RGB and optical flow features by filtering redundant information and\nadjusting their feature distribution to better align with the WS-TAL task.\nAdditionally, the proposed GUEF adaptively eliminates the interference of\nbackground noise by fusing snippet-level evidences to refine uncertainty\nmeasurement and select superior foreground feature information, which enables\nthe model to concentrate on integral action instances to achieve better action\nlocalization and classification performance. Experimental results conducted on\nthe THUMOS14 dataset demonstrate that our method outperforms state-of-the-art\nmethods. Our code is available in\n\\url{https://github.com/heyuanpengpku/GUEF/tree/main}.\n", "date": "2024-12-27", "categories": []}
{"id": "2402.02431", "title": "Learning Mutual Excitation for Hand-to-Hand and Human-to-Human\n  Interaction Recognition", "abstract": "  Recognizing interactive actions, including hand-to-hand interaction and\nhuman-to-human interaction, has attracted increasing attention for various\napplications in the field of video analysis and human-robot interaction.\nConsidering the success of graph convolution in modeling topology-aware\nfeatures from skeleton data, recent methods commonly operate graph convolution\non separate entities and use late fusion for interactive action recognition,\nwhich can barely model the mutual semantic relationships between pairwise\nentities. To this end, we propose a mutual excitation graph convolutional\nnetwork (me-GCN) by stacking mutual excitation graph convolution (me-GC)\nlayers. Specifically, me-GC uses a mutual topology excitation module to firstly\nextract adjacency matrices from individual entities and then adaptively model\nthe mutual constraints between them. Moreover, me-GC extends the above idea and\nfurther uses a mutual feature excitation module to extract and merge deep\nfeatures from pairwise entities. Compared with graph convolution, our proposed\nme-GC gradually learns mutual information in each layer and each stage of graph\nconvolution operations. Extensive experiments on a challenging hand-to-hand\ninteraction dataset, i.e., the Assembely101 dataset, and two large-scale\nhuman-to-human interaction datasets, i.e., NTU60-Interaction and\nNTU120-Interaction consistently verify the superiority of our proposed method,\nwhich outperforms the state-of-the-art GCN-based and Transformer-based methods.\n", "date": "2024-02-04", "categories": []}
{"id": "2411.07118", "title": "ConvMixFormer- A Resource-efficient Convolution Mixer for\n  Transformer-based Dynamic Hand Gesture Recognition", "abstract": "  Transformer models have demonstrated remarkable success in many domains such\nas natural language processing (NLP) and computer vision. With the growing\ninterest in transformer-based architectures, they are now utilized for gesture\nrecognition. So, we also explore and devise a novel ConvMixFormer architecture\nfor dynamic hand gestures. The transformers use quadratic scaling of the\nattention features with the sequential data, due to which these models are\ncomputationally complex and heavy. We have considered this drawback of the\ntransformer and designed a resource-efficient model that replaces the\nself-attention in the transformer with the simple convolutional layer-based\ntoken mixer. The computational cost and the parameters used for the\nconvolution-based mixer are comparatively less than the quadratic\nself-attention. Convolution-mixer helps the model capture the local spatial\nfeatures that self-attention struggles to capture due to their sequential\nprocessing nature. Further, an efficient gate mechanism is employed instead of\na conventional feed-forward network in the transformer to help the model\ncontrol the flow of features within different stages of the proposed model.\nThis design uses fewer learnable parameters which is nearly half the vanilla\ntransformer that helps in fast and efficient training. The proposed method is\nevaluated on NVidia Dynamic Hand Gesture and Briareo datasets and our model has\nachieved state-of-the-art results on single and multimodal inputs. We have also\nshown the parameter efficiency of the proposed ConvMixFormer model compared to\nother methods. The source code is available at\nhttps://github.com/mallikagarg/ConvMixFormer.\n", "date": "2024-11-11", "categories": []}
{"id": "2412.07072", "title": "Stable Mean Teacher for Semi-supervised Video Action Detection", "abstract": "  In this work, we focus on semi-supervised learning for video action\ndetection. Video action detection requires spatiotemporal localization in\naddition to classification, and a limited amount of labels makes the model\nprone to unreliable predictions. We present Stable Mean Teacher, a simple\nend-to-end teacher-based framework that benefits from improved and temporally\nconsistent pseudo labels. It relies on a novel Error Recovery (EoR) module,\nwhich learns from students' mistakes on labeled samples and transfers this\nknowledge to the teacher to improve pseudo labels for unlabeled samples.\nMoreover, existing spatiotemporal losses do not take temporal coherency into\naccount and are prone to temporal inconsistencies. To address this, we present\nDifference of Pixels (DoP), a simple and novel constraint focused on temporal\nconsistency, leading to coherent temporal detections. We evaluate our approach\non four different spatiotemporal detection benchmarks: UCF101-24, JHMDB21, AVA,\nand YouTube-VOS. Our approach outperforms the supervised baselines for action\ndetection by an average margin of 23.5% on UCF101-24, 16% on JHMDB21, and 3.3%\non AVA. Using merely 10% and 20% of data, it provides competitive performance\ncompared to the supervised baseline trained on 100% annotations on UCF101-24\nand JHMDB21, respectively. We further evaluate its effectiveness on AVA for\nscaling to large-scale datasets and YouTube-VOS for video object segmentation,\ndemonstrating its generalization capability to other tasks in the video domain.\nCode and models are publicly available.\n", "date": "2024-12-10", "categories": []}
{"id": "2412.16454", "title": "FACTS: Fine-Grained Action Classification for Tactical Sports", "abstract": "  Classifying fine-grained actions in fast-paced, close-combat sports such as\nfencing and boxing presents unique challenges due to the complexity, speed, and\nnuance of movements. Traditional methods reliant on pose estimation or fancy\nsensor data often struggle to capture these dynamics accurately. We introduce\nFACTS, a novel transformer-based approach for fine-grained action recognition\nthat processes raw video data directly, eliminating the need for pose\nestimation and the use of cumbersome body markers and sensors. FACTS achieves\nstate-of-the-art performance, with 90% accuracy on fencing actions and 83.25%\non boxing actions. Additionally, we present a new publicly available dataset\nfeaturing 8 detailed fencing actions, addressing critical gaps in sports\nanalytics resources. Our findings enhance training, performance analysis, and\nspectator engagement, setting a new benchmark for action classification in\ntactical sports.\n", "date": "2024-12-21", "categories": []}
{"id": "2412.16946", "title": "Video Domain Incremental Learning for Human Action Recognition in Home\n  Environments", "abstract": "  It is significantly challenging to recognize daily human actions in homes due\nto the diversity and dynamic changes in unconstrained home environments. It\nspurs the need to continually adapt to various users and scenes. Fine-tuning\ncurrent video understanding models on newly encountered domains often leads to\ncatastrophic forgetting, where the models lose their ability to perform well on\npreviously learned scenarios. To address this issue, we formalize the problem\nof Video Domain Incremental Learning (VDIL), which enables models to learn\ncontinually from different domains while maintaining a fixed set of action\nclasses. Existing continual learning research primarily focuses on\nclass-incremental learning, while the domain incremental learning has been\nlargely overlooked in video understanding. In this work, we introduce a novel\nbenchmark of domain incremental human action recognition for unconstrained home\nenvironments. We design three domain split types (user, scene, hybrid) to\nsystematically assess the challenges posed by domain shifts in real-world home\nsettings. Furthermore, we propose a baseline learning strategy based on replay\nand reservoir sampling techniques without domain labels to handle scenarios\nwith limited memory and task agnosticism. Extensive experimental results\ndemonstrate that our simple sampling and replay strategy outperforms most\nexisting continual learning methods across the three proposed benchmarks.\n", "date": "2024-12-22", "categories": []}
{"id": "2412.17517", "title": "Dataset for Real-World Human Action Detection Using FMCW mmWave Radar", "abstract": "  Human action detection using privacy-preserving mmWave radar sensors is\nstudied for its applications in healthcare and home automation. Unlike existing\nresearch, limited to simulations in controlled environments, we present a\nreal-world mmWave radar dataset with baseline results for human action\ndetection.\n", "date": "2024-12-23", "categories": []}
{"id": "2412.12990", "title": "Future Aspects in Human Action Recognition: Exploring Emerging\n  Techniques and Ethical Influences", "abstract": "  Visual-based human action recognition can be found in various application\nfields, e.g., surveillance systems, sports analytics, medical assistive\ntechnologies, or human-robot interaction frameworks, and it concerns the\nidentification and classification of individuals' activities within a video.\nSince actions typically occur over a sequence of consecutive images, it is\nparticularly challenging due to the inclusion of temporal analysis, which\nintroduces an extra layer of complexity. However, although multiple approaches\ntry to handle temporal analysis, there are still difficulties because of their\ncomputational cost and lack of adaptability. Therefore, different types of\nvision data, containing transition information between consecutive images,\nprovided by next-generation hardware sensors will guide the robotics community\nin tackling the problem of human action recognition. On the other hand, while\nthere is a plethora of still-image datasets, that researchers can adopt to\ntrain new artificial intelligence models, videos representing human activities\nare of limited capabilities, e.g., small and unbalanced datasets or selected\nwithout control from multiple sources. To this end, generating new and\nrealistic synthetic videos is possible since labeling is performed throughout\nthe data creation process, while reinforcement learning techniques can permit\nthe avoidance of considerable dataset dependence. At the same time, human\nfactors' involvement raises ethical issues for the research community, as\ndoubts and concerns about new technologies already exist.\n", "date": "2024-12-17", "categories": []}
{"id": "2407.07024", "title": "Exploring Scalability of Self-Training for Open-Vocabulary Temporal\n  Action Localization", "abstract": "  The vocabulary size in temporal action localization (TAL) is limited by the\nscarcity of large-scale annotated datasets. To overcome this, recent works\nintegrate vision-language models (VLMs), such as CLIP, for open-vocabulary TAL\n(OV-TAL). However, despite the success of VLMs trained on extensive datasets,\nexisting OV-TAL methods still rely on human-labeled TAL datasets of limited\nsize to train action localizers, limiting their generalizability. In this\npaper, we explore the scalability of self-training with unlabeled YouTube\nvideos for OV-TAL. Our approach consists of two stages: (1) a class-agnostic\naction localizer is trained on a human-labeled TAL dataset to generate\npseudo-labels for unlabeled videos, and (2) the large-scale pseudo-labeled\ndataset is then used to train the localizer. Extensive experiments demonstrate\nthat leveraging web-scale videos in self-training significantly enhances the\ngeneralizability of an action localizer. Additionally, we identify limitations\nin existing OV-TAL evaluation schemes and propose a new benchmark for thorough\nassessment. Finally, we showcase the TAL performance of the large multimodal\nmodel Gemini-1.5 on our new benchmark. Code is released at\nhttps://github.com/HYUNJS/STOV-TAL.\n", "date": "2024-07-09", "categories": []}
{"id": "2408.16729", "title": "Prediction-Feedback DETR for Temporal Action Detection", "abstract": "  Temporal Action Detection (TAD) is fundamental yet challenging for real-world\nvideo applications. Leveraging the unique benefits of transformers, various\nDETR-based approaches have been adopted in TAD. However, it has recently been\nidentified that the attention collapse in self-attention causes the performance\ndegradation of DETR for TAD. Building upon previous research, this paper newly\naddresses the attention collapse problem in cross-attention within DETR-based\nTAD methods. Moreover, our findings reveal that cross-attention exhibits\npatterns distinct from predictions, indicating a short-cut phenomenon. To\nresolve this, we propose a new framework, Prediction-Feedback DETR (Pred-DETR),\nwhich utilizes predictions to restore the collapse and align the cross- and\nself-attention with predictions. Specifically, we devise novel\nprediction-feedback objectives using guidance from the relations of the\npredictions. As a result, Pred-DETR significantly alleviates the collapse and\nachieves state-of-the-art performance among DETR-based methods on various\nchallenging benchmarks including THUMOS14, ActivityNet-v1.3, HACS, and\nFineAction.\n", "date": "2024-08-29", "categories": []}
{"id": "2407.16727", "title": "A study of animal action segmentation algorithms across supervised,\n  unsupervised, and semi-supervised learning paradigms", "abstract": "  Action segmentation of behavioral videos is the process of labeling each\nframe as belonging to one or more discrete classes, and is a crucial component\nof many studies that investigate animal behavior. A wide range of algorithms\nexist to automatically parse discrete animal behavior, encompassing supervised,\nunsupervised, and semi-supervised learning paradigms. These algorithms -- which\ninclude tree-based models, deep neural networks, and graphical models -- differ\nwidely in their structure and assumptions on the data. Using four datasets\nspanning multiple species -- fly, mouse, and human -- we systematically study\nhow the outputs of these various algorithms align with manually annotated\nbehaviors of interest. Along the way, we introduce a semi-supervised action\nsegmentation model that bridges the gap between supervised deep neural networks\nand unsupervised graphical models. We find that fully supervised temporal\nconvolutional networks with the addition of temporal information in the\nobservations perform the best on our supervised metrics across all datasets.\n", "date": "2024-07-23", "categories": []}
{"id": "2412.13594", "title": "Generalizable Sensor-Based Activity Recognition via Categorical Concept\n  Invariant Learning", "abstract": "  Human Activity Recognition (HAR) aims to recognize activities by training\nmodels on massive sensor data. In real-world deployment, a crucial aspect of\nHAR that has been largely overlooked is that the test sets may have different\ndistributions from training sets due to inter-subject variability including\nage, gender, behavioral habits, etc., which leads to poor generalization\nperformance. One promising solution is to learn domain-invariant\nrepresentations to enable a model to generalize on an unseen distribution.\nHowever, most existing methods only consider the feature-invariance of the\npenultimate layer for domain-invariant learning, which leads to suboptimal\nresults. In this paper, we propose a Categorical Concept Invariant Learning\n(CCIL) framework for generalizable activity recognition, which introduces a\nconcept matrix to regularize the model in the training stage by simultaneously\nconcentrating on feature-invariance and logit-invariance. Our key idea is that\nthe concept matrix for samples belonging to the same activity category should\nbe similar. Extensive experiments on four public HAR benchmarks demonstrate\nthat our CCIL substantially outperforms the state-of-the-art approaches under\ncross-person, cross-dataset, cross-position, and one-person-to-another\nsettings.\n", "date": "2024-12-18", "categories": []}
{"id": "2412.12829", "title": "2by2: Weakly-Supervised Learning for Global Action Segmentation", "abstract": "  This paper presents a simple yet effective approach for the poorly\ninvestigated task of global action segmentation, aiming at grouping frames\ncapturing the same action across videos of different activities. Unlike the\ncase of videos depicting all the same activity, the temporal order of actions\nis not roughly shared among all videos, making the task even more challenging.\nWe propose to use activity labels to learn, in a weakly-supervised fashion,\naction representations suitable for global action segmentation. For this\npurpose, we introduce a triadic learning approach for video pairs, to ensure\nintra-video action discrimination, as well as inter-video and inter-activity\naction association. For the backbone architecture, we use a Siamese network\nbased on sparse transformers that takes as input video pairs and determine\nwhether they belong to the same activity. The proposed approach is validated on\ntwo challenging benchmark datasets: Breakfast and YouTube Instructions,\noutperforming state-of-the-art methods.\n", "date": "2024-12-17", "categories": []}
{"id": "2412.10846", "title": "Detecting Activities of Daily Living in Egocentric Video to\n  Contextualize Hand Use at Home in Outpatient Neurorehabilitation Settings", "abstract": "  Wearable egocentric cameras and machine learning have the potential to\nprovide clinicians with a more nuanced understanding of patient hand use at\nhome after stroke and spinal cord injury (SCI). However, they require detailed\ncontextual information (i.e., activities and object interactions) to\neffectively interpret metrics and meaningfully guide therapy planning. We\ndemonstrate that an object-centric approach, focusing on what objects patients\ninteract with rather than how they move, can effectively recognize Activities\nof Daily Living (ADL) in real-world rehabilitation settings. We evaluated our\nmodels on a complex dataset collected in the wild comprising 2261 minutes of\negocentric video from 16 participants with impaired hand function. By\nleveraging pre-trained object detection and hand-object interaction models, our\nsystem achieves robust performance across different impairment levels and\nenvironments, with our best model achieving a mean weighted F1-score of 0.78\n+/- 0.12 and maintaining an F1-score > 0.5 for all participants using\nleave-one-subject-out cross validation. Through qualitative analysis, we\nobserve that this approach generates clinically interpretable information about\nfunctional object use while being robust to patient-specific movement\nvariations, making it particularly suitable for rehabilitation contexts with\nprevalent upper limb impairment.\n", "date": "2024-12-14", "categories": []}
{"id": "2412.11228", "title": "Uni-AdaFocus: Spatial-temporal Dynamic Computation for Video Recognition", "abstract": "  This paper presents a comprehensive exploration of the phenomenon of data\nredundancy in video understanding, with the aim to improve computational\nefficiency. Our investigation commences with an examination of spatial\nredundancy, which refers to the observation that the most informative region in\neach video frame usually corresponds to a small image patch, whose shape, size\nand location shift smoothly across frames. Motivated by this phenomenon, we\nformulate the patch localization problem as a dynamic decision task, and\nintroduce a spatially adaptive video recognition approach, termed AdaFocus. In\nspecific, a lightweight encoder is first employed to quickly process the full\nvideo sequence, whose features are then utilized by a policy network to\nidentify the most task-relevant regions. Subsequently, the selected patches are\ninferred by a high-capacity deep network for the final prediction. The full\nmodel can be trained in end-to-end conveniently. Furthermore, AdaFocus can be\nextended by further considering temporal and sample-wise redundancies, i.e.,\nallocating the majority of computation to the most task-relevant frames, and\nminimizing the computation spent on relatively \"easier\" videos. Our resulting\napproach, Uni-AdaFocus, establishes a comprehensive framework that seamlessly\nintegrates spatial, temporal, and sample-wise dynamic computation, while it\npreserves the merits of AdaFocus in terms of efficient end-to-end training and\nhardware friendliness. In addition, Uni-AdaFocus is general and flexible as it\nis compatible with off-the-shelf efficient backbones (e.g., TSM and X3D), which\ncan be readily deployed as our feature extractor, yielding a significantly\nimproved computational efficiency. Empirically, extensive experiments based on\nseven benchmark datasets and three application scenarios substantiate that\nUni-AdaFocus is considerably more efficient than the competitive baselines.\n", "date": "2024-12-15", "categories": []}
{"id": "2412.09202", "title": "Temporal Action Localization with Cross Layer Task Decoupling and\n  Refinement", "abstract": "  Temporal action localization (TAL) involves dual tasks to classify and\nlocalize actions within untrimmed videos. However, the two tasks often have\nconflicting requirements for features. Existing methods typically employ\nseparate heads for classification and localization tasks but share the same\ninput feature, leading to suboptimal performance. To address this issue, we\npropose a novel TAL method with Cross Layer Task Decoupling and Refinement\n(CLTDR). Based on the feature pyramid of video, CLTDR strategy integrates\nsemantically strong features from higher pyramid layers and detailed\nboundary-aware boundary features from lower pyramid layers to effectively\ndisentangle the action classification and localization tasks. Moreover, the\nmultiple features from cross layers are also employed to refine and align the\ndisentangled classification and regression results. At last, a lightweight\nGated Multi-Granularity (GMG) module is proposed to comprehensively extract and\naggregate video features at instant, local, and global temporal granularities.\nBenefiting from the CLTDR and GMG modules, our method achieves state-of-the-art\nperformance on five challenging benchmarks: THUMOS14, MultiTHUMOS,\nEPIC-KITCHENS-100, ActivityNet-1.3, and HACS. Our code and pre-trained models\nare publicly available at: https://github.com/LiQiang0307/CLTDR-GMG.\n", "date": "2024-12-12", "categories": []}
{"id": "2412.09037", "title": "Beyond Confusion: A Fine-grained Dialectical Examination of Human\n  Activity Recognition Benchmark Datasets", "abstract": "  The research of machine learning (ML) algorithms for human activity\nrecognition (HAR) has made significant progress with publicly available\ndatasets. However, most research prioritizes statistical metrics over examining\nnegative sample details. While recent models like transformers have been\napplied to HAR datasets with limited success from the benchmark metrics, their\ncounterparts have effectively solved problems on similar levels with near 100%\naccuracy. This raises questions about the limitations of current approaches.\nThis paper aims to address these open questions by conducting a fine-grained\ninspection of six popular HAR benchmark datasets. We identified for some parts\nof the data, none of the six chosen state-of-the-art ML methods can correctly\nclassify, denoted as the intersect of false classifications (IFC). Analysis of\nthe IFC reveals several underlying problems, including ambiguous annotations,\nirregularities during recording execution, and misaligned transition periods.\nWe contribute to the field by quantifying and characterizing annotated data\nambiguities, providing a trinary categorization mask for dataset patching, and\nstressing potential improvements for future data collections.\n", "date": "2024-12-12", "categories": []}
{"id": "2410.21108", "title": "LiGAR: LiDAR-Guided Hierarchical Transformer for Multi-Modal Group\n  Activity Recognition", "abstract": "  Group Activity Recognition (GAR) remains challenging in computer vision due\nto the complex nature of multi-agent interactions. This paper introduces LiGAR,\na LIDAR-Guided Hierarchical Transformer for Multi-Modal Group Activity\nRecognition. LiGAR leverages LiDAR data as a structural backbone to guide the\nprocessing of visual and textual information, enabling robust handling of\nocclusions and complex spatial arrangements. Our framework incorporates a\nMulti-Scale LIDAR Transformer, Cross-Modal Guided Attention, and an Adaptive\nFusion Module to integrate multi-modal data at different semantic levels\neffectively. LiGAR's hierarchical architecture captures group activities at\nvarious granularities, from individual actions to scene-level dynamics.\nExtensive experiments on the JRDB-PAR, Volleyball, and NBA datasets demonstrate\nLiGAR's superior performance, achieving state-of-the-art results with\nimprovements of up to 10.6% in F1-score on JRDB-PAR and 5.9% in Mean Per Class\nAccuracy on the NBA dataset. Notably, LiGAR maintains high performance even\nwhen LiDAR data is unavailable during inference, showcasing its adaptability.\nOur ablation studies highlight the significant contributions of each component\nand the effectiveness of our multi-modal, multi-scale approach in advancing the\nfield of group activity recognition.\n", "date": "2024-10-28", "categories": []}
{"id": "2404.01282", "title": "LoSA: Long-Short-range Adapter for Scaling End-to-End Temporal Action\n  Localization", "abstract": "  Temporal Action Localization (TAL) involves localizing and classifying action\nsnippets in an untrimmed video. The emergence of large video foundation models\nhas led RGB-only video backbones to outperform previous methods needing both\nRGB and optical flow modalities. Leveraging these large models is often limited\nto training only the TAL head due to the prohibitively large GPU memory\nrequired to adapt the video backbone for TAL. To overcome this limitation, we\nintroduce LoSA, the first memory-and-parameter-efficient backbone adapter\ndesigned specifically for TAL to handle untrimmed videos. LoSA specializes for\nTAL by introducing Long-Short-range Adapters that adapt the intermediate layers\nof the video backbone over different temporal ranges. These adapters run\nparallel to the video backbone to significantly reduce memory footprint. LoSA\nalso includes Long-Short-range Gated Fusion that strategically combines the\noutput of these adapters from the video backbone layers to enhance the video\nfeatures provided to the TAL head. Experiments show that LoSA significantly\noutperforms all existing methods on standard TAL benchmarks, THUMOS-14 and\nActivityNet-v1.3, by scaling end-to-end backbone adaptation to\nbillion-parameter-plus models like VideoMAEv2~(ViT-g) and leveraging them\nbeyond head-only transfer learning.\n", "date": "2024-04-01", "categories": []}
{"id": "2408.15996", "title": "Spatio-Temporal Context Prompting for Zero-Shot Action Detection", "abstract": "  Spatio-temporal action detection encompasses the tasks of localizing and\nclassifying individual actions within a video. Recent works aim to enhance this\nprocess by incorporating interaction modeling, which captures the relationship\nbetween people and their surrounding context. However, these approaches have\nprimarily focused on fully-supervised learning, and the current limitation lies\nin the lack of generalization capability to recognize unseen action categories.\nIn this paper, we aim to adapt the pretrained image-language models to detect\nunseen actions. To this end, we propose a method which can effectively leverage\nthe rich knowledge of visual-language models to perform Person-Context\nInteraction. Meanwhile, our Context Prompting module will utilize contextual\ninformation to prompt labels, thereby enhancing the generation of more\nrepresentative text features. Moreover, to address the challenge of recognizing\ndistinct actions by multiple people at the same timestamp, we design the\nInterest Token Spotting mechanism which employs pretrained visual knowledge to\nfind each person's interest context tokens, and then these tokens will be used\nfor prompting to generate text features tailored to each individual. To\nevaluate the ability to detect unseen actions, we propose a comprehensive\nbenchmark on J-HMDB, UCF101-24, and AVA datasets. The experiments show that our\nmethod achieves superior results compared to previous approaches and can be\nfurther extended to multi-action videos, bringing it closer to real-world\napplications. The code and data can be found in\nhttps://webber2933.github.io/ST-CLIP-project-page.\n", "date": "2024-08-28", "categories": []}
{"id": "2412.03903", "title": "Using SlowFast Networks for Near-Miss Incident Analysis in Dashcam\n  Videos", "abstract": "  This paper classifies near-miss traffic videos using the SlowFast deep neural\nnetwork that mimics the characteristics of the slow and fast visual information\nprocessed by two different streams from the M (Magnocellular) and P\n(Parvocellular) cells of the human brain. The approach significantly improves\nthe accuracy of the traffic near-miss video analysis and presents insights into\nhuman visual perception in traffic scenarios. Moreover, it contributes to\ntraffic safety enhancements and provides novel perspectives on the potential\ncognitive errors in traffic accidents.\n", "date": "2024-12-05", "categories": []}
{"id": "2412.04353", "title": "ActFusion: a Unified Diffusion Model for Action Segmentation and\n  Anticipation", "abstract": "  Temporal action segmentation and long-term action anticipation are two\npopular vision tasks for the temporal analysis of actions in videos. Despite\napparent relevance and potential complementarity, these two problems have been\ninvestigated as separate and distinct tasks. In this work, we tackle these two\nproblems, action segmentation and action anticipation, jointly using a unified\ndiffusion model dubbed ActFusion. The key idea to unification is to train the\nmodel to effectively handle both visible and invisible parts of the sequence in\nan integrated manner; the visible part is for temporal segmentation, and the\ninvisible part is for future anticipation. To this end, we introduce a new\nanticipative masking strategy during training in which a late part of the video\nframes is masked as invisible, and learnable tokens replace these frames to\nlearn to predict the invisible future. Experimental results demonstrate the\nbi-directional benefits between action segmentation and anticipation. ActFusion\nachieves the state-of-the-art performance across the standard benchmarks of 50\nSalads, Breakfast, and GTEA, outperforming task-specific models in both of the\ntwo tasks with a single unified model through joint learning.\n", "date": "2024-12-05", "categories": []}
{"id": "2412.02127", "title": "Streamlining Video Analysis for Efficient Violence Detection", "abstract": "  This paper addresses the challenge of automated violence detection in video\nframes captured by surveillance cameras, specifically focusing on classifying\nscenes as \"fight\" or \"non-fight.\" This task is critical for enhancing unmanned\nsecurity systems, online content filtering, and related applications. We\npropose an approach using a 3D Convolutional Neural Network (3D CNN)-based\nmodel named X3D to tackle this problem. Our approach incorporates\npre-processing steps such as tube extraction, volume cropping, and frame\naggregation, combined with clustering techniques, to accurately localize and\nclassify fight scenes. Extensive experimentation demonstrates the effectiveness\nof our method in distinguishing violent from non-violent events, providing\nvaluable insights for advancing practical violence detection systems.\n", "date": "2024-11-29", "categories": []}
{"id": "2412.01267", "title": "EdgeOAR: Real-time Online Action Recognition On Edge Devices", "abstract": "  This paper addresses the challenges of Online Action Recognition (OAR), a\nframework that involves instantaneous analysis and classification of behaviors\nin video streams. OAR must operate under stringent latency constraints, making\nit an indispensable component for real-time feedback for edge computing.\nExisting methods, which typically rely on the processing of entire video clips,\nfall short in scenarios requiring immediate recognition. To address this, we\ndesigned EdgeOAR, a novel framework specifically designed for OAR on edge\ndevices. EdgeOAR includes the Early Exit-oriented Task-specific Feature\nEnhancement Module (TFEM), which comprises lightweight submodules to optimize\nfeatures in both temporal and spatial dimensions. We design an iterative\ntraining method to enable TFEM learning features from the beginning of the\nvideo. Additionally, EdgeOAR includes an Inverse Information Entropy (IIE) and\nModality Consistency (MC)-driven fusion module to fuse features and make better\nexit decisions. This design overcomes the two main challenges: robust modeling\nof spatio-temporal action representations with limited initial frames in online\nvideo streams and balancing accuracy and efficiency on resource-constrained\nedge devices. Experiments show that on the UCF-101 dataset, our method EdgeOAR\nreduces latency by 99.23% and energy consumption by 99.28% compared to\nstate-of-the-art (SOTA) method. And achieves an adequate accuracy on edge\ndevices.\n", "date": "2024-12-02", "categories": []}
{"id": "2407.19497", "title": "Skeleton-based Group Activity Recognition via Spatial-Temporal Panoramic\n  Graph", "abstract": "  Group Activity Recognition aims to understand collective activities from\nvideos. Existing solutions primarily rely on the RGB modality, which encounters\nchallenges such as background variations, occlusions, motion blurs, and\nsignificant computational overhead. Meanwhile, current keypoint-based methods\noffer a lightweight and informative representation of human motions but\nnecessitate accurate individual annotations and specialized interaction\nreasoning modules. To address these limitations, we design a panoramic graph\nthat incorporates multi-person skeletons and objects to encapsulate group\nactivity, offering an effective alternative to RGB video. This panoramic graph\nenables Graph Convolutional Network (GCN) to unify intra-person, inter-person,\nand person-object interactive modeling through spatial-temporal graph\nconvolutions. In practice, we develop a novel pipeline that extracts skeleton\ncoordinates using pose estimation and tracking algorithms and employ\nMulti-person Panoramic GCN (MP-GCN) to predict group activities. Extensive\nexperiments on Volleyball and NBA datasets demonstrate that the MP-GCN achieves\nstate-of-the-art performance in both accuracy and efficiency. Notably, our\nmethod outperforms RGB-based approaches by using only estimated 2D keypoints as\ninput. Code is available at https://github.com/mgiant/MP-GCN\n", "date": "2024-07-28", "categories": []}
{"id": "2411.01769", "title": "ARN-LSTM: A Multi-Stream Fusion Model for Skeleton-based Action\n  Recognition", "abstract": "  This paper presents the ARN-LSTM architecture, a novel multi-stream action\nrecognition model designed to address the challenge of simultaneously capturing\nspatial motion and temporal dynamics in action sequences. Traditional methods\noften focus solely on spatial or temporal features, limiting their ability to\ncomprehend complex human activities fully. Our proposed model integrates joint,\nmotion, and temporal information through a multi-stream fusion architecture.\nSpecifically, it comprises a jointstream for extracting skeleton features, a\ntemporal stream for capturing dynamic temporal features, and an ARN-LSTM block\nthat utilizes Time-Distributed Long Short-Term Memory (TD-LSTM) layers followed\nby an Attention Relation Network (ARN) to model temporal relations. The outputs\nfrom these streams are fused in a fully connected layer to provide the final\naction prediction. Evaluations on the NTU RGB+D 60 and NTU RGB+D 120 datasets\noutperform the superior performance of our model, particularly in group\nactivity recognition.\n", "date": "2024-11-04", "categories": []}
{"id": "2411.19544", "title": "SkelMamba: A State Space Model for Efficient Skeleton Action Recognition\n  of Neurological Disorders", "abstract": "  We introduce a novel state-space model (SSM)-based framework for\nskeleton-based human action recognition, with an anatomically-guided\narchitecture that improves state-of-the-art performance in both clinical\ndiagnostics and general action recognition tasks. Our approach decomposes\nskeletal motion analysis into spatial, temporal, and spatio-temporal streams,\nusing channel partitioning to capture distinct movement characteristics\nefficiently. By implementing a structured, multi-directional scanning strategy\nwithin SSMs, our model captures local joint interactions and global motion\npatterns across multiple anatomical body parts. This anatomically-aware\ndecomposition enhances the ability to identify subtle motion patterns critical\nin medical diagnosis, such as gait anomalies associated with neurological\nconditions. On public action recognition benchmarks, i.e., NTU RGB+D, NTU RGB+D\n120, and NW-UCLA, our model outperforms current state-of-the-art methods,\nachieving accuracy improvements up to $3.2\\%$ with lower computational\ncomplexity than previous leading transformer-based models. We also introduce a\nnovel medical dataset for motion-based patient neurological disorder analysis\nto validate our method's potential in automated disease diagnosis.\n", "date": "2024-11-29", "categories": []}
{"id": "2411.18002", "title": "An End-to-End Two-Stream Network Based on RGB Flow and Representation\n  Flow for Human Action Recognition", "abstract": "  With the rapid advancements in deep learning, computer vision tasks have seen\nsignificant improvements, making two-stream neural networks a popular focus for\nvideo based action recognition. Traditional models using RGB and optical flow\nstreams achieve strong performance but at a high computational cost. To address\nthis, we introduce a representation flow algorithm to replace the optical flow\nbranch in the egocentric action recognition model, enabling end-to-end training\nwhile reducing computational cost and prediction time. Our model, designed for\negocentric action recognition, uses class activation maps (CAMs) to improve\naccuracy and ConvLSTM for spatio temporal encoding with spatial attention. When\nevaluated on the GTEA61, EGTEA GAZE+, and HMDB datasets, our model matches the\naccuracy of the original model on GTEA61 and exceeds it by 0.65% and 0.84% on\nEGTEA GAZE+ and HMDB, respectively. Prediction runtimes are significantly\nreduced to 0.1881s, 0.1503s, and 0.1459s, compared to the original model's\n101.6795s, 25.3799s, and 203.9958s. Ablation studies were also conducted to\nstudy the impact of different parameters on model performance.\n  Keywords: two-stream, egocentric, action recognition, CAM, representation\nflow, CAM, ConvLSTM\n", "date": "2024-11-27", "categories": []}
{"id": "2411.11288", "title": "Neuron: Learning Context-Aware Evolving Representations for Zero-Shot\n  Skeleton Action Recognition", "abstract": "  Zero-shot skeleton action recognition is a non-trivial task that requires\nrobust unseen generalization with prior knowledge from only seen classes and\nshared semantics. Existing methods typically build the skeleton-semantics\ninteractions by uncontrollable mappings and conspicuous representations,\nthereby can hardly capture the intricate and fine-grained relationship for\neffective cross-modal transferability. To address these issues, we propose a\nnovel dyNamically Evolving dUal skeleton-semantic syneRgistic framework with\nthe guidance of cOntext-aware side informatioN (dubbed Neuron), to explore more\nfine-grained cross-modal correspondence from micro to macro perspectives at\nboth spatial and temporal levels, respectively. Concretely, 1) we first\nconstruct the spatial-temporal evolving micro-prototypes and integrate dynamic\ncontext-aware side information to capture the intricate and synergistic\nskeleton-semantic correlations step-by-step, progressively refining cross-model\nalignment; and 2) we introduce the spatial compression and temporal memory\nmechanisms to guide the growth of spatial-temporal micro-prototypes, enabling\nthem to absorb structure-related spatial representations and\nregularity-dependent temporal patterns. Notably, such processes are analogous\nto the learning and growth of neurons, equipping the framework with the\ncapacity to generalize to novel unseen action categories. Extensive experiments\non various benchmark datasets demonstrated the superiority of the proposed\nmethod.\n", "date": "2024-11-18", "categories": []}
{"id": "2411.17584", "title": "Pre-training for Action Recognition with Automatically Generated Fractal\n  Datasets", "abstract": "  In recent years, interest in synthetic data has grown, particularly in the\ncontext of pre-training the image modality to support a range of computer\nvision tasks, including object classification, medical imaging etc. Previous\nwork has demonstrated that synthetic samples, automatically produced by various\ngenerative processes, can replace real counterparts and yield strong visual\nrepresentations. This approach resolves issues associated with real data such\nas collection and labeling costs, copyright and privacy.\n  We extend this trend to the video domain applying it to the task of action\nrecognition. Employing fractal geometry, we present methods to automatically\nproduce large-scale datasets of short synthetic video clips, which can be\nutilized for pre-training neural models. The generated video clips are\ncharacterized by notable variety, stemmed by the innate ability of fractals to\ngenerate complex multi-scale structures. To narrow the domain gap, we further\nidentify key properties of real videos and carefully emulate them during\npre-training. Through thorough ablations, we determine the attributes that\nstrengthen downstream results and offer general guidelines for pre-training\nwith synthetic videos. The proposed approach is evaluated by fine-tuning\npre-trained models on established action recognition datasets HMDB51 and UCF101\nas well as four other video benchmarks related to group action recognition,\nfine-grained action recognition and dynamic scenes. Compared to standard\nKinetics pre-training, our reported results come close and are even superior on\na portion of downstream datasets. Code and samples of synthetic videos are\navailable at https://github.com/davidsvy/fractal_video .\n", "date": "2024-11-26", "categories": []}
{"id": "2411.15284", "title": "When Spatial meets Temporal in Action Recognition", "abstract": "  Video action recognition has made significant strides, but challenges remain\nin effectively using both spatial and temporal information. While existing\nmethods often focus on either spatial features (e.g., object appearance) or\ntemporal dynamics (e.g., motion), they rarely address the need for a\ncomprehensive integration of both. Capturing the rich temporal evolution of\nvideo frames, while preserving their spatial details, is crucial for improving\naccuracy. In this paper, we introduce the Temporal Integration and Motion\nEnhancement (TIME) layer, a novel preprocessing technique designed to\nincorporate temporal information. The TIME layer generates new video frames by\nrearranging the original sequence, preserving temporal order while embedding\n$N^2$ temporally evolving frames into a single spatial grid of size $N \\times\nN$. This transformation creates new frames that balance both spatial and\ntemporal information, making them compatible with existing video models. When\n$N=1$, the layer captures rich spatial details, similar to existing methods. As\n$N$ increases ($N\\geq2$), temporal information becomes more prominent, while\nthe spatial information decreases to ensure compatibility with model inputs. We\ndemonstrate the effectiveness of the TIME layer by integrating it into popular\naction recognition models, such as ResNet-50, Vision Transformer, and Video\nMasked Autoencoders, for both RGB and depth video data. Our experiments show\nthat the TIME layer enhances recognition accuracy, offering valuable insights\nfor video processing tasks.\n", "date": "2024-11-22", "categories": []}
{"id": "2305.06310", "title": "SoGAR: Self-supervised Spatiotemporal Attention-based Social Group\n  Activity Recognition", "abstract": "  This paper introduces a novel approach to Social Group Activity Recognition\n(SoGAR) using Self-supervised Transformers network that can effectively utilize\nunlabeled video data. To extract spatio-temporal information, we created local\nand global views with varying frame rates. Our self-supervised objective\nensures that features extracted from contrasting views of the same video were\nconsistent across spatio-temporal domains. Our proposed approach is efficient\nin using transformer-based encoders to alleviate the weakly supervised setting\nof group activity recognition. By leveraging the benefits of transformer\nmodels, our approach can model long-term relationships along spatio-temporal\ndimensions. Our proposed SoGAR method achieved state-of-the-art results on\nthree group activity recognition benchmarks, namely JRDB-PAR, NBA, and\nVolleyball datasets, surpassing the current numbers in terms of F1-score, MCA,\nand MPCA metrics.\n", "date": "2023-04-27", "categories": []}
{"id": "2411.12525", "title": "Rethinking Top Probability from Multi-view for Distracted Driver\n  Behaviour Localization", "abstract": "  Naturalistic driving action localization task aims to recognize and\ncomprehend human behaviors and actions from video data captured during\nreal-world driving scenarios. Previous studies have shown great action\nlocalization performance by applying a recognition model followed by\nprobability-based post-processing. Nevertheless, the probabilities provided by\nthe recognition model frequently contain confused information causing challenge\nfor post-processing. In this work, we adopt an action recognition model based\non self-supervise learning to detect distracted activities and give potential\naction probabilities. Subsequently, a constraint ensemble strategy takes\nadvantages of multi-camera views to provide robust predictions. Finally, we\nintroduce a conditional post-processing operation to locate distracted\nbehaviours and action temporal boundaries precisely. Experimenting on test set\nA2, our method obtains the sixth position on the public leaderboard of track 3\nof the 2024 AI City Challenge.\n", "date": "2024-11-19", "categories": []}
{"id": "2411.10922", "title": "Exploiting VLM Localizability and Semantics for Open Vocabulary Action\n  Detection", "abstract": "  Action detection aims to detect (recognize and localize) human actions\nspatially and temporally in videos. Existing approaches focus on the closed-set\nsetting where an action detector is trained and tested on videos from a fixed\nset of action categories. However, this constrained setting is not viable in an\nopen world where test videos inevitably come beyond the trained action\ncategories. In this paper, we address the practical yet challenging\nOpen-Vocabulary Action Detection (OVAD) problem. It aims to detect any action\nin test videos while training a model on a fixed set of action categories. To\nachieve such an open-vocabulary capability, we propose a novel method OpenMixer\nthat exploits the inherent semantics and localizability of large\nvision-language models (VLM) within the family of query-based detection\ntransformers (DETR). Specifically, the OpenMixer is developed by spatial and\ntemporal OpenMixer blocks (S-OMB and T-OMB), and a dynamically fused alignment\n(DFA) module. The three components collectively enjoy the merits of strong\ngeneralization from pre-trained VLMs and end-to-end learning from DETR design.\nMoreover, we established OVAD benchmarks under various settings, and the\nexperimental results show that the OpenMixer performs the best over baselines\nfor detecting seen and unseen actions. We release the codes, models, and\ndataset splits at https://github.com/Cogito2012/OpenMixer.\n", "date": "2024-11-17", "categories": []}
{"id": "2411.11335", "title": "Video-to-Task Learning via Motion-Guided Attention for Few-Shot Action\n  Recognition", "abstract": "  In recent years, few-shot action recognition has achieved remarkable\nperformance through spatio-temporal relation modeling. Although a wide range of\nspatial and temporal alignment modules have been proposed, they primarily\naddress spatial or temporal misalignments at the video level, while the\nspatio-temporal relationships across different videos at the task level remain\nunderexplored. Recent studies utilize class prototypes to learn task-specific\nfeatures but overlook the spatio-temporal relationships across different videos\nat the task level, especially in the spatial dimension, where these\nrelationships provide rich information. In this paper, we propose a novel Dual\nMotion-Guided Attention Learning method (called DMGAL) for few-shot action\nrecognition, aiming to learn the spatio-temporal relationships from the\nvideo-specific to the task-specific level. To achieve this, we propose a\ncarefully designed Motion-Guided Attention (MGA) method to identify and\ncorrelate motion-related region features from the video level to the task\nlevel. Specifically, the Self Motion-Guided Attention module (S-MGA) achieves\nspatio-temporal relation modeling at the video level by identifying and\ncorrelating motion-related region features between different frames within a\nvideo. The Cross Motion-Guided Attention module (C-MGA) identifies and\ncorrelates motion-related region features between frames of different videos\nwithin a specific task to achieve spatio-temporal relationships at the task\nlevel. This approach enables the model to construct class prototypes that fully\nincorporate spatio-temporal relationships from the video-specific level to the\ntask-specific level. We validate the effectiveness of our DMGAL method by\nemploying both fully fine-tuning and adapter-tuning paradigms. The models\ndeveloped using these paradigms are termed DMGAL-FT and DMGAL-Adapter,\nrespectively.\n", "date": "2024-11-18", "categories": []}
{"id": "2411.11548", "title": "Real-Time Fitness Exercise Classification and Counting from Video Frames", "abstract": "  This paper introduces a novel method for real-time exercise classification\nusing a Bidirectional Long Short-Term Memory (BiLSTM) neural network. Existing\nexercise recognition approaches often rely on synthetic datasets, raw\ncoordinate inputs sensitive to user and camera variations, and fail to fully\nexploit the temporal dependencies in exercise movements. These issues limit\ntheir generalizability and robustness in real-world conditions, where lighting,\ncamera angles, and user body types vary.\n  To address these challenges, we propose a BiLSTM-based model that leverages\ninvariant features, such as joint angles, alongside raw coordinates. By using\nboth angles and (x, y, z) coordinates, the model adapts to changes in\nperspective, user positioning, and body differences, improving generalization.\nTraining on 30-frame sequences enables the BiLSTM to capture the temporal\ncontext of exercises and recognize patterns evolving over time.\n  We compiled a dataset combining synthetic data from the InfiniteRep dataset\nand real-world videos from Kaggle and other sources. This dataset includes four\ncommon exercises: squat, push-up, shoulder press, and bicep curl. The model was\ntrained and validated on these diverse datasets, achieving an accuracy of over\n99% on the test set. To assess generalizability, the model was tested on 2\nseparate test sets representative of typical usage conditions. Comparisons with\nthe previous approach from the literature are present in the result section\nshowing that the proposed model is the best-performing one.\n  The classifier is integrated into a web application providing real-time\nexercise classification and repetition counting without manual exercise\nselection.\n  Demo and datasets are available at the following GitHub Repository:\nhttps://github.com/RiccardoRiccio/Fitness-AI-Trainer-With-Automatic-Exercise-Recognition-and-Counting.\n", "date": "2024-11-18", "categories": []}
{"id": "2411.08814", "title": "Process-aware Human Activity Recognition", "abstract": "  Humans naturally follow distinct patterns when conducting their daily\nactivities, which are driven by established practices and processes, such as\nproduction workflows, social norms and daily routines. Human activity\nrecognition (HAR) algorithms usually use neural networks or machine learning\ntechniques to analyse inherent relationships within the data. However, these\napproaches often overlook the contextual information in which the data are\ngenerated, potentially limiting their effectiveness. We propose a novel\napproach that incorporates process information from context to enhance the HAR\nperformance. Specifically, we align probabilistic events generated by machine\nlearning models with process models derived from contextual information. This\nalignment adaptively weighs these two sources of information to optimise HAR\naccuracy. Our experiments demonstrate that our approach achieves better\naccuracy and Macro F1-score compared to baseline models.\n", "date": "2024-11-13", "categories": []}
{"id": "2411.08878", "title": "A Short Note on Evaluating RepNet for Temporal Repetition Counting in\n  Videos", "abstract": "  We discuss some consistent issues on how RepNet has been evaluated in various\npapers. As a way to mitigate these issues, we report RepNet performance results\non different datasets, and release evaluation code and the RepNet checkpoint to\nobtain these results. Code URL:\nhttps://github.com/google-research/google-research/blob/master/repnet/\n", "date": "2024-11-13", "categories": []}
{"id": "2411.05636", "title": "Video RWKV:Video Action Recognition Based RWKV", "abstract": "  To address the challenges of high computational costs and long-distance\ndependencies in exist ing video understanding methods, such as CNNs and\nTransformers, this work introduces RWKV to the video domain in a novel way. We\npropose a LSTM CrossRWKV (LCR) framework, designed for spatiotemporal\nrepresentation learning to tackle the video understanding task. Specifically,\nthe proposed linear complexity LCR incorporates a novel Cross RWKV gate to\nfacilitate interaction be tween current frame edge information and past\nfeatures, enhancing the focus on the subject through edge features and globally\naggregating inter-frame features over time. LCR stores long-term mem ory for\nvideo processing through an enhanced LSTM recurrent execution mechanism. By\nleveraging the Cross RWKV gate and recurrent execution, LCR effectively\ncaptures both spatial and temporal features. Additionally, the edge information\nserves as a forgetting gate for LSTM, guiding long-term memory management.Tube\nmasking strategy reduces redundant information in food and reduces\noverfitting.These advantages enable LSTM CrossRWKV to set a new benchmark in\nvideo under standing, offering a scalable and efficient solution for\ncomprehensive video analysis. All code and models are publicly available.\n", "date": "2024-11-08", "categories": []}
{"id": "2403.12385", "title": "Benchmarking Badminton Action Recognition with a New Fine-Grained\n  Dataset", "abstract": "  In the dynamic and evolving field of computer vision, action recognition has\nbecome a key focus, especially with the advent of sophisticated methodologies\nlike Convolutional Neural Networks (CNNs), Convolutional 3D, Transformer, and\nspatial-temporal feature fusion. These technologies have shown promising\nresults on well-established benchmarks but face unique challenges in real-world\napplications, particularly in sports analysis, where the precise decomposition\nof activities and the distinction of subtly different actions are crucial.\nExisting datasets like UCF101, HMDB51, and Kinetics have offered a diverse\nrange of video data for various scenarios. However, there's an increasing need\nfor fine-grained video datasets that capture detailed categorizations and\nnuances within broader action categories. In this paper, we introduce the\nVideoBadminton dataset derived from high-quality badminton footage. Through an\nexhaustive evaluation of leading methodologies on this dataset, this study aims\nto advance the field of action recognition, particularly in badminton sports.\nThe introduction of VideoBadminton could not only serve for badminton action\nrecognition but also provide a dataset for recognizing fine-grained actions.\nThe insights gained from these evaluations are expected to catalyze further\nresearch in action comprehension, especially within sports contexts.\n", "date": "2024-03-19", "categories": []}
{"id": "2411.00883", "title": "Technical Report for ActivityNet Challenge 2022 -- Temporal Action\n  Localization", "abstract": "  In the task of temporal action localization of ActivityNet-1.3 datasets, we\npropose to locate the temporal boundaries of each action and predict action\nclass in untrimmed videos. We first apply VideoSwinTransformer as feature\nextractor to extract different features. Then we apply a unified network\nfollowing Faster-TAD to simultaneously obtain proposals and semantic labels.\nLast, we ensemble the results of different temporal action detection models\nwhich complement each other. Faster-TAD simplifies the pipeline of TAD and gets\nremarkable performance, obtaining comparable results as those of multi-step\napproaches.\n", "date": "2024-10-31", "categories": []}
{"id": "2411.01122", "title": "OnlineTAS: An Online Baseline for Temporal Action Segmentation", "abstract": "  Temporal context plays a significant role in temporal action segmentation. In\nan offline setting, the context is typically captured by the segmentation\nnetwork after observing the entire sequence. However, capturing and using such\ncontext information in an online setting remains an under-explored problem.\nThis work presents the an online framework for temporal action segmentation. At\nthe core of the framework is an adaptive memory designed to accommodate dynamic\nchanges in context over time, alongside a feature augmentation module that\nenhances the frames with the memory. In addition, we propose a post-processing\napproach to mitigate the severe over-segmentation in the online setting. On\nthree common segmentation benchmarks, our approach achieves state-of-the-art\nperformance.\n", "date": "2024-11-02", "categories": []}
{"id": "2411.01348", "title": "Optimizing Violence Detection in Video Classification Accuracy through\n  3D Convolutional Neural Networks", "abstract": "  As violent crimes continue to happen, it becomes necessary to have security\ncameras that can rapidly identify moments of violence with excellent accuracy.\nThe purpose of this study is to identify how many frames should be analyzed at\na time in order to optimize a violence detection model's accuracy as a\nparameter of the depth of a 3D convolutional network. Previous violence\nclassification models have been created, but their application to live footage\nmay be flawed. In this project, a convolutional neural network was created to\nanalyze optical flow frames of each video. The number of frames analyzed at a\ntime would vary with one, two, three, ten, and twenty frames, and each model\nwould be trained for 20 epochs. The greatest validation accuracy was 94.87% and\noccurred with the model that analyzed three frames at a time. This means that\nmachine learning models to detect violence may function better when analyzing\nthree frames at a time for this dataset. The methodology used to identify the\noptimal number of frames to analyze at a time could be used in other\napplications of video classification, especially those of complex or abstract\nactions, such as violence.\n", "date": "2024-11-02", "categories": []}
{"id": "2411.02065", "title": "AM Flow: Adapters for Temporal Processing in Action Recognition", "abstract": "  Deep learning models, in particular \\textit{image} models, have recently\ngained generalisability and robustness. %are becoming more general and robust\nby the day. In this work, we propose to exploit such advances in the realm of\n\\textit{video} classification. Video foundation models suffer from the\nrequirement of extensive pretraining and a large training time. Towards\nmitigating such limitations, we propose \"\\textit{Attention Map (AM) Flow}\" for\nimage models, a method for identifying pixels relevant to motion in each input\nvideo frame. In this context, we propose two methods to compute AM flow,\ndepending on camera motion. AM flow allows the separation of spatial and\ntemporal processing, while providing improved results over combined\nspatio-temporal processing (as in video models). Adapters, one of the popular\ntechniques in parameter efficient transfer learning, facilitate the\nincorporation of AM flow into pretrained image models, mitigating the need for\nfull-finetuning. We extend adapters to \"\\textit{temporal processing adapters}\"\nby incorporating a temporal processing unit into the adapters. Our work\nachieves faster convergence, therefore reducing the number of epochs needed for\ntraining. Moreover, we endow an image model with the ability to achieve\nstate-of-the-art results on popular action recognition datasets. This reduces\ntraining time and simplifies pretraining. We present experiments on\nKinetics-400, Something-Something v2, and Toyota Smarthome datasets, showcasing\nstate-of-the-art or comparable results.\n", "date": "2024-11-04", "categories": []}
{"id": "2410.23806", "title": "Human Action Recognition (HAR) Using Skeleton-based Spatial Temporal\n  Relative Transformer Network: ST-RTR", "abstract": "  Human Action Recognition (HAR) is an interesting research area in\nhuman-computer interaction used to monitor the activities of elderly and\ndisabled individuals affected by physical and mental health. In the recent era,\nskeleton-based HAR has received much attention because skeleton data has shown\nthat it can handle changes in striking, body size, camera views, and complex\nbackgrounds. One key characteristic of ST-GCN is automatically learning spatial\nand temporal patterns from skeleton sequences. It has some limitations, as this\nmethod only works for short-range correlation due to its limited receptive\nfield. Consequently, understanding human action requires long-range\ninterconnection. To address this issue, we developed a spatial-temporal\nrelative transformer ST-RTR model. The ST-RTR includes joint and relay nodes,\nwhich allow efficient communication and data transmission within the network.\nThese nodes help to break the inherent spatial and temporal skeleton\ntopologies, which enables the model to understand long-range human action\nbetter. Furthermore, we combine ST-RTR with a fusion model for further\nperformance improvements. To assess the performance of the ST-RTR method, we\nconducted experiments on three skeleton-based HAR benchmarks: NTU RGB+D 60, NTU\nRGB+D 120, and UAV-Human. It boosted CS and CV by 2.11 % and 1.45% on NTU RGB+D\n60, 1.25% and 1.05% on NTU RGB+D 120. On UAV-Human datasets, accuracy improved\nby 2.54%. The experimental outcomes explain that the proposed ST-RTR model\nsignificantly improves action recognition associated with the standard ST-GCN\nmethod.\n", "date": "2024-10-31", "categories": []}
{"id": "2410.23641", "title": "Recovering Complete Actions for Cross-dataset Skeleton Action\n  Recognition", "abstract": "  Despite huge progress in skeleton-based action recognition, its\ngeneralizability to different domains remains a challenging issue. In this\npaper, to solve the skeleton action generalization problem, we present a\nrecover-and-resample augmentation framework based on a novel complete action\nprior. We observe that human daily actions are confronted with temporal\nmismatch across different datasets, as they are usually partial observations of\ntheir complete action sequences. By recovering complete actions and resampling\nfrom these full sequences, we can generate strong augmentations for unseen\ndomains. At the same time, we discover the nature of general action\ncompleteness within large datasets, indicated by the per-frame diversity over\ntime. This allows us to exploit two assets of transferable knowledge that can\nbe shared across action samples and be helpful for action completion: boundary\nposes for determining the action start, and linear temporal transforms for\ncapturing global action patterns. Therefore, we formulate the recovering stage\nas a two-step stochastic action completion with boundary pose-conditioned\nextrapolation followed by smooth linear transforms. Both the boundary poses and\nlinear transforms can be efficiently learned from the whole dataset via\nclustering. We validate our approach on a cross-dataset setting with three\nskeleton action datasets, outperforming other domain generalization approaches\nby a considerable margin.\n", "date": "2024-10-31", "categories": []}
{"id": "2410.24166", "title": "Approaches to human activity recognition via passive radar", "abstract": "  The thesis explores novel methods for Human Activity Recognition (HAR) using\npassive radar with a focus on non-intrusive Wi-Fi Channel State Information\n(CSI) data. Traditional HAR approaches often use invasive sensors like cameras\nor wearables, raising privacy issues. This study leverages the non-intrusive\nnature of CSI, using Spiking Neural Networks (SNN) to interpret signal\nvariations caused by human movements. These networks, integrated with symbolic\nreasoning frameworks such as DeepProbLog, enhance the adaptability and\ninterpretability of HAR systems. SNNs offer reduced power consumption, ideal\nfor privacy-sensitive applications. Experimental results demonstrate SNN-based\nneurosymbolic models achieve high accuracy making them a promising alternative\nfor HAR across various domains.\n", "date": "2024-10-31", "categories": []}
{"id": "2410.23092", "title": "First Place Solution to the ECCV 2024 ROAD++ Challenge @ ROAD++ Atomic\n  Activity Recognition 2024", "abstract": "  This report presents our team's technical solution for participating in Track\n3 of the 2024 ECCV ROAD++ Challenge. The task of Track 3 is atomic activity\nrecognition, which aims to identify 64 types of atomic activities in road\nscenes based on video content. Our approach primarily addresses the challenges\nof small objects, discriminating between single object and a group of objects,\nas well as model overfitting in this task. Firstly, we construct a multi-branch\nactivity recognition framework that not only separates different object\ncategories but also the tasks of single object and object group recognition,\nthereby enhancing recognition accuracy. Subsequently, we develop various model\nensembling strategies, including integrations of multiple frame sampling\nsequences, different frame sampling sequence lengths, multiple training epochs,\nand different backbone networks. Furthermore, we propose an atomic activity\nrecognition data augmentation method, which greatly expands the sample space by\nflipping video frames and road topology, effectively mitigating model\noverfitting. Our methods rank first in the test set of Track 3 for the ROAD++\nChallenge 2024, and achieve 69% mAP.\n", "date": "2024-10-30", "categories": []}
{"id": "2410.21300", "title": "Contrastive Learning with Auxiliary User Detection for Identifying\n  Activities", "abstract": "  Human Activity Recognition (HAR) is essential in ubiquitous computing, with\nfar-reaching real-world applications. While recent SOTA HAR research has\ndemonstrated impressive performance, some key aspects remain under-explored.\nFirstly, HAR can be both highly contextualized and personalized. However, prior\nwork has predominantly focused on being Context-Aware (CA) while largely\nignoring the necessity of being User-Aware (UA). We argue that addressing the\nimpact of innate user action-performing differences is equally crucial as\nconsidering external contextual environment settings in HAR tasks. Secondly,\nbeing user-aware makes the model acknowledge user discrepancies but does not\nnecessarily guarantee mitigation of these discrepancies, i.e., unified\npredictions under the same activities. There is a need for a methodology that\nexplicitly enforces closer (different user, same activity) representations. To\nbridge this gap, we introduce CLAUDIA, a novel framework designed to address\nthese issues. Specifically, we expand the contextual scope of the CA-HAR task\nby integrating User Identification (UI) within the CA-HAR framework, jointly\npredicting both CA-HAR and UI in a new task called User and Context-Aware HAR\n(UCA-HAR). This approach enriches personalized and contextual understanding by\njointly learning user-invariant and user-specific patterns. Inspired by SOTA\ndesigns in the visual domain, we introduce a supervised contrastive loss\nobjective on instance-instance pairs to enhance model efficacy and improve\nlearned feature quality. Evaluation across three real-world CA-HAR datasets\nreveals substantial performance enhancements, with average improvements ranging\nfrom 5.8% to 14.1% in Matthew's Correlation Coefficient and 3.0% to 7.2% in\nMacro F1 score.\n", "date": "2024-10-21", "categories": []}
{"id": "2209.00990", "title": "Self-Supervised Human Activity Recognition with Localized Time-Frequency\n  Contrastive Representation Learning", "abstract": "  In this paper, we propose a self-supervised learning solution for human\nactivity recognition with smartphone accelerometer data. We aim to develop a\nmodel that learns strong representations from accelerometer signals, in order\nto perform robust human activity classification, while reducing the model's\nreliance on class labels. Specifically, we intend to enable cross-dataset\ntransfer learning such that our network pre-trained on a particular dataset can\nperform effective activity classification on other datasets (successive to a\nsmall amount of fine-tuning). To tackle this problem, we design our solution\nwith the intention of learning as much information from the accelerometer\nsignals as possible. As a result, we design two separate pipelines, one that\nlearns the data in time-frequency domain, and the other in time-domain alone.\nIn order to address the issues mentioned above in regards to cross-dataset\ntransfer learning, we use self-supervised contrastive learning to train each of\nthese streams. Next, each stream is fine-tuned for final classification, and\neventually the two are fused to provide the final results. We evaluate the\nperformance of the proposed solution on three datasets, namely MotionSense,\nHAPT, and HHAR, and demonstrate that our solution outperforms prior works in\nthis field. We further evaluate the performance of the method in learning\ngeneralized features, by using MobiAct dataset for pre-training and the\nremaining three datasets for the downstream classification task, and show that\nthe proposed solution achieves better performance in comparison with other\nself-supervised methods in cross-dataset transfer learning.\n", "date": "2022-08-26", "categories": []}
{"id": "2303.02982", "title": "CLIP-guided Prototype Modulating for Few-shot Action Recognition", "abstract": "  Learning from large-scale contrastive language-image pre-training like CLIP\nhas shown remarkable success in a wide range of downstream tasks recently, but\nit is still under-explored on the challenging few-shot action recognition\n(FSAR) task. In this work, we aim to transfer the powerful multimodal knowledge\nof CLIP to alleviate the inaccurate prototype estimation issue due to data\nscarcity, which is a critical problem in low-shot regimes. To this end, we\npresent a CLIP-guided prototype modulating framework called CLIP-FSAR, which\nconsists of two key components: a video-text contrastive objective and a\nprototype modulation. Specifically, the former bridges the task discrepancy\nbetween CLIP and the few-shot video task by contrasting videos and\ncorresponding class text descriptions. The latter leverages the transferable\ntextual concepts from CLIP to adaptively refine visual prototypes with a\ntemporal Transformer. By this means, CLIP-FSAR can take full advantage of the\nrich semantic priors in CLIP to obtain reliable prototypes and achieve accurate\nfew-shot classification. Extensive experiments on five commonly used benchmarks\ndemonstrate the effectiveness of our proposed method, and CLIP-FSAR\nsignificantly outperforms existing state-of-the-art methods under various\nsettings. The source code and models will be publicly available at\nhttps://github.com/alibaba-mmai-research/CLIP-FSAR.\n", "date": "2023-03-06", "categories": []}
{"id": "2305.02673", "title": "Modelling Spatio-Temporal Interactions For Compositional Action\n  Recognition", "abstract": "  Humans have the natural ability to recognize actions even if the objects\ninvolved in the action or the background are changed. Humans can abstract away\nthe action from the appearance of the objects which is referred to as\ncompositionality of actions. We focus on this compositional aspect of action\nrecognition to impart human-like generalization abilities to video\naction-recognition models. First, we propose an interaction model that captures\nboth fine-grained and long-range interactions between hands and objects.\nFrame-wise hand-object interactions capture fine-grained movements, while\nlong-range interactions capture broader context and disambiguate actions across\ntime. Second, in order to provide additional contextual cues to differentiate\nsimilar actions, we infuse the interaction tokens with global motion\ninformation from video tokens. The final global motion refined interaction\ntokens are used for compositional action recognition. We show the effectiveness\nof our interaction-centric approach on the compositional Something-Else dataset\nwhere we obtain a new state-of-the-art result outperforming recent\nobject-centric methods by a significant margin.\n", "date": "2023-05-04", "categories": []}
{"id": "2310.10547", "title": "InfoGCN++: Learning Representation by Predicting the Future for Online\n  Human Skeleton-based Action Recognition", "abstract": "  Skeleton-based action recognition has made significant advancements recently,\nwith models like InfoGCN showcasing remarkable accuracy. However, these models\nexhibit a key limitation: they necessitate complete action observation prior to\nclassification, which constrains their applicability in real-time situations\nsuch as surveillance and robotic systems. To overcome this barrier, we\nintroduce InfoGCN++, an innovative extension of InfoGCN, explicitly developed\nfor online skeleton-based action recognition. InfoGCN++ augments the abilities\nof the original InfoGCN model by allowing real-time categorization of action\ntypes, independent of the observation sequence's length. It transcends\nconventional approaches by learning from current and anticipated future\nmovements, thereby creating a more thorough representation of the entire\nsequence. Our approach to prediction is managed as an extrapolation issue,\ngrounded on observed actions. To enable this, InfoGCN++ incorporates Neural\nOrdinary Differential Equations, a concept that lets it effectively model the\ncontinuous evolution of hidden states. Following rigorous evaluations on three\nskeleton-based action recognition benchmarks, InfoGCN++ demonstrates\nexceptional performance in online action recognition. It consistently equals or\nexceeds existing techniques, highlighting its significant potential to reshape\nthe landscape of real-time action recognition applications. Consequently, this\nwork represents a major leap forward from InfoGCN, pushing the limits of what's\npossible in online, skeleton-based action recognition. The code for InfoGCN++\nis publicly available at https://github.com/stnoah1/infogcn2 for further\nexploration and validation.\n", "date": "2023-10-16", "categories": []}
{"id": "2410.19553", "title": "On Occlusions in Video Action Detection: Benchmark Datasets And Training\n  Recipes", "abstract": "  This paper explores the impact of occlusions in video action detection. We\nfacilitate this study by introducing five new benchmark datasets namely O-UCF\nand O-JHMDB consisting of synthetically controlled static/dynamic occlusions,\nOVIS-UCF and OVIS-JHMDB consisting of occlusions with realistic motions and\nReal-OUCF for occlusions in realistic-world scenarios. We formally confirm an\nintuitive expectation: existing models suffer a lot as occlusion severity is\nincreased and exhibit different behaviours when occluders are static vs when\nthey are moving. We discover several intriguing phenomenon emerging in neural\nnets: 1) transformers can naturally outperform CNN models which might have even\nused occlusion as a form of data augmentation during training 2) incorporating\nsymbolic-components like capsules to such backbones allows them to bind to\noccluders never even seen during training and 3) Islands of agreement can\nemerge in realistic images/videos without instance-level supervision,\ndistillation or contrastive-based objectives2(eg. video-textual training). Such\nemergent properties allow us to derive simple yet effective training recipes\nwhich lead to robust occlusion models inductively satisfying the first two\nstages of the binding mechanism (grouping/segregation). Models leveraging these\nrecipes outperform existing video action-detectors under occlusion by 32.3% on\nO-UCF, 32.7% on O-JHMDB & 2.6% on Real-OUCF in terms of the vMAP metric. The\ncode for this work has been released at\nhttps://github.com/rajatmodi62/OccludedActionBenchmark.\n", "date": "2024-10-25", "categories": []}
{"id": "2309.17327", "title": "Telling Stories for Common Sense Zero-Shot Action Recognition", "abstract": "  Video understanding has long suffered from reliance on large labeled\ndatasets, motivating research into zero-shot learning. Recent progress in\nlanguage modeling presents opportunities to advance zero-shot video analysis,\nbut constructing an effective semantic space relating action classes remains\nchallenging. We address this by introducing a novel dataset, Stories, which\ncontains rich textual descriptions for diverse action classes extracted from\nWikiHow articles. For each class, we extract multi-sentence narratives\ndetailing the necessary steps, scenes, objects, and verbs that characterize the\naction. This contextual data enables modeling of nuanced relationships between\nactions, paving the way for zero-shot transfer. We also propose an approach\nthat harnesses Stories to improve feature generation for training zero-shot\nclassification. Without any target dataset fine-tuning, our method achieves new\nstate-of-the-art on multiple benchmarks, improving top-1 accuracy by up to\n6.1%. We believe Stories provides a valuable resource that can catalyze\nprogress in zero-shot action recognition. The textual narratives forge\nconnections between seen and unseen classes, overcoming the bottleneck of\nlabeled data that has long impeded advancements in this exciting domain. The\ndata can be found here: https://github.com/kini5gowda/Stories .\n", "date": "2023-09-29", "categories": []}
{"id": "2309.09534", "title": "Selective Volume Mixup for Video Action Recognition", "abstract": "  The recent advances in Convolutional Neural Networks (CNNs) and Vision\nTransformers have convincingly demonstrated high learning capability for video\naction recognition on large datasets. Nevertheless, deep models often suffer\nfrom the overfitting effect on small-scale datasets with a limited number of\ntraining videos. A common solution is to exploit the existing image\naugmentation strategies for each frame individually including Mixup, Cutmix,\nand RandAugment, which are not particularly optimized for video data. In this\npaper, we propose a novel video augmentation strategy named Selective Volume\nMixup (SV-Mix) to improve the generalization ability of deep models with\nlimited training videos. SV-Mix devises a learnable selective module to choose\nthe most informative volumes from two videos and mixes the volumes up to\nachieve a new training video. Technically, we propose two new modules, i.e., a\nspatial selective module to select the local patches for each spatial position,\nand a temporal selective module to mix the entire frames for each timestamp and\nmaintain the spatial pattern. At each time, we randomly choose one of the two\nmodules to expand the diversity of training samples. The selective modules are\njointly optimized with the video action recognition framework to find the\noptimal augmentation strategy. We empirically demonstrate the merits of the\nSV-Mix augmentation on a wide range of video action recognition benchmarks and\nconsistently boot the performances of both CNN-based and transformer-based\nmodels.\n", "date": "2023-09-18", "categories": []}
{"id": "2312.00188", "title": "REACT: Recognize Every Action Everywhere All At Once", "abstract": "  Group Activity Recognition (GAR) is a fundamental problem in computer vision,\nwith diverse applications in sports video analysis, video surveillance, and\nsocial scene understanding. Unlike conventional action recognition, GAR aims to\nclassify the actions of a group of individuals as a whole, requiring a deep\nunderstanding of their interactions and spatiotemporal relationships. To\naddress the challenges in GAR, we present REACT (\\textbf{R}ecognize\n\\textbf{E}very \\textbf{Act}ion Everywhere All At Once), a novel architecture\ninspired by the transformer encoder-decoder model explicitly designed to model\ncomplex contextual relationships within videos, including multi-modality and\nspatio-temporal features. Our architecture features a cutting-edge\nVision-Language Encoder block for integrated temporal, spatial, and multi-modal\ninteraction modeling. This component efficiently encodes spatiotemporal\ninteractions, even with sparsely sampled frames, and recovers essential local\ninformation. Our Action Decoder Block refines the joint understanding of text\nand video data, allowing us to precisely retrieve bounding boxes, enhancing the\nlink between semantics and visual reality. At the core, our Actor Fusion Block\norchestrates a fusion of actor-specific data and textual features, striking a\nbalance between specificity and context. Our method outperforms\nstate-of-the-art GAR approaches in extensive experiments, demonstrating\nsuperior accuracy in recognizing and understanding group activities. Our\narchitecture's potential extends to diverse real-world applications, offering\nempirical evidence of its performance gains. This work significantly advances\nthe field of group activity recognition, providing a robust framework for\nnuanced scene comprehension.\n", "date": "2023-11-27", "categories": []}
{"id": "2410.16746", "title": "SpikMamba: When SNN meets Mamba in Event-based Human Action Recognition", "abstract": "  Human action recognition (HAR) plays a key role in various applications such\nas video analysis, surveillance, autonomous driving, robotics, and healthcare.\nMost HAR algorithms are developed from RGB images, which capture detailed\nvisual information. However, these algorithms raise concerns in\nprivacy-sensitive environments due to the recording of identifiable features.\nEvent cameras offer a promising solution by capturing scene brightness changes\nsparsely at the pixel level, without capturing full images. Moreover, event\ncameras have high dynamic ranges that can effectively handle scenarios with\ncomplex lighting conditions, such as low light or high contrast environments.\nHowever, using event cameras introduces challenges in modeling the spatially\nsparse and high temporal resolution event data for HAR. To address these\nissues, we propose the SpikMamba framework, which combines the energy\nefficiency of spiking neural networks and the long sequence modeling capability\nof Mamba to efficiently capture global features from spatially sparse and high\na temporal resolution event data. Additionally, to improve the locality of\nmodeling, a spiking window-based linear attention mechanism is used. Extensive\nexperiments show that SpikMamba achieves remarkable recognition performance,\nsurpassing the previous state-of-the-art by 1.45%, 7.22%, 0.15%, and 3.92% on\nthe PAF, HARDVS, DVS128, and E-FAction datasets, respectively. The code is\navailable at https://github.com/Typistchen/SpikMamba.\n", "date": "2024-10-22", "categories": []}
{"id": "2410.06353", "title": "Language-Assisted Human Part Motion Learning for Skeleton-Based Temporal\n  Action Segmentation", "abstract": "  Skeleton-based Temporal Action Segmentation involves the dense action\nclassification of variable-length skeleton sequences. Current approaches\nprimarily apply graph-based networks to extract framewise, whole-body-level\nmotion representations, and use one-hot encoded labels for model optimization.\nHowever, whole-body motion representations do not capture fine-grained\npart-level motion representations and the one-hot encoded labels neglect the\nintrinsic semantic relationships within the language-based action definitions.\nTo address these limitations, we propose a novel method named Language-assisted\nHuman Part Motion Representation Learning (LPL), which contains a Disentangled\nPart Motion Encoder (DPE) to extract dual-level (i.e., part and whole-body)\nmotion representations and a Language-assisted Distribution Alignment (LDA)\nstrategy for optimizing spatial relations within representations. Specifically,\nafter part-aware skeleton encoding via DPE, LDA generates dual-level action\ndescriptions to construct a textual embedding space with the help of a\nlarge-scale language model. Then, LDA motivates the alignment of the embedding\nspace between text descriptions and motions. This alignment allows LDA not only\nto enhance intra-class compactness but also to transfer the language-encoded\nsemantic correlations among actions to skeleton-based motion learning.\nMoreover, we propose a simple yet efficient Semantic Offset Adapter to smooth\nthe cross-domain misalignment. Our experiments indicate that LPL achieves\nstate-of-the-art performance across various datasets (e.g., +4.4\\% Accuracy,\n+5.6\\% F1 on the PKU-MMD dataset). Moreover, LDA is compatible with existing\nmethods and improves their performance (e.g., +4.8\\% Accuracy, +4.3\\% F1 on the\nLARa dataset) without additional inference costs.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.15279", "title": "ContextDet: Temporal Action Detection with Adaptive Context Aggregation", "abstract": "  Temporal action detection (TAD), which locates and recognizes action\nsegments, remains a challenging task in video understanding due to variable\nsegment lengths and ambiguous boundaries. Existing methods treat neighboring\ncontexts of an action segment indiscriminately, leading to imprecise boundary\npredictions. We introduce a single-stage ContextDet framework, which makes use\nof large-kernel convolutions in TAD for the first time. Our model features a\npyramid adaptive context aggragation (ACA) architecture, capturing long context\nand improving action discriminability. Each ACA level consists of two novel\nmodules. The context attention module (CAM) identifies salient contextual\ninformation, encourages context diversity, and preserves context integrity\nthrough a context gating block (CGB). The long context module (LCM) makes use\nof a mixture of large- and small-kernel convolutions to adaptively gather\nlong-range context and fine-grained local features. Additionally, by varying\nthe length of these large kernels across the ACA pyramid, our model provides\nlightweight yet effective context aggregation and action discrimination. We\nconducted extensive experiments and compared our model with a number of\nadvanced TAD methods on six challenging TAD benchmarks: MultiThumos, Charades,\nFineAction, EPIC-Kitchens 100, Thumos14, and HACS, demonstrating superior\naccuracy at reduced inference speed.\n", "date": "2024-10-20", "categories": []}
{"id": "2410.16037", "title": "Improving the Multi-label Atomic Activity Recognition by Robust Visual\n  Feature and Advanced Attention @ ROAD++ Atomic Activity Recognition 2024", "abstract": "  Road++ Track3 proposes a multi-label atomic activity recognition task in\ntraffic scenarios, which can be standardized as a 64-class multi-label video\naction recognition task. In the multi-label atomic activity recognition task,\nthe robustness of visual feature extraction remains a key challenge, which\ndirectly affects the model performance and generalization ability. To cope with\nthese issues, our team optimized three aspects: data processing, model and\npost-processing. Firstly, the appropriate resolution and video sampling\nstrategy are selected, and a fixed sampling strategy is set on the validation\nand test sets. Secondly, in terms of model training, the team selects a variety\nof visual backbone networks for feature extraction, and then introduces the\naction-slot model, which is trained on the training and validation sets, and\nreasoned on the test set. Finally, for post-processing, the team combined the\nstrengths and weaknesses of different models for weighted fusion, and the final\nmAP on the test set was 58%, which is 4% higher than the challenge baseline.\n", "date": "2024-10-21", "categories": []}
{"id": "2410.03302", "title": "Action Selection Learning for Multi-label Multi-view Action Recognition", "abstract": "  Multi-label multi-view action recognition aims to recognize multiple\nconcurrent or sequential actions from untrimmed videos captured by multiple\ncameras. Existing work has focused on multi-view action recognition in a narrow\narea with strong labels available, where the onset and offset of each action\nare labeled at the frame-level. This study focuses on real-world scenarios\nwhere cameras are distributed to capture a wide-range area with only weak\nlabels available at the video-level. We propose the method named Multi-view\nAction Selection Learning (MultiASL), which leverages action selection learning\nto enhance view fusion by selecting the most useful information from different\nviewpoints. The proposed method includes a Multi-view Spatial-Temporal\nTransformer video encoder to extract spatial and temporal features from\nmulti-viewpoint videos. Action Selection Learning is employed at the\nframe-level, using pseudo ground-truth obtained from weak labels at the\nvideo-level, to identify the most relevant frames for action recognition.\nExperiments in a real-world office environment using the MM-Office dataset\ndemonstrate the superior performance of the proposed method compared to\nexisting methods. The source code is available at\nhttps://github.com/thanhhff/MultiASL/.\n", "date": "2024-10-04", "categories": []}
{"id": "2410.14238", "title": "Storyboard guided Alignment for Fine-grained Video Action Recognition", "abstract": "  Fine-grained video action recognition can be conceptualized as a video-text\nmatching problem. Previous approaches often rely on global video semantics to\nconsolidate video embeddings, which can lead to misalignment in video-text\npairs due to a lack of understanding of action semantics at an atomic\ngranularity level. To tackle this challenge, we propose a multi-granularity\nframework based on two observations: (i) videos with different global semantics\nmay share similar atomic actions or appearances, and (ii) atomic actions within\na video can be momentary, slow, or even non-directly related to the global\nvideo semantics. Inspired by the concept of storyboarding, which disassembles a\nscript into individual shots, we enhance global video semantics by generating\nfine-grained descriptions using a pre-trained large language model. These\ndetailed descriptions capture common atomic actions depicted in videos. A\nfiltering metric is proposed to select the descriptions that correspond to the\natomic actions present in both the videos and the descriptions. By employing\nglobal semantics and fine-grained descriptions, we can identify key frames in\nvideos and utilize them to aggregate embeddings, thereby making the embedding\nmore accurate. Extensive experiments on various video action recognition\ndatasets demonstrate superior performance of our proposed method in supervised,\nfew-shot, and zero-shot settings.\n", "date": "2024-10-18", "categories": []}
{"id": "2301.10931", "title": "Towards Continual Egocentric Activity Recognition: A Multi-modal\n  Egocentric Activity Dataset for Continual Learning", "abstract": "  With the rapid development of wearable cameras, a massive collection of\negocentric video for first-person visual perception becomes available. Using\negocentric videos to predict first-person activity faces many challenges,\nincluding limited field of view, occlusions, and unstable motions. Observing\nthat sensor data from wearable devices facilitates human activity recognition,\nmulti-modal activity recognition is attracting increasing attention. However,\nthe deficiency of related dataset hinders the development of multi-modal deep\nlearning for egocentric activity recognition. Nowadays, deep learning in real\nworld has led to a focus on continual learning that often suffers from\ncatastrophic forgetting. But the catastrophic forgetting problem for egocentric\nactivity recognition, especially in the context of multiple modalities, remains\nunexplored due to unavailability of dataset. In order to assist this research,\nwe present a multi-modal egocentric activity dataset for continual learning\nnamed UESTC-MMEA-CL, which is collected by self-developed glasses integrating a\nfirst-person camera and wearable sensors. It contains synchronized data of\nvideos, accelerometers, and gyroscopes, for 32 types of daily activities,\nperformed by 10 participants. Its class types and scale are compared with other\npublicly available datasets. The statistical analysis of the sensor data is\ngiven to show the auxiliary effects for different behaviors. And results of\negocentric activity recognition are reported when using separately, and\njointly, three modalities: RGB, acceleration, and gyroscope, on a base network\narchitecture. To explore the catastrophic forgetting in continual learning\ntasks, four baseline methods are extensively evaluated with different\nmulti-modal combinations. We hope the UESTC-MMEA-CL can promote future studies\non continual learning for first-person activity recognition in wearable\napplications.\n", "date": "2023-01-26", "categories": []}
{"id": "2406.01662", "title": "Few-Shot Classification of Interactive Activities of Daily Living\n  (InteractADL)", "abstract": "  Understanding Activities of Daily Living (ADLs) is a crucial step for\ndifferent applications including assistive robots, smart homes, and healthcare.\nHowever, to date, few benchmarks and methods have focused on complex ADLs,\nespecially those involving multi-person interactions in home environments. In\nthis paper, we propose a new dataset and benchmark, InteractADL, for\nunderstanding complex ADLs that involve interaction between humans (and\nobjects). Furthermore, complex ADLs occurring in home environments comprise a\nchallenging long-tailed distribution due to the rarity of multi-person\ninteractions, and pose fine-grained visual recognition tasks due to the\npresence of semantically and visually similar classes. To address these issues,\nwe propose a novel method for fine-grained few-shot video classification called\nName Tuning that enables greater semantic separability by learning optimal\nclass name vectors. We show that Name Tuning can be combined with existing\nprompt tuning strategies to learn the entire input text (rather than only\nlearning the prompt or class names) and demonstrate improved performance for\nfew-shot classification on InteractADL and 4 other fine-grained visual\nclassification benchmarks. For transparency and reproducibility, we release our\ncode at https://github.com/zanedurante/vlm_benchmark.\n", "date": "2024-06-03", "categories": []}
{"id": "2410.05771", "title": "Cefdet: Cognitive Effectiveness Network Based on Fuzzy Inference for\n  Action Detection", "abstract": "  Action detection and understanding provide the foundation for the generation\nand interaction of multimedia content. However, existing methods mainly focus\non constructing complex relational inference networks, overlooking the judgment\nof detection effectiveness. Moreover, these methods frequently generate\ndetection results with cognitive abnormalities. To solve the above problems,\nthis study proposes a cognitive effectiveness network based on fuzzy inference\n(Cefdet), which introduces the concept of \"cognition-based detection\" to\nsimulate human cognition. First, a fuzzy-driven cognitive effectiveness\nevaluation module (FCM) is established to introduce fuzzy inference into action\ndetection. FCM is combined with human action features to simulate the\ncognition-based detection process, which clearly locates the position of frames\nwith cognitive abnormalities. Then, a fuzzy cognitive update strategy (FCS) is\nproposed based on the FCM, which utilizes fuzzy logic to re-detect the\ncognition-based detection results and effectively update the results with\ncognitive abnormalities. Experimental results demonstrate that Cefdet exhibits\nsuperior performance against several mainstream algorithms on the public\ndatasets, validating its effectiveness and superiority. Code is available at\nhttps://github.com/12sakura/Cefdet.\n", "date": "2024-10-08", "categories": []}
{"id": "2304.05088", "title": "WEAR: An Outdoor Sports Dataset for Wearable and Egocentric Activity\n  Recognition", "abstract": "  Research has shown the complementarity of camera- and inertial-based data for\nmodeling human activities, yet datasets with both egocentric video and\ninertial-based sensor data remain scarce. In this paper, we introduce WEAR, an\noutdoor sports dataset for both vision- and inertial-based human activity\nrecognition (HAR). Data from 22 participants performing a total of 18 different\nworkout activities was collected with synchronized inertial (acceleration) and\ncamera (egocentric video) data recorded at 11 different outside locations. WEAR\nprovides a challenging prediction scenario in changing outdoor environments\nusing a sensor placement, in line with recent trends in real-world\napplications. Benchmark results show that through our sensor placement, each\nmodality interestingly offers complementary strengths and weaknesses in their\nprediction performance. Further, in light of the recent success of single-stage\nTemporal Action Localization (TAL) models, we demonstrate their versatility of\nnot only being trained using visual data, but also using raw inertial data and\nbeing capable to fuse both modalities by means of simple concatenation. The\ndataset and code to reproduce experiments is publicly available via:\nmariusbock.github.io/wear/.\n", "date": "2023-04-11", "categories": []}
{"id": "2410.11404", "title": "MoChat: Joints-Grouped Spatio-Temporal Grounding LLM for Multi-Turn\n  Motion Comprehension and Description", "abstract": "  Despite continuous advancements in deep learning for understanding human\nmotion, existing models often struggle to accurately identify action timing and\nspecific body parts, typically supporting only single-round interaction. Such\nlimitations in capturing fine-grained motion details reduce their effectiveness\nin motion understanding tasks. In this paper, we propose MoChat, a multimodal\nlarge language model capable of spatio-temporal grounding of human motion and\nunderstanding multi-turn dialogue context. To achieve these capabilities, we\ngroup the spatial information of each skeleton frame based on human anatomical\nstructure and then apply them with Joints-Grouped Skeleton Encoder, whose\noutputs are combined with LLM embeddings to create spatio-aware and\ntemporal-aware embeddings separately. Additionally, we develop a pipeline for\nextracting timestamps from skeleton sequences based on textual annotations, and\nconstruct multi-turn dialogues for spatially grounding. Finally, various task\ninstructions are generated for jointly training. Experimental results\ndemonstrate that MoChat achieves state-of-the-art performance across multiple\nmetrics in motion understanding tasks, making it as the first model capable of\nfine-grained spatio-temporal grounding of human motion.\n", "date": "2024-10-15", "categories": []}
{"id": "2311.15831", "title": "Temporal Action Localization for Inertial-based Human Activity\n  Recognition", "abstract": "  As of today, state-of-the-art activity recognition from wearable sensors\nrelies on algorithms being trained to classify fixed windows of data. In\ncontrast, video-based Human Activity Recognition, known as Temporal Action\nLocalization (TAL), has followed a segment-based prediction approach,\nlocalizing activity segments in a timeline of arbitrary length. This paper is\nthe first to systematically demonstrate the applicability of state-of-the-art\nTAL models for both offline and near-online Human Activity Recognition (HAR)\nusing raw inertial data as well as pre-extracted latent features as input.\nOffline prediction results show that TAL models are able to outperform popular\ninertial models on a multitude of HAR benchmark datasets, with improvements\nreaching as much as 26% in F1-score. We show that by analyzing timelines as a\nwhole, TAL models can produce more coherent segments and achieve higher\nNULL-class accuracy across all datasets. We demonstrate that TAL is less suited\nfor the immediate classification of small-sized windows of data, yet offers an\ninteresting perspective on inertial-based HAR -- alleviating the need for\nfixed-size windows and enabling algorithms to recognize activities of arbitrary\nlength. With design choices and training concepts yet to be explored, we argue\nthat TAL architectures could be of significant value to the inertial-based HAR\ncommunity. The code and data download to reproduce experiments is publicly\navailable via github.com/mariusbock/tal_for_har.\n", "date": "2023-11-27", "categories": []}
{"id": "2403.18074", "title": "Every Shot Counts: Using Exemplars for Repetition Counting in Videos", "abstract": "  Video repetition counting infers the number of repetitions of recurring\nactions or motion within a video. We propose an exemplar-based approach that\ndiscovers visual correspondence of video exemplars across repetitions within\ntarget videos. Our proposed Every Shot Counts (ESCounts) model is an\nattention-based encoder-decoder that encodes videos of varying lengths\nalongside exemplars from the same and different videos. In training, ESCounts\nregresses locations of high correspondence to the exemplars within the video.\nIn tandem, our method learns a latent that encodes representations of general\nrepetitive motions, which we use for exemplar-free, zero-shot inference.\nExtensive experiments over commonly used datasets (RepCount, Countix, and\nUCFRep) showcase ESCounts obtaining state-of-the-art performance across all\nthree datasets. Detailed ablations further demonstrate the effectiveness of our\nmethod.\n", "date": "2024-03-26", "categories": []}
{"id": "2410.09088", "title": "The Solution for Temporal Action Localisation Task of Perception Test\n  Challenge 2024", "abstract": "  This report presents our method for Temporal Action Localisation (TAL), which\nfocuses on identifying and classifying actions within specific time intervals\nthroughout a video sequence. We employ a data augmentation technique by\nexpanding the training dataset using overlapping labels from the\nSomething-SomethingV2 dataset, enhancing the model's ability to generalize\nacross various action classes. For feature extraction, we utilize\nstate-of-the-art models, including UMT, VideoMAEv2 for video features, and\nBEATs and CAV-MAE for audio features. Our approach involves training both\nmultimodal (video and audio) and unimodal (video only) models, followed by\ncombining their predictions using the Weighted Box Fusion (WBF) method. This\nfusion strategy ensures robust action localisation. our overall approach\nachieves a score of 0.5498, securing first place in the competition.\n", "date": "2024-10-08", "categories": []}
{"id": "2410.09902", "title": "Multi class activity classification in videos using Motion History Image\n  generation", "abstract": "  Human action recognition has been a topic of interest across multiple fields\nranging from security to entertainment systems. Tracking the motion and\nidentifying the action being performed on a real time basis is necessary for\ncritical security systems. In entertainment, especially gaming, the need for\nimmediate responses for actions and gestures are paramount for the success of\nthat system. We show that Motion History image has been a well established\nframework to capture the temporal and activity information in multi dimensional\ndetail enabling various usecases including classification. We utilize MHI to\nproduce sample data to train a classifier and demonstrate its effectiveness for\naction classification across six different activities in a single multi-action\nvideo. We analyze the classifier performance and identify usecases where MHI\nstruggles to generate the appropriate activity image and discuss mechanisms and\nfuture work to overcome those limitations.\n", "date": "2024-10-13", "categories": []}
{"id": "2410.09954", "title": "EITNet: An IoT-Enhanced Framework for Real-Time Basketball Action\n  Recognition", "abstract": "  Integrating IoT technology into basketball action recognition enhances sports\nanalytics, providing crucial insights into player performance and game\nstrategy. However, existing methods often fall short in terms of accuracy and\nefficiency, particularly in complex, real-time environments where player\nmovements are frequently occluded or involve intricate interactions. To\novercome these challenges, we propose the EITNet model, a deep learning\nframework that combines EfficientDet for object detection, I3D for\nspatiotemporal feature extraction, and TimeSformer for temporal analysis, all\nintegrated with IoT technology for seamless real-time data collection and\nprocessing. Our contributions include developing a robust architecture that\nimproves recognition accuracy to 92\\%, surpassing the baseline EfficientDet\nmodel's 87\\%, and reducing loss to below 5.0 compared to EfficientDet's 9.0\nover 50 epochs. Furthermore, the integration of IoT technology enhances\nreal-time data processing, providing adaptive insights into player performance\nand strategy. The paper details the design and implementation of EITNet,\nexperimental validation, and a comprehensive evaluation against existing\nmodels. The results demonstrate EITNet's potential to significantly advance\nautomated sports analysis and optimize data utilization for player performance\nand strategy improvement.\n", "date": "2024-10-13", "categories": []}
{"id": "2410.10497", "title": "Continual Learning Improves Zero-Shot Action Recognition", "abstract": "  Zero-shot action recognition requires a strong ability to generalize from\npre-training and seen classes to novel unseen classes. Similarly, continual\nlearning aims to develop models that can generalize effectively and learn new\ntasks without forgetting the ones previously learned. The generalization goals\nof zero-shot and continual learning are closely aligned, however techniques\nfrom continual learning have not been applied to zero-shot action recognition.\nIn this paper, we propose a novel method based on continual learning to address\nzero-shot action recognition. This model, which we call {\\em Generative\nIterative Learning} (GIL) uses a memory of synthesized features of past\nclasses, and combines these synthetic features with real ones from novel\nclasses. The memory is used to train a classification model, ensuring a\nbalanced exposure to both old and new classes. Experiments demonstrate that\n{\\em GIL} improves generalization in unseen classes, achieving a new\nstate-of-the-art in zero-shot recognition across multiple benchmarks.\nImportantly, {\\em GIL} also boosts performance in the more challenging\ngeneralized zero-shot setting, where models need to retain knowledge about\nclasses seen before fine-tuning.\n", "date": "2024-10-14", "categories": []}
{"id": "2410.10589", "title": "MoTE: Reconciling Generalization with Specialization for Visual-Language\n  to Video Knowledge Transfer", "abstract": "  Transferring visual-language knowledge from large-scale foundation models for\nvideo recognition has proved to be effective. To bridge the domain gap,\nadditional parametric modules are added to capture the temporal information.\nHowever, zero-shot generalization diminishes with the increase in the number of\nspecialized parameters, making existing works a trade-off between zero-shot and\nclose-set performance. In this paper, we present MoTE, a novel framework that\nenables generalization and specialization to be balanced in one unified model.\nOur approach tunes a mixture of temporal experts to learn multiple task views\nwith various degrees of data fitting. To maximally preserve the knowledge of\neach expert, we propose \\emph{Weight Merging Regularization}, which regularizes\nthe merging process of experts in weight space. Additionally with temporal\nfeature modulation to regularize the contribution of temporal feature during\ntest. We achieve a sound balance between zero-shot and close-set video\nrecognition tasks and obtain state-of-the-art or competitive results on various\ndatasets, including Kinetics-400 \\& 600, UCF, and HMDB. Code is available at\n\\url{https://github.com/ZMHH-H/MoTE}.\n", "date": "2024-10-14", "categories": []}
{"id": "2404.06741", "title": "An Animation-based Augmentation Approach for Action Recognition from\n  Discontinuous Video", "abstract": "  Action recognition, an essential component of computer vision, plays a\npivotal role in multiple applications. Despite significant improvements brought\nby Convolutional Neural Networks (CNNs), these models suffer performance\ndeclines when trained with discontinuous video frames, which is a frequent\nscenario in real-world settings. This decline primarily results from the loss\nof temporal continuity, which is crucial for understanding the semantics of\nhuman actions. To overcome this issue, we introduce the 4A (Action\nAnimation-based Augmentation Approach) pipeline, which employs a series of\nsophisticated techniques: starting with 2D human pose estimation from RGB\nvideos, followed by Quaternion-based Graph Convolution Network for joint\norientation and trajectory prediction, and Dynamic Skeletal Interpolation for\ncreating smoother, diversified actions using game engine technology. This\ninnovative approach generates realistic animations in varied game environments,\nviewed from multiple viewpoints. In this way, our method effectively bridges\nthe domain gap between virtual and real-world data. In experimental\nevaluations, the 4A pipeline achieves comparable or even superior performance\nto traditional training approaches using real-world data, while requiring only\n10% of the original data volume. Additionally, our approach demonstrates\nenhanced performance on In-the-wild videos, marking a significant advancement\nin the field of action recognition.\n", "date": "2024-04-10", "categories": []}
{"id": "2410.08410", "title": "Human Stone Toolmaking Action Grammar (HSTAG): A Challenging Benchmark\n  for Fine-grained Motor Behavior Recognition", "abstract": "  Action recognition has witnessed the development of a growing number of novel\nalgorithms and datasets in the past decade. However, the majority of public\nbenchmarks were constructed around activities of daily living and annotated at\na rather coarse-grained level, which lacks diversity in domain-specific\ndatasets, especially for rarely seen domains. In this paper, we introduced\nHuman Stone Toolmaking Action Grammar (HSTAG), a meticulously annotated video\ndataset showcasing previously undocumented stone toolmaking behaviors, which\ncan be used for investigating the applications of advanced artificial\nintelligence techniques in understanding a rapid succession of complex\ninteractions between two hand-held objects. HSTAG consists of 18,739 video\nclips that record 4.5 hours of experts' activities in stone toolmaking. Its\nunique features include (i) brief action durations and frequent transitions,\nmirroring the rapid changes inherent in many motor behaviors; (ii) multiple\nangles of view and switches among multiple tools, increasing intra-class\nvariability; (iii) unbalanced class distributions and high similarity among\ndifferent action sequences, adding difficulty in capturing distinct patterns\nfor each action. Several mainstream action recognition models are used to\nconduct experimental analysis, which showcases the challenges and uniqueness of\nHSTAG https://nyu.databrary.org/volume/1697.\n", "date": "2024-10-10", "categories": []}
{"id": "2401.05426", "title": "CoSS: Co-optimizing Sensor and Sampling Rate for Data-Efficient AI in\n  Human Activity Recognition", "abstract": "  Recent advancements in Artificial Neural Networks have significantly improved\nhuman activity recognition using multiple time-series sensors. While employing\nnumerous sensors with high-frequency sampling rates usually improves the\nresults, it often leads to data inefficiency and unnecessary expansion of the\nANN, posing a challenge for their practical deployment on edge devices.\nAddressing these issues, our work introduces a pragmatic framework for\ndata-efficient utilization in HAR tasks, considering the optimization of both\nsensor modalities and sampling rate simultaneously. Central to our approach are\nthe designed trainable parameters, termed 'Weight Scores,' which assess the\nsignificance of each sensor modality and sampling rate during the training\nphase. These scores guide the sensor modalities and sampling rate selection.\nThe pruning method allows users to make a trade-off between computational\nbudgets and performance by selecting the sensor modalities and sampling rates\naccording to the weight score ranking. We tested our framework's effectiveness\nin optimizing sensor modality and sampling rate selection using three public\nHAR benchmark datasets. The results show that the sensor and sampling rate\ncombination selected via CoSS achieves similar classification performance to\nconfigurations using the highest sampling rate with all sensors but at a\nreduced hardware cost.\n", "date": "2024-01-03", "categories": []}
{"id": "2410.07917", "title": "Understanding Human Activity with Uncertainty Measure for Novelty in\n  Graph Convolutional Networks", "abstract": "  Understanding human activity is a crucial aspect of developing intelligent\nrobots, particularly in the domain of human-robot collaboration. Nevertheless,\nexisting systems encounter challenges such as over-segmentation, attributed to\nerrors in the up-sampling process of the decoder. In response, we introduce a\npromising solution: the Temporal Fusion Graph Convolutional Network. This\ninnovative approach aims to rectify the inadequate boundary estimation of\nindividual actions within an activity stream and mitigate the issue of\nover-segmentation in the temporal dimension.\n  Moreover, systems leveraging human activity recognition frameworks for\ndecision-making necessitate more than just the identification of actions. They\nrequire a confidence value indicative of the certainty regarding the\ncorrespondence between observations and training examples. This is crucial to\nprevent overly confident responses to unforeseen scenarios that were not part\nof the training data and may have resulted in mismatches due to weak similarity\nmeasures within the system. To address this, we propose the incorporation of a\nSpectral Normalized Residual connection aimed at enhancing efficient estimation\nof novelty in observations. This innovative approach ensures the preservation\nof input distance within the feature space by imposing constraints on the\nmaximum gradients of weight updates. By limiting these gradients, we promote a\nmore robust handling of novel situations, thereby mitigating the risks\nassociated with overconfidence. Our methodology involves the use of a Gaussian\nprocess to quantify the distance in feature space.\n", "date": "2024-10-10", "categories": []}
{"id": "2308.01532", "title": "MA-FSAR: Multimodal Adaptation of CLIP for Few-Shot Action Recognition", "abstract": "  Applying large-scale vision-language pre-trained models like CLIP to few-shot\naction recognition (FSAR) can significantly enhance both performance and\nefficiency. While several studies have recognized this advantage, most of them\nresort to full-parameter fine-tuning to make CLIP's visual encoder adapt to the\nFSAR data, which not only costs high computations but also overlooks the\npotential of the visual encoder to engage in temporal modeling and focus on\ntargeted semantics directly. To tackle these issues, we introduce MA-FSAR, a\nframework that employs the Parameter-Efficient Fine-Tuning (PEFT) technique to\nenhance the CLIP visual encoder in terms of action-related temporal and\nsemantic representations. Our solution involves a Fine-grained Multimodal\nAdaptation, which is different from the previous attempts of PEFT in regular\naction recognition. Specifically, we first insert a Global Temporal Adaptation\nthat only receives the class token to capture global motion cues efficiently.\nThen these outputs integrate with visual tokens to enhance local temporal\ndynamics by a Local Multimodal Adaptation, which incorporates text features\nunique to the FSAR support set branch to highlight fine-grained semantics\nrelated to actions. In addition to these token-level designs, we propose a\nprototype-level text-guided construction module to further enrich the temporal\nand semantic characteristics of video prototypes. Extensive experiments\ndemonstrate our superior performance in various tasks using minor trainable\nparameters.\n", "date": "2023-08-03", "categories": []}
{"id": "2410.03546", "title": "Multidimensional Human Activity Recognition With Large Language Model: A\n  Conceptual Framework", "abstract": "  In high-stake environments like emergency response or elder care, the\nintegration of large language model (LLM), revolutionize risk assessment,\nresource allocation, and emergency responses in Human Activity Recognition\n(HAR) systems by leveraging data from various wearable sensors. We propose a\nconceptual framework that utilizes various wearable devices, each considered as\na single dimension, to support a multidimensional learning approach within HAR\nsystems. By integrating and processing data from these diverse sources, LLMs\ncan process and translate complex sensor inputs into actionable insights. This\nintegration mitigates the inherent uncertainties and complexities associated\nwith them, and thus enhancing the responsiveness and effectiveness of emergency\nservices. This paper sets the stage for exploring the transformative potential\nof LLMs within HAR systems in empowering emergency workers to navigate the\nunpredictable and risky environments they encounter in their critical roles.\n", "date": "2024-09-16", "categories": []}
{"id": "2304.04023", "title": "Attack-Augmentation Mixing-Contrastive Skeletal Representation Learning", "abstract": "  Contrastive learning, relying on effective positive and negative sample\npairs, is beneficial to learn informative skeleton representations in\nunsupervised skeleton-based action recognition. To achieve these positive and\nnegative pairs, existing weak/strong data augmentation methods have to randomly\nchange the appearance of skeletons for indirectly pursuing semantic\nperturbations. However, such approaches have two limitations: i) solely\nperturbing appearance cannot well capture the intrinsic semantic information of\nskeletons, and ii) randomly perturbation may change the original\npositive/negative pairs to soft positive/negative ones. To address the above\ndilemma, we start the first attempt to explore an attack-based augmentation\nscheme that additionally brings in direct semantic perturbation, for\nconstructing hard positive pairs and further assisting in constructing hard\nnegative pairs. In particular, we propose a novel Attack-Augmentation\nMixing-Contrastive skeletal representation learning (A$^2$MC) to contrast hard\npositive features and hard negative features for learning more robust skeleton\nrepresentations. In A$^2$MC, Attack-Augmentation (Att-Aug) is designed to\ncollaboratively perform targeted and untargeted perturbations of skeletons via\nattack and augmentation respectively, for generating high-quality hard positive\nfeatures. Meanwhile, Positive-Negative Mixer (PNM) is presented to mix hard\npositive features and negative features for generating hard negative features,\nwhich are adopted for updating the mixed memory banks. Extensive experiments on\nthree public datasets demonstrate that A$^2$MC is competitive with the\nstate-of-the-art methods. The code will be accessible on A$^2$MC\n(https://github.com/1xbq1/A2MC).\n", "date": "2023-04-08", "categories": []}
{"id": "2407.03179", "title": "Motion meets Attention: Video Motion Prompts", "abstract": "  Videos contain rich spatio-temporal information. Traditional methods for\nextracting motion, used in tasks such as action recognition, often rely on\nvisual contents rather than precise motion features. This phenomenon is\nreferred to as 'blind motion extraction' behavior, which proves inefficient in\ncapturing motions of interest due to a lack of motion-guided cues. Recently,\nattention mechanisms have enhanced many computer vision tasks by effectively\nhighlighting salient visual areas. Inspired by this, we propose a modified\nSigmoid function with learnable slope and shift parameters as an attention\nmechanism to modulate motion signals from frame differencing maps. This\napproach generates a sequence of attention maps that enhance the processing of\nmotion-related video content. To ensure temporal continuity and smoothness of\nthe attention maps, we apply pair-wise temporal attention variation\nregularization to remove unwanted motions (e.g., noise) while preserving\nimportant ones. We then perform Hadamard product between each pair of attention\nmaps and the original video frames to highlight the evolving motions of\ninterest over time. These highlighted motions, termed video motion prompts, are\nsubsequently used as inputs to the model instead of the original video frames.\nWe formalize this process as a motion prompt layer and incorporate the\nregularization term into the loss function to learn better motion prompts. This\nlayer serves as an adapter between the model and the video data, bridging the\ngap between traditional 'blind motion extraction' and the extraction of\nrelevant motions of interest. We show that our lightweight, plug-and-play\nmotion prompt layer seamlessly integrates into models like SlowFast, X3D, and\nTimeSformer, enhancing performance on benchmarks such as FineGym and MPII\nCooking 2.\n", "date": "2024-07-03", "categories": []}
{"id": "2307.07791", "title": "Cross-Model Cross-Stream Learning for Self-Supervised Human Action\n  Recognition", "abstract": "  Considering the instance-level discriminative ability, contrastive learning\nmethods, including MoCo and SimCLR, have been adapted from the original image\nrepresentation learning task to solve the self-supervised skeleton-based action\nrecognition task. These methods usually use multiple data streams (i.e., joint,\nmotion, and bone) for ensemble learning, meanwhile, how to construct a\ndiscriminative feature space within a single stream and effectively aggregate\nthe information from multiple streams remains an open problem. To this end,\nthis paper first applies a new contrastive learning method called BYOL to learn\nfrom skeleton data, and then formulate SkeletonBYOL as a simple yet effective\nbaseline for self-supervised skeleton-based action recognition. Inspired by\nSkeletonBYOL, this paper further presents a Cross-Model and Cross-Stream (CMCS)\nframework. This framework combines Cross-Model Adversarial Learning (CMAL) and\nCross-Stream Collaborative Learning (CSCL). Specifically, CMAL learns\nsingle-stream representation by cross-model adversarial loss to obtain more\ndiscriminative features. To aggregate and interact with multi-stream\ninformation, CSCL is designed by generating similarity pseudo label of ensemble\nlearning as supervision and guiding feature generation for individual streams.\nExtensive experiments on three datasets verify the complementary properties\nbetween CMAL and CSCL and also verify that the proposed method can achieve\nbetter results than state-of-the-art methods using various evaluation\nprotocols.\n", "date": "2023-07-15", "categories": []}
{"id": "2409.18408", "title": "Query matching for spatio-temporal action detection with query-based\n  object detector", "abstract": "  In this paper, we propose a method that extends the query-based object\ndetection model, DETR, to spatio-temporal action detection, which requires\nmaintaining temporal consistency in videos. Our proposed method applies DETR to\neach frame and uses feature shift to incorporate temporal information. However,\nDETR's object queries in each frame may correspond to different objects, making\na simple feature shift ineffective. To overcome this issue, we propose query\nmatching across different frames, ensuring that queries for the same object are\nmatched and used for the feature shift. Experimental results show that\nperformance on the JHMDB21 dataset improves significantly when query features\nare shifted using the proposed query matching.\n", "date": "2024-09-27", "categories": []}
{"id": "2409.17483", "title": "Heterogeneous Hyper-Graph Neural Networks for Context-aware Human\n  Activity Recognition", "abstract": "  Context-aware Human Activity Recognition (CHAR) is challenging due to the\nneed to recognize the user's current activity from signals that vary\nsignificantly with contextual factors such as phone placements and the varied\nstyles with which different users perform the same activity. In this paper, we\nargue that context-aware activity visit patterns in realistic in-the-wild data\ncan equivocally be considered as a general graph representation learning task.\nWe posit that exploiting underlying graphical patterns in CHAR data can improve\nCHAR task performance and representation learning. Building on the intuition\nthat certain activities are frequently performed with the phone placed in\ncertain positions, we focus on the context-aware human activity problem of\nrecognizing the <Activity, Phone Placement> tuple. We demonstrate that CHAR\ndata has an underlying graph structure that can be viewed as a heterogenous\nhypergraph that has multiple types of nodes and hyperedges (an edge connecting\nmore than two nodes). Subsequently, learning <Activity, Phone Placement>\nrepresentations becomes a graph node representation learning problem. After\ntask transformation, we further propose a novel Heterogeneous HyperGraph Neural\nNetwork architecture for Context-aware Human Activity Recognition (HHGNN-CHAR),\nwith three types of heterogeneous nodes (user, phone placement, and activity).\nConnections between all types of nodes are represented by hyperedges. Rigorous\nevaluation demonstrated that on an unscripted, in-the-wild CHAR dataset, our\nproposed framework significantly outperforms state-of-the-art (SOTA) baselines\nincluding CHAR models that do not exploit graphs, and GNN variants that do not\nincorporate heterogeneous nodes or hyperedges with overall improvements 14.04%\non Matthews Correlation Coefficient (MCC) and 7.01% on Macro F1 scores.\n", "date": "2024-09-26", "categories": []}
{"id": "2409.17951", "title": "Spatial Hierarchy and Temporal Attention Guided Cross Masking for\n  Self-supervised Skeleton-based Action Recognition", "abstract": "  In self-supervised skeleton-based action recognition, the mask reconstruction\nparadigm is gaining interest in enhancing model refinement and robustness\nthrough effective masking. However, previous works primarily relied on a single\nmasking criterion, resulting in the model overfitting specific features and\noverlooking other effective information. In this paper, we introduce a\nhierarchy and attention guided cross-masking framework (HA-CM) that applies\nmasking to skeleton sequences from both spatial and temporal perspectives.\nSpecifically, in spatial graphs, we utilize hyperbolic space to maintain joint\ndistinctions and effectively preserve the hierarchical structure of\nhigh-dimensional skeletons, employing joint hierarchy as the masking criterion.\nIn temporal flows, we substitute traditional distance metrics with the global\nattention of joints for masking, addressing the convergence of distances in\nhigh-dimensional space and the lack of a global perspective. Additionally, we\nincorporate cross-contrast loss based on the cross-masking framework into the\nloss function to enhance the model's learning of instance-level features. HA-CM\nshows efficiency and universality on three public large-scale datasets, NTU-60,\nNTU-120, and PKU-MMD. The source code of our HA-CM is available at\nhttps://github.com/YinxPeng/HA-CM-main.\n", "date": "2024-09-26", "categories": []}
{"id": "2403.08214", "title": "P2LHAP:Wearable sensor-based human activity recognition, segmentation\n  and forecast through Patch-to-Label Seq2Seq Transformer", "abstract": "  Traditional deep learning methods struggle to simultaneously segment,\nrecognize, and forecast human activities from sensor data. This limits their\nusefulness in many fields such as healthcare and assisted living, where\nreal-time understanding of ongoing and upcoming activities is crucial. This\npaper introduces P2LHAP, a novel Patch-to-Label Seq2Seq framework that tackles\nall three tasks in a efficient single-task model. P2LHAP divides sensor data\nstreams into a sequence of \"patches\", served as input tokens, and outputs a\nsequence of patch-level activity labels including the predicted future\nactivities. A unique smoothing technique based on surrounding patch labels, is\nproposed to identify activity boundaries accurately. Additionally, P2LHAP\nlearns patch-level representation by sensor signal channel-independent\nTransformer encoders and decoders. All channels share embedding and Transformer\nweights across all sequences. Evaluated on three public datasets, P2LHAP\nsignificantly outperforms the state-of-the-art in all three tasks,\ndemonstrating its effectiveness and potential for real-world applications.\n", "date": "2024-03-13", "categories": []}
{"id": "2409.14205", "title": "Egocentric zone-aware action recognition across environments", "abstract": "  Human activities exhibit a strong correlation between actions and the places\nwhere these are performed, such as washing something at a sink. More\nspecifically, in daily living environments we may identify particular\nlocations, hereinafter named activity-centric zones, which may afford a set of\nhomogeneous actions. Their knowledge can serve as a prior to favor vision\nmodels to recognize human activities. However, the appearance of these zones is\nscene-specific, limiting the transferability of this prior information to\nunfamiliar areas and domains. This problem is particularly relevant in\negocentric vision, where the environment takes up most of the image, making it\neven more difficult to separate the action from the context. In this paper, we\ndiscuss the importance of decoupling the domain-specific appearance of\nactivity-centric zones from their universal, domain-agnostic representations,\nand show how the latter can improve the cross-domain transferability of\nEgocentric Action Recognition (EAR) models. We validate our solution on the\nEPIC-Kitchens-100 and Argo1M datasets\n", "date": "2024-09-21", "categories": []}
{"id": "2409.13091", "title": "Interpretable Action Recognition on Hard to Classify Actions", "abstract": "  We investigate a human-like interpretable model of video understanding.\nHumans recognise complex activities in video by recognising critical\nspatio-temporal relations among explicitly recognised objects and parts, for\nexample, an object entering the aperture of a container. To mimic this we build\non a model which uses positions of objects and hands, and their motions, to\nrecognise the activity taking place. To improve this model we focussed on three\nof the most confused classes (for this model) and identified that the lack of\n3D information was the major problem. To address this we extended our basic\nmodel by adding 3D awareness in two ways: (1) A state-of-the-art object\ndetection model was fine-tuned to determine the difference between \"Container\"\nand \"NotContainer\" in order to integrate object shape information into the\nexisting object features. (2) A state-of-the-art depth estimation model was\nused to extract depth values for individual objects and calculate depth\nrelations to expand the existing relations used our interpretable model. These\n3D extensions to our basic model were evaluated on a subset of three\nsuperficially similar \"Putting\" actions from the Something-Something-v2\ndataset. The results showed that the container detector did not improve\nperformance, but the addition of depth relations made a significant improvement\nto performance.\n", "date": "2024-09-19", "categories": []}
{"id": "2408.03612", "title": "JARViS: Detecting Actions in Video Using Unified Actor-Scene Context\n  Relation Modeling", "abstract": "  Video action detection (VAD) is a formidable vision task that involves the\nlocalization and classification of actions within the spatial and temporal\ndimensions of a video clip. Among the myriad VAD architectures, two-stage VAD\nmethods utilize a pre-trained person detector to extract the region of interest\nfeatures, subsequently employing these features for action detection. However,\nthe performance of two-stage VAD methods has been limited as they depend solely\non localized actor features to infer action semantics. In this study, we\npropose a new two-stage VAD framework called Joint Actor-scene context Relation\nmodeling based on Visual Semantics (JARViS), which effectively consolidates\ncross-modal action semantics distributed globally across spatial and temporal\ndimensions using Transformer attention. JARViS employs a person detector to\nproduce densely sampled actor features from a keyframe. Concurrently, it uses a\nvideo backbone to create spatio-temporal scene features from a video clip.\nFinally, the fine-grained interactions between actors and scenes are modeled\nthrough a Unified Action-Scene Context Transformer to directly output the final\nset of actions in parallel. Our experimental results demonstrate that JARViS\noutperforms existing methods by significant margins and achieves\nstate-of-the-art performance on three popular VAD datasets, including AVA,\nUCF101-24, and JHMDB51-21.\n", "date": "2024-08-07", "categories": []}
{"id": "2211.01783", "title": "Quantifying and Learning Static vs. Dynamic Information in Deep\n  Spatiotemporal Networks", "abstract": "  There is limited understanding of the information captured by deep\nspatiotemporal models in their intermediate representations. For example, while\nevidence suggests that action recognition algorithms are heavily influenced by\nvisual appearance in single frames, no quantitative methodology exists for\nevaluating such static bias in the latent representation compared to bias\ntoward dynamics. We tackle this challenge by proposing an approach for\nquantifying the static and dynamic biases of any spatiotemporal model, and\napply our approach to three tasks, action recognition, automatic video object\nsegmentation (AVOS) and video instance segmentation (VIS). Our key findings\nare: (i) Most examined models are biased toward static information. (ii) Some\ndatasets that are assumed to be biased toward dynamics are actually biased\ntoward static information. (iii) Individual channels in an architecture can be\nbiased toward static, dynamic or a combination of the two. (iv) Most models\nconverge to their culminating biases in the first half of training. We then\nexplore how these biases affect performance on dynamically biased datasets. For\naction recognition, we propose StaticDropout, a semantically guided dropout\nthat debiases a model from static information toward dynamics. For AVOS, we\ndesign a better combination of fusion and cross connection layers compared with\nprevious architectures.\n", "date": "2022-11-03", "categories": []}
{"id": "2405.20606", "title": "Vision-Language Meets the Skeleton: Progressively Distillation with\n  Cross-Modal Knowledge for 3D Action Representation Learning", "abstract": "  Skeleton-based action representation learning aims to interpret and\nunderstand human behaviors by encoding the skeleton sequences, which can be\ncategorized into two primary training paradigms: supervised learning and\nself-supervised learning. However, the former one-hot classification requires\nlabor-intensive predefined action categories annotations, while the latter\ninvolves skeleton transformations (e.g., cropping) in the pretext tasks that\nmay impair the skeleton structure. To address these challenges, we introduce a\nnovel skeleton-based training framework (C$^2$VL) based on Cross-modal\nContrastive learning that uses the progressive distillation to learn\ntask-agnostic human skeleton action representation from the Vision-Language\nknowledge prompts. Specifically, we establish the vision-language action\nconcept space through vision-language knowledge prompts generated by\npre-trained large multimodal models (LMMs), which enrich the fine-grained\ndetails that the skeleton action space lacks. Moreover, we propose the\nintra-modal self-similarity and inter-modal cross-consistency softened targets\nin the cross-modal representation learning process to progressively control and\nguide the degree of pulling vision-language knowledge prompts and corresponding\nskeletons closer. These soft instance discrimination and self-knowledge\ndistillation strategies contribute to the learning of better skeleton-based\naction representations from the noisy skeleton-vision-language pairs. During\nthe inference phase, our method requires only the skeleton data as the input\nfor action recognition and no longer for vision-language prompts. Extensive\nexperiments on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD datasets demonstrate\nthat our method outperforms the previous methods and achieves state-of-the-art\nresults. Code is available at: https://github.com/cseeyangchen/C2VL.\n", "date": "2024-05-31", "categories": []}
{"id": "2409.09611", "title": "Integrating Audio Narrations to Strengthen Domain Generalization in\n  Multimodal First-Person Action Recognition", "abstract": "  First-person activity recognition is rapidly growing due to the widespread\nuse of wearable cameras but faces challenges from domain shifts across\ndifferent environments, such as varying objects or background scenes. We\npropose a multimodal framework that improves domain generalization by\nintegrating motion, audio, and appearance features. Key contributions include\nanalyzing the resilience of audio and motion features to domain shifts, using\naudio narrations for enhanced audio-text alignment, and applying consistency\nratings between audio and visual narrations to optimize the impact of audio in\nrecognition during training. Our approach achieves state-of-the-art performance\non the ARGO1M dataset, effectively generalizing across unseen scenarios and\nlocations.\n", "date": "2024-09-15", "categories": []}
{"id": "2409.09678", "title": "A Comprehensive Methodological Survey of Human Activity Recognition\n  Across Divers Data Modalities", "abstract": "  Human Activity Recognition (HAR) systems aim to understand human behaviour\nand assign a label to each action, attracting significant attention in computer\nvision due to their wide range of applications. HAR can leverage various data\nmodalities, such as RGB images and video, skeleton, depth, infrared, point\ncloud, event stream, audio, acceleration, and radar signals. Each modality\nprovides unique and complementary information suited to different application\nscenarios. Consequently, numerous studies have investigated diverse approaches\nfor HAR using these modalities. This paper presents a comprehensive survey of\nthe latest advancements in HAR from 2014 to 2024, focusing on machine learning\n(ML) and deep learning (DL) approaches categorized by input data modalities. We\nreview both single-modality and multi-modality techniques, highlighting\nfusion-based and co-learning frameworks. Additionally, we cover advancements in\nhand-crafted action features, methods for recognizing human-object\ninteractions, and activity detection. Our survey includes a detailed dataset\ndescription for each modality and a summary of the latest HAR systems, offering\ncomparative results on benchmark datasets. Finally, we provide insightful\nobservations and propose effective future research directions in HAR.\n", "date": "2024-09-15", "categories": []}
{"id": "2409.08853", "title": "Using The Concept Hierarchy for Household Action Recognition", "abstract": "  We propose a method to systematically represent both the static and the\ndynamic components of environments, i.e. objects and agents, as well as the\nchanges that are happening in the environment, i.e. the actions and skills\nperformed by agents. Our approach, the Concept Hierarchy, provides the\nnecessary information for autonomous systems to represent environment states,\nperform action modeling and recognition, and plan the execution of tasks.\nAdditionally, the hierarchical structure supports generalization and knowledge\ntransfer to environments. We rigorously define tasks, actions, skills, and\naffordances that enable human-understandable action and skill recognition.\n", "date": "2024-09-13", "categories": []}
{"id": "2408.15679", "title": "DEAR: Depth-Enhanced Action Recognition", "abstract": "  Detecting actions in videos, particularly within cluttered scenes, poses\nsignificant challenges due to the limitations of 2D frame analysis from a\ncamera perspective. Unlike human vision, which benefits from 3D understanding,\nrecognizing actions in such environments can be difficult. This research\nintroduces a novel approach integrating 3D features and depth maps alongside\nRGB features to enhance action recognition accuracy. Our method involves\nprocessing estimated depth maps through a separate branch from the RGB feature\nencoder and fusing the features to understand the scene and actions\ncomprehensively. Using the Side4Video framework and VideoMamba, which employ\nCLIP and VisionMamba for spatial feature extraction, our approach outperformed\nour implementation of the Side4Video network on the Something-Something V2\ndataset. Our code is available at: https://github.com/SadeghRahmaniB/DEAR\n", "date": "2024-08-28", "categories": []}
{"id": "2409.07581", "title": "Violence detection in videos using deep recurrent and convolutional\n  neural networks", "abstract": "  Violence and abnormal behavior detection research have known an increase of\ninterest in recent years, due mainly to a rise in crimes in large cities\nworldwide. In this work, we propose a deep learning architecture for violence\ndetection which combines both recurrent neural networks (RNNs) and\n2-dimensional convolutional neural networks (2D CNN). In addition to video\nframes, we use optical flow computed using the captured sequences. CNN extracts\nspatial characteristics in each frame, while RNN extracts temporal\ncharacteristics. The use of optical flow allows to encode the movements in the\nscenes. The proposed approaches reach the same level as the state-of-the-art\ntechniques and sometime surpass them. It was validated on 3 databases achieving\ngood results.\n", "date": "2024-09-11", "categories": []}
{"id": "2409.07588", "title": "2D bidirectional gated recurrent unit convolutional Neural networks for\n  end-to-end violence detection In videos", "abstract": "  Abnormal behavior detection, action recognition, fight and violence detection\nin videos is an area that has attracted a lot of interest in recent years. In\nthis work, we propose an architecture that combines a Bidirectional Gated\nRecurrent Unit (BiGRU) and a 2D Convolutional Neural Network (CNN) to detect\nviolence in video sequences. A CNN is used to extract spatial characteristics\nfrom each frame, while the BiGRU extracts temporal and local motion\ncharacteristics using CNN extracted features from multiple frames. The proposed\nend-to-end deep learning network is tested in three public datasets with\nvarying scene complexities. The proposed network achieves accuracies up to 98%.\nThe obtained results are promising and show the performance of the proposed\nend-to-end approach.\n", "date": "2024-09-11", "categories": []}
{"id": "2407.19698", "title": "Classification Matters: Improving Video Action Detection with\n  Class-Specific Attention", "abstract": "  Video action detection (VAD) aims to detect actors and classify their actions\nin a video. We figure that VAD suffers more from classification rather than\nlocalization of actors. Hence, we analyze how prevailing methods form features\nfor classification and find that they prioritize actor regions, yet often\noverlooking the essential contextual information necessary for accurate\nclassification. Accordingly, we propose to reduce the bias toward actor and\nencourage paying attention to the context that is relevant to each action\nclass. By assigning a class-dedicated query to each action class, our model can\ndynamically determine where to focus for effective classification. The proposed\nmodel demonstrates superior performance on three challenging benchmarks with\nsignificantly fewer parameters and less computation.\n", "date": "2024-07-29", "categories": []}
{"id": "2408.16412", "title": "Text-Enhanced Zero-Shot Action Recognition: A training-free approach", "abstract": "  Vision-language models (VLMs) have demonstrated remarkable performance across\nvarious visual tasks, leveraging joint learning of visual and textual\nrepresentations. While these models excel in zero-shot image tasks, their\napplication to zero-shot video action recognition (ZSVAR) remains challenging\ndue to the dynamic and temporal nature of actions. Existing methods for ZS-VAR\ntypically require extensive training on specific datasets, which can be\nresource-intensive and may introduce domain biases. In this work, we propose\nText-Enhanced Action Recognition (TEAR), a simple approach to ZS-VAR that is\ntraining-free and does not require the availability of training data or\nextensive computational resources. Drawing inspiration from recent findings in\nvision and language literature, we utilize action descriptors for decomposition\nand contextual information to enhance zero-shot action recognition. Through\nexperiments on UCF101, HMDB51, and Kinetics-600 datasets, we showcase the\neffectiveness and applicability of our proposed approach in addressing the\nchallenges of ZS-VAR.\n", "date": "2024-08-29", "categories": []}
{"id": "2310.02523", "title": "A Spatio-Temporal Attention-Based Method for Detecting Student Classroom\n  Behaviors", "abstract": "  Accurately detecting student behavior from classroom videos is beneficial for\nanalyzing their classroom status and improving teaching efficiency. However,\nlow accuracy in student classroom behavior detection is a prevalent issue. To\naddress this issue, we propose a Spatio-Temporal Attention-Based Method for\nDetecting Student Classroom Behaviors (BDSTA). Firstly, the SlowFast network is\nused to generate motion and environmental information feature maps from the\nvideo. Then, the spatio-temporal attention module is applied to the feature\nmaps, including information aggregation, compression and stimulation processes.\nSubsequently, attention maps in the time, channel and space dimensions are\nobtained, and multi-label behavior classification is performed based on these\nattention maps. To solve the long-tail data problem that exists in student\nclassroom behavior datasets, we use an improved focal loss function to assign\nmore weight to the tail class data during training. Experimental results are\nconducted on a self-made student classroom behavior dataset named STSCB.\nCompared with the SlowFast model, the average accuracy of student behavior\nclassification detection improves by 8.94\\% using BDSTA.\n", "date": "2023-10-04", "categories": []}
{"id": "2310.16267", "title": "Student Classroom Behavior Detection based on Spatio-Temporal Network\n  and Multi-Model Fusion", "abstract": "  Using deep learning methods to detect students' classroom behavior\nautomatically is a promising approach for analyzing their class performance and\nimproving teaching effectiveness. However, the lack of publicly available\nspatio-temporal datasets on student behavior, as well as the high cost of\nmanually labeling such datasets, pose significant challenges for researchers in\nthis field. To address this issue, we proposed a method for extending the\nspatio-temporal behavior dataset in Student Classroom Scenarios\n(SCB-ST-Dataset4) through image dataset. Our SCB-ST-Dataset4 comprises 757265\nimages with 25810 labels, focusing on 3 behaviors: hand-raising, reading,\nwriting. Our proposed method can rapidly generate spatio-temporal behavior\ndatasets without requiring extra manual labeling. Furthermore, we proposed a\nBehavior Similarity Index (BSI) to explore the similarity of behaviors. We\nevaluated the dataset using the YOLOv5, YOLOv7, YOLOv8, and SlowFast\nalgorithms, achieving a mean average precision (map) of up to 82.3%. Last, we\nfused multiple models to generate student behavior-related data from various\nperspectives. The experiment further demonstrates the effectiveness of our\nmethod. And SCB-ST-Dataset4 provides a robust foundation for future research in\nstudent behavior detection, potentially contributing to advancements in this\nfield. The SCB-ST-Dataset4 is available for download at:\nhttps://github.com/Whiffe/SCB-dataset.\n", "date": "2023-10-25", "categories": []}
{"id": "2311.17366", "title": "Generative Hierarchical Temporal Transformer for Hand Pose and Action\n  Modeling", "abstract": "  We present a novel unified framework that concurrently tackles recognition\nand future prediction for human hand pose and action modeling. Previous works\ngenerally provide isolated solutions for either recognition or prediction,\nwhich not only increases the complexity of integration in practical\napplications, but more importantly, cannot exploit the synergy of both sides\nand suffer suboptimal performances in their respective domains. To address this\nproblem, we propose a generative Transformer VAE architecture to model hand\npose and action, where the encoder and decoder capture recognition and\nprediction respectively, and their connection through the VAE bottleneck\nmandates the learning of consistent hand motion from the past to the future and\nvice versa. Furthermore, to faithfully model the semantic dependency and\ndifferent temporal granularity of hand pose and action, we decompose the\nframework into two cascaded VAE blocks: the first and latter blocks\nrespectively model the short-span poses and long-span action, and are connected\nby a mid-level feature representing a sub-second series of hand poses. This\ndecomposition into block cascades facilitates capturing both short-term and\nlong-term temporal regularity in pose and action modeling, and enables training\ntwo blocks separately to fully utilize datasets with annotations of different\ntemporal granularities. We train and evaluate our framework across multiple\ndatasets; results show that our joint modeling of recognition and prediction\nimproves over isolated solutions, and that our semantic and temporal hierarchy\nfacilitates long-term pose and action modeling.\n", "date": "2023-11-29", "categories": []}
{"id": "2408.13152", "title": "Long-term Pre-training for Temporal Action Detection with Transformers", "abstract": "  Temporal action detection (TAD) is challenging, yet fundamental for\nreal-world video applications. Recently, DETR-based models for TAD have been\nprevailing thanks to their unique benefits. However, transformers demand a huge\ndataset, and unfortunately data scarcity in TAD causes a severe degeneration.\nIn this paper, we identify two crucial problems from data scarcity: attention\ncollapse and imbalanced performance. To this end, we propose a new pre-training\nstrategy, Long-Term Pre-training (LTP), tailored for transformers. LTP has two\nmain components: 1) class-wise synthesis, 2) long-term pretext tasks. Firstly,\nwe synthesize long-form video features by merging video snippets of a target\nclass and non-target classes. They are analogous to untrimmed data used in TAD,\ndespite being created from trimmed data. In addition, we devise two types of\nlong-term pretext tasks to learn long-term dependency. They impose long-term\nconditions such as finding second-to-fourth or short-duration actions. Our\nextensive experiments show state-of-the-art performances in DETR-based methods\non ActivityNet-v1.3 and THUMOS14 by a large margin. Moreover, we demonstrate\nthat LTP significantly relieves the data scarcity issues in TAD.\n", "date": "2024-08-23", "categories": []}
{"id": "2409.05749", "title": "ReL-SAR: Representation Learning for Skeleton Action Recognition with\n  Convolutional Transformers and BYOL", "abstract": "  To extract robust and generalizable skeleton action recognition features,\nlarge amounts of well-curated data are typically required, which is a\nchallenging task hindered by annotation and computation costs. Therefore,\nunsupervised representation learning is of prime importance to leverage\nunlabeled skeleton data. In this work, we investigate unsupervised\nrepresentation learning for skeleton action recognition. For this purpose, we\ndesigned a lightweight convolutional transformer framework, named ReL-SAR,\nexploiting the complementarity of convolutional and attention layers for\njointly modeling spatial and temporal cues in skeleton sequences. We also use a\nSelection-Permutation strategy for skeleton joints to ensure more informative\ndescriptions from skeletal data. Finally, we capitalize on Bootstrap Your Own\nLatent (BYOL) to learn robust representations from unlabeled skeleton sequence\ndata. We achieved very competitive results on limited-size datasets: MCAD,\nIXMAS, JHMDB, and NW-UCLA, showing the effectiveness of our proposed method\nagainst state-of-the-art methods in terms of both performance and computational\nefficiency. To ensure reproducibility and reusability, the source code\nincluding all implementation parameters is provided at:\nhttps://github.com/SafwenNaimi/Representation-Learning-for-Skeleton-Action-Recognition-with-Convolutional-Transformers-and-BYOL\n", "date": "2024-09-09", "categories": []}
{"id": "2206.11011", "title": "Weakly-Supervised Temporal Action Localization by Progressive\n  Complementary Learning", "abstract": "  Weakly Supervised Temporal Action Localization (WSTAL) aims to localize and\nclassify action instances in long untrimmed videos with only video-level\ncategory labels. Due to the lack of snippet-level supervision for indicating\naction boundaries, previous methods typically assign pseudo labels for\nunlabeled snippets. However, since some action instances of different\ncategories are visually similar, it is non-trivial to exactly label the\n(usually) one action category for a snippet, and incorrect pseudo labels would\nimpair the localization performance. To address this problem, we propose a\nnovel method from a category exclusion perspective, named Progressive\nComplementary Learning (ProCL), which gradually enhances the snippet-level\nsupervision. Our method is inspired by the fact that video-level labels\nprecisely indicate the categories that all snippets surely do not belong to,\nwhich is ignored by previous works. Accordingly, we first exclude these surely\nnon-existent categories by a complementary learning loss. And then, we\nintroduce the background-aware pseudo complementary labeling in order to\nexclude more categories for snippets of less ambiguity. Furthermore, for the\nremaining ambiguous snippets, we attempt to reduce the ambiguity by\ndistinguishing foreground actions from the background. Extensive experimental\nresults show that our method achieves new state-of-the-art performance on two\npopular benchmarks, namely THUMOS14 and ActivityNet1.3.\n", "date": "2022-06-22", "categories": []}
{"id": "2312.00826", "title": "DEVIAS: Learning Disentangled Video Representations of Action and Scene", "abstract": "  Video recognition models often learn scene-biased action representation due\nto the spurious correlation between actions and scenes in the training data.\nSuch models show poor performance when the test data consists of videos with\nunseen action-scene combinations. Although scene-debiased action recognition\nmodels might address the issue, they often overlook valuable scene information\nin the data. To address this challenge, we propose to learn DisEntangled VIdeo\nrepresentations of Action and Scene (DEVIAS), for more holistic video\nunderstanding. We propose an encoder-decoder architecture to learn disentangled\naction and scene representations with a single model. The architecture consists\nof a disentangling encoder (DE), an action mask decoder (AMD), and a prediction\nhead. The key to achieving the disentanglement is employing both DE and AMD\nduring training time. The DE uses the slot attention mechanism to learn\ndisentangled action and scene representations. For further disentanglement, an\nAMD learns to predict action masks, given an action slot. With the resulting\ndisentangled representations, we can achieve robust performance across diverse\nscenarios, including both seen and unseen action-scene combinations. We\nrigorously validate the proposed method on the UCF-101, Kinetics-400, and HVU\ndatasets for the seen, and the SCUBA, HAT, and HVU datasets for unseen\naction-scene combination scenarios. Furthermore, DEVIAS provides flexibility to\nadjust the emphasis on action or scene information depending on dataset\ncharacteristics for downstream tasks. DEVIAS shows favorable performance in\nvarious downstream tasks: Diving48, Something-Something-V2, UCF-101, and\nActivityNet. The code is available at https://github.com/KHU-VLL/DEVIAS.\n", "date": "2023-11-30", "categories": []}
{"id": "2409.04205", "title": "Introducing Gating and Context into Temporal Action Detection", "abstract": "  Temporal Action Detection (TAD), the task of localizing and classifying\nactions in untrimmed video, remains challenging due to action overlaps and\nvariable action durations. Recent findings suggest that TAD performance is\ndependent on the structural design of transformers rather than on the\nself-attention mechanism. Building on this insight, we propose a refined\nfeature extraction process through lightweight, yet effective operations.\nFirst, we employ a local branch that employs parallel convolutions with varying\nwindow sizes to capture both fine-grained and coarse-grained temporal features.\nThis branch incorporates a gating mechanism to select the most relevant\nfeatures. Second, we introduce a context branch that uses boundary frames as\nkey-value pairs to analyze their relationship with the central frame through\ncross-attention. The proposed method captures temporal dependencies and\nimproves contextual understanding. Evaluations of the gating mechanism and\ncontext branch on challenging datasets (THUMOS14 and EPIC-KITCHEN 100) show a\nconsistent improvement over the baseline and existing methods.\n", "date": "2024-09-06", "categories": []}
{"id": "2407.08572", "title": "Boosting Adversarial Transferability for Skeleton-based Action\n  Recognition via Exploring the Model Posterior Space", "abstract": "  Skeletal motion plays a pivotal role in human activity recognition (HAR).\nRecently, attack methods have been proposed to identify the universal\nvulnerability of skeleton-based HAR(S-HAR). However, the research of\nadversarial transferability on S-HAR is largely missing. More importantly,\nexisting attacks all struggle in transfer across unknown S-HAR models. We\nobserved that the key reason is that the loss landscape of the action\nrecognizers is rugged and sharp. Given the established correlation in prior\nstudies~\\cite{qin2022boosting,wu2020towards} between loss landscape and\nadversarial transferability, we assume and empirically validate that smoothing\nthe loss landscape could potentially improve adversarial transferability on\nS-HAR. This is achieved by proposing a new post-train Dual Bayesian strategy,\nwhich can effectively explore the model posterior space for a collection of\nsurrogates without the need for re-training. Furthermore, to craft adversarial\nexamples along the motion manifold, we incorporate the attack gradient with\ninformation of the motion dynamics in a Bayesian manner. Evaluated on benchmark\ndatasets, e.g. HDM05 and NTU 60, the average transfer success rate can reach as\nhigh as 35.9\\% and 45.5\\% respectively. In comparison, current state-of-the-art\nskeletal attacks achieve only 3.6\\% and 9.8\\%. The high adversarial\ntransferability remains consistent across various surrogate, victim, and even\ndefense models. Through a comprehensive analysis of the results, we provide\ninsights on what surrogates are more likely to exhibit transferability, to shed\nlight on future research.\n", "date": "2024-07-11", "categories": []}
{"id": "2409.02385", "title": "Unified Framework with Consistency across Modalities for Human Activity\n  Recognition", "abstract": "  Recognizing human activities in videos is challenging due to the\nspatio-temporal complexity and context-dependence of human interactions. Prior\nstudies often rely on single input modalities, such as RGB or skeletal data,\nlimiting their ability to exploit the complementary advantages across\nmodalities. Recent studies focus on combining these two modalities using simple\nfeature fusion techniques. However, due to the inherent disparities in\nrepresentation between these input modalities, designing a unified neural\nnetwork architecture to effectively leverage their complementary information\nremains a significant challenge. To address this, we propose a comprehensive\nmultimodal framework for robust video-based human activity recognition. Our key\ncontribution is the introduction of a novel compositional query machine, called\nCOMPUTER ($\\textbf{COMP}ositional h\\textbf{U}man-cen\\textbf{T}ric\nqu\\textbf{ER}y$ machine), a generic neural architecture that models the\ninteractions between a human of interest and its surroundings in both space and\ntime. Thanks to its versatile design, COMPUTER can be leveraged to distill\ndistinctive representations for various input modalities. Additionally, we\nintroduce a consistency loss that enforces agreement in prediction between\nmodalities, exploiting the complementary information from multimodal inputs for\nrobust human movement recognition. Through extensive experiments on action\nlocalization and group activity recognition tasks, our approach demonstrates\nsuperior performance when compared with state-of-the-art methods. Our code is\navailable at: https://github.com/tranxuantuyen/COMPUTER.\n", "date": "2024-09-04", "categories": []}
{"id": "2409.02910", "title": "SITAR: Semi-supervised Image Transformer for Action Recognition", "abstract": "  Recognizing actions from a limited set of labeled videos remains a challenge\nas annotating visual data is not only tedious but also can be expensive due to\nclassified nature. Moreover, handling spatio-temporal data using deep $3$D\ntransformers for this can introduce significant computational complexity. In\nthis paper, our objective is to address video action recognition in a\nsemi-supervised setting by leveraging only a handful of labeled videos along\nwith a collection of unlabeled videos in a compute efficient manner.\nSpecifically, we rearrange multiple frames from the input videos in row-column\nform to construct super images. Subsequently, we capitalize on the vast pool of\nunlabeled samples and employ contrastive learning on the encoded super images.\nOur proposed approach employs two pathways to generate representations for\ntemporally augmented super images originating from the same video.\nSpecifically, we utilize a 2D image-transformer to generate representations and\napply a contrastive loss function to minimize the similarity between\nrepresentations from different videos while maximizing the representations of\nidentical videos. Our method demonstrates superior performance compared to\nexisting state-of-the-art approaches for semi-supervised action recognition\nacross various benchmark datasets, all while significantly reducing\ncomputational costs.\n", "date": "2024-09-04", "categories": []}
{"id": "2311.18825", "title": "CAST: Cross-Attention in Space and Time for Video Action Recognition", "abstract": "  Recognizing human actions in videos requires spatial and temporal\nunderstanding. Most existing action recognition models lack a balanced\nspatio-temporal understanding of videos. In this work, we propose a novel\ntwo-stream architecture, called Cross-Attention in Space and Time (CAST), that\nachieves a balanced spatio-temporal understanding of videos using only RGB\ninput. Our proposed bottleneck cross-attention mechanism enables the spatial\nand temporal expert models to exchange information and make synergistic\npredictions, leading to improved performance. We validate the proposed method\nwith extensive experiments on public benchmarks with different characteristics:\nEPIC-KITCHENS-100, Something-Something-V2, and Kinetics-400. Our method\nconsistently shows favorable performance across these datasets, while the\nperformance of existing methods fluctuates depending on the dataset\ncharacteristics.\n", "date": "2023-11-30", "categories": []}
{"id": "2401.11790", "title": "Deep Learning for Computer Vision based Activity Recognition and Fall\n  Detection of the Elderly: a Systematic Review", "abstract": "  As the percentage of elderly people in developed countries increases\nworldwide, the healthcare of this collective is a worrying matter, especially\nif it includes the preservation of their autonomy. In this direction, many\nstudies are being published on Ambient Assisted Living (AAL) systems, which\nhelp to reduce the preoccupations raised by the independent living of the\nelderly. In this study, a systematic review of the literature is presented on\nfall detection and Human Activity Recognition (HAR) for the elderly, as the two\nmain tasks to solve to guarantee the safety of elderly people living alone. To\naddress the current tendency to perform these two tasks, the review focuses on\nthe use of Deep Learning (DL) based approaches on computer vision data. In\naddition, different collections of data like DL models, datasets or hardware\n(e.g. depth or thermal cameras) are gathered from the reviewed studies and\nprovided for reference in future studies. Strengths and weaknesses of existing\napproaches are also discussed and, based on them, our recommendations for\nfuture works are provided.\n", "date": "2024-01-22", "categories": []}
{"id": "2409.00349", "title": "ToddlerAct: A Toddler Action Recognition Dataset for Gross Motor\n  Development Assessment", "abstract": "  Assessing gross motor development in toddlers is crucial for understanding\ntheir physical development and identifying potential developmental delays or\ndisorders. However, existing datasets for action recognition primarily focus on\nadults, lacking the diversity and specificity required for accurate assessment\nin toddlers. In this paper, we present ToddlerAct, a toddler gross motor action\nrecognition dataset, aiming to facilitate research in early childhood\ndevelopment. The dataset consists of video recordings capturing a variety of\ngross motor activities commonly observed in toddlers aged under three years\nold. We describe the data collection process, annotation methodology, and\ndataset characteristics. Furthermore, we benchmarked multiple state-of-the-art\nmethods including image-based and skeleton-based action recognition methods on\nour datasets. Our findings highlight the importance of domain-specific datasets\nfor accurate assessment of gross motor development in toddlers and lay the\nfoundation for future research in this critical area. Our dataset will be\navailable at https://github.com/ipl-uw/ToddlerAct.\n", "date": "2024-08-31", "categories": []}
{"id": "2409.00731", "title": "A Critical Analysis on Machine Learning Techniques for Video-based Human\n  Activity Recognition of Surveillance Systems: A Review", "abstract": "  Upsurging abnormal activities in crowded locations such as airports, train\nstations, bus stops, shopping malls, etc., urges the necessity for an\nintelligent surveillance system. An intelligent surveillance system can\ndifferentiate between normal and suspicious activities from real-time video\nanalysis that will enable to take appropriate measures regarding the level of\nan anomaly instantaneously and efficiently. Video-based human activity\nrecognition has intrigued many researchers with its pressing issues and a\nvariety of applications ranging from simple hand gesture recognition to crucial\nbehavior recognition in a surveillance system. This paper provides a critical\nsurvey of video-based Human Activity Recognition (HAR) techniques beginning\nwith an examination of basic approaches for detecting and recognizing\nsuspicious behavior followed by a critical analysis of machine learning and\ndeep learning techniques such as Convolutional Neural Network (CNN), Recurrent\nNeural Network (RNN), Hidden Markov Model (HMM), K-means Clustering etc. A\ndetailed investigation and comparison are done on these learning techniques on\nthe basis of feature extraction techniques, parameter initialization, and\noptimization algorithms, accuracy, etc. The purpose of this review is to\nprioritize positive schemes and to assist researchers with emerging\nadvancements in this field's future endeavors. This paper also pragmatically\ndiscusses existing challenges in the field of HAR and examines the prospects in\nthe field.\n", "date": "2024-09-01", "categories": []}
{"id": "2409.01448", "title": "FinePseudo: Improving Pseudo-Labelling through Temporal-Alignablity for\n  Semi-Supervised Fine-Grained Action Recognition", "abstract": "  Real-life applications of action recognition often require a fine-grained\nunderstanding of subtle movements, e.g., in sports analytics, user interactions\nin AR/VR, and surgical videos. Although fine-grained actions are more costly to\nannotate, existing semi-supervised action recognition has mainly focused on\ncoarse-grained action recognition. Since fine-grained actions are more\nchallenging due to the absence of scene bias, classifying these actions\nrequires an understanding of action-phases. Hence, existing coarse-grained\nsemi-supervised methods do not work effectively. In this work, we for the first\ntime thoroughly investigate semi-supervised fine-grained action recognition\n(FGAR). We observe that alignment distances like dynamic time warping (DTW)\nprovide a suitable action-phase-aware measure for comparing fine-grained\nactions, a concept previously unexploited in FGAR. However, since regular DTW\ndistance is pairwise and assumes strict alignment between pairs, it is not\ndirectly suitable for classifying fine-grained actions. To utilize such\nalignment distances in a limited-label setting, we propose an\nAlignability-Verification-based Metric learning technique to effectively\ndiscriminate between fine-grained action pairs. Our learnable alignability\nscore provides a better phase-aware measure, which we use to refine the\npseudo-labels of the primary video encoder. Our collaborative\npseudo-labeling-based framework `\\textit{FinePseudo}' significantly outperforms\nprior methods on four fine-grained action recognition datasets: Diving48,\nFineGym99, FineGym288, and FineDiving, and shows improvement on existing\ncoarse-grained datasets: Kinetics400 and Something-SomethingV2. We also\ndemonstrate the robustness of our collaborative pseudo-labeling in handling\nnovel unlabeled classes in open-world semi-supervised setups. Project Page:\nhttps://daveishan.github.io/finepsuedo-webpage/.\n", "date": "2024-09-02", "categories": []}
{"id": "2408.16442", "title": "Integrating Features for Recognizing Human Activities through Optimized\n  Parameters in Graph Convolutional Networks and Transformer Architectures", "abstract": "  Human activity recognition is a major field of study that employs computer\nvision, machine vision, and deep learning techniques to categorize human\nactions. The field of deep learning has made significant progress, with\narchitectures that are extremely effective at capturing human dynamics. This\nstudy emphasizes the influence of feature fusion on the accuracy of activity\nrecognition. This technique addresses the limitation of conventional models,\nwhich face difficulties in identifying activities because of their limited\ncapacity to understand spatial and temporal features. The technique employs\nsensory data obtained from four publicly available datasets: HuGaDB, PKU-MMD,\nLARa, and TUG. The accuracy and F1-score of two deep learning models,\nspecifically a Transformer model and a Parameter-Optimized Graph Convolutional\nNetwork (PO-GCN), were evaluated using these datasets. The feature fusion\ntechnique integrated the final layer features from both models and inputted\nthem into a classifier. Empirical evidence demonstrates that PO-GCN outperforms\nstandard models in activity recognition. HuGaDB demonstrated a 2.3% improvement\nin accuracy and a 2.2% increase in F1-score. TUG showed a 5% increase in\naccuracy and a 0.5% rise in F1-score. On the other hand, LARa and PKU-MMD\nachieved lower accuracies of 64% and 69% respectively. This indicates that the\nintegration of features enhanced the performance of both the Transformer model\nand PO-GCN.\n", "date": "2024-08-29", "categories": []}
{"id": "2408.16638", "title": "3D Pose-Based Temporal Action Segmentation for Figure Skating: A\n  Fine-Grained and Jump Procedure-Aware Annotation Approach", "abstract": "  Understanding human actions from videos is essential in many domains,\nincluding sports. In figure skating, technical judgments are performed by\nwatching skaters' 3D movements, and its part of the judging procedure can be\nregarded as a Temporal Action Segmentation (TAS) task. TAS tasks in figure\nskating that automatically assign temporal semantics to video are actively\nresearched. However, there is a lack of datasets and effective methods for TAS\ntasks requiring 3D pose data. In this study, we first created the FS-Jump3D\ndataset of complex and dynamic figure skating jumps using optical markerless\nmotion capture. We also propose a new fine-grained figure skating jump TAS\ndataset annotation method with which TAS models can learn jump procedures. In\nthe experimental results, we validated the usefulness of 3D pose features as\ninput and the fine-grained dataset for the TAS model in figure skating.\nFS-Jump3D Dataset is available at https://github.com/ryota-skating/FS-Jump3D.\n", "date": "2024-08-29", "categories": []}
{"id": "2305.12437", "title": "SCP: Soft Conditional Prompt Learning for Aerial Video Action\n  Recognition", "abstract": "  We present a new learning approach, Soft Conditional Prompt Learning (SCP),\nwhich leverages the strengths of prompt learning for aerial video action\nrecognition. Our approach is designed to predict the action of each agent by\nhelping the models focus on the descriptions or instructions associated with\nactions in the input videos for aerial/robot visual perception. Our formulation\nsupports various prompts, including learnable prompts, auxiliary visual\ninformation, and large vision models to improve the recognition performance. We\npresent a soft conditional prompt method that learns to dynamically generate\nprompts from a pool of prompt experts under different video inputs. By sharing\nthe same objective with the task, our proposed SCP can optimize prompts that\nguide the model's predictions while explicitly learning input-invariant (prompt\nexperts pool) and input-specific (data-dependent) prompt knowledge. In\npractice, we observe a 3.17-10.2% accuracy improvement on the aerial video\ndatasets (Okutama, NECDrone), which consist of scenes with single-agent and\nmulti-agent actions. We further evaluate our approach on ground camera videos\nto verify the effectiveness and generalization and achieve a 1.0-3.6%\nimprovement on dataset SSV2. We integrate our method into the ROS2 as well.\n", "date": "2023-05-21", "categories": []}
{"id": "2401.12471", "title": "Training-Free Action Recognition and Goal Inference with Dynamic Frame\n  Selection", "abstract": "  We introduce VidTFS, a Training-free, open-vocabulary video goal and action\ninference framework that combines the frozen vision foundational model (VFM)\nand large language model (LLM) with a novel dynamic Frame Selection module. Our\nexperiments demonstrate that the proposed frame selection module improves the\nperformance of the framework significantly. We validate the performance of the\nproposed VidTFS on four widely used video datasets, including CrossTask, COIN,\nUCF101, and ActivityNet, covering goal inference and action recognition tasks\nunder open-vocabulary settings without requiring any training or fine-tuning.\nThe results show that VidTFS outperforms pretrained and instruction-tuned\nmultimodal language models that directly stack LLM and VFM for downstream video\ninference tasks. Our VidTFS with its adaptability shows the future potential\nfor generalizing to new training-free video inference tasks.\n", "date": "2024-01-23", "categories": []}
{"id": "2408.14146", "title": "TSAK: Two-Stage Semantic-Aware Knowledge Distillation for Efficient\n  Wearable Modality and Model Optimization in Manufacturing Lines", "abstract": "  Smaller machine learning models, with less complex architectures and sensor\ninputs, can benefit wearable sensor-based human activity recognition (HAR)\nsystems in many ways, from complexity and cost to battery life. In the specific\ncase of smart factories, optimizing human-robot collaboration hinges on the\nimplementation of cutting-edge, human-centric AI systems. To this end, workers'\nactivity recognition enables accurate quantification of performance metrics,\nimproving efficiency holistically. We present a two-stage semantic-aware\nknowledge distillation (KD) approach, TSAK, for efficient, privacy-aware, and\nwearable HAR in manufacturing lines, which reduces the input sensor modalities\nas well as the machine learning model size, while reaching similar recognition\nperformance as a larger multi-modal and multi-positional teacher model. The\nfirst stage incorporates a teacher classifier model encoding attention, causal,\nand combined representations. The second stage encompasses a semantic\nclassifier merging the three representations from the first stage. To evaluate\nTSAK, we recorded a multi-modal dataset at a smart factory testbed with\nwearable and privacy-aware sensors (IMU and capacitive) located on both\nworkers' hands. In addition, we evaluated our approach on OpenPack, the only\navailable open dataset mimicking the wearable sensor placements on both hands\nin the manufacturing HAR scenario. We compared several KD strategies with\ndifferent representations to regulate the training process of a smaller student\nmodel. Compared to the larger teacher model, the student model takes fewer\nsensor channels from a single hand, has 79% fewer parameters, runs 8.88 times\nfaster, and requires 96.6% less computing power (FLOPS).\n", "date": "2024-08-26", "categories": []}
{"id": "2408.14659", "title": "Comparative Analysis: Violence Recognition from Videos using Transfer\n  Learning", "abstract": "  Action recognition has become a hot topic in computer vision. However, the\nmain applications of computer vision in video processing have focused on\ndetection of relatively simple actions while complex events such as violence\ndetection have been comparatively less investigated. This study focuses on the\nbenchmarking of various deep learning techniques on a complex dataset. Next, a\nlarger dataset is utilized to test the uplift from increasing volume of data.\nThe dataset size increase from 500 to 1,600 videos resulted in a notable\naverage accuracy improvement of 6% across four models.\n", "date": "2024-08-26", "categories": []}
{"id": "2305.15699", "title": "Cross-view Action Recognition Understanding From Exocentric to\n  Egocentric Perspective", "abstract": "  Understanding action recognition in egocentric videos has emerged as a vital\nresearch topic with numerous practical applications. With the limitation in the\nscale of egocentric data collection, learning robust deep learning-based action\nrecognition models remains difficult. Transferring knowledge learned from the\nlarge-scale exocentric data to the egocentric data is challenging due to the\ndifference in videos across views. Our work introduces a novel cross-view\nlearning approach to action recognition (CVAR) that effectively transfers\nknowledge from the exocentric to the selfish view. First, we present a novel\ngeometric-based constraint into the self-attention mechanism in Transformer\nbased on analyzing the camera positions between two views. Then, we propose a\nnew cross-view self-attention loss learned on unpaired cross-view data to\nenforce the self-attention mechanism learning to transfer knowledge across\nviews. Finally, to further improve the performance of our cross-view learning\napproach, we present the metrics to measure the correlations in videos and\nattention maps effectively. Experimental results on standard egocentric action\nrecognition benchmarks, i.e., Charades-Ego, EPIC-Kitchens-55, and\nEPIC-Kitchens-100, have shown our approach's effectiveness and state-of-the-art\nperformance.\n", "date": "2023-05-25", "categories": []}
{"id": "2311.17428", "title": "SigFormer: Sparse Signal-Guided Transformer for Multi-Modal Human Action\n  Segmentation", "abstract": "  Multi-modal human action segmentation is a critical and challenging task with\na wide range of applications. Nowadays, the majority of approaches concentrate\non the fusion of dense signals (i.e., RGB, optical flow, and depth maps).\nHowever, the potential contributions of sparse IoT sensor signals, which can be\ncrucial for achieving accurate recognition, have not been fully explored. To\nmake up for this, we introduce a Sparse signalguided Transformer (SigFormer) to\ncombine both dense and sparse signals. We employ mask attention to fuse\nlocalized features by constraining cross-attention within the regions where\nsparse signals are valid. However, since sparse signals are discrete, they lack\nsufficient information about the temporal action boundaries. Therefore, in\nSigFormer, we propose to emphasize the boundary information at two stages to\nalleviate this problem. In the first feature extraction stage, we introduce an\nintermediate bottleneck module to jointly learn both category and boundary\nfeatures of each dense modality through the inner loss functions. After the\nfusion of dense modalities and sparse signals, we then devise a two-branch\narchitecture that explicitly models the interrelationship between action\ncategory and temporal boundary. Experimental results demonstrate that SigFormer\noutperforms the state-of-the-art approaches on a multi-modal action\nsegmentation dataset from real industrial environments, reaching an outstanding\nF1 score of 0.958. The codes and pre-trained models have been available at\nhttps://github.com/LIUQI-creat/SigFormer.\n", "date": "2023-11-29", "categories": []}
{"id": "2408.13463", "title": "HabitAction: A Video Dataset for Human Habitual Behavior Recognition", "abstract": "  Human Action Recognition (HAR) is a very crucial task in computer vision. It\nhelps to carry out a series of downstream tasks, like understanding human\nbehaviors. Due to the complexity of human behaviors, many highly valuable\nbehaviors are not yet encompassed within the available datasets for HAR, e.g.,\nhuman habitual behaviors (HHBs). HHBs hold significant importance for analyzing\na person's personality, habits, and psychological changes. To solve these\nproblems, in this work, we build a novel video dataset to demonstrate various\nHHBs. These behaviors in the proposed dataset are able to reflect internal\nmental states and specific emotions of the characters, e.g., crossing arms\nsuggests to shield oneself from perceived threats. The dataset contains 30\ncategories of habitual behaviors including more than 300,000 frames and 6,899\naction instances. Since these behaviors usually appear at small local parts of\nhuman action videos, it is difficult for existing action recognition methods to\nhandle these local features. Therefore, we also propose a two-stream model\nusing both human skeletons and RGB appearances. Experimental results\ndemonstrate that our proposed method has much better performance in action\nrecognition than the existing methods on the proposed dataset.\n", "date": "2024-08-24", "categories": []}
{"id": "2408.13765", "title": "FMI-TAL: Few-shot Multiple Instances Temporal Action Localization by\n  Probability Distribution Learning and Interval Cluster Refinement", "abstract": "  The present few-shot temporal action localization model can't handle the\nsituation where videos contain multiple action instances. So the purpose of\nthis paper is to achieve manifold action instances localization in a lengthy\nuntrimmed query video using limited trimmed support videos. To address this\nchallenging problem effectively, we proposed a novel solution involving a\nspatial-channel relation transformer with probability learning and cluster\nrefinement. This method can accurately identify the start and end boundaries of\nactions in the query video, utilizing only a limited number of labeled videos.\nOur proposed method is adept at capturing both temporal and spatial contexts to\neffectively classify and precisely locate actions in videos, enabling a more\ncomprehensive utilization of these crucial details. The selective cosine\npenalization algorithm is designed to suppress temporal boundaries that do not\ninclude action scene switches. The probability learning combined with the label\ngeneration algorithm alleviates the problem of action duration diversity and\nenhances the model's ability to handle fuzzy action boundaries. The interval\ncluster can help us get the final results with multiple instances situations in\nfew-shot temporal action localization. Our model achieves competitive\nperformance through meticulous experimentation utilizing the benchmark datasets\nActivityNet1.3 and THUMOS14. Our code is readily available at\nhttps://github.com/ycwfs/FMI-TAL.\n", "date": "2024-08-25", "categories": []}
{"id": "2408.13777", "title": "Towards Completeness: A Generalizable Action Proposal Generator for\n  Zero-Shot Temporal Action Localization", "abstract": "  To address the zero-shot temporal action localization (ZSTAL) task, existing\nworks develop models that are generalizable to detect and classify actions from\nunseen categories. They typically develop a category-agnostic action detector\nand combine it with the Contrastive Language-Image Pre-training (CLIP) model to\nsolve ZSTAL. However, these methods suffer from incomplete action proposals\ngenerated for \\textit{unseen} categories, since they follow a frame-level\nprediction paradigm and require hand-crafted post-processing to generate action\nproposals. To address this problem, in this work, we propose a novel model\nnamed Generalizable Action Proposal generator (GAP), which can interface\nseamlessly with CLIP and generate action proposals in a holistic way. Our GAP\nis built in a query-based architecture and trained with a proposal-level\nobjective, enabling it to estimate proposal completeness and eliminate the\nhand-crafted post-processing. Based on this architecture, we propose an\nAction-aware Discrimination loss to enhance the category-agnostic dynamic\ninformation of actions. Besides, we introduce a Static-Dynamic Rectifying\nmodule that incorporates the generalizable static information from CLIP to\nrefine the predicted proposals, which improves proposal completeness in a\ngeneralizable manner. Our experiments show that our GAP achieves\nstate-of-the-art performance on two challenging ZSTAL benchmarks, i.e.,\nThumos14 and ActivityNet1.3. Specifically, our model obtains significant\nperformance improvement over previous works on the two benchmarks, i.e., +3.2%\nand +3.4% average mAP, respectively.\n", "date": "2024-08-25", "categories": []}
{"id": "2408.09527", "title": "ALS-HAR: Harnessing Wearable Ambient Light Sensors to Enhance IMU-based\n  Human Activity Recogntion", "abstract": "  Despite the widespread integration of ambient light sensors (ALS) in smart\ndevices commonly used for screen brightness adaptation, their application in\nhuman activity recognition (HAR), primarily through body-worn ALS, is largely\nunexplored. In this work, we developed ALS-HAR, a robust wearable light-based\nmotion activity classifier. Although ALS-HAR achieves comparable accuracy to\nother modalities, its natural sensitivity to external disturbances, such as\nchanges in ambient light, weather conditions, or indoor lighting, makes it\nchallenging for daily use. To address such drawbacks, we introduce strategies\nto enhance environment-invariant IMU-based activity classifications through\naugmented multi-modal and contrastive classifications by transferring the\nknowledge extracted from the ALS. Our experiments on a real-world activity\ndataset for three different scenarios demonstrate that while ALS-HAR's accuracy\nstrongly relies on external lighting conditions, cross-modal information can\nstill improve other HAR systems, such as IMU-based classifiers.Even in\nscenarios where ALS performs insufficiently, the additional knowledge enables\nimproved accuracy and macro F1 score by up to 4.2 % and 6.4 %, respectively,\nfor IMU-based classifiers and even surpasses multi-modal sensor fusion models\nin two of our three experiment scenarios. Our research highlights the untapped\npotential of ALS integration in advancing sensor-based HAR technology, paving\nthe way for practical and efficient wearable ALS-based activity recognition\nsystems with potential applications in healthcare, sports monitoring, and smart\nindoor environments.\n", "date": "2024-08-18", "categories": []}
{"id": "2408.12211", "title": "Computer-Aided Fall Recognition Using a Three-Stream Spatial-Temporal\n  GCN Model with Adaptive Feature Aggregation", "abstract": "  The prevention of falls is paramount in modern healthcare, particularly for\nthe elderly, as falls can lead to severe injuries or even fatalities.\nAdditionally, the growing incidence of falls among the elderly, coupled with\nthe urgent need to prevent suicide attempts resulting from medication overdose,\nunderscores the critical importance of accurate and efficient fall detection\nmethods. In this scenario, a computer-aided fall detection system is inevitable\nto save elderly people's lives worldwide. Many researchers have been working to\ndevelop fall detection systems. However, the existing fall detection systems\noften struggle with issues such as unsatisfactory performance accuracy, limited\nrobustness, high computational complexity, and sensitivity to environmental\nfactors due to a lack of effective features. In response to these challenges,\nthis paper proposes a novel three-stream spatial-temporal feature-based fall\ndetection system. Our system incorporates joint skeleton-based spatial and\ntemporal Graph Convolutional Network (GCN) features, joint motion-based spatial\nand temporal GCN features, and residual connections-based features. Each stream\nemploys adaptive graph-based feature aggregation and consecutive separable\nconvolutional neural networks (Sep-TCN), significantly reducing computational\ncomplexity and model parameters compared to prior systems. Experimental results\nacross multiple datasets demonstrate the superior effectiveness and efficiency\nof our proposed system, with accuracies of 99.51\\%, 99.15\\%, 99.79\\% and 99.85\n\\% achieved on the ImViA, UR-Fall, Fall-UP and FU-Kinect datasets,\nrespectively. The remarkable performance of our system highlights its\nsuperiority, efficiency, and generalizability in real-world fall detection\nscenarios, offering significant advancements in healthcare and societal\nwell-being.\n", "date": "2024-08-22", "categories": []}
{"id": "2408.12475", "title": "Frame Order Matters: A Temporal Sequence-Aware Model for Few-Shot Action\n  Recognition", "abstract": "  In this paper, we propose a novel Temporal Sequence-Aware Model (TSAM) for\nfew-shot action recognition (FSAR), which incorporates a sequential perceiver\nadapter into the pre-training framework, to integrate both the spatial\ninformation and the sequential temporal dynamics into the feature embeddings.\nDifferent from the existing fine-tuning approaches that capture temporal\ninformation by exploring the relationships among all the frames, our\nperceiver-based adapter recurrently captures the sequential dynamics alongside\nthe timeline, which could perceive the order change. To obtain the\ndiscriminative representations for each class, we extend a textual corpus for\neach class derived from the large language models (LLMs) and enrich the visual\nprototypes by integrating the contextual semantic information. Besides, We\nintroduce an unbalanced optimal transport strategy for feature matching that\nmitigates the impact of class-unrelated features, thereby facilitating more\neffective decision-making. Experimental results on five FSAR datasets\ndemonstrate that our method set a new benchmark, beating the second-best\ncompetitors with large margins.\n", "date": "2024-08-22", "categories": []}
{"id": "2209.05044", "title": "Predicting the Next Action by Modeling the Abstract Goal", "abstract": "  The problem of anticipating human actions is an inherently uncertain one.\nHowever, we can reduce this uncertainty if we have a sense of the goal that the\nactor is trying to achieve. Here, we present an action anticipation model that\nleverages goal information for the purpose of reducing the uncertainty in\nfuture predictions. Since we do not possess goal information or the observed\nactions during inference, we resort to visual representation to encapsulate\ninformation about both actions and goals. Through this, we derive a novel\nconcept called abstract goal which is conditioned on observed sequences of\nvisual features for action anticipation. We design the abstract goal as a\ndistribution whose parameters are estimated using a variational recurrent\nnetwork. We sample multiple candidates for the next action and introduce a goal\nconsistency measure to determine the best candidate that follows from the\nabstract goal. Our method obtains impressive results on the very challenging\nEpic-Kitchens55 (EK55), EK100, and EGTEA Gaze+ datasets. We obtain absolute\nimprovements of +13.69, +11.24, and +5.19 for Top-1 verb, Top-1 noun, and Top-1\naction anticipation accuracy respectively over prior state-of-the-art methods\nfor seen kitchens (S1) of EK55. Similarly, we also obtain significant\nimprovements in the unseen kitchens (S2) set for Top-1 verb (+10.75), noun\n(+5.84) and action (+2.87) anticipation. Similar trend is observed for EGTEA\nGaze+ dataset, where absolute improvement of +9.9, +13.1 and +6.8 is obtained\nfor noun, verb, and action anticipation. It is through the submission of this\npaper that our method is currently the new state-of-the-art for action\nanticipation in EK55 and EGTEA Gaze+\nhttps://competitions.codalab.org/competitions/20071#results Code available at\nhttps://github.com/debadityaroy/Abstract_Goal\n", "date": "2022-09-12", "categories": []}
{"id": "2408.11552", "title": "Explainable Deep Learning Framework for Human Activity Recognition", "abstract": "  In the realm of human activity recognition (HAR), the integration of\nexplainable Artificial Intelligence (XAI) emerges as a critical necessity to\nelucidate the decision-making processes of complex models, fostering\ntransparency and trust. Traditional explanatory methods like Class Activation\nMapping (CAM) and attention mechanisms, although effective in highlighting\nregions vital for decisions in various contexts, prove inadequate for HAR. This\ninadequacy stems from the inherently abstract nature of HAR data, rendering\nthese explanations obscure. In contrast, state-of-th-art post-hoc\ninterpretation techniques for time series can explain the model from other\nperspectives. However, this requires extra effort. It usually takes 10 to 20\nseconds to generate an explanation. To overcome these challenges, we proposes a\nnovel, model-agnostic framework that enhances both the interpretability and\nefficacy of HAR models through the strategic use of competitive data\naugmentation. This innovative approach does not rely on any particular model\narchitecture, thereby broadening its applicability across various HAR models.\nBy implementing competitive data augmentation, our framework provides intuitive\nand accessible explanations of model decisions, thereby significantly advancing\nthe interpretability of HAR systems without compromising on performance.\n", "date": "2024-08-21", "categories": []}
{"id": "2408.01766", "title": "MultiFuser: Multimodal Fusion Transformer for Enhanced Driver Action\n  Recognition", "abstract": "  Driver action recognition, aiming to accurately identify drivers' behaviours,\nis crucial for enhancing driver-vehicle interactions and ensuring driving\nsafety. Unlike general action recognition, drivers' environments are often\nchallenging, being gloomy and dark, and with the development of sensors,\nvarious cameras such as IR and depth cameras have emerged for analyzing\ndrivers' behaviors. Therefore, in this paper, we propose a novel multimodal\nfusion transformer, named MultiFuser, which identifies cross-modal\ninterrelations and interactions among multimodal car cabin videos and\nadaptively integrates different modalities for improved representations.\nSpecifically, MultiFuser comprises layers of Bi-decomposed Modules to model\nspatiotemporal features, with a modality synthesizer for multimodal features\nintegration. Each Bi-decomposed Module includes a Modal Expertise ViT block for\nextracting modality-specific features and a Patch-wise Adaptive Fusion block\nfor efficient cross-modal fusion. Extensive experiments are conducted on\nDrive&Act dataset and the results demonstrate the efficacy of our proposed\napproach.\n", "date": "2024-08-03", "categories": []}
{"id": "2408.09220", "title": "Flatten: Video Action Recognition is an Image Classification task", "abstract": "  In recent years, video action recognition, as a fundamental task in the field\nof video understanding, has been deeply explored by numerous researchers.Most\ntraditional video action recognition methods typically involve converting\nvideos into three-dimensional data that encapsulates both spatial and temporal\ninformation, subsequently leveraging prevalent image understanding models to\nmodel and analyze these data. However,these methods have significant drawbacks.\nFirstly, when delving into video action recognition tasks, image understanding\nmodels often need to be adapted accordingly in terms of model architecture and\npreprocessing for these spatiotemporal tasks; Secondly, dealing with\nhigh-dimensional data often poses greater challenges and incurs higher time\ncosts compared to its lower-dimensional counterparts.To bridge the gap between\nimage-understanding and video-understanding tasks while simplifying the\ncomplexity of video comprehension, we introduce a novel video representation\narchitecture, Flatten, which serves as a plug-and-play module that can be\nseamlessly integrated into any image-understanding network for efficient and\neffective 3D temporal data modeling.Specifically, by applying specific\nflattening operations (e.g., row-major transform), 3D spatiotemporal data is\ntransformed into 2D spatial information, and then ordinary image understanding\nmodels are used to capture temporal dynamic and spatial semantic information,\nwhich in turn accomplishes effective and efficient video action recognition.\nExtensive experiments on commonly used datasets (Kinetics-400,\nSomething-Something v2, and HMDB-51) and three classical image classification\nmodels (Uniformer, SwinV2, and ResNet), have demonstrated that embedding\nFlatten provides a significant performance improvements over original model.\n", "date": "2024-08-17", "categories": []}
{"id": "2408.09354", "title": "Boundary-Recovering Network for Temporal Action Detection", "abstract": "  Temporal action detection (TAD) is challenging, yet fundamental for\nreal-world video applications. Large temporal scale variation of actions is one\nof the most primary difficulties in TAD. Naturally, multi-scale features have\npotential in localizing actions of diverse lengths as widely used in object\ndetection. Nevertheless, unlike objects in images, actions have more ambiguity\nin their boundaries. That is, small neighboring objects are not considered as a\nlarge one while short adjoining actions can be misunderstood as a long one. In\nthe coarse-to-fine feature pyramid via pooling, these vague action boundaries\ncan fade out, which we call 'vanishing boundary problem'. To this end, we\npropose Boundary-Recovering Network (BRN) to address the vanishing boundary\nproblem. BRN constructs scale-time features by introducing a new axis called\nscale dimension by interpolating multi-scale features to the same temporal\nlength. On top of scale-time features, scale-time blocks learn to exchange\nfeatures across scale levels, which can effectively settle down the issue. Our\nextensive experiments demonstrate that our model outperforms the\nstate-of-the-art on the two challenging benchmarks, ActivityNet-v1.3 and\nTHUMOS14, with remarkably reduced degree of the vanishing boundary problem.\n", "date": "2024-08-18", "categories": []}
{"id": "2408.09356", "title": "Joint Temporal Pooling for Improving Skeleton-based Action Recognition", "abstract": "  In skeleton-based human action recognition, temporal pooling is a critical\nstep for capturing spatiotemporal relationship of joint dynamics. Conventional\npooling methods overlook the preservation of motion information and treat each\nframe equally. However, in an action sequence, only a few segments of frames\ncarry discriminative information related to the action. This paper presents a\nnovel Joint Motion Adaptive Temporal Pooling (JMAP) method for improving\nskeleton-based action recognition. Two variants of JMAP, frame-wise pooling and\njoint-wise pooling, are introduced. The efficacy of JMAP has been validated\nthrough experiments on the popular NTU RGB+D 120 and PKU-MMD datasets.\n", "date": "2024-08-18", "categories": []}
{"id": "2408.09764", "title": "Event Stream based Human Action Recognition: A High-Definition Benchmark\n  Dataset and Algorithms", "abstract": "  Human Action Recognition (HAR) stands as a pivotal research domain in both\ncomputer vision and artificial intelligence, with RGB cameras dominating as the\npreferred tool for investigation and innovation in this field. However, in\nreal-world applications, RGB cameras encounter numerous challenges, including\nlight conditions, fast motion, and privacy concerns. Consequently, bio-inspired\nevent cameras have garnered increasing attention due to their advantages of low\nenergy consumption, high dynamic range, etc. Nevertheless, most existing\nevent-based HAR datasets are low resolution ($346 \\times 260$). In this paper,\nwe propose a large-scale, high-definition ($1280 \\times 800$) human action\nrecognition dataset based on the CeleX-V event camera, termed CeleX-HAR. It\nencompasses 150 commonly occurring action categories, comprising a total of\n124,625 video sequences. Various factors such as multi-view, illumination,\naction speed, and occlusion are considered when recording these data. To build\na more comprehensive benchmark dataset, we report over 20 mainstream HAR models\nfor future works to compare. In addition, we also propose a novel Mamba vision\nbackbone network for event stream based HAR, termed EVMamba, which equips the\nspatial plane multi-directional scanning and novel voxel temporal scanning\nmechanism. By encoding and mining the spatio-temporal information of event\nstreams, our EVMamba has achieved favorable results across multiple datasets.\nBoth the dataset and source code will be released on\n\\url{https://github.com/Event-AHU/CeleX-HAR}\n", "date": "2024-08-19", "categories": []}
{"id": "2408.09919", "title": "Long-Tail Temporal Action Segmentation with Group-wise Temporal Logit\n  Adjustment", "abstract": "  Procedural activity videos often exhibit a long-tailed action distribution\ndue to varying action frequencies and durations. However, state-of-the-art\ntemporal action segmentation methods overlook the long tail and fail to\nrecognize tail actions. Existing long-tail methods make class-independent\nassumptions and struggle to identify tail classes when applied to temporal\nsegmentation frameworks. This work proposes a novel group-wise temporal logit\nadjustment~(G-TLA) framework that combines a group-wise softmax formulation\nwhile leveraging activity information and action ordering for logit adjustment.\nThe proposed framework significantly improves in segmenting tail actions\nwithout any performance loss on head actions.\n", "date": "2024-08-19", "categories": []}
{"id": "2407.06162", "title": "RNNs, CNNs and Transformers in Human Action Recognition: A Survey and a\n  Hybrid Model", "abstract": "  Human Action Recognition (HAR) encompasses the task of monitoring human\nactivities across various domains, including but not limited to medical,\neducational, entertainment, visual surveillance, video retrieval, and the\nidentification of anomalous activities. Over the past decade, the field of HAR\nhas witnessed substantial progress by leveraging Convolutional Neural Networks\n(CNNs) to effectively extract and comprehend intricate information, thereby\nenhancing the overall performance of HAR systems. Recently, the domain of\ncomputer vision has witnessed the emergence of Vision Transformers (ViTs) as a\npotent solution. The efficacy of transformer architecture has been validated\nbeyond the confines of image analysis, extending their applicability to diverse\nvideo-related tasks. Notably, within this landscape, the research community has\nshown keen interest in HAR, acknowledging its manifold utility and widespread\nadoption across various domains. This article aims to present an encompassing\nsurvey that focuses on CNNs and the evolution of Recurrent Neural Networks\n(RNNs) to ViTs given their importance in the domain of HAR. By conducting a\nthorough examination of existing literature and exploring emerging trends, this\nstudy undertakes a critical analysis and synthesis of the accumulated knowledge\nin this field. Additionally, it investigates the ongoing efforts to develop\nhybrid approaches. Following this direction, this article presents a novel\nhybrid model that seeks to integrate the inherent strengths of CNNs and ViTs.\n", "date": "2024-06-02", "categories": []}
{"id": "2407.15706", "title": "Multi-Modality Co-Learning for Efficient Skeleton-based Action\n  Recognition", "abstract": "  Skeleton-based action recognition has garnered significant attention due to\nthe utilization of concise and resilient skeletons. Nevertheless, the absence\nof detailed body information in skeletons restricts performance, while other\nmultimodal methods require substantial inference resources and are inefficient\nwhen using multimodal data during both training and inference stages. To\naddress this and fully harness the complementary multimodal features, we\npropose a novel multi-modality co-learning (MMCL) framework by leveraging the\nmultimodal large language models (LLMs) as auxiliary networks for efficient\nskeleton-based action recognition, which engages in multi-modality co-learning\nduring the training stage and keeps efficiency by employing only concise\nskeletons in inference. Our MMCL framework primarily consists of two modules.\nFirst, the Feature Alignment Module (FAM) extracts rich RGB features from video\nframes and aligns them with global skeleton features via contrastive learning.\nSecond, the Feature Refinement Module (FRM) uses RGB images with temporal\ninformation and text instruction to generate instructive features based on the\npowerful generalization of multimodal LLMs. These instructive text features\nwill further refine the classification scores and the refined scores will\nenhance the model's robustness and generalization in a manner similar to soft\nlabels. Extensive experiments on NTU RGB+D, NTU RGB+D 120 and Northwestern-UCLA\nbenchmarks consistently verify the effectiveness of our MMCL, which outperforms\nthe existing skeleton-based action recognition methods. Meanwhile, experiments\non UTD-MHAD and SYSU-Action datasets demonstrate the commendable generalization\nof our MMCL in zero-shot and domain-adaptive action recognition. Our code is\npublicly available at: https://github.com/liujf69/MMCL-Action.\n", "date": "2024-07-22", "categories": []}
{"id": "2403.09805", "title": "On the Utility of 3D Hand Poses for Action Recognition", "abstract": "  3D hand pose is an underexplored modality for action recognition. Poses are\ncompact yet informative and can greatly benefit applications with limited\ncompute budgets. However, poses alone offer an incomplete understanding of\nactions, as they cannot fully capture objects and environments with which\nhumans interact. We propose HandFormer, a novel multimodal transformer, to\nefficiently model hand-object interactions. HandFormer combines 3D hand poses\nat a high temporal resolution for fine-grained motion modeling with sparsely\nsampled RGB frames for encoding scene semantics. Observing the unique\ncharacteristics of hand poses, we temporally factorize hand modeling and\nrepresent each joint by its short-term trajectories. This factorized pose\nrepresentation combined with sparse RGB samples is remarkably efficient and\nhighly accurate. Unimodal HandFormer with only hand poses outperforms existing\nskeleton-based methods at 5x fewer FLOPs. With RGB, we achieve new\nstate-of-the-art performance on Assembly101 and H2O with significant\nimprovements in egocentric action recognition.\n", "date": "2024-03-14", "categories": []}
{"id": "2408.07282", "title": "Consistency Based Weakly Self-Supervised Learning for Human Activity\n  Recognition with Wearables", "abstract": "  While the widely available embedded sensors in smartphones and other wearable\ndevices make it easier to obtain data of human activities, recognizing\ndifferent types of human activities from sensor-based data remains a difficult\nresearch topic in ubiquitous computing. One reason for this is that most of the\ncollected data is unlabeled. However, many current human activity recognition\n(HAR) systems are based on supervised methods, which heavily rely on the labels\nof the data. We describe a weakly self-supervised approach in this paper that\nconsists of two stages: (1) In stage one, the model learns from the nature of\nhuman activities by projecting the data into an embedding space where similar\nactivities are grouped together; (2) In stage two, the model is fine-tuned\nusing similarity information in a few-shot learning fashion using the\nsimilarity information of the data. This allows downstream classification or\nclustering tasks to benefit from the embeddings. Experiments on three benchmark\ndatasets demonstrate the framework's effectiveness and show that our approach\ncan help the clustering algorithm achieve comparable performance in identifying\nand categorizing the underlying human activities as pure supervised techniques\napplied directly to a corresponding fully labeled data set.\n", "date": "2024-07-29", "categories": []}
{"id": "2404.16609", "title": "SFMViT: SlowFast Meet ViT in Chaotic World", "abstract": "  The task of spatiotemporal action localization in chaotic scenes is a\nchallenging task toward advanced video understanding. Paving the way with\nhigh-quality video feature extraction and enhancing the precision of\ndetector-predicted anchors can effectively improve model performance. To this\nend, we propose a high-performance dual-stream spatiotemporal feature\nextraction network SFMViT with an anchor pruning strategy. The backbone of our\nSFMViT is composed of ViT and SlowFast with prior knowledge of spatiotemporal\naction localization, which fully utilizes ViT's excellent global feature\nextraction capabilities and SlowFast's spatiotemporal sequence modeling\ncapabilities. Secondly, we introduce the confidence maximum heap to prune the\nanchors detected in each frame of the picture to filter out the effective\nanchors. These designs enable our SFMViT to achieve a mAP of 26.62% in the\nChaotic World dataset, far exceeding existing models. Code is available at\nhttps://github.com/jfightyr/SlowFast-Meet-ViT.\n", "date": "2024-04-25", "categories": []}
{"id": "2408.06437", "title": "HAT: History-Augmented Anchor Transformer for Online Temporal Action\n  Localization", "abstract": "  Online video understanding often relies on individual frames, leading to\nframe-by-frame predictions. Recent advancements such as Online Temporal Action\nLocalization (OnTAL), extend this approach to instance-level predictions.\nHowever, existing methods mainly focus on short-term context, neglecting\nhistorical information. To address this, we introduce the History-Augmented\nAnchor Transformer (HAT) Framework for OnTAL. By integrating historical\ncontext, our framework enhances the synergy between long-term and short-term\ninformation, improving the quality of anchor features crucial for\nclassification and localization. We evaluate our model on both procedural\negocentric (PREGO) datasets (EGTEA and EPIC) and standard non-PREGO OnTAL\ndatasets (THUMOS and MUSES). Results show that our model outperforms\nstate-of-the-art approaches significantly on PREGO datasets and achieves\ncomparable or slightly superior performance on non-PREGO datasets, underscoring\nthe importance of leveraging long-term history, especially in procedural and\negocentric action scenarios. Code is available at:\nhttps://github.com/sakibreza/ECCV24-HAT/\n", "date": "2024-08-12", "categories": []}
{"id": "2207.08816", "title": "Discovering Behavioral Predispositions in Data to Improve Human Activity\n  Recognition", "abstract": "  The automatic, sensor-based assessment of challenging behavior of persons\nwith dementia is an important task to support the selection of interventions.\nHowever, predicting behaviors like apathy and agitation is challenging due to\nthe large inter- and intra-patient variability. Goal of this paper is to\nimprove the recognition performance by making use of the observation that\npatients tend to show specific behaviors at certain times of the day or week.\nWe propose to identify such segments of similar behavior via clustering the\ndistributions of annotations of the time segments. All time segments within a\ncluster then consist of similar behaviors and thus indicate a behavioral\npredisposition (BPD). We utilize BPDs by training a classifier for each BPD.\nEmpirically, we demonstrate that when the BPD per time segment is known,\nactivity recognition performance can be substantially improved.\n", "date": "2022-07-18", "categories": []}
{"id": "2306.10858", "title": "Multi-Granularity Hand Action Detection", "abstract": "  Detecting hand actions in videos is crucial for understanding video content\nand has diverse real-world applications. Existing approaches often focus on\nwhole-body actions or coarse-grained action categories, lacking fine-grained\nhand-action localization information. To fill this gap, we introduce the\nFHA-Kitchens (Fine-Grained Hand Actions in Kitchen Scenes) dataset, providing\nboth coarse- and fine-grained hand action categories along with localization\nannotations. This dataset comprises 2,377 video clips and 30,047 frames,\nannotated with approximately 200k bounding boxes and 880 action categories.\nEvaluation of existing action detection methods on FHA-Kitchens reveals varying\ngeneralization capabilities across different granularities. To handle\nmulti-granularity in hand actions, we propose MG-HAD, an End-to-End\nMulti-Granularity Hand Action Detection method. It incorporates two new\ndesigns: Multi-dimensional Action Queries and Coarse-Fine Contrastive\nDenoising. Extensive experiments demonstrate MG-HAD's effectiveness for\nmulti-granularity hand action detection, highlighting the significance of\nFHA-Kitchens for future research and real-world applications. The dataset and\nsource code are available at https://github.com/superZ678/MG-HAD.\n", "date": "2023-06-19", "categories": []}
{"id": "2408.05955", "title": "Probabilistic Vision-Language Representation for Weakly Supervised\n  Temporal Action Localization", "abstract": "  Weakly supervised temporal action localization (WTAL) aims to detect action\ninstances in untrimmed videos using only video-level annotations. Since many\nexisting works optimize WTAL models based on action classification labels, they\nencounter the task discrepancy problem (i.e., localization-by-classification).\nTo tackle this issue, recent studies have attempted to utilize action category\nnames as auxiliary semantic knowledge through vision-language pre-training\n(VLP). However, there are still areas where existing research falls short.\nPrevious approaches primarily focused on leveraging textual information from\nlanguage models but overlooked the alignment of dynamic human action and VLP\nknowledge in a joint space. Furthermore, the deterministic representation\nemployed in previous studies struggles to capture fine-grained human motions.\nTo address these problems, we propose a novel framework that aligns human\naction knowledge and VLP knowledge in a probabilistic embedding space.\nMoreover, we propose intra- and inter-distribution contrastive learning to\nenhance the probabilistic embedding space based on statistical similarities.\nExtensive experiments and ablation studies reveal that our method significantly\noutperforms all previous state-of-the-art methods. Code is available at\nhttps://github.com/sejong-rcv/PVLR.\n", "date": "2024-08-12", "categories": []}
{"id": "2408.06158", "title": "OmniCLIP: Adapting CLIP for Video Recognition with Spatial-Temporal\n  Omni-Scale Feature Learning", "abstract": "  Recent Vision-Language Models (VLMs) \\textit{e.g.} CLIP have made great\nprogress in video recognition. Despite the improvement brought by the strong\nvisual backbone in extracting spatial features, CLIP still falls short in\ncapturing and integrating spatial-temporal features which is essential for\nvideo recognition. In this paper, we propose OmniCLIP, a framework that adapts\nCLIP for video recognition by focusing on learning comprehensive features\nencompassing spatial, temporal, and dynamic spatial-temporal scales, which we\nrefer to as omni-scale features. This is achieved through the design of\nspatial-temporal blocks that include parallel temporal adapters (PTA), enabling\nefficient temporal modeling. Additionally, we introduce a self-prompt generator\n(SPG) module to capture dynamic object spatial features. The synergy between\nPTA and SPG allows OmniCLIP to discern varying spatial information across\nframes and assess object scales over time. We have conducted extensive\nexperiments in supervised video recognition, few-shot video recognition, and\nzero-shot recognition tasks. The results demonstrate the effectiveness of our\nmethod, especially with OmniCLIP achieving a top-1 accuracy of 74.30\\% on\nHMDB51 in a 16-shot setting, surpassing the recent MotionPrompt approach even\nwith full training data. The code is available at\n\\url{https://github.com/XiaoBuL/OmniCLIP}.\n", "date": "2024-08-12", "categories": []}
{"id": "2408.02623", "title": "YOWOv3: An Efficient and Generalized Framework for Human Action\n  Detection and Recognition", "abstract": "  In this paper, we propose a new framework called YOWOv3, which is an improved\nversion of YOWOv2, designed specifically for the task of Human Action Detection\nand Recognition. This framework is designed to facilitate extensive\nexperimentation with different configurations and supports easy customization\nof various components within the model, reducing efforts required for\nunderstanding and modifying the code. YOWOv3 demonstrates its superior\nperformance compared to YOWOv2 on two widely used datasets for Human Action\nDetection and Recognition: UCF101-24 and AVAv2.2. Specifically, the predecessor\nmodel YOWOv2 achieves an mAP of 85.2% and 20.3% on UCF101-24 and AVAv2.2,\nrespectively, with 109.7M parameters and 53.6 GFLOPs. In contrast, our model -\nYOWOv3, with only 59.8M parameters and 39.8 GFLOPs, achieves an mAP of 88.33%\nand 20.31% on UCF101-24 and AVAv2.2, respectively. The results demonstrate that\nYOWOv3 significantly reduces the number of parameters and GFLOPs while still\nachieving comparable performance.\n", "date": "2024-08-05", "categories": []}
{"id": "2408.05169", "title": "Weak-Annotation of HAR Datasets using Vision Foundation Models", "abstract": "  As wearable-based data annotation remains, to date, a tedious, time-consuming\ntask requiring researchers to dedicate substantial time, benchmark datasets\nwithin the field of Human Activity Recognition in lack richness and size\ncompared to datasets available within related fields. Recently, vision\nfoundation models such as CLIP have gained significant attention, helping the\nvision community advance in finding robust, generalizable feature\nrepresentations. With the majority of researchers within the wearable community\nrelying on vision modalities to overcome the limited expressiveness of wearable\ndata and accurately label their to-be-released benchmark datasets offline, we\npropose a novel, clustering-based annotation pipeline to significantly reduce\nthe amount of data that needs to be annotated by a human annotator. We show\nthat using our approach, the annotation of centroid clips suffices to achieve\naverage labelling accuracies close to 90% across three publicly available HAR\nbenchmark datasets. Using the weakly annotated datasets, we further demonstrate\nthat we can match the accuracy scores of fully-supervised deep learning\nclassifiers across all three benchmark datasets. Code as well as supplementary\nfigures and results are publicly downloadable via\ngithub.com/mariusbock/weak_har.\n", "date": "2024-08-09", "categories": []}
{"id": "2303.00952", "title": "Towards Activated Muscle Group Estimation in the Wild", "abstract": "  In this paper, we tackle the new task of video-based Activated Muscle Group\nEstimation (AMGE) aiming at identifying active muscle regions during physical\nactivity in the wild. To this intent, we provide the MuscleMap dataset\nfeaturing >15K video clips with 135 different activities and 20 labeled muscle\ngroups. This dataset opens the vistas to multiple video-based applications in\nsports and rehabilitation medicine under flexible environment constraints. The\nproposed MuscleMap dataset is constructed with YouTube videos, specifically\ntargeting High-Intensity Interval Training (HIIT) physical exercise in the\nwild. To make the AMGE model applicable in real-life situations, it is crucial\nto ensure that the model can generalize well to numerous types of physical\nactivities not present during training and involving new combinations of\nactivated muscles. To achieve this, our benchmark also covers an evaluation\nsetting where the model is exposed to activity types excluded from the training\nset. Our experiments reveal that the generalizability of existing architectures\nadapted for the AMGE task remains a challenge. Therefore, we also propose a new\napproach, TransM3E, which employs a multi-modality feature fusion mechanism\nbetween both the video transformer model and the skeleton-based graph\nconvolution model with novel cross-modal knowledge distillation executed on\nmulti-classification tokens. The proposed method surpasses all popular video\nclassification models when dealing with both, previously seen and new types of\nphysical activities. The database and code can be found at\nhttps://github.com/KPeng9510/MuscleMap.\n", "date": "2023-03-02", "categories": []}
{"id": "2403.09975", "title": "Skeleton-Based Human Action Recognition with Noisy Labels", "abstract": "  Understanding human actions from body poses is critical for assistive robots\nsharing space with humans in order to make informed and safe decisions about\nthe next interaction. However, precise temporal localization and annotation of\nactivity sequences is time-consuming and the resulting labels are often noisy.\nIf not effectively addressed, label noise negatively affects the model's\ntraining, resulting in lower recognition quality. Despite its importance,\naddressing label noise for skeleton-based action recognition has been\noverlooked so far. In this study, we bridge this gap by implementing a\nframework that augments well-established skeleton-based human action\nrecognition methods with label-denoising strategies from various research areas\nto serve as the initial benchmark. Observations reveal that these baselines\nyield only marginal performance when dealing with sparse skeleton data.\nConsequently, we introduce a novel methodology, NoiseEraSAR, which integrates\nglobal sample selection, co-teaching, and Cross-Modal Mixture-of-Experts\n(CM-MOE) strategies, aimed at mitigating the adverse impacts of label noise.\nOur proposed approach demonstrates better performance on the established\nbenchmark, setting new state-of-the-art standards. The source code for this\nstudy is accessible at https://github.com/xuyizdby/NoiseEraSAR.\n", "date": "2024-03-15", "categories": []}
{"id": "2407.03291", "title": "VCHAR:Variance-Driven Complex Human Activity Recognition framework with\n  Generative Representation", "abstract": "  Complex human activity recognition (CHAR) remains a pivotal challenge within\nubiquitous computing, especially in the context of smart environments. Existing\nstudies typically require meticulous labeling of both atomic and complex\nactivities, a task that is labor-intensive and prone to errors due to the\nscarcity and inaccuracies of available datasets. Most prior research has\nfocused on datasets that either precisely label atomic activities or, at\nminimum, their sequence approaches that are often impractical in real world\nsettings.In response, we introduce VCHAR (Variance-Driven Complex Human\nActivity Recognition), a novel framework that treats the outputs of atomic\nactivities as a distribution over specified intervals. Leveraging generative\nmethodologies, VCHAR elucidates the reasoning behind complex activity\nclassifications through video-based explanations, accessible to users without\nprior machine learning expertise. Our evaluation across three publicly\navailable datasets demonstrates that VCHAR enhances the accuracy of complex\nactivity recognition without necessitating precise temporal or sequential\nlabeling of atomic activities. Furthermore, user studies confirm that VCHAR's\nexplanations are more intelligible compared to existing methods, facilitating a\nbroader understanding of complex activity recognition among non-experts.\n", "date": "2024-07-03", "categories": []}
{"id": "2408.02957", "title": "Online Temporal Action Localization with Memory-Augmented Transformer", "abstract": "  Online temporal action localization (On-TAL) is the task of identifying\nmultiple action instances given a streaming video. Since existing methods take\nas input only a video segment of fixed size per iteration, they are limited in\nconsidering long-term context and require tuning the segment size carefully. To\novercome these limitations, we propose memory-augmented transformer (MATR).\nMATR utilizes the memory queue that selectively preserves the past segment\nfeatures, allowing to leverage long-term context for inference. We also propose\na novel action localization method that observes the current input segment to\npredict the end time of the ongoing action and accesses the memory queue to\nestimate the start time of the action. Our method outperformed existing methods\non two datasets, THUMOS14 and MUSES, surpassing not only TAL methods in the\nonline setting but also some offline TAL methods.\n", "date": "2024-08-06", "categories": []}
{"id": "2408.03097", "title": "Prototype Learning for Micro-gesture Classification", "abstract": "  In this paper, we briefly introduce the solution developed by our team,\nHFUT-VUT, for the track of Micro-gesture Classification in the MiGA challenge\nat IJCAI 2024. The task of micro-gesture classification task involves\nrecognizing the category of a given video clip, which focuses on more\nfine-grained and subtle body movements compared to typical action recognition\ntasks. Given the inherent complexity of micro-gesture recognition, which\nincludes large intra-class variability and minimal inter-class differences, we\nutilize two innovative modules, i.e., the cross-modal fusion module and\nprototypical refinement module, to improve the discriminative ability of MG\nfeatures, thereby improving the classification accuracy. Our solution achieved\nsignificant success, ranking 1st in the track of Micro-gesture Classification.\nWe surpassed the performance of last year's leading team by a substantial\nmargin, improving Top-1 accuracy by 6.13%.\n", "date": "2024-08-06", "categories": []}
{"id": "2405.02961", "title": "JOSENet: A Joint Stream Embedding Network for Violence Detection in\n  Surveillance Videos", "abstract": "  The increasing proliferation of video surveillance cameras and the escalating\ndemand for crime prevention have intensified interest in the task of violence\ndetection within the research community. Compared to other action recognition\ntasks, violence detection in surveillance videos presents additional issues,\nsuch as the wide variety of real fight scenes. Unfortunately, existing datasets\nfor violence detection are relatively small in comparison to those for other\naction recognition tasks. Moreover, surveillance footage often features\ndifferent individuals in each video and varying backgrounds for each camera. In\naddition, fast detection of violent actions in real-life surveillance videos is\ncrucial to prevent adverse outcomes, thus necessitating models that are\noptimized for reduced memory usage and computational costs. These challenges\ncomplicate the application of traditional action recognition methods. To tackle\nall these issues, we introduce JOSENet, a novel self-supervised framework that\nprovides outstanding performance for violence detection in surveillance videos.\nThe proposed model processes two spatiotemporal video streams, namely RGB\nframes and optical flows, and incorporates a new regularized self-supervised\nlearning approach for videos. JOSENet demonstrates improved performance\ncompared to state-of-the-art methods, while utilizing only one-fourth of the\nframes per video segment and operating at a reduced frame rate. The source code\nis available at https://github.com/ispamm/JOSENet.\n", "date": "2024-05-05", "categories": []}
{"id": "2407.18289", "title": "MARINE: A Computer Vision Model for Detecting Rare Predator-Prey\n  Interactions in Animal Videos", "abstract": "  Encounters between predator and prey play an essential role in ecosystems,\nbut their rarity makes them difficult to detect in video recordings. Although\nadvances in action recognition (AR) and temporal action detection (AD),\nespecially transformer-based models and vision foundation models, have achieved\nhigh performance on human action datasets, animal videos remain relatively\nunder-researched. This thesis addresses this gap by proposing the model MARINE,\nwhich utilizes motion-based frame selection designed for fast animal actions\nand DINOv2 feature extraction with a trainable classification head for action\nrecognition. MARINE outperforms VideoMAE in identifying predator attacks in\nvideos of fish, both on a small and specific coral reef dataset (81.53\\%\nagainst 52.64\\% accuracy), and on a subset of the more extensive Animal Kingdom\ndataset (94.86\\% against 83.14\\% accuracy). In a multi-label setting on a\nrepresentative sample of Animal Kingdom, MARINE achieves 23.79\\% mAP,\npositioning it mid-field among existing benchmarks. Furthermore, in an AD task\non the coral reef dataset, MARINE achieves 80.78\\% AP (against VideoMAE's\n34.89\\%) although at a lowered t-IoU threshold of 25\\%. Therefore, despite room\nfor improvement, MARINE offers an effective starter framework to apply to AR\nand AD tasks on animal recordings and thus contribute to the study of natural\necosystems.\n", "date": "2024-07-25", "categories": []}
{"id": "2408.02024", "title": "Faster Diffusion Action Segmentation", "abstract": "  Temporal Action Segmentation (TAS) is an essential task in video analysis,\naiming to segment and classify continuous frames into distinct action segments.\nHowever, the ambiguous boundaries between actions pose a significant challenge\nfor high-precision segmentation. Recent advances in diffusion models have\ndemonstrated substantial success in TAS tasks due to their stable training\nprocess and high-quality generation capabilities. However, the heavy sampling\nsteps required by diffusion models pose a substantial computational burden,\nlimiting their practicality in real-time applications. Additionally, most\nrelated works utilize Transformer-based encoder architectures. Although these\narchitectures excel at capturing long-range dependencies, they incur high\ncomputational costs and face feature-smoothing issues when processing long\nvideo sequences. To address these challenges, we propose EffiDiffAct, an\nefficient and high-performance TAS algorithm. Specifically, we develop a\nlightweight temporal feature encoder that reduces computational overhead and\nmitigates the rank collapse phenomenon associated with traditional\nself-attention mechanisms. Furthermore, we introduce an adaptive skip strategy\nthat allows for dynamic adjustment of timestep lengths based on computed\nsimilarity metrics during inference, thereby further enhancing computational\nefficiency. Comprehensive experiments on the 50Salads, Breakfast, and GTEA\ndatasets demonstrated the effectiveness of the proposed algorithm.\n", "date": "2024-08-04", "categories": []}
{"id": "2408.02033", "title": "Enhancing Human Action Recognition and Violence Detection Through Deep\n  Learning Audiovisual Fusion", "abstract": "  This paper proposes a hybrid fusion-based deep learning approach based on two\ndifferent modalities, audio and video, to improve human activity recognition\nand violence detection in public places. To take advantage of audiovisual\nfusion, late fusion, intermediate fusion, and hybrid fusion-based deep learning\n(HFBDL) are used and compared. Since the objective is to detect and recognize\nhuman violence in public places, Real-life violence situation (RLVS) dataset is\nexpanded and used. Simulating results of HFBDL show 96.67\\% accuracy on\nvalidation data, which is more accurate than the other state-of-the-art methods\non this dataset. To showcase our model's ability in real-world scenarios,\nanother dataset of 54 sounded videos of both violent and non-violent situations\nwas recorded. The model could successfully detect 52 out of 54 videos\ncorrectly. The proposed method shows a promising performance on real scenarios.\nThus, it can be used for human action recognition and violence detection in\npublic places for security purposes.\n", "date": "2024-08-04", "categories": []}
{"id": "2408.00249", "title": "Task-Adapter: Task-specific Adaptation of Image Models for Few-shot\n  Action Recognition", "abstract": "  Existing works in few-shot action recognition mostly fine-tune a pre-trained\nimage model and design sophisticated temporal alignment modules at feature\nlevel. However, simply fully fine-tuning the pre-trained model could cause\noverfitting due to the scarcity of video samples. Additionally, we argue that\nthe exploration of task-specific information is insufficient when relying\nsolely on well extracted abstract features. In this work, we propose a simple\nbut effective task-specific adaptation method (Task-Adapter) for few-shot\naction recognition. By introducing the proposed Task-Adapter into the last\nseveral layers of the backbone and keeping the parameters of the original\npre-trained model frozen, we mitigate the overfitting problem caused by full\nfine-tuning and advance the task-specific mechanism into the process of feature\nextraction. In each Task-Adapter, we reuse the frozen self-attention layer to\nperform task-specific self-attention across different videos within the given\ntask to capture both distinctive information among classes and shared\ninformation within classes, which facilitates task-specific adaptation and\nenhances subsequent metric measurement between the query feature and support\nprototypes. Experimental results consistently demonstrate the effectiveness of\nour proposed Task-Adapter on four standard few-shot action recognition\ndatasets. Especially on temporal challenging SSv2 dataset, our method\noutperforms the state-of-the-art methods by a large margin.\n", "date": "2024-08-01", "categories": []}
{"id": "2408.00420", "title": "MPT-PAR:Mix-Parameters Transformer for Panoramic Activity Recognition", "abstract": "  The objective of the panoramic activity recognition task is to identify\nbehaviors at various granularities within crowded and complex environments,\nencompassing individual actions, social group activities, and global\nactivities. Existing methods generally use either parameter-independent modules\nto capture task-specific features or parameter-sharing modules to obtain common\nfeatures across all tasks. However, there is often a strong interrelatedness\nand complementary effect between tasks of different granularities that previous\nmethods have yet to notice. In this paper, we propose a model called MPT-PAR\nthat considers both the unique characteristics of each task and the synergies\nbetween different tasks simultaneously, thereby maximizing the utilization of\nfeatures across multi-granularity activity recognition. Furthermore, we\nemphasize the significance of temporal and spatial information by introducing a\nspatio-temporal relation-enhanced module and a scene representation learning\nmodule, which integrate the the spatio-temporal context of action and global\nscene into the feature map of each granularity. Our method achieved an overall\nF1 score of 47.5\\% on the JRDB-PAR dataset, significantly outperforming all the\nstate-of-the-art methods.\n", "date": "2024-08-01", "categories": []}
{"id": "2311.08094", "title": "SkelVIT: Consensus of Vision Transformers for a Lightweight\n  Skeleton-Based Action Recognition System", "abstract": "  Skeleton-based action recognition receives the attention of many researchers\nas it is robust to viewpoint and illumination changes, and its processing is\nmuch more efficient than the processing of video frames. With the emergence of\ndeep learning models, it has become very popular to represent the skeleton data\nin pseudo-image form and apply CNN for action recognition. Thereafter, studies\nconcentrated on finding effective methods for forming pseudo-images. Recently,\nattention networks, more specifically transformers have provided promising\nresults in various vision problems. In this study, the effectiveness of VIT for\nskeleton-based action recognition is examined and its robustness on the\npseudo-image representation scheme is investigated. To this end, a three-level\narchitecture, SkelVit is proposed, which forms a set of pseudo images, applies\na classifier on each of the representations, and combines their results to find\nthe final action class. The performance of SkelVit is examined thoroughly via a\nset of experiments. First, the sensitivity of the system to representation is\ninvestigated by comparing it with two of the state-of-the-art pseudo-image\nrepresentation methods. Then, the classifiers of SkelVit are realized in two\nexperimental setups by CNNs and VITs, and their performances are compared. In\nthe final experimental setup, the contribution of combining classifiers is\nexamined by applying the model with a different number of classifiers.\nExperimental studies reveal that the proposed system with its lightweight\nrepresentation scheme achieves better results than the state-of-the-art\nmethods. It is also observed that the vision transformer is less sensitive to\nthe initial pseudo-image representation compared to CNN. Nevertheless, even\nwith the vision transformer, the recognition performance can be further\nimproved by the consensus of classifiers.\n", "date": "2023-11-14", "categories": []}
{"id": "2407.21525", "title": "Skeleton-Based Action Recognition with Spatial-Structural Graph\n  Convolution", "abstract": "  Human Activity Recognition (HAR) is a field of study that focuses on\nidentifying and classifying human activities. Skeleton-based Human Activity\nRecognition has received much attention in recent years, where Graph\nConvolutional Network (GCN) based method is widely used and has achieved\nremarkable results. However, the representation of skeleton data and the issue\nof over-smoothing in GCN still need to be studied. 1). Compared to central\nnodes, edge nodes can only aggregate limited neighbor information, and\ndifferent edge nodes of the human body are always structurally related.\nHowever, the information from edge nodes is crucial for fine-grained activity\nrecognition. 2). The Graph Convolutional Network suffers from a significant\nover-smoothing issue, causing nodes to become increasingly similar as the\nnumber of network layers increases. Based on these two ideas, we propose a\ntwo-stream graph convolution method called Spatial-Structural GCN (SpSt-GCN).\nSpatial GCN performs information aggregation based on the topological structure\nof the human body, and structural GCN performs differentiation based on the\nsimilarity of edge node sequences. The spatial connection is fixed, and the\nhuman skeleton naturally maintains this topology regardless of the actions\nperformed by humans. However, the structural connection is dynamic and depends\non the type of movement the human body is performing. Based on this idea, we\nalso propose an entirely data-driven structural connection, which greatly\nincreases flexibility. We evaluate our method on two large-scale datasets,\ni.e., NTU RGB+D and NTU RGB+D 120. The proposed method achieves good results\nwhile being efficient.\n", "date": "2024-07-31", "categories": []}
{"id": "2407.12322", "title": "Frequency Guidance Matters: Skeletal Action Recognition by\n  Frequency-Aware Mixed Transformer", "abstract": "  Recently, transformers have demonstrated great potential for modeling\nlong-term dependencies from skeleton sequences and thereby gained\never-increasing attention in skeleton action recognition. However, the existing\ntransformer-based approaches heavily rely on the naive attention mechanism for\ncapturing the spatiotemporal features, which falls short in learning\ndiscriminative representations that exhibit similar motion patterns. To address\nthis challenge, we introduce the Frequency-aware Mixed Transformer\n(FreqMixFormer), specifically designed for recognizing similar skeletal actions\nwith subtle discriminative motions. First, we introduce a frequency-aware\nattention module to unweave skeleton frequency representations by embedding\njoint features into frequency attention maps, aiming to distinguish the\ndiscriminative movements based on their frequency coefficients. Subsequently,\nwe develop a mixed transformer architecture to incorporate spatial features\nwith frequency features to model the comprehensive frequency-spatial patterns.\nAdditionally, a temporal transformer is proposed to extract the global\ncorrelations across frames. Extensive experiments show that FreqMiXFormer\noutperforms SOTA on 3 popular skeleton action recognition datasets, including\nNTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets.\n", "date": "2024-07-17", "categories": []}
{"id": "2303.18177", "title": "STMT: A Spatial-Temporal Mesh Transformer for MoCap-Based Action\n  Recognition", "abstract": "  We study the problem of human action recognition using motion capture (MoCap)\nsequences. Unlike existing techniques that take multiple manual steps to derive\nstandardized skeleton representations as model input, we propose a novel\nSpatial-Temporal Mesh Transformer (STMT) to directly model the mesh sequences.\nThe model uses a hierarchical transformer with intra-frame off-set attention\nand inter-frame self-attention. The attention mechanism allows the model to\nfreely attend between any two vertex patches to learn non-local relationships\nin the spatial-temporal domain. Masked vertex modeling and future frame\nprediction are used as two self-supervised tasks to fully activate the\nbi-directional and auto-regressive attention in our hierarchical transformer.\nThe proposed method achieves state-of-the-art performance compared to\nskeleton-based and point-cloud-based models on common MoCap benchmarks. Code is\navailable at https://github.com/zgzxy001/STMT.\n", "date": "2023-03-31", "categories": []}
{"id": "2406.01316", "title": "Enhancing Inertial Hand based HAR through Joint Representation of\n  Language, Pose and Synthetic IMUs", "abstract": "  Due to the scarcity of labeled sensor data in HAR, prior research has turned\nto video data to synthesize Inertial Measurement Units (IMU) data, capitalizing\non its rich activity annotations. However, generating IMU data from videos\npresents challenges for HAR in real-world settings, attributed to the poor\nquality of synthetic IMU data and its limited efficacy in subtle, fine-grained\nmotions. In this paper, we propose Multi$^3$Net, our novel multi-modal,\nmultitask, and contrastive-based framework approach to address the issue of\nlimited data. Our pretraining procedure uses videos from online repositories,\naiming to learn joint representations of text, pose, and IMU simultaneously. By\nemploying video data and contrastive learning, our method seeks to enhance\nwearable HAR performance, especially in recognizing subtle activities.Our\nexperimental findings validate the effectiveness of our approach in improving\nHAR performance with IMU data. We demonstrate that models trained with\nsynthetic IMU data generated from videos using our method surpass existing\napproaches in recognizing fine-grained activities.\n", "date": "2024-06-03", "categories": []}
{"id": "2407.09924", "title": "Region-aware Image-based Human Action Retrieval with Transformers", "abstract": "  Human action understanding is a fundamental and challenging task in computer\nvision. Although there exists tremendous research on this area, most works\nfocus on action recognition, while action retrieval has received less\nattention. In this paper, we focus on the neglected but important task of\nimage-based action retrieval which aims to find images that depict the same\naction as a query image. We establish benchmarks for this task and set up\nimportant baseline methods for fair comparison. We present an end-to-end model\nthat learns rich action representations from three aspects: the anchored\nperson, contextual regions, and the global image. A novel fusion transformer\nmodule is designed to model the relationships among different features and\neffectively fuse them into an action representation. Experiments on the\nStanford-40 and PASCAL VOC 2012 Action datasets show that the proposed method\nsignificantly outperforms previous approaches for image-based action retrieval.\n", "date": "2024-07-13", "categories": []}
{"id": "2407.19820", "title": "ActivityCLIP: Enhancing Group Activity Recognition by Mining\n  Complementary Information from Text to Supplement Image Modality", "abstract": "  Previous methods usually only extract the image modality's information to\nrecognize group activity. However, mining image information is approaching\nsaturation, making it difficult to extract richer information. Therefore,\nextracting complementary information from other modalities to supplement image\ninformation has become increasingly important. In fact, action labels provide\nclear text information to express the action's semantics, which existing\nmethods often overlook. Thus, we propose ActivityCLIP, a plug-and-play method\nfor mining the text information contained in the action labels to supplement\nthe image information for enhancing group activity recognition. ActivityCLIP\nconsists of text and image branches, where the text branch is plugged into the\nimage branch (The off-the-shelf image-based method). The text branch includes\nImage2Text and relation modeling modules. Specifically, we propose the\nknowledge transfer module, Image2Text, which adapts image information into text\ninformation extracted by CLIP via knowledge distillation. Further, to keep our\nmethod convenient, we add fewer trainable parameters based on the relation\nmodule of the image branch to model interaction relation in the text branch. To\nshow our method's generality, we replicate three representative methods by\nActivityCLIP, which adds only limited trainable parameters, achieving favorable\nperformance improvements for each method. We also conduct extensive ablation\nstudies and compare our method with state-of-the-art methods to demonstrate the\neffectiveness of ActivityCLIP.\n", "date": "2024-07-29", "categories": []}
{"id": "2407.17792", "title": "Harnessing Temporal Causality for Advanced Temporal Action Detection", "abstract": "  As a fundamental task in long-form video understanding, temporal action\ndetection (TAD) aims to capture inherent temporal relations in untrimmed videos\nand identify candidate actions with precise boundaries. Over the years, various\nnetworks, including convolutions, graphs, and transformers, have been explored\nfor effective temporal modeling for TAD. However, these modules typically treat\npast and future information equally, overlooking the crucial fact that changes\nin action boundaries are essentially causal events. Inspired by this insight,\nwe propose leveraging the temporal causality of actions to enhance TAD\nrepresentation by restricting the model's access to only past or future\ncontext. We introduce CausalTAD, which combines causal attention and causal\nMamba to achieve state-of-the-art performance on multiple benchmarks. Notably,\nwith CausalTAD, we ranked 1st in the Action Recognition, Action Detection, and\nAudio-Based Interaction Detection tracks at the EPIC-Kitchens Challenge 2024,\nas well as 1st in the Moment Queries track at the Ego4D Challenge 2024. Our\ncode is available at https://github.com/sming256/OpenTAD/.\n", "date": "2024-07-25", "categories": []}
{"id": "2407.07673", "title": "Towards Adaptive Pseudo-label Learning for Semi-Supervised Temporal\n  Action Localization", "abstract": "  Alleviating noisy pseudo labels remains a key challenge in Semi-Supervised\nTemporal Action Localization (SS-TAL). Existing methods often filter pseudo\nlabels based on strict conditions, but they typically assess classification and\nlocalization quality separately, leading to suboptimal pseudo-label ranking and\nselection. In particular, there might be inaccurate pseudo labels within\nselected positives, alongside reliable counterparts erroneously assigned to\nnegatives. To tackle these problems, we propose a novel Adaptive Pseudo-label\nLearning (APL) framework to facilitate better pseudo-label selection.\nSpecifically, to improve the ranking quality, Adaptive Label Quality Assessment\n(ALQA) is proposed to jointly learn classification confidence and localization\nreliability, followed by dynamically selecting pseudo labels based on the joint\nscore. Additionally, we propose an Instance-level Consistency Discriminator\n(ICD) for eliminating ambiguous positives and mining potential positives\nsimultaneously based on inter-instance intrinsic consistency, thereby leading\nto a more precise selection. We further introduce a general unsupervised\nAction-aware Contrastive Pre-training (ACP) to enhance the discrimination both\nwithin actions and between actions and backgrounds, which benefits SS-TAL.\nExtensive experiments on THUMOS14 and ActivityNet v1.3 demonstrate that our\nmethod achieves state-of-the-art performance under various semi-supervised\nsettings.\n", "date": "2024-07-10", "categories": []}
{"id": "2407.18249", "title": "Trajectory-aligned Space-time Tokens for Few-shot Action Recognition", "abstract": "  We propose a simple yet effective approach for few-shot action recognition,\nemphasizing the disentanglement of motion and appearance representations. By\nharnessing recent progress in tracking, specifically point trajectories and\nself-supervised representation learning, we build trajectory-aligned tokens\n(TATs) that capture motion and appearance information. This approach\nsignificantly reduces the data requirements while retaining essential\ninformation. To process these representations, we use a Masked Space-time\nTransformer that effectively learns to aggregate information to facilitate\nfew-shot action recognition. We demonstrate state-of-the-art results on\nfew-shot action recognition across multiple datasets. Our project page is\navailable at https://www.cs.umd.edu/~pulkit/tats\n", "date": "2024-07-25", "categories": []}
{"id": "2404.09308", "title": "In My Perspective, In My Hands: Accurate Egocentric 2D Hand Pose and\n  Action Recognition", "abstract": "  Action recognition is essential for egocentric video understanding, allowing\nautomatic and continuous monitoring of Activities of Daily Living (ADLs)\nwithout user effort. Existing literature focuses on 3D hand pose input, which\nrequires computationally intensive depth estimation networks or wearing an\nuncomfortable depth sensor. In contrast, there has been insufficient research\nin understanding 2D hand pose for egocentric action recognition, despite the\navailability of user-friendly smart glasses in the market capable of capturing\na single RGB image. Our study aims to fill this research gap by exploring the\nfield of 2D hand pose estimation for egocentric action recognition, making two\ncontributions. Firstly, we introduce two novel approaches for 2D hand pose\nestimation, namely EffHandNet for single-hand estimation and EffHandEgoNet,\ntailored for an egocentric perspective, capturing interactions between hands\nand objects. Both methods outperform state-of-the-art models on H2O and FPHA\npublic benchmarks. Secondly, we present a robust action recognition\narchitecture from 2D hand and object poses. This method incorporates\nEffHandEgoNet, and a transformer-based action recognition method. Evaluated on\nH2O and FPHA datasets, our architecture has a faster inference time and\nachieves an accuracy of 91.32% and 94.43%, respectively, surpassing state of\nthe art, including 3D-based methods. Our work demonstrates that using 2D\nskeletal data is a robust approach for egocentric action understanding.\nExtensive evaluation and ablation studies show the impact of the hand pose\nestimation approach, and how each input affects the overall performance.\n", "date": "2024-04-14", "categories": []}
{"id": "2404.09490", "title": "Leveraging Temporal Contextualization for Video Action Recognition", "abstract": "  We propose a novel framework for video understanding, called Temporally\nContextualized CLIP (TC-CLIP), which leverages essential temporal information\nthrough global interactions in a spatio-temporal domain within a video. To be\nspecific, we introduce Temporal Contextualization (TC), a layer-wise temporal\ninformation infusion mechanism for videos, which 1) extracts core information\nfrom each frame, 2) connects relevant information across frames for the\nsummarization into context tokens, and 3) leverages the context tokens for\nfeature encoding. Furthermore, the Video-conditional Prompting (VP) module\nprocesses context tokens to generate informative prompts in the text modality.\nExtensive experiments in zero-shot, few-shot, base-to-novel, and\nfully-supervised action recognition validate the effectiveness of our model.\nAblation studies for TC and VP support our design choices. Our project page\nwith the source code is available at https://github.com/naver-ai/tc-clip\n", "date": "2024-04-15", "categories": []}
{"id": "2407.14655", "title": "LORTSAR: Low-Rank Transformer for Skeleton-based Action Recognition", "abstract": "  The complexity of state-of-the-art Transformer-based models for\nskeleton-based action recognition poses significant challenges in terms of\ncomputational efficiency and resource utilization. In this paper, we explore\nthe application of Singular Value Decomposition (SVD) to effectively reduce the\nmodel sizes of these pre-trained models, aiming to minimize their resource\nconsumption while preserving accuracy. Our method, LORTSAR (LOw-Rank\nTransformer for Skeleton-based Action Recognition), also includes a fine-tuning\nstep to compensate for any potential accuracy degradation caused by model\ncompression, and is applied to two leading Transformer-based models,\n\"Hyperformer\" and \"STEP-CATFormer\". Experimental results on the \"NTU RGB+D\" and\n\"NTU RGB+D 120\" datasets show that our method can reduce the number of model\nparameters substantially with negligible degradation or even performance\nincrease in recognition accuracy. This confirms that SVD combined with\npost-compression fine-tuning can boost model efficiency, paving the way for\nmore sustainable, lightweight, and high-performance technologies in human\naction recognition.\n", "date": "2024-07-19", "categories": []}
{"id": "2407.14811", "title": "Decoupled Prompt-Adapter Tuning for Continual Activity Recognition", "abstract": "  Action recognition technology plays a vital role in enhancing security\nthrough surveillance systems, enabling better patient monitoring in healthcare,\nproviding in-depth performance analysis in sports, and facilitating seamless\nhuman-AI collaboration in domains such as manufacturing and assistive\ntechnologies. The dynamic nature of data in these areas underscores the need\nfor models that can continuously adapt to new video data without losing\npreviously acquired knowledge, highlighting the critical role of advanced\ncontinual action recognition. To address these challenges, we propose Decoupled\nPrompt-Adapter Tuning (DPAT), a novel framework that integrates adapters for\ncapturing spatial-temporal information and learnable prompts for mitigating\ncatastrophic forgetting through a decoupled training strategy. DPAT uniquely\nbalances the generalization benefits of prompt tuning with the plasticity\nprovided by adapters in pretrained vision models, effectively addressing the\nchallenge of maintaining model performance amidst continuous data evolution\nwithout necessitating extensive finetuning. DPAT consistently achieves\nstate-of-the-art performance across several challenging action recognition\nbenchmarks, thus demonstrating the effectiveness of our model in the domain of\ncontinual action recognition.\n", "date": "2024-07-20", "categories": []}
{"id": "2407.14834", "title": "Can VLMs be used on videos for action recognition? LLMs are Visual\n  Reasoning Coordinators", "abstract": "  Recent advancements have introduced multiple vision-language models (VLMs)\ndemonstrating impressive commonsense reasoning across various domains. Despite\ntheir individual capabilities, the potential of synergizing these complementary\nVLMs remains underexplored. The Cola Framework addresses this by showcasing how\na large language model (LLM) can efficiently coordinate multiple VLMs through\nnatural language communication, leveraging their distinct strengths. We have\nverified this claim on the challenging A-OKVQA dataset, confirming the\neffectiveness of such coordination. Building on this, our study investigates\nwhether the same methodology can be applied to surveillance videos for action\nrecognition. Specifically, we explore if leveraging the combined knowledge base\nof VLMs and LLM can effectively deduce actions from a video when presented with\nonly a few selectively important frames and minimal temporal information. Our\nexperiments demonstrate that LLM, when coordinating different VLMs, can\nsuccessfully recognize patterns and deduce actions in various scenarios despite\nthe weak temporal signals. However, our findings suggest that to enhance this\napproach as a viable alternative solution, integrating a stronger temporal\nsignal and exposing the models to slightly more frames would be beneficial.\n", "date": "2024-07-20", "categories": []}
{"id": "2407.15605", "title": "Probing Fine-Grained Action Understanding and Cross-View Generalization\n  of Foundation Models", "abstract": "  Foundation models (FMs) are large neural networks trained on broad datasets,\nexcelling in downstream tasks with minimal fine-tuning. Human activity\nrecognition in video has advanced with FMs, driven by competition among\ndifferent architectures. However, high accuracies on standard benchmarks can\ndraw an artificially rosy picture, as they often overlook real-world factors\nlike changing camera perspectives. Popular benchmarks, mostly from YouTube or\nmovies, offer diverse views but only coarse actions, which are insufficient for\nuse-cases needing fine-grained, domain-specific actions. Domain-specific\ndatasets (e.g., for industrial assembly) typically use data from limited static\nperspectives. This paper empirically evaluates how perspective changes affect\ndifferent FMs in fine-grained human activity recognition. We compare multiple\nbackbone architectures and design choices, including image- and video- based\nmodels, and various strategies for temporal information fusion, including\ncommonly used score averaging and more novel attention-based temporal\naggregation mechanisms. This is the first systematic study of different\nfoundation models and specific design choices for human activity recognition\nfrom unknown views, conducted with the goal to provide guidance for backbone-\nand temporal- fusion scheme selection. Code and models will be made publicly\navailable to the community.\n", "date": "2024-07-22", "categories": []}
{"id": "2312.12347", "title": "SMC-NCA: Semantic-guided Multi-level Contrast for Semi-supervised\n  Temporal Action Segmentation", "abstract": "  Semi-supervised temporal action segmentation (SS-TA) aims to perform\nframe-wise classification in long untrimmed videos, where only a fraction of\nvideos in the training set have labels. Recent studies have shown the potential\nof contrastive learning in unsupervised representation learning using\nunlabelled data. However, learning the representation of each frame by\nunsupervised contrastive learning for action segmentation remains an open and\nchallenging problem. In this paper, we propose a novel Semantic-guided\nMulti-level Contrast scheme with a Neighbourhood-Consistency-Aware unit\n(SMC-NCA) to extract strong frame-wise representations for SS-TAS.\nSpecifically, for representation learning, SMC is first used to explore intra-\nand inter-information variations in a unified and contrastive way, based on\naction-specific semantic information and temporal information highlighting\nrelations between actions. Then, the NCA module, which is responsible for\nenforcing spatial consistency between neighbourhoods centered at different\nframes to alleviate over-segmentation issues, works alongside SMC for\nsemi-supervised learning (SSL). Our SMC outperforms the other state-of-the-art\nmethods on three benchmarks, offering improvements of up to 17.8% and 12.6% in\nterms of Edit distance and accuracy, respectively. Additionally, the NCA unit\nresults in significantly better segmentation performance in the presence of\nonly 5% labelled videos. We also demonstrate the generalizability and\neffectiveness of the proposed method on our Parkinson Disease's Mouse Behaviour\n(PDMB) dataset. Code is available at https://github.com/FeixiangZhou/SMC-NCA.\n", "date": "2023-12-19", "categories": []}
{"id": "2407.06113", "title": "C2C: Component-to-Composition Learning for Zero-Shot Compositional\n  Action Recognition", "abstract": "  Compositional actions consist of dynamic (verbs) and static (objects)\nconcepts. Humans can easily recognize unseen compositions using the learned\nconcepts. For machines, solving such a problem requires a model to recognize\nunseen actions composed of previously observed verbs and objects, thus\nrequiring so-called compositional generalization ability. To facilitate this\nresearch, we propose a novel Zero-Shot Compositional Action Recognition\n(ZS-CAR) task. For evaluating the task, we construct a new benchmark,\nSomething-composition (Sth-com), based on the widely used Something-Something\nV2 dataset. We also propose a novel Component-to-Composition (C2C) learning\nmethod to solve the new ZS-CAR task. C2C includes an independent component\nlearning module and a composition inference module. Last, we devise an enhanced\ntraining strategy to address the challenges of component variations between\nseen and unseen compositions and to handle the subtle balance between learning\nseen and unseen actions. The experimental results demonstrate that the proposed\nframework significantly surpasses the existing compositional generalization\nmethods and sets a new state-of-the-art. The new Sth-com benchmark and code are\navailable at https://github.com/RongchangLi/ZSCAR_C2C.\n", "date": "2024-07-08", "categories": []}
{"id": "2407.12483", "title": "Towards AI-Powered Video Assistant Referee System (VARS) for Association\n  Football", "abstract": "  Over the past decade, the technology used by referees in football has\nimproved substantially, enhancing the fairness and accuracy of decisions. This\nprogress has culminated in the implementation of the Video Assistant Referee\n(VAR), an innovation that enables backstage referees to review incidents on the\npitch from multiple points of view. However, the VAR is currently limited to\nprofessional leagues due to its expensive infrastructure and the lack of\nreferees worldwide. In this paper, we present the semi-automated Video\nAssistant Referee System (VARS) that leverages the latest findings in\nmulti-view video analysis. VARS sets a new state-of-the-art on the\nSoccerNet-MVFoul dataset, a multi-view video dataset of football fouls. Our\nVARS achieves a new state-of-the-art on the SoccerNet-MVFoul dataset by\nrecognizing the type of foul in 50% of instances and the appropriate sanction\nin 46% of cases. Finally, we conducted a comparative study to investigate human\nperformance in classifying fouls and their corresponding severity and compared\nthese findings to our VARS. The results of our study highlight the potential of\nour VARS to reach human performance and support football refereeing across all\nlevels of professional and amateur federations.\n", "date": "2024-07-17", "categories": []}
{"id": "2407.12987", "title": "ActionSwitch: Class-agnostic Detection of Simultaneous Actions in\n  Streaming Videos", "abstract": "  Online Temporal Action Localization (On-TAL) is a critical task that aims to\ninstantaneously identify action instances in untrimmed streaming videos as soon\nas an action concludes -- a major leap from frame-based Online Action Detection\n(OAD). Yet, the challenge of detecting overlapping actions is often overlooked\neven though it is a common scenario in streaming videos. Current methods that\ncan address concurrent actions depend heavily on class information, limiting\ntheir flexibility. This paper introduces ActionSwitch, the first class-agnostic\nOn-TAL framework capable of detecting overlapping actions. By obviating the\nreliance on class information, ActionSwitch provides wider applicability to\nvarious situations, including overlapping actions of the same class or\nscenarios where class information is unavailable. This approach is complemented\nby the proposed \"conservativeness loss\", which directly embeds a conservative\ndecision-making principle into the loss function for On-TAL. Our ActionSwitch\nachieves state-of-the-art performance in complex datasets, including\nEpic-Kitchens 100 targeting the challenging egocentric view and FineAction\nconsisting of fine-grained actions.\n", "date": "2024-07-17", "categories": []}
{"id": "2407.13078", "title": "Enhancing Temporal Action Localization: Advanced S6 Modeling with\n  Recurrent Mechanism", "abstract": "  Temporal Action Localization (TAL) is a critical task in video analysis,\nidentifying precise start and end times of actions. Existing methods like CNNs,\nRNNs, GCNs, and Transformers have limitations in capturing long-range\ndependencies and temporal causality. To address these challenges, we propose a\nnovel TAL architecture leveraging the Selective State Space Model (S6). Our\napproach integrates the Feature Aggregated Bi-S6 block, Dual Bi-S6 structure,\nand a recurrent mechanism to enhance temporal and channel-wise dependency\nmodeling without increasing parameter complexity. Extensive experiments on\nbenchmark datasets demonstrate state-of-the-art results with mAP scores of\n74.2% on THUMOS-14, 42.9% on ActivityNet, 29.6% on FineAction, and 45.8% on\nHACS. Ablation studies validate our method's effectiveness, showing that the\nDual structure in the Stem module and the recurrent mechanism outperform\ntraditional approaches. Our findings demonstrate the potential of S6-based\nmodels in TAL tasks, paving the way for future research.\n", "date": "2024-07-18", "categories": []}
{"id": "2407.13216", "title": "QuIIL at T3 challenge: Towards Automation in Life-Saving Intervention\n  Procedures from First-Person View", "abstract": "  In this paper, we present our solutions for a spectrum of automation tasks in\nlife-saving intervention procedures within the Trauma THOMPSON (T3) Challenge,\nencompassing action recognition, action anticipation, and Visual Question\nAnswering (VQA). For action recognition and anticipation, we propose a\npre-processing strategy that samples and stitches multiple inputs into a single\nimage and then incorporates momentum- and attention-based knowledge\ndistillation to improve the performance of the two tasks. For training, we\npresent an action dictionary-guided design, which consistently yields the most\nfavorable results across our experiments. In the realm of VQA, we leverage\nobject-level features and deploy co-attention networks to train both object and\nquestion features. Notably, we introduce a novel frame-question cross-attention\nmechanism at the network's core for enhanced performance. Our solutions achieve\nthe $2^{nd}$ rank in action recognition and anticipation tasks and $1^{st}$\nrank in the VQA task.\n", "date": "2024-07-18", "categories": []}
{"id": "2407.13460", "title": "SA-DVAE: Improving Zero-Shot Skeleton-Based Action Recognition by\n  Disentangled Variational Autoencoders", "abstract": "  Existing zero-shot skeleton-based action recognition methods utilize\nprojection networks to learn a shared latent space of skeleton features and\nsemantic embeddings. The inherent imbalance in action recognition datasets,\ncharacterized by variable skeleton sequences yet constant class labels,\npresents significant challenges for alignment. To address the imbalance, we\npropose SA-DVAE -- Semantic Alignment via Disentangled Variational\nAutoencoders, a method that first adopts feature disentanglement to separate\nskeleton features into two independent parts -- one is semantic-related and\nanother is irrelevant -- to better align skeleton and semantic features. We\nimplement this idea via a pair of modality-specific variational autoencoders\ncoupled with a total correction penalty. We conduct experiments on three\nbenchmark datasets: NTU RGB+D, NTU RGB+D 120 and PKU-MMD, and our experimental\nresults show that SA-DAVE produces improved performance over existing methods.\nThe code is available at https://github.com/pha123661/SA-DVAE.\n", "date": "2024-07-18", "categories": []}
{"id": "2403.09508", "title": "SkateFormer: Skeletal-Temporal Transformer for Human Action Recognition", "abstract": "  Skeleton-based action recognition, which classifies human actions based on\nthe coordinates of joints and their connectivity within skeleton data, is\nwidely utilized in various scenarios. While Graph Convolutional Networks (GCNs)\nhave been proposed for skeleton data represented as graphs, they suffer from\nlimited receptive fields constrained by joint connectivity. To address this\nlimitation, recent advancements have introduced transformer-based methods.\nHowever, capturing correlations between all joints in all frames requires\nsubstantial memory resources. To alleviate this, we propose a novel approach\ncalled Skeletal-Temporal Transformer (SkateFormer) that partitions joints and\nframes based on different types of skeletal-temporal relation (Skate-Type) and\nperforms skeletal-temporal self-attention (Skate-MSA) within each partition. We\ncategorize the key skeletal-temporal relations for action recognition into a\ntotal of four distinct types. These types combine (i) two skeletal relation\ntypes based on physically neighboring and distant joints, and (ii) two temporal\nrelation types based on neighboring and distant frames. Through this\npartition-specific attention strategy, our SkateFormer can selectively focus on\nkey joints and frames crucial for action recognition in an action-adaptive\nmanner with efficient computation. Extensive experiments on various benchmark\ndatasets validate that our SkateFormer outperforms recent state-of-the-art\nmethods.\n", "date": "2024-03-14", "categories": []}
{"id": "2407.12312", "title": "Shap-Mix: Shapley Value Guided Mixing for Long-Tailed Skeleton Based\n  Action Recognition", "abstract": "  In real-world scenarios, human actions often fall into a long-tailed\ndistribution. It makes the existing skeleton-based action recognition works,\nwhich are mostly designed based on balanced datasets, suffer from a sharp\nperformance degradation. Recently, many efforts have been madeto image/video\nlong-tailed learning. However, directly applying them to skeleton data can be\nsub-optimal due to the lack of consideration of the crucial spatial-temporal\nmotion patterns, especially for some modality-specific methodologies such as\ndata augmentation. To this end, considering the crucial role of the body parts\nin the spatially concentrated human actions, we attend to the mixing\naugmentations and propose a novel method, Shap-Mix, which improves long-tailed\nlearning by mining representative motion patterns for tail categories.\nSpecifically, we first develop an effective spatial-temporal mixing strategy\nfor the skeleton to boost representation quality. Then, the employed saliency\nguidance method is presented, consisting of the saliency estimation based on\nShapley value and a tail-aware mixing policy. It preserves the salient motion\nparts of minority classes in mixed data, explicitly establishing the\nrelationships between crucial body structure cues and high-level semantics.\nExtensive experiments on three large-scale skeleton datasets show our\nremarkable performance improvement under both long-tailed and balanced\nsettings. Our project is publicly available at:\nhttps://jhang2020.github.io/Projects/Shap-Mix/Shap-Mix.html.\n", "date": "2024-07-17", "categories": []}
{"id": "2312.02638", "title": "Synchronization is All You Need: Exocentric-to-Egocentric Transfer for\n  Temporal Action Segmentation with Unlabeled Synchronized Video Pairs", "abstract": "  We consider the problem of transferring a temporal action segmentation system\ninitially designed for exocentric (fixed) cameras to an egocentric scenario,\nwhere wearable cameras capture video data. The conventional supervised approach\nrequires the collection and labeling of a new set of egocentric videos to adapt\nthe model, which is costly and time-consuming. Instead, we propose a novel\nmethodology which performs the adaptation leveraging existing labeled\nexocentric videos and a new set of unlabeled, synchronized\nexocentric-egocentric video pairs, for which temporal action segmentation\nannotations do not need to be collected. We implement the proposed methodology\nwith an approach based on knowledge distillation, which we investigate both at\nthe feature and Temporal Action Segmentation model level. Experiments on\nAssembly101 and EgoExo4D demonstrate the effectiveness of the proposed method\nagainst classic unsupervised domain adaptation and temporal alignment\napproaches. Without bells and whistles, our best model performs on par with\nsupervised approaches trained on labeled egocentric data, without ever seeing a\nsingle egocentric label, achieving a +15.99 improvement in the edit score\n(28.59 vs 12.60) on the Assembly101 dataset compared to a baseline model\ntrained solely on exocentric data. In similar settings, our method also\nimproves edit score by +3.32 on the challenging EgoExo4D benchmark. Code is\navailable here: https://github.com/fpv-iplab/synchronization-is-all-you-need.\n", "date": "2023-12-05", "categories": []}
{"id": "2402.08875", "title": "Advancing Human Action Recognition with Foundation Models trained on\n  Unlabeled Public Videos", "abstract": "  The increasing variety and quantity of tagged multimedia content on a variety\nof online platforms offer a unique opportunity to advance the field of human\naction recognition. In this study, we utilize 283,582 unique, unlabeled TikTok\nvideo clips, categorized into 386 hashtags, to train a domain-specific\nfoundation model for action recognition. We employ VideoMAE V2, an advanced\nmodel integrating Masked Autoencoders (MAE) with Vision Transformers (ViT),\npre-trained on this diverse collection of unstructured videos. Our model,\nfine-tuned on established action recognition benchmarks such as UCF101 and\nHMDB51, achieves state-of-the-art results: 99.05% on UCF101, 86.08% on HMDB51,\n85.51% on Kinetics-400, and 74.27% on Something-Something V2 using the\nViT-giant backbone. These results highlight the potential of using unstructured\nand unlabeled videos as a valuable source of diverse and dynamic content for\ntraining foundation models. Our investigation confirms that while initial\nincreases in pre-training data volume significantly enhance model performance,\nthe gains diminish as the dataset size continues to expand. Our findings\nemphasize two critical axioms in self-supervised learning for computer vision:\n(1) additional pre-training data can yield diminishing benefits for some\ndatasets and (2) quality is more important than quantity in self-supervised\nlearning, especially when building foundation models.\n", "date": "2024-02-14", "categories": []}
{"id": "2406.08877", "title": "EgoExo-Fitness: Towards Egocentric and Exocentric Full-Body Action\n  Understanding", "abstract": "  We present EgoExo-Fitness, a new full-body action understanding dataset,\nfeaturing fitness sequence videos recorded from synchronized egocentric and\nfixed exocentric (third-person) cameras. Compared with existing full-body\naction understanding datasets, EgoExo-Fitness not only contains videos from\nfirst-person perspectives, but also provides rich annotations. Specifically,\ntwo-level temporal boundaries are provided to localize single action videos\nalong with sub-steps of each action. More importantly, EgoExo-Fitness\nintroduces innovative annotations for interpretable action judgement--including\ntechnical keypoint verification, natural language comments on action execution,\nand action quality scores. Combining all of these, EgoExo-Fitness provides new\nresources to study egocentric and exocentric full-body action understanding\nacross dimensions of \"what\", \"when\", and \"how well\". To facilitate research on\negocentric and exocentric full-body action understanding, we construct\nbenchmarks on a suite of tasks (i.e., action classification, action\nlocalization, cross-view sequence verification, cross-view skill determination,\nand a newly proposed task of guidance-based execution verification), together\nwith detailed analysis. Code and data will be available at\nhttps://github.com/iSEE-Laboratory/EgoExo-Fitness/tree/main.\n", "date": "2024-06-13", "categories": []}
{"id": "2407.05106", "title": "DailyDVS-200: A Comprehensive Benchmark Dataset for Event-Based Action\n  Recognition", "abstract": "  Neuromorphic sensors, specifically event cameras, revolutionize visual data\nacquisition by capturing pixel intensity changes with exceptional dynamic\nrange, minimal latency, and energy efficiency, setting them apart from\nconventional frame-based cameras. The distinctive capabilities of event cameras\nhave ignited significant interest in the domain of event-based action\nrecognition, recognizing their vast potential for advancement. However, the\ndevelopment in this field is currently slowed by the lack of comprehensive,\nlarge-scale datasets, which are critical for developing robust recognition\nframeworks. To bridge this gap, we introduces DailyDVS-200, a meticulously\ncurated benchmark dataset tailored for the event-based action recognition\ncommunity. DailyDVS-200 is extensive, covering 200 action categories across\nreal-world scenarios, recorded by 47 participants, and comprises more than\n22,000 event sequences. This dataset is designed to reflect a broad spectrum of\naction types, scene complexities, and data acquisition diversity. Each sequence\nin the dataset is annotated with 14 attributes, ensuring a detailed\ncharacterization of the recorded actions. Moreover, DailyDVS-200 is structured\nto facilitate a wide range of research paths, offering a solid foundation for\nboth validating existing approaches and inspiring novel methodologies. By\nsetting a new benchmark in the field, we challenge the current limitations of\nneuromorphic data processing and invite a surge of new approaches in\nevent-based action recognition techniques, which paves the way for future\nexplorations in neuromorphic computing and beyond. The dataset and source code\nare available at https://github.com/QiWang233/DailyDVS-200.\n", "date": "2024-07-06", "categories": []}
{"id": "2407.09520", "title": "Exploring the Impact of Hand Pose and Shadow on Hand-washing Action\n  Recognition", "abstract": "  In the real world, camera-based application systems can face many challenges,\nincluding environmental factors and distribution shift. In this paper, we\ninvestigate how pose and shadow impact a classifier's performance, using the\nspecific application of handwashing action recognition. To accomplish this, we\ngenerate synthetic data with desired variations to introduce controlled\ndistribution shift. Using our synthetic dataset, we define a classifier's\nbreakdown points to be where the system's performance starts to degrade\nsharply, and we show these are heavily impacted by pose and shadow conditions.\nIn particular, heavier and larger shadows create earlier breakdown points.\nAlso, it is intriguing to observe model accuracy drop to almost zero with\nbigger changes in pose. Moreover, we propose a simple mitigation strategy for\npose-induced breakdown points by utilizing additional training data from\nnon-canonical poses. Results show that the optimal choices of additional\ntraining poses are those with moderate deviations from the canonical poses with\n50-60 degrees of rotation.\n", "date": "2024-06-19", "categories": []}
{"id": "2407.09529", "title": "Towards LLM-Powered Ambient Sensor Based Multi-Person Human Activity\n  Recognition", "abstract": "  Human Activity Recognition (HAR) is one of the central problems in fields\nsuch as healthcare, elderly care, and security at home. However, traditional\nHAR approaches face challenges including data scarcity, difficulties in model\ngeneralization, and the complexity of recognizing activities in multi-person\nscenarios. This paper proposes a system framework called LAHAR, based on large\nlanguage models. Utilizing prompt engineering techniques, LAHAR addresses HAR\nin multi-person scenarios by enabling subject separation and action-level\ndescriptions of events occurring in the environment. We validated our approach\non the ARAS dataset, and the results demonstrate that LAHAR achieves comparable\naccuracy to the state-of-the-art method at higher resolutions and maintains\nrobustness in multi-person scenarios.\n", "date": "2024-06-25", "categories": []}
{"id": "2407.08971", "title": "Full-Stage Pseudo Label Quality Enhancement for Weakly-supervised\n  Temporal Action Localization", "abstract": "  Weakly-supervised Temporal Action Localization (WSTAL) aims to localize\nactions in untrimmed videos using only video-level supervision. Latest WSTAL\nmethods introduce pseudo label learning framework to bridge the gap between\nclassification-based training and inferencing targets at localization, and\nachieve cutting-edge results. In these frameworks, a classification-based model\nis used to generate pseudo labels for a regression-based student model to learn\nfrom. However, the quality of pseudo labels in the framework, which is a key\nfactor to the final result, is not carefully studied. In this paper, we propose\na set of simple yet efficient pseudo label quality enhancement mechanisms to\nbuild our FuSTAL framework. FuSTAL enhances pseudo label quality at three\nstages: cross-video contrastive learning at proposal Generation-Stage,\nprior-based filtering at proposal Selection-Stage and EMA-based distillation at\nTraining-Stage. These designs enhance pseudo label quality at different stages\nin the framework, and help produce more informative, less false and smoother\naction proposals. With the help of these comprehensive designs at all stages,\nFuSTAL achieves an average mAP of 50.8% on THUMOS'14, outperforming the\nprevious best method by 1.2%, and becomes the first method to reach the\nmilestone of 50%.\n", "date": "2024-07-12", "categories": []}
{"id": "2404.04933", "title": "UniMD: Towards Unifying Moment Retrieval and Temporal Action Detection", "abstract": "  Temporal Action Detection (TAD) focuses on detecting pre-defined actions,\nwhile Moment Retrieval (MR) aims to identify the events described by open-ended\nnatural language within untrimmed videos. Despite that they focus on different\nevents, we observe they have a significant connection. For instance, most\ndescriptions in MR involve multiple actions from TAD. In this paper, we aim to\ninvestigate the potential synergy between TAD and MR. Firstly, we propose a\nunified architecture, termed Unified Moment Detection (UniMD), for both TAD and\nMR. It transforms the inputs of the two tasks, namely actions for TAD or events\nfor MR, into a common embedding space, and utilizes two novel query-dependent\ndecoders to generate a uniform output of classification score and temporal\nsegments. Secondly, we explore the efficacy of two task fusion learning\napproaches, pre-training and co-training, in order to enhance the mutual\nbenefits between TAD and MR. Extensive experiments demonstrate that the\nproposed task fusion learning scheme enables the two tasks to help each other\nand outperform the separately trained counterparts. Impressively, UniMD\nachieves state-of-the-art results on three paired datasets Ego4D, Charades-STA,\nand ActivityNet. Our code is available at https://github.com/yingsen1/UniMD.\n", "date": "2024-04-07", "categories": []}
{"id": "2407.01872", "title": "Referring Atomic Video Action Recognition", "abstract": "  We introduce a new task called Referring Atomic Video Action Recognition\n(RAVAR), aimed at identifying atomic actions of a particular person based on a\ntextual description and the video data of this person. This task differs from\ntraditional action recognition and localization, where predictions are\ndelivered for all present individuals. In contrast, we focus on recognizing the\ncorrect atomic action of a specific individual, guided by text. To explore this\ntask, we present the RefAVA dataset, containing 36,630 instances with manually\nannotated textual descriptions of the individuals. To establish a strong\ninitial benchmark, we implement and validate baselines from various domains,\ne.g., atomic action localization, video question answering, and text-video\nretrieval. Since these existing methods underperform on RAVAR, we introduce\nRefAtomNet -- a novel cross-stream attention-driven method specialized for the\nunique challenges of RAVAR: the need to interpret a textual referring\nexpression for the targeted individual, utilize this reference to guide the\nspatial localization and harvest the prediction of the atomic actions for the\nreferring person. The key ingredients are: (1) a multi-stream architecture that\nconnects video, text, and a new location-semantic stream, and (2) cross-stream\nagent attention fusion and agent token fusion which amplify the most relevant\ninformation across these streams and consistently surpasses standard\nattention-based fusion on RAVAR. Extensive experiments demonstrate the\neffectiveness of RefAtomNet and its building blocks for recognizing the action\nof the described individual. The dataset and code will be made publicly\navailable at https://github.com/KPeng9510/RAVAR.\n", "date": "2024-07-02", "categories": []}
{"id": "2310.11950", "title": "Too Good To Be True: performance overestimation in (re)current practices\n  for Human Activity Recognition", "abstract": "  Today, there are standard and well established procedures within the Human\nActivity Recognition (HAR) pipeline. However, some of these conventional\napproaches lead to accuracy overestimation. In particular, sliding windows for\ndata segmentation followed by standard random k-fold cross validation, produce\nbiased results. An analysis of previous literature and present-day studies,\nsurprisingly, shows that these are common approaches in state-of-the-art\nstudies on HAR. It is important to raise awareness in the scientific community\nabout this problem, whose negative effects are being overlooked. Otherwise,\npublications of biased results lead to papers that report lower accuracies,\nwith correct unbiased methods, harder to publish. Several experiments with\ndifferent types of datasets and different types of classification models allow\nus to exhibit the problem and show it persists independently of the method or\ndataset.\n", "date": "2023-10-18", "categories": []}
{"id": "2310.02011", "title": "Decoding Human Activities: Analyzing Wearable Accelerometer and\n  Gyroscope Data for Activity Recognition", "abstract": "  A person's movement or relative positioning can be effectively captured by\ndifferent types of sensors and corresponding sensor output can be utilized in\nvarious manipulative techniques for the classification of different human\nactivities. This letter proposes an effective scheme for human activity\nrecognition, which introduces two unique approaches within a multi-structural\narchitecture, named FusionActNet. The first approach aims to capture the static\nand dynamic behavior of a particular action by using two dedicated residual\nnetworks and the second approach facilitates the final decision-making process\nby introducing a guidance module. A two-stage training process is designed\nwhere at the first stage, residual networks are pre-trained separately by using\nstatic (where the human body is immobile) and dynamic (involving movement of\nthe human body) data. In the next stage, the guidance module along with the\npre-trained static or dynamic models are used to train the given sensor data.\nHere the guidance module learns to emphasize the most relevant prediction\nvector obtained from the static or dynamic models, which helps to effectively\nclassify different human activities. The proposed scheme is evaluated using two\nbenchmark datasets and compared with state-of-the-art methods. The results\nclearly demonstrate that our method outperforms existing approaches in terms of\naccuracy, precision, recall, and F1 score, achieving 97.35% and 95.35% accuracy\non the UCI HAR and Motion-Sense datasets, respectively which highlights both\nthe effectiveness and stability of the proposed scheme.\n", "date": "2023-10-03", "categories": []}
{"id": "2407.06628", "title": "Masked Video and Body-worn IMU Autoencoder for Egocentric Action\n  Recognition", "abstract": "  Compared with visual signals, Inertial Measurement Units (IMUs) placed on\nhuman limbs can capture accurate motion signals while being robust to lighting\nvariation and occlusion. While these characteristics are intuitively valuable\nto help egocentric action recognition, the potential of IMUs remains\nunder-explored. In this work, we present a novel method for action recognition\nthat integrates motion data from body-worn IMUs with egocentric video. Due to\nthe scarcity of labeled multimodal data, we design an MAE-based self-supervised\npretraining method, obtaining strong multi-modal representations via modeling\nthe natural correlation between visual and motion signals. To model the complex\nrelation of multiple IMU devices placed across the body, we exploit the\ncollaborative dynamics in multiple IMU devices and propose to embed the\nrelative motion features of human joints into a graph structure. Experiments\nshow our method can achieve state-of-the-art performance on multiple public\ndatasets. The effectiveness of our MAE-based pretraining and graph-based IMU\nmodeling are further validated by experiments in more challenging scenarios,\nincluding partially missing IMU devices and video quality corruption, promoting\nmore flexible usages in the real world.\n", "date": "2024-07-09", "categories": []}
{"id": "2407.05657", "title": "DMSD-CDFSAR: Distillation from Mixed-Source Domain for Cross-Domain\n  Few-shot Action Recognition", "abstract": "  Few-shot action recognition is an emerging field in computer vision,\nprimarily focused on meta-learning within the same domain. However, challenges\narise in real-world scenario deployment, as gathering extensive labeled data\nwithin a specific domain is laborious and time-intensive. Thus, attention\nshifts towards cross-domain few-shot action recognition, requiring the model to\ngeneralize across domains with significant deviations. Therefore, we propose a\nnovel approach, ``Distillation from Mixed-Source Domain\", tailored to address\nthis conundrum. Our method strategically integrates insights from both labeled\ndata of the source domain and unlabeled data of the target domain during the\ntraining. The ResNet18 is used as the backbone to extract spatial features from\nthe source and target domains. We design two branches for meta-training: the\noriginal-source and the mixed-source branches. In the first branch, a Domain\nTemporal Encoder is employed to capture temporal features for both the source\nand target domains. Additionally, a Domain Temporal Decoder is employed to\nreconstruct all extracted features. In the other branch, a Domain Mixed Encoder\nis used to handle labeled source domain data and unlabeled target domain data,\ngenerating mixed-source domain features. We incorporate a pre-training stage\nbefore meta-training, featuring a network architecture similar to that of the\nfirst branch. Lastly, we introduce a dual distillation mechanism to refine the\nclassification probabilities of source domain features, aligning them with\nthose of mixed-source domain features. This iterative process enriches the\ninsights of the original-source branch with knowledge from the mixed-source\nbranch, thereby enhancing the model's generalization capabilities. Our code is\navailable at URL: \\url{https://xxxx/xxxx/xxxx.git}\n", "date": "2024-07-08", "categories": []}
{"id": "2407.04490", "title": "Micro-gesture Online Recognition using Learnable Query Points", "abstract": "  In this paper, we briefly introduce the solution developed by our team,\nHFUT-VUT, for the Micro-gesture Online Recognition track in the MiGA challenge\nat IJCAI 2024. The Micro-gesture Online Recognition task involves identifying\nthe category and locating the start and end times of micro-gestures in video\nclips. Compared to the typical Temporal Action Detection task, the\nMicro-gesture Online Recognition task focuses more on distinguishing between\nmicro-gestures and pinpointing the start and end times of actions. Our solution\nranks 2nd in the Micro-gesture Online Recognition track.\n", "date": "2024-07-05", "categories": []}
{"id": "2407.02713", "title": "Advancing Compressed Video Action Recognition through Progressive\n  Knowledge Distillation", "abstract": "  Compressed video action recognition classifies video samples by leveraging\nthe different modalities in compressed videos, namely motion vectors,\nresiduals, and intra-frames. For this purpose, three neural networks are\ndeployed, each dedicated to processing one modality. Our observations indicate\nthat the network processing intra-frames tend to converge to a flatter minimum\nthan the network processing residuals, which in turn converges to a flatter\nminimum than the motion vector network. This hierarchy in convergence motivates\nour strategy for knowledge transfer among modalities to achieve flatter minima,\nwhich are generally associated with better generalization. With this insight,\nwe propose Progressive Knowledge Distillation (PKD), a technique that\nincrementally transfers knowledge across the modalities. This method involves\nattaching early exits (Internal Classifiers - ICs) to the three networks. PKD\ndistills knowledge starting from the motion vector network, followed by the\nresidual, and finally, the intra-frame network, sequentially improving IC\naccuracy. Further, we propose the Weighted Inference with Scaled Ensemble\n(WISE), which combines outputs from the ICs using learned weights, boosting\naccuracy during inference. Our experiments demonstrate the effectiveness of\ntraining the ICs with PKD compared to standard cross-entropy-based training,\nshowing IC accuracy improvements of up to 5.87% and 11.42% on the UCF-101 and\nHMDB-51 datasets, respectively. Additionally, WISE improves accuracy by up to\n4.28% and 9.30% on UCF-101 and HMDB-51, respectively.\n", "date": "2024-07-02", "categories": []}
{"id": "2407.02934", "title": "PosMLP-Video: Spatial and Temporal Relative Position Encoding for\n  Efficient Video Recognition", "abstract": "  In recent years, vision Transformers and MLPs have demonstrated remarkable\nperformance in image understanding tasks. However, their inherently dense\ncomputational operators, such as self-attention and token-mixing layers, pose\nsignificant challenges when applied to spatio-temporal video data. To address\nthis gap, we propose PosMLP-Video, a lightweight yet powerful MLP-like backbone\nfor video recognition. Instead of dense operators, we use efficient relative\npositional encoding (RPE) to build pairwise token relations, leveraging\nsmall-sized parameterized relative position biases to obtain each relation\nscore. Specifically, to enable spatio-temporal modeling, we extend the image\nPosMLP's positional gating unit to temporal, spatial, and spatio-temporal\nvariants, namely PoTGU, PoSGU, and PoSTGU, respectively. These gating units can\nbe feasibly combined into three types of spatio-temporal factorized positional\nMLP blocks, which not only decrease model complexity but also maintain good\nperformance. Additionally, we enrich relative positional relationships by using\nchannel grouping. Experimental results on three video-related tasks demonstrate\nthat PosMLP-Video achieves competitive speed-accuracy trade-offs compared to\nthe previous state-of-the-art models. In particular, PosMLP-Video pre-trained\non ImageNet1K achieves 59.0%/70.3% top-1 accuracy on Something-Something V1/V2\nand 82.1% top-1 accuracy on Kinetics-400 while requiring much fewer parameters\nand FLOPs than other models. The code is released at\nhttps://github.com/zhouds1918/PosMLP_Video.\n", "date": "2024-07-03", "categories": []}
{"id": "2407.03197", "title": "DyFADet: Dynamic Feature Aggregation for Temporal Action Detection", "abstract": "  Recent proposed neural network-based Temporal Action Detection (TAD) models\nare inherently limited to extracting the discriminative representations and\nmodeling action instances with various lengths from complex scenes by\nshared-weights detection heads. Inspired by the successes in dynamic neural\nnetworks, in this paper, we build a novel dynamic feature aggregation (DFA)\nmodule that can simultaneously adapt kernel weights and receptive fields at\ndifferent timestamps. Based on DFA, the proposed dynamic encoder layer\naggregates the temporal features within the action time ranges and guarantees\nthe discriminability of the extracted representations. Moreover, using DFA\nhelps to develop a Dynamic TAD head (DyHead), which adaptively aggregates the\nmulti-scale features with adjusted parameters and learned receptive fields\nbetter to detect the action instances with diverse ranges from videos. With the\nproposed encoder layer and DyHead, a new dynamic TAD model, DyFADet, achieves\npromising performance on a series of challenging TAD benchmarks, including\nHACS-Segment, THUMOS14, ActivityNet-1.3, Epic-Kitchen 100, Ego4D-Moment\nQueriesV1.0, and FineAction. Code is released to\nhttps://github.com/yangle15/DyFADet-pytorch.\n", "date": "2024-07-03", "categories": []}
{"id": "2406.04609", "title": "Diverse Intra- and Inter-Domain Activity Style Fusion for Cross-Person\n  Generalization in Activity Recognition", "abstract": "  Existing domain generalization (DG) methods for cross-person generalization\ntasks often face challenges in capturing intra- and inter-domain style\ndiversity, resulting in domain gaps with the target domain. In this study, we\nexplore a novel perspective to tackle this problem, a process conceptualized as\ndomain padding. This proposal aims to enrich the domain diversity by\nsynthesizing intra- and inter-domain style data while maintaining robustness to\nclass labels. We instantiate this concept using a conditional diffusion model\nand introduce a style-fused sampling strategy to enhance data generation\ndiversity. In contrast to traditional condition-guided sampling, our\nstyle-fused sampling strategy allows for the flexible use of one or more random\nstyles to guide data synthesis. This feature presents a notable advancement: it\nallows for the maximum utilization of possible permutations and combinations\namong existing styles to generate a broad spectrum of new style instances.\nEmpirical evaluations on a broad range of datasets demonstrate that our\ngenerated data achieves remarkable diversity within the domain space. Both\nintra- and inter-domain generated data have proven to be significant and\nvaluable, contributing to varying degrees of performance enhancements. Notably,\nour approach outperforms state-of-the-art DG methods in all human activity\nrecognition tasks.\n", "date": "2024-06-07", "categories": []}
{"id": "2407.01265", "title": "OSL-ActionSpotting: A Unified Library for Action Spotting in Sports\n  Videos", "abstract": "  Action spotting is crucial in sports analytics as it enables the precise\nidentification and categorization of pivotal moments in sports matches,\nproviding insights that are essential for performance analysis and tactical\ndecision-making. The fragmentation of existing methodologies, however, impedes\nthe progression of sports analytics, necessitating a unified codebase to\nsupport the development and deployment of action spotting for video analysis.\nIn this work, we introduce OSL-ActionSpotting, a Python library that unifies\ndifferent action spotting algorithms to streamline research and applications in\nsports video analytics. OSL-ActionSpotting encapsulates various\nstate-of-the-art techniques into a singular, user-friendly framework, offering\nstandardized processes for action spotting and analysis across multiple\ndatasets. We successfully integrated three cornerstone action spotting methods\ninto OSL-ActionSpotting, achieving performance metrics that match those of the\noriginal, disparate codebases. This unification within a single library\npreserves the effectiveness of each method and enhances usability and\naccessibility for researchers and practitioners in sports analytics. By\nbridging the gaps between various action spotting techniques,\nOSL-ActionSpotting significantly contributes to the field of sports video\nanalysis, fostering enhanced analytical capabilities and collaborative research\nopportunities. The scalable and modularized design of the library ensures its\nlong-term relevance and adaptability to future technological advancements in\nthe domain.\n", "date": "2024-07-01", "categories": []}
{"id": "2407.01397", "title": "Mask and Compress: Efficient Skeleton-based Action Recognition in\n  Continual Learning", "abstract": "  The use of skeletal data allows deep learning models to perform action\nrecognition efficiently and effectively. Herein, we believe that exploring this\nproblem within the context of Continual Learning is crucial. While numerous\nstudies focus on skeleton-based action recognition from a traditional offline\nperspective, only a handful venture into online approaches. In this respect, we\nintroduce CHARON (Continual Human Action Recognition On skeletoNs), which\nmaintains consistent performance while operating within an efficient framework.\nThrough techniques like uniform sampling, interpolation, and a memory-efficient\ntraining stage based on masking, we achieve improved recognition accuracy while\nminimizing computational overhead. Our experiments on Split NTU-60 and the\nproposed Split NTU-120 datasets demonstrate that CHARON sets a new benchmark in\nthis domain. The code is available at https://github.com/Sperimental3/CHARON.\n", "date": "2024-07-01", "categories": []}
{"id": "2406.18011", "title": "Expressive Keypoints for Skeleton-based Action Recognition via Skeleton\n  Transformation", "abstract": "  In the realm of skeleton-based action recognition, the traditional methods\nwhich rely on coarse body keypoints fall short of capturing subtle human\nactions. In this work, we propose Expressive Keypoints that incorporates hand\nand foot details to form a fine-grained skeletal representation, improving the\ndiscriminative ability for existing models in discerning intricate actions. To\nefficiently model Expressive Keypoints, the Skeleton Transformation strategy is\npresented to gradually downsample the keypoints and prioritize prominent joints\nby allocating the importance weights. Additionally, a plug-and-play Instance\nPooling module is exploited to extend our approach to multi-person scenarios\nwithout surging computation costs. Extensive experimental results over seven\ndatasets present the superiority of our method compared to the state-of-the-art\nfor skeleton-based human action recognition. Code is available at\nhttps://github.com/YijieYang23/SkeleT-GCN.\n", "date": "2024-06-26", "categories": []}
{"id": "2406.16886", "title": "Sensor Data Augmentation from Skeleton Pose Sequences for Improving\n  Human Activity Recognition", "abstract": "  The proliferation of deep learning has significantly advanced various fields,\nyet Human Activity Recognition (HAR) has not fully capitalized on these\ndevelopments, primarily due to the scarcity of labeled datasets. Despite the\nintegration of advanced Inertial Measurement Units (IMUs) in ubiquitous\nwearable devices like smartwatches and fitness trackers, which offer\nself-labeled activity data from users, the volume of labeled data remains\ninsufficient compared to domains where deep learning has achieved remarkable\nsuccess. Addressing this gap, in this paper, we propose a novel approach to\nimprove wearable sensor-based HAR by introducing a pose-to-sensor network model\nthat generates sensor data directly from 3D skeleton pose sequences. our method\nsimultaneously trains the pose-to-sensor network and a human activity\nclassifier, optimizing both data reconstruction and activity recognition. Our\ncontributions include the integration of simultaneous training, direct\npose-to-sensor generation, and a comprehensive evaluation on the MM-Fit\ndataset. Experimental results demonstrate the superiority of our framework with\nsignificant performance improvements over baseline methods.\n", "date": "2024-04-25", "categories": []}
{"id": "2406.17443", "title": "Using joint angles based on the international biomechanical standards\n  for human action recognition and related tasks", "abstract": "  Keypoint data has received a considerable amount of attention in machine\nlearning for tasks like action detection and recognition. However, human\nexperts in movement such as doctors, physiotherapists, sports scientists and\ncoaches use a notion of joint angles standardised by the International Society\nof Biomechanics to precisely and efficiently communicate static body poses and\nmovements. In this paper, we introduce the basic biomechanical notions and show\nhow they can be used to convert common keypoint data into joint angles that\nuniquely describe the given pose and have various desirable mathematical\nproperties, such as independence of both the camera viewpoint and the person\nperforming the action. We experimentally demonstrate that the joint angle\nrepresentation of keypoint data is suitable for machine learning applications\nand can in some cases bring an immediate performance gain. The use of joint\nangles as a human meaningful representation of kinematic data is in particular\npromising for applications where interpretability and dialog with human experts\nis important, such as many sports and medical applications. To facilitate\nfurther research in this direction, we will release a python package to convert\nkeypoint data into joint angles as outlined in this paper.\n", "date": "2024-06-25", "categories": []}
{"id": "2406.15556", "title": "Open-Vocabulary Temporal Action Localization using Multimodal Guidance", "abstract": "  Open-Vocabulary Temporal Action Localization (OVTAL) enables a model to\nrecognize any desired action category in videos without the need to explicitly\ncurate training data for all categories. However, this flexibility poses\nsignificant challenges, as the model must recognize not only the action\ncategories seen during training but also novel categories specified at\ninference. Unlike standard temporal action localization, where training and\ntest categories are predetermined, OVTAL requires understanding contextual cues\nthat reveal the semantics of novel categories. To address these challenges, we\nintroduce OVFormer, a novel open-vocabulary framework extending ActionFormer\nwith three key contributions. First, we employ task-specific prompts as input\nto a large language model to obtain rich class-specific descriptions for action\ncategories. Second, we introduce a cross-attention mechanism to learn the\nalignment between class representations and frame-level video features,\nfacilitating the multimodal guided features. Third, we propose a two-stage\ntraining strategy which includes training with a larger vocabulary dataset and\nfinetuning to downstream data to generalize to novel categories. OVFormer\nextends existing TAL methods to open-vocabulary settings. Comprehensive\nevaluations on the THUMOS14 and ActivityNet-1.3 benchmarks demonstrate the\neffectiveness of our method. Code and pretrained models will be publicly\nreleased.\n", "date": "2024-06-21", "categories": []}
{"id": "2406.16638", "title": "Feature Fusion for Human Activity Recognition using Parameter-Optimized\n  Multi-Stage Graph Convolutional Network and Transformer Models", "abstract": "  Human activity recognition (HAR) is a crucial area of research that involves\nunderstanding human movements using computer and machine vision technology.\nDeep learning has emerged as a powerful tool for this task, with models such as\nConvolutional Neural Networks (CNNs) and Transformers being employed to capture\nvarious aspects of human motion. One of the key contributions of this work is\nthe demonstration of the effectiveness of feature fusion in improving HAR\naccuracy by capturing spatial and temporal features, which has important\nimplications for the development of more accurate and robust activity\nrecognition systems. The study uses sensory data from HuGaDB, PKU-MMD, LARa,\nand TUG datasets. Two model, the PO-MS-GCN and a Transformer were trained and\nevaluated, with PO-MS-GCN outperforming state-of-the-art models. HuGaDB and TUG\nachieved high accuracies and f1-scores, while LARa and PKU-MMD had lower\nscores. Feature fusion improved results across datasets.\n", "date": "2024-06-24", "categories": []}
{"id": "2206.02909", "title": "Self-supervised Learning for Human Activity Recognition Using 700,000\n  Person-days of Wearable Data", "abstract": "  Advances in deep learning for human activity recognition have been relatively\nlimited due to the lack of large labelled datasets. In this study, we leverage\nself-supervised learning techniques on the UK-Biobank activity tracker\ndataset--the largest of its kind to date--containing more than 700,000\nperson-days of unlabelled wearable sensor data. Our resulting activity\nrecognition model consistently outperformed strong baselines across seven\nbenchmark datasets, with an F1 relative improvement of 2.5%-100% (median\n18.4%), the largest improvements occurring in the smaller datasets. In contrast\nto previous studies, our results generalise across external datasets, devices,\nand environments. Our open-source model will help researchers and developers to\nbuild customisable and generalisable activity classifiers with high\nperformance.\n", "date": "2022-06-06", "categories": []}
{"id": "2309.06219", "title": "Human Action Co-occurrence in Lifestyle Vlogs using Graph Link\n  Prediction", "abstract": "  We introduce the task of automatic human action co-occurrence identification,\ni.e., determine whether two human actions can co-occur in the same interval of\ntime. We create and make publicly available the ACE (Action Co-occurrencE)\ndataset, consisting of a large graph of ~12k co-occurring pairs of visual\nactions and their corresponding video clips. We describe graph link prediction\nmodels that leverage visual and textual information to automatically infer if\ntwo actions are co-occurring. We show that graphs are particularly well suited\nto capture relations between human actions, and the learned graph\nrepresentations are effective for our task and capture novel and relevant\ninformation across different data domains. The ACE dataset and the code\nintroduced in this paper are publicly available at\nhttps://github.com/MichiganNLP/vlog_action_co-occurrence.\n", "date": "2023-09-12", "categories": []}
{"id": "2406.13327", "title": "Part-aware Unified Representation of Language and Skeleton for Zero-shot\n  Action Recognition", "abstract": "  While remarkable progress has been made on supervised skeleton-based action\nrecognition, the challenge of zero-shot recognition remains relatively\nunexplored. In this paper, we argue that relying solely on aligning label-level\nsemantics and global skeleton features is insufficient to effectively transfer\nlocally consistent visual knowledge from seen to unseen classes. To address\nthis limitation, we introduce Part-aware Unified Representation between\nLanguage and Skeleton (PURLS) to explore visual-semantic alignment at both\nlocal and global scales. PURLS introduces a new prompting module and a novel\npartitioning module to generate aligned textual and visual representations\nacross different levels. The former leverages a pre-trained GPT-3 to infer\nrefined descriptions of the global and local (body-part-based and\ntemporal-interval-based) movements from the original action labels. The latter\nemploys an adaptive sampling strategy to group visual features from all body\njoint movements that are semantically relevant to a given description. Our\napproach is evaluated on various skeleton/language backbones and three\nlarge-scale datasets, i.e., NTU-RGB+D 60, NTU-RGB+D 120, and a newly curated\ndataset Kinetics-skeleton 200. The results showcase the universality and\nsuperior performance of PURLS, surpassing prior skeleton-based solutions and\nstandard baselines from other domains. The source codes can be accessed at\nhttps://github.com/azzh1/PURLS.\n", "date": "2024-06-19", "categories": []}
{"id": "2406.14446", "title": "Maintenance Required: Updating and Extending Bootstrapped Human Activity\n  Recognition Systems for Smart Homes", "abstract": "  Developing human activity recognition (HAR) systems for smart homes is not\nstraightforward due to varied layouts of the homes and their personalized\nsettings, as well as idiosyncratic behaviors of residents. As such,\noff-the-shelf HAR systems are effective in limited capacity for an individual\nhome, and HAR systems often need to be derived \"from scratch\", which comes with\nsubstantial efforts and often is burdensome to the resident. Previous work has\nsuccessfully targeted the initial phase. At the end of this initial phase, we\nidentify seed points. We build on bootstrapped HAR systems and introduce an\neffective updating and extension procedure for continuous improvement of HAR\nsystems with the aim of keeping up with ever changing life circumstances. Our\nmethod makes use of the seed points identified at the end of the initial\nbootstrapping phase. A contrastive learning framework is trained using these\nseed points and labels obtained for the same. This model is then used to\nimprove the segmentation accuracy of the identified prominent activities.\nImprovements in the activity recognition system through this procedure help\nmodel the majority of the routine activities in the smart home. We demonstrate\nthe effectiveness of our procedure through experiments on the CASAS datasets\nthat show the practical value of our approach.\n", "date": "2024-06-20", "categories": []}
{"id": "2406.14472", "title": "Self-supervised Multi-actor Social Activity Understanding in Streaming\n  Videos", "abstract": "  This work addresses the problem of Social Activity Recognition (SAR), a\ncritical component in real-world tasks like surveillance and assistive\nrobotics. Unlike traditional event understanding approaches, SAR necessitates\nmodeling individual actors' appearance and motions and contextualizing them\nwithin their social interactions. Traditional action localization methods fall\nshort due to their single-actor, single-action assumption. Previous SAR\nresearch has relied heavily on densely annotated data, but privacy concerns\nlimit their applicability in real-world settings. In this work, we propose a\nself-supervised approach based on multi-actor predictive learning for SAR in\nstreaming videos. Using a visual-semantic graph structure, we model social\ninteractions, enabling relational reasoning for robust performance with minimal\nlabeled data. The proposed framework achieves competitive performance on\nstandard group activity recognition benchmarks. Evaluation on three publicly\navailable action localization benchmarks demonstrates its generalizability to\narbitrary action localization.\n", "date": "2024-06-20", "categories": []}
{"id": "2406.12440", "title": "Deep self-supervised learning with visualisation for automatic gesture\n  recognition", "abstract": "  Gesture is an important mean of non-verbal communication, with visual\nmodality allows human to convey information during interaction, facilitating\npeoples and human-machine interactions. However, it is considered difficult to\nautomatically recognise gestures. In this work, we explore three different\nmeans to recognise hand signs using deep learning: supervised learning based\nmethods, self-supervised methods and visualisation based techniques applied to\n3D moving skeleton data. Self-supervised learning used to train fully\nconnected, CNN and LSTM method. Then, reconstruction method is applied to\nunlabelled data in simulated settings using CNN as a backbone where we use the\nlearnt features to perform the prediction in the remaining labelled data.\nLastly, Grad-CAM is applied to discover the focus of the models. Our\nexperiments results show that supervised learning method is capable to\nrecognise gesture accurately, with self-supervised learning increasing the\naccuracy in simulated settings. Finally, Grad-CAM visualisation shows that\nindeed the models focus on relevant skeleton joints on the associated gesture.\n", "date": "2024-06-18", "categories": []}
{"id": "2406.07191", "title": "MeMSVD: Long-Range Temporal Structure Capturing Using Incremental SVD", "abstract": "  This paper is on long-term video understanding where the goal is to recognise\nhuman actions over long temporal windows (up to minutes long). In prior work,\nlong temporal context is captured by constructing a long-term memory bank\nconsisting of past and future video features which are then integrated into\nstandard (short-term) video recognition backbones through the use of attention\nmechanisms. Two well-known problems related to this approach are the quadratic\ncomplexity of the attention operation and the fact that the whole feature bank\nmust be stored in memory for inference. To address both issues, we propose an\nalternative to attention-based schemes which is based on a low-rank\napproximation of the memory obtained using Singular Value Decomposition. Our\nscheme has two advantages: (a) it reduces complexity by more than an order of\nmagnitude, and (b) it is amenable to an efficient implementation for the\ncalculation of the memory bases in an incremental fashion which does not\nrequire the storage of the whole feature bank in memory. The proposed scheme\nmatches or surpasses the accuracy achieved by attention-based mechanisms while\nbeing memory-efficient. Through extensive experiments, we demonstrate that our\nframework generalises to different architectures and tasks, outperforming the\nstate-of-the-art in three datasets.\n", "date": "2024-06-11", "categories": []}
{"id": "2406.05722", "title": "ALGO: Object-Grounded Visual Commonsense Reasoning for Open-World\n  Egocentric Action Recognition", "abstract": "  Learning to infer labels in an open world, i.e., in an environment where the\ntarget \"labels\" are unknown, is an important characteristic for achieving\nautonomy. Foundation models pre-trained on enormous amounts of data have shown\nremarkable generalization skills through prompting, particularly in zero-shot\ninference. However, their performance is restricted to the correctness of the\ntarget label's search space. In an open world, this target search space can be\nunknown or exceptionally large, which severely restricts the performance of\nsuch models. To tackle this challenging problem, we propose a neuro-symbolic\nframework called ALGO - Action Learning with Grounded Object recognition that\nuses symbolic knowledge stored in large-scale knowledge bases to infer\nactivities in egocentric videos with limited supervision using two steps.\nFirst, we propose a neuro-symbolic prompting approach that uses object-centric\nvision-language models as a noisy oracle to ground objects in the video through\nevidence-based reasoning. Second, driven by prior commonsense knowledge, we\ndiscover plausible activities through an energy-based symbolic pattern theory\nframework and learn to ground knowledge-based action (verb) concepts in the\nvideo. Extensive experiments on four publicly available datasets\n(EPIC-Kitchens, GTEA Gaze, GTEA Gaze Plus) demonstrate its performance on\nopen-world activity inference.\n", "date": "2024-06-09", "categories": []}
{"id": "2406.06187", "title": "An Effective-Efficient Approach for Dense Multi-Label Action Detection", "abstract": "  Unlike the sparse label action detection task, where a single action occurs\nin each timestamp of a video, in a dense multi-label scenario, actions can\noverlap. To address this challenging task, it is necessary to simultaneously\nlearn (i) temporal dependencies and (ii) co-occurrence action relationships.\nRecent approaches model temporal information by extracting multi-scale features\nthrough hierarchical transformer-based networks. However, the self-attention\nmechanism in transformers inherently loses temporal positional information. We\nargue that combining this with multiple sub-sampling processes in hierarchical\ndesigns can lead to further loss of positional information. Preserving this\ninformation is essential for accurate action detection. In this paper, we\naddress this issue by proposing a novel transformer-based network that (a)\nemploys a non-hierarchical structure when modelling different ranges of\ntemporal dependencies and (b) embeds relative positional encoding in its\ntransformer layers. Furthermore, to model co-occurrence action relationships,\ncurrent methods explicitly embed class relations into the transformer network.\nHowever, these approaches are not computationally efficient, as the network\nneeds to compute all possible pair action class relations. We also overcome\nthis challenge by introducing a novel learning paradigm that allows the network\nto benefit from explicitly modelling temporal co-occurrence action dependencies\nwithout imposing their additional computational costs during inference. We\nevaluate the performance of our proposed approach on two challenging dense\nmulti-label benchmark datasets and show that our method improves the current\nstate-of-the-art results.\n", "date": "2024-06-10", "categories": []}
{"id": "2310.06403", "title": "Boundary Discretization and Reliable Classification Network for Temporal\n  Action Detection", "abstract": "  Temporal action detection aims to recognize the action category and determine\neach action instance's starting and ending time in untrimmed videos. The mixed\nmethods have achieved remarkable performance by seamlessly merging anchor-based\nand anchor-free approaches. Nonetheless, there are still two crucial issues\nwithin the mixed framework: (1) Brute-force merging and handcrafted anchor\ndesign hinder the substantial potential and practicality of the mixed methods.\n(2) Within-category predictions show a significant abundance of false\npositives. In this paper, we propose a novel Boundary Discretization and\nReliable Classification Network (BDRC-Net) that addresses the issues above by\nintroducing boundary discretization and reliable classification modules.\nSpecifically, the boundary discretization module (BDM) elegantly merges\nanchor-based and anchor-free approaches in the form of boundary discretization,\neliminating the need for the traditional handcrafted anchor design.\nFurthermore, the reliable classification module (RCM) predicts reliable global\naction categories to reduce false positives. Extensive experiments conducted on\ndifferent benchmarks demonstrate that our proposed method achieves competitive\ndetection performance. The code will be released at\nhttps://github.com/zhenyingfang/BDRC-Net.\n", "date": "2023-10-10", "categories": []}
{"id": "2406.04649", "title": "SMART: Scene-motion-aware human action recognition framework for mental\n  disorder group", "abstract": "  Patients with mental disorders often exhibit risky abnormal actions, such as\nclimbing walls or hitting windows, necessitating intelligent video behavior\nmonitoring for smart healthcare with the rising Internet of Things (IoT)\ntechnology. However, the development of vision-based Human Action Recognition\n(HAR) for these actions is hindered by the lack of specialized algorithms and\ndatasets. In this paper, we innovatively propose to build a vision-based HAR\ndataset including abnormal actions often occurring in the mental disorder group\nand then introduce a novel Scene-Motion-aware Action Recognition Technology\nframework, named SMART, consisting of two technical modules. First, we propose\na scene perception module to extract human motion trajectory and human-scene\ninteraction features, which introduces additional scene information for a\nsupplementary semantic representation of the above actions. Second, the\nmulti-stage fusion module fuses the skeleton motion, motion trajectory, and\nhuman-scene interaction features, enhancing the semantic association between\nthe skeleton motion and the above supplementary representation, thus generating\na comprehensive representation with both human motion and scene information.\nThe effectiveness of our proposed method has been validated on our\nself-collected HAR dataset (MentalHAD), achieving 94.9% and 93.1% accuracy in\nun-seen subjects and scenes and outperforming state-of-the-art approaches by\n6.5% and 13.2%, respectively. The demonstrated subject- and scene-\ngeneralizability makes it possible for SMART's migration to practical\ndeployment in smart healthcare systems for mental disorder patients in medical\nsettings. The code and dataset will be released publicly for further research:\nhttps://github.com/Inowlzy/SMART.git.\n", "date": "2024-06-07", "categories": []}
{"id": "2310.13585", "title": "POTLoc: Pseudo-Label Oriented Transformer for Point-Supervised Temporal\n  Action Localization", "abstract": "  This paper tackles the challenge of point-supervised temporal action\ndetection, wherein only a single frame is annotated for each action instance in\nthe training set. Most of the current methods, hindered by the sparse nature of\nannotated points, struggle to effectively represent the continuous structure of\nactions or the inherent temporal and semantic dependencies within action\ninstances. Consequently, these methods frequently learn merely the most\ndistinctive segments of actions, leading to the creation of incomplete action\nproposals. This paper proposes POTLoc, a Pseudo-label Oriented Transformer for\nweakly-supervised Action Localization utilizing only point-level annotation.\nPOTLoc is designed to identify and track continuous action structures via a\nself-training strategy. The base model begins by generating action proposals\nsolely with point-level supervision. These proposals undergo refinement and\nregression to enhance the precision of the estimated action boundaries, which\nsubsequently results in the production of `pseudo-labels' to serve as\nsupplementary supervisory signals. The architecture of the model integrates a\ntransformer with a temporal feature pyramid to capture video snippet\ndependencies and model actions of varying duration. The pseudo-labels,\nproviding information about the coarse locations and boundaries of actions,\nassist in guiding the transformer for enhanced learning of action dynamics.\nPOTLoc outperforms the state-of-the-art point-supervised methods on THUMOS'14\nand ActivityNet-v1.2 datasets.\n", "date": "2023-10-20", "categories": []}
{"id": "2402.12706", "title": "Learning Causal Domain-Invariant Temporal Dynamics for Few-Shot Action\n  Recognition", "abstract": "  Few-shot action recognition aims at quickly adapting a pre-trained model to\nthe novel data with a distribution shift using only a limited number of\nsamples. Key challenges include how to identify and leverage the transferable\nknowledge learned by the pre-trained model. We therefore propose CDTD, or\nCausal Domain-Invariant Temporal Dynamics for knowledge transfer. To identify\nthe temporally invariant and variant representations, we employ the causal\nrepresentation learning methods for unsupervised pertaining, and then tune the\nclassifier with supervisions in next stage. Specifically, we assume the domain\ninformation can be well estimated and the pre-trained image decoder and\ntransition models can be well transferred. During adaptation, we fix the\ntransferable temporal dynamics and update the image encoder and domain\nestimator. The efficacy of our approach is revealed by the superior accuracy of\nCDTD over leading alternatives across standard few-shot action recognition\ndatasets.\n", "date": "2024-02-20", "categories": []}
{"id": "2406.02468", "title": "DL-KDD: Dual-Light Knowledge Distillation for Action Recognition in the\n  Dark", "abstract": "  Human action recognition in dark videos is a challenging task for computer\nvision. Recent research focuses on applying dark enhancement methods to improve\nthe visibility of the video. However, such video processing results in the loss\nof critical information in the original (un-enhanced) video. Conversely,\ntraditional two-stream methods are capable of learning information from both\noriginal and processed videos, but it can lead to a significant increase in the\ncomputational cost during the inference phase in the task of video\nclassification. To address these challenges, we propose a novel teacher-student\nvideo classification framework, named Dual-Light KnowleDge Distillation for\nAction Recognition in the Dark (DL-KDD). This framework enables the model to\nlearn from both original and enhanced video without introducing additional\ncomputational cost during inference. Specifically, DL-KDD utilizes the strategy\nof knowledge distillation during training. The teacher model is trained with\nenhanced video, and the student model is trained with both the original video\nand the soft target generated by the teacher model. This teacher-student\nframework allows the student model to predict action using only the original\ninput video during inference. In our experiments, the proposed DL-KDD framework\noutperforms state-of-the-art methods on the ARID, ARID V1.5, and Dark-48\ndatasets. We achieve the best performance on each dataset and up to a 4.18%\nimprovement on Dark-48, using only original video inputs, thus avoiding the use\nof two-stream framework or enhancement modules for inference. We further\nvalidate the effectiveness of the distillation strategy in ablative\nexperiments. The results highlight the advantages of our knowledge distillation\nframework in dark human action recognition.\n", "date": "2024-06-04", "categories": []}
{"id": "2403.05234", "title": "Benchmarking Micro-action Recognition: Dataset, Methods, and\n  Applications", "abstract": "  Micro-action is an imperceptible non-verbal behaviour characterised by\nlow-intensity movement. It offers insights into the feelings and intentions of\nindividuals and is important for human-oriented applications such as emotion\nrecognition and psychological assessment. However, the identification,\ndifferentiation, and understanding of micro-actions pose challenges due to the\nimperceptible and inaccessible nature of these subtle human behaviors in\neveryday life. In this study, we innovatively collect a new micro-action\ndataset designated as Micro-action-52 (MA-52), and propose a benchmark named\nmicro-action network (MANet) for micro-action recognition (MAR) task. Uniquely,\nMA-52 provides the whole-body perspective including gestures, upper- and\nlower-limb movements, attempting to reveal comprehensive micro-action cues. In\ndetail, MA-52 contains 52 micro-action categories along with seven body part\nlabels, and encompasses a full array of realistic and natural micro-actions,\naccounting for 205 participants and 22,422 video instances collated from the\npsychological interviews. Based on the proposed dataset, we assess MANet and\nother nine prevalent action recognition methods. MANet incorporates squeeze-and\nexcitation (SE) and temporal shift module (TSM) into the ResNet architecture\nfor modeling the spatiotemporal characteristics of micro-actions. Then a\njoint-embedding loss is designed for semantic matching between video and action\nlabels; the loss is used to better distinguish between visually similar yet\ndistinct micro-action categories. The extended application in emotion\nrecognition has demonstrated one of the important values of our proposed\ndataset and method. In the future, further exploration of human behaviour,\nemotion, and psychological assessment will be conducted in depth. The dataset\nand source code are released at https://github.com/VUT-HFUT/Micro-Action.\n", "date": "2024-03-08", "categories": []}
{"id": "2406.00639", "title": "An Information Compensation Framework for Zero-Shot Skeleton-based\n  Action Recognition", "abstract": "  Zero-shot human skeleton-based action recognition aims to construct a model\nthat can recognize actions outside the categories seen during training.\nPrevious research has focused on aligning sequences' visual and semantic\nspatial distributions. However, these methods extract semantic features simply.\nThey ignore that proper prompt design for rich and fine-grained action cues can\nprovide robust representation space clustering. In order to alleviate the\nproblem of insufficient information available for skeleton sequences, we design\nan information compensation learning framework from an information-theoretic\nperspective to improve zero-shot action recognition accuracy with a\nmulti-granularity semantic interaction mechanism. Inspired by ensemble\nlearning, we propose a multi-level alignment (MLA) approach to compensate\ninformation for action classes. MLA aligns multi-granularity embeddings with\nvisual embedding through a multi-head scoring mechanism to distinguish\nsemantically similar action names and visually similar actions. Furthermore, we\nintroduce a new loss function sampling method to obtain a tight and robust\nrepresentation. Finally, these multi-granularity semantic embeddings are\nsynthesized to form a proper decision surface for classification. Significant\naction recognition performance is achieved when evaluated on the challenging\nNTU RGB+D, NTU RGB+D 120, and PKU-MMD benchmarks and validate that\nmulti-granularity semantic features facilitate the differentiation of action\nclusters with similar visual features.\n", "date": "2024-06-02", "categories": []}
{"id": "2406.01073", "title": "Understanding the Cross-Domain Capabilities of Video-Based Few-Shot\n  Action Recognition Models", "abstract": "  Few-shot action recognition (FSAR) aims to learn a model capable of\nidentifying novel actions in videos using only a few examples. In assuming the\nbase dataset seen during meta-training and novel dataset used for evaluation\ncan come from different domains, cross-domain few-shot learning alleviates data\ncollection and annotation costs required by methods with greater supervision\nand conventional (single-domain) few-shot methods. While this form of learning\nhas been extensively studied for image classification, studies in cross-domain\nFSAR (CD-FSAR) are limited to proposing a model, rather than first\nunderstanding the cross-domain capabilities of existing models. To this end, we\nsystematically evaluate existing state-of-the-art single-domain,\ntransfer-based, and cross-domain FSAR methods on new cross-domain tasks with\nincreasing difficulty, measured based on the domain shift between the base and\nnovel set. Our empirical meta-analysis reveals a correlation between domain\ndifference and downstream few-shot performance, and uncovers several important\ninsights into which model aspects are effective for CD-FSAR and which need\nfurther development. Namely, we find that as the domain difference increases,\nthe simple transfer-learning approach outperforms other methods by over 12\npercentage points, and under these more challenging cross-domain settings, the\nspecialised cross-domain model achieves the lowest performance. We also witness\nstate-of-the-art single-domain FSAR models which use temporal alignment\nachieving similar or worse performance than earlier methods which do not,\nsuggesting existing temporal alignment techniques fail to generalise on unseen\ndomains. To the best of our knowledge, we are the first to systematically study\nthe CD-FSAR problem in-depth. We hope the insights and challenges revealed in\nour study inspires and informs future work in these directions.\n", "date": "2024-06-03", "categories": []}
{"id": "2406.01079", "title": "Object Aware Egocentric Online Action Detection", "abstract": "  Advancements in egocentric video datasets like Ego4D, EPIC-Kitchens, and\nEgo-Exo4D have enriched the study of first-person human interactions, which is\ncrucial for applications in augmented reality and assisted living. Despite\nthese advancements, current Online Action Detection methods, which efficiently\ndetect actions in streaming videos, are predominantly designed for exocentric\nviews and thus fail to capitalize on the unique perspectives inherent to\negocentric videos. To address this gap, we introduce an Object-Aware Module\nthat integrates egocentric-specific priors into existing OAD frameworks,\nenhancing first-person footage interpretation. Utilizing object-specific\ndetails and temporal dynamics, our module improves scene understanding in\ndetecting actions. Validated extensively on the Epic-Kitchens 100 dataset, our\nwork can be seamlessly integrated into existing models with minimal overhead\nand bring consistent performance enhancements, marking an important step\nforward in adapting action detection systems to egocentric video analysis.\n", "date": "2024-06-03", "categories": []}
{"id": "2405.20892", "title": "MALT: Multi-scale Action Learning Transformer for Online Action\n  Detection", "abstract": "  Online action detection (OAD) aims to identify ongoing actions from streaming\nvideo in real-time, without access to future frames. Since these actions\nmanifest at varying scales of granularity, ranging from coarse to fine,\nprojecting an entire set of action frames to a single latent encoding may\nresult in a lack of local information, necessitating the acquisition of action\nfeatures across multiple scales. In this paper, we propose a multi-scale action\nlearning transformer (MALT), which includes a novel recurrent decoder (used for\nfeature fusion) that includes fewer parameters and can be trained more\nefficiently. A hierarchical encoder with multiple encoding branches is further\nproposed to capture multi-scale action features. The output from the preceding\nbranch is then incrementally input to the subsequent branch as part of a\ncross-attention calculation. In this way, output features transition from\ncoarse to fine as the branches deepen. We also introduce an explicit frame\nscoring mechanism employing sparse attention, which filters irrelevant frames\nmore efficiently, without requiring an additional network. The proposed method\nachieved state-of-the-art performance on two benchmark datasets (THUMOS'14 and\nTVSeries), outperforming all existing models used for comparison, with an mAP\nof 0.2% for THUMOS'14 and an mcAP of 0.1% for TVseries.\n", "date": "2024-05-31", "categories": []}
{"id": "2306.17574", "title": "SpATr: MoCap 3D Human Action Recognition based on Spiral Auto-encoder\n  and Transformer Network", "abstract": "  Recent technological advancements have significantly expanded the potential\nof human action recognition through harnessing the power of 3D data. This data\nprovides a richer understanding of actions, including depth information that\nenables more accurate analysis of spatial and temporal characteristics. In this\ncontext, We study the challenge of 3D human action recognition.Unlike prior\nmethods, that rely on sampling 2D depth images, skeleton points, or point\nclouds, often leading to substantial memory requirements and the ability to\nhandle only short sequences, we introduce a novel approach for 3D human action\nrecognition, denoted as SpATr (Spiral Auto-encoder and Transformer Network),\nspecifically designed for fixed-topology mesh sequences. The SpATr model\ndisentangles space and time in the mesh sequences. A lightweight auto-encoder,\nbased on spiral convolutions, is employed to extract spatial geometrical\nfeatures from each 3D mesh. These convolutions are lightweight and specifically\ndesigned for fix-topology mesh data. Subsequently, a temporal transformer,\nbased on self-attention, captures the temporal context within the feature\nsequence. The self-attention mechanism enables long-range dependencies\ncapturing and parallel processing, ensuring scalability for long sequences. The\nproposed method is evaluated on three prominent 3D human action datasets:\nBabel, MoVi, and BMLrub, from the Archive of Motion Capture As Surface Shapes\n(AMASS). Our results analysis demonstrates the competitive performance of our\nSpATr model in 3D human action recognition while maintaining efficient memory\nusage. The code and the training results will soon be made publicly available\nat https://github.com/h-bouzid/spatr.\n", "date": "2023-06-30", "categories": []}
{"id": "2405.17729", "title": "Hierarchical Action Recognition: A Contrastive Video-Language Approach\n  with Hierarchical Interactions", "abstract": "  Video recognition remains an open challenge, requiring the identification of\ndiverse content categories within videos. Mainstream approaches often perform\nflat classification, overlooking the intrinsic hierarchical structure relating\ncategories. To address this, we formalize the novel task of hierarchical video\nrecognition, and propose a video-language learning framework tailored for\nhierarchical recognition. Specifically, our framework encodes dependencies\nbetween hierarchical category levels, and applies a top-down constraint to\nfilter recognition predictions. We further construct a new fine-grained dataset\nbased on medical assessments for rehabilitation of stroke patients, serving as\na challenging benchmark for hierarchical recognition. Through extensive\nexperiments, we demonstrate the efficacy of our approach for hierarchical\nrecognition, significantly outperforming conventional methods, especially for\nfine-grained subcategories. The proposed framework paves the way for\nhierarchical modeling in video understanding tasks, moving beyond flat\ncategorization.\n", "date": "2024-05-28", "categories": []}
{"id": "2405.17855", "title": "A Deep Neural Network Approach to Fare Evasion", "abstract": "  Fare evasion is a problem for public transport companies, with LSTM models\nthis issue can help companies get an analytical insight into where this issue\noccurs the most, to prevent capital loss. In addition to the financial burden\nthis problem causes, having more inspectors is not enough to alleviate the\nproblem. The purpose of this study is to find a different way to predict fare\nevasion in the public transport sector. Through the use of keypoint extractions\nof passengers in video footage, an LSTM model is trained on those keypoints to\nhelp predict the actions of passengers between payments and evasions. The\nresults were promising when it came to predicting the actions of passengers on\nreal-time footage. Thus a sophisticated approach can help to decrease the fare\nevasion problem. A ReID model can be used alongside the LSTM model for better\naccuracy, as there is always the chance that a person might only pay for the\nfare at a later stage. With both models, it is possible for public transport\ncompanies to start narrowing down where the root of their fare evasion problems\nemerges.\n", "date": "2024-05-28", "categories": []}
{"id": "2405.18012", "title": "Flow-Assisted Motion Learning Network for Weakly-Supervised Group\n  Activity Recognition", "abstract": "  Weakly-Supervised Group Activity Recognition (WSGAR) aims to understand the\nactivity performed together by a group of individuals with the video-level\nlabel and without actor-level labels. We propose Flow-Assisted Motion Learning\nNetwork (Flaming-Net) for WSGAR, which consists of the motion-aware actor\nencoder to extract actor features and the two-pathways relation module to infer\nthe interaction among actors and their activity. Flaming-Net leverages an\nadditional optical flow modality in the training stage to enhance its motion\nawareness when finding locally active actors. The first pathway of the relation\nmodule, the actor-centric path, initially captures the temporal dynamics of\nindividual actors and then constructs inter-actor relationships. In parallel,\nthe group-centric path starts by building spatial connections between actors\nwithin the same timeframe and then captures simultaneous spatio-temporal\ndynamics among them. We demonstrate that Flaming-Net achieves new\nstate-of-the-art WSGAR results on two benchmarks, including a 2.8%p higher MPCA\nscore on the NBA dataset. Importantly, we use the optical flow modality only\nfor training and not for inference.\n", "date": "2024-05-28", "categories": []}
{"id": "2403.15212", "title": "GCN-DevLSTM: Path Development for Skeleton-Based Action Recognition", "abstract": "  Skeleton-based action recognition (SAR) in videos is an important but\nchallenging task in computer vision. The recent state-of-the-art (SOTA) models\nfor SAR are primarily based on graph convolutional neural networks (GCNs),\nwhich are powerful in extracting the spatial information of skeleton data.\nHowever, it is yet clear that such GCN-based models can effectively capture the\ntemporal dynamics of human action sequences. To this end, we propose the G-Dev\nlayer, which exploits the path development -- a principled and parsimonious\nrepresentation for sequential data by leveraging the Lie group structure. By\nintegrating the G-Dev layer, the hybrid G-DevLSTM module enhances the\ntraditional LSTM to reduce the time dimension while retaining high-frequency\ninformation. It can be conveniently applied to any temporal graph data,\ncomplementing existing advanced GCN-based models. Our empirical studies on the\nNTU60, NTU120 and Chalearn2013 datasets demonstrate that our proposed\nGCN-DevLSTM network consistently improves the strong GCN baseline models and\nachieves SOTA results with superior robustness in SAR tasks. The code is\navailable at https://github.com/DeepIntoStreams/GCN-DevLSTM.\n", "date": "2024-03-22", "categories": []}
{"id": "2405.15813", "title": "From CNNs to Transformers in Multimodal Human Action Recognition: A\n  Survey", "abstract": "  Due to its widespread applications, human action recognition is one of the\nmost widely studied research problems in Computer Vision. Recent studies have\nshown that addressing it using multimodal data leads to superior performance as\ncompared to relying on a single data modality. During the adoption of deep\nlearning for visual modelling in the last decade, action recognition approaches\nhave mainly relied on Convolutional Neural Networks (CNNs). However, the recent\nrise of Transformers in visual modelling is now also causing a paradigm shift\nfor the action recognition task. This survey captures this transition while\nfocusing on Multimodal Human Action Recognition (MHAR). Unique to the induction\nof multimodal computational models is the process of \"fusing\" the features of\nthe individual data modalities. Hence, we specifically focus on the fusion\ndesign aspects of the MHAR approaches. We analyze the classic and emerging\ntechniques in this regard, while also highlighting the popular trends in the\nadaption of CNN and Transformer building blocks for the overall problem. In\nparticular, we emphasize on recent design choices that have led to more\nefficient MHAR models. Unlike existing reviews, which discuss Human Action\nRecognition from a broad perspective, this survey is specifically aimed at\npushing the boundaries of MHAR research by identifying promising architectural\nand fusion design choices to train practicable models. We also provide an\noutlook of the multimodal datasets from their scale and evaluation viewpoint.\nFinally, building on the reviewed literature, we discuss the challenges and\nfuture avenues for MHAR.\n", "date": "2024-05-22", "categories": []}
{"id": "2405.15962", "title": "Wearable-based behaviour interpolation for semi-supervised human\n  activity recognition", "abstract": "  While traditional feature engineering for Human Activity Recognition (HAR)\ninvolves a trial-anderror process, deep learning has emerged as a preferred\nmethod for high-level representations of sensor-based human activities.\nHowever, most deep learning-based HAR requires a large amount of labelled data\nand extracting HAR features from unlabelled data for effective deep learning\ntraining remains challenging. We, therefore, introduce a deep semi-supervised\nHAR approach, MixHAR, which concurrently uses labelled and unlabelled\nactivities. Our MixHAR employs a linear interpolation mechanism to blend\nlabelled and unlabelled activities while addressing both inter- and\nintra-activity variability. A unique challenge identified is the\nactivityintrusion problem during mixing, for which we propose a mixing\ncalibration mechanism to mitigate it in the feature embedding space.\nAdditionally, we rigorously explored and evaluated the five\nconventional/popular deep semi-supervised technologies on HAR, acting as the\nbenchmark of deep semi-supervised HAR. Our results demonstrate that MixHAR\nsignificantly improves performance, underscoring the potential of deep\nsemi-supervised techniques in HAR.\n", "date": "2024-05-24", "categories": []}
{"id": "2405.15995", "title": "Efficient Temporal Action Segmentation via Boundary-aware Query Voting", "abstract": "  Although the performance of Temporal Action Segmentation (TAS) has improved\nin recent years, achieving promising results often comes with a high\ncomputational cost due to dense inputs, complex model structures, and\nresource-intensive post-processing requirements. To improve the efficiency\nwhile keeping the performance, we present a novel perspective centered on\nper-segment classification. By harnessing the capabilities of Transformers, we\ntokenize each video segment as an instance token, endowed with intrinsic\ninstance segmentation. To realize efficient action segmentation, we introduce\nBaFormer, a boundary-aware Transformer network. It employs instance queries for\ninstance segmentation and a global query for class-agnostic boundary\nprediction, yielding continuous segment proposals. During inference, BaFormer\nemploys a simple yet effective voting strategy to classify boundary-wise\nsegments based on instance segmentation. Remarkably, as a single-stage\napproach, BaFormer significantly reduces the computational costs, utilizing\nonly 6% of the running time compared to state-of-the-art method DiffAct, while\nproducing better or comparable accuracy over several popular benchmarks. The\ncode for this project is publicly available at\nhttps://github.com/peiyao-w/BaFormer.\n", "date": "2024-05-25", "categories": []}
{"id": "2405.17188", "title": "The SkatingVerse Workshop & Challenge: Methods and Results", "abstract": "  The SkatingVerse Workshop & Challenge aims to encourage research in\ndeveloping novel and accurate methods for human action understanding. The\nSkatingVerse dataset used for the SkatingVerse Challenge has been publicly\nreleased. There are two subsets in the dataset, i.e., the training subset and\ntesting subset. The training subsets consists of 19,993 RGB video sequences,\nand the testing subsets consists of 8,586 RGB video sequences. Around 10\nparticipating teams from the globe competed in the SkatingVerse Challenge. In\nthis paper, we provide a brief summary of the SkatingVerse Workshop & Challenge\nincluding brief introductions to the top three methods. The submission\nleaderboard will be reopened for researchers that are interested in the human\naction understanding challenge. The benchmark dataset and other information can\nbe found at: https://skatingverse.github.io/.\n", "date": "2024-05-27", "categories": []}
{"id": "2311.17118", "title": "Towards Weakly Supervised End-to-end Learning for Long-video Action\n  Recognition", "abstract": "  Developing end-to-end action recognition models on long videos is fundamental\nand crucial for long-video action understanding. Due to the unaffordable cost\nof end-to-end training on the whole long videos, existing works generally train\nmodels on short clips trimmed from long videos. However, this\n``trimming-then-training'' practice requires action interval annotations for\nclip-level supervision, i.e., knowing which actions are trimmed into the clips.\nUnfortunately, collecting such annotations is very expensive and prevents model\ntraining at scale. To this end, this work aims to build a weakly supervised\nend-to-end framework for training recognition models on long videos, with only\nvideo-level action category labels. Without knowing the precise temporal\nlocations of actions in long videos, our proposed weakly supervised framework,\nnamely AdaptFocus, estimates where and how likely the actions will occur to\nadaptively focus on informative action clips for end-to-end training. The\neffectiveness of the proposed AdaptFocus framework is demonstrated on three\nlong-video datasets. Furthermore, for downstream long-video tasks, our\nAdaptFocus framework provides a weakly supervised feature extraction pipeline\nfor extracting more robust long-video features, such that the state-of-the-art\nmethods on downstream tasks are significantly advanced. We will release the\ncode and models.\n", "date": "2023-11-28", "categories": []}
{"id": "2403.01560", "title": "Rethinking CLIP-based Video Learners in Cross-Domain Open-Vocabulary\n  Action Recognition", "abstract": "  Building upon the impressive success of CLIP (Contrastive Language-Image\nPretraining), recent pioneer works have proposed to adapt the powerful CLIP to\nvideo data, leading to efficient and effective video learners for\nopen-vocabulary action recognition. Inspired by that humans perform actions in\ndiverse environments, our work delves into an intriguing question: Can\nCLIP-based video learners effectively generalize to video domains they have not\nencountered during training? To answer this, we establish a CROSS-domain\nOpen-Vocabulary Action recognition benchmark named XOV-Action, and conduct a\ncomprehensive evaluation of five state-of-the-art CLIP-based video learners\nunder various types of domain gaps. The evaluation demonstrates that previous\nmethods exhibit limited action recognition performance in unseen video domains,\nrevealing potential challenges of the cross-domain open-vocabulary action\nrecognition task. In this paper, we focus on one critical challenge of the\ntask, namely scene bias, and accordingly contribute a novel scene-aware\nvideo-text alignment method. Our key idea is to distinguish video\nrepresentations apart from scene-encoded text representations, aiming to learn\nscene-agnostic video representations for recognizing actions across domains.\nExtensive experiments demonstrate the effectiveness of our method. The\nbenchmark and code will be available at\nhttps://github.com/KunyuLin/XOV-Action/.\n", "date": "2024-03-03", "categories": []}
{"id": "2309.15683", "title": "End-to-End Streaming Video Temporal Action Segmentation with Reinforce\n  Learning", "abstract": "  The streaming temporal action segmentation (STAS) task, a supplementary task\nof temporal action segmentation (TAS), has not received adequate attention in\nthe field of video understanding. Existing TAS methods are constrained to\noffline scenarios due to their heavy reliance on multimodal features and\ncomplete contextual information. The STAS task requires the model to classify\neach frame of the entire untrimmed video sequence clip by clip in time, thereby\nextending the applicability of TAS methods to online scenarios. However,\ndirectly applying existing TAS methods to SATS tasks results in significantly\npoor segmentation outcomes. In this paper, we thoroughly analyze the\nfundamental differences between STAS tasks and TAS tasks, attributing the\nsevere performance degradation when transferring models to model bias and\noptimization dilemmas. We introduce an end-to-end streaming video temporal\naction segmentation model with reinforcement learning (SVTAS-RL). The\nend-to-end modeling method mitigates the modeling bias introduced by the change\nin task nature and enhances the feasibility of online solutions. Reinforcement\nlearning is utilized to alleviate the optimization dilemma. Through extensive\nexperiments, the SVTAS-RL model significantly outperforms existing STAS models\nand achieves competitive performance to the state-of-the-art TAS model on\nmultiple datasets under the same evaluation criteria, demonstrating notable\nadvantages on the ultra-long video dataset EGTEA. Code is available at\nhttps://github.com/Thinksky5124/SVTAS.\n", "date": "2023-09-27", "categories": []}
{"id": "2312.17686", "title": "Multiscale Vision Transformers meet Bipartite Matching for efficient\n  single-stage Action Localization", "abstract": "  Action Localization is a challenging problem that combines detection and\nrecognition tasks, which are often addressed separately. State-of-the-art\nmethods rely on off-the-shelf bounding box detections pre-computed at high\nresolution, and propose transformer models that focus on the classification\ntask alone. Such two-stage solutions are prohibitive for real-time deployment.\nOn the other hand, single-stage methods target both tasks by devoting part of\nthe network (generally the backbone) to sharing the majority of the workload,\ncompromising performance for speed. These methods build on adding a DETR head\nwith learnable queries that after cross- and self-attention can be sent to\ncorresponding MLPs for detecting a person's bounding box and action. However,\nDETR-like architectures are challenging to train and can incur in big\ncomplexity.\n  In this paper, we observe that \\textbf{a straight bipartite matching loss can\nbe applied to the output tokens of a vision transformer}. This results in a\nbackbone + MLP architecture that can do both tasks without the need of an extra\nencoder-decoder head and learnable queries. We show that a single MViTv2-S\narchitecture trained with bipartite matching to perform both tasks surpasses\nthe same MViTv2-S when trained with RoI align on pre-computed bounding boxes.\nWith a careful design of token pooling and the proposed training pipeline, our\nBipartite-Matching Vision Transformer model, \\textbf{BMViT}, achieves +3 mAP on\nAVA2.2. w.r.t. the two-stage MViTv2-S counterpart. Code is available at\n\\href{https://github.com/IoannaNti/BMViT}{https://github.com/IoannaNti/BMViT}\n", "date": "2023-12-29", "categories": []}
{"id": "2212.04873", "title": "Multimodal Prototype-Enhanced Network for Few-Shot Action Recognition", "abstract": "  Current methods for few-shot action recognition mainly fall into the metric\nlearning framework following ProtoNet, which demonstrates the importance of\nprototypes. Although they achieve relatively good performance, the effect of\nmultimodal information is ignored, e.g. label texts. In this work, we propose a\nnovel MultimOdal PRototype-ENhanced Network (MORN), which uses the semantic\ninformation of label texts as multimodal information to enhance prototypes. A\nCLIP visual encoder and a frozen CLIP text encoder are introduced to obtain\nfeatures with good multimodal initialization. Then in the visual flow, visual\nprototypes are computed by a visual prototype-computed module. In the text\nflow, a semantic-enhanced (SE) module and an inflating operation are used to\nobtain text prototypes. The final multimodal prototypes are then computed by a\nmultimodal prototype-enhanced (MPE) module. Besides, we define a PRototype\nSImilarity DiffErence (PRIDE) to evaluate the quality of prototypes, which is\nused to verify our improvement on the prototype level and effectiveness of\nMORN. We conduct extensive experiments on four popular few-shot action\nrecognition datasets: HMDB51, UCF101, Kinetics and SSv2, and MORN achieves\nstate-of-the-art results. When plugging PRIDE into the training stage, the\nperformance can be further improved.\n", "date": "2022-12-09", "categories": []}
{"id": "2405.11511", "title": "Online Action Representation using Change Detection and Symbolic\n  Programming", "abstract": "  This paper addresses the critical need for online action representation,\nwhich is essential for various applications like rehabilitation, surveillance,\netc. The task can be defined as representation of actions as soon as they\nhappen in a streaming video without access to video frames in the future. Most\nof the existing methods use predefined window sizes for video segments, which\nis a restrictive assumption on the dynamics. The proposed method employs a\nchange detection algorithm to automatically segment action sequences, which\nform meaningful sub-actions and subsequently fit symbolic generative motion\nprograms to the clipped segments. We determine the start time and end time of\nsegments using change detection followed by a piece-wise linear fit algorithm\non joint angle and bone length sequences. Domain-specific symbolic primitives\nare fit to pose keypoint trajectories of those extracted segments in order to\nobtain a higher level semantic representation. Since this representation is\npart-based, it is complementary to the compositional nature of human actions,\ni.e., a complex activity can be broken down into elementary sub-actions. We\nshow the effectiveness of this representation in the downstream task of class\nagnostic repetition detection. We propose a repetition counting algorithm based\non consecutive similarity matching of primitives, which can do online\nrepetition counting. We also compare the results with a similar but offline\nrepetition counting algorithm. The results of the experiments demonstrate that,\ndespite operating online, the proposed method performs better or on par with\nthe existing method.\n", "date": "2024-05-19", "categories": []}
{"id": "2405.10832", "title": "Open-Vocabulary Spatio-Temporal Action Detection", "abstract": "  Spatio-temporal action detection (STAD) is an important fine-grained video\nunderstanding task. Current methods require box and label supervision for all\naction classes in advance. However, in real-world applications, it is very\nlikely to come across new action classes not seen in training because the\naction category space is large and hard to enumerate. Also, the cost of data\nannotation and model training for new classes is extremely high for traditional\nmethods, as we need to perform detailed box annotations and re-train the whole\nnetwork from scratch. In this paper, we propose a new challenging setting by\nperforming open-vocabulary STAD to better mimic the situation of action\ndetection in an open world. Open-vocabulary spatio-temporal action detection\n(OV-STAD) requires training a model on a limited set of base classes with box\nand label supervision, which is expected to yield good generalization\nperformance on novel action classes. For OV-STAD, we build two benchmarks based\non the existing STAD datasets and propose a simple but effective method based\non pretrained video-language models (VLM). To better adapt the holistic VLM for\nthe fine-grained action detection task, we carefully fine-tune it on the\nlocalized video region-text pairs. This customized fine-tuning endows the VLM\nwith better motion understanding, thus contributing to a more accurate\nalignment between video regions and texts. Local region feature and global\nvideo feature fusion before alignment is adopted to further improve the action\ndetection performance by providing global context. Our method achieves a\npromising performance on novel classes.\n", "date": "2024-05-17", "categories": []}
{"id": "2307.11973", "title": "Two-stream Multi-level Dynamic Point Transformer for Two-person\n  Interaction Recognition", "abstract": "  As a fundamental aspect of human life, two-person interactions contain\nmeaningful information about people's activities, relationships, and social\nsettings. Human action recognition serves as the foundation for many smart\napplications, with a strong focus on personal privacy. However, recognizing\ntwo-person interactions poses more challenges due to increased body occlusion\nand overlap compared to single-person actions. In this paper, we propose a\npoint cloud-based network named Two-stream Multi-level Dynamic Point\nTransformer for two-person interaction recognition. Our model addresses the\nchallenge of recognizing two-person interactions by incorporating local-region\nspatial information, appearance information, and motion information. To achieve\nthis, we introduce a designed frame selection method named Interval Frame\nSampling (IFS), which efficiently samples frames from videos, capturing more\ndiscriminative information in a relatively short processing time. Subsequently,\na frame features learning module and a two-stream multi-level feature\naggregation module extract global and partial features from the sampled frames,\neffectively representing the local-region spatial information, appearance\ninformation, and motion information related to the interactions. Finally, we\napply a transformer to perform self-attention on the learned features for the\nfinal classification. Extensive experiments are conducted on two large-scale\ndatasets, the interaction subsets of NTU RGB+D 60 and NTU RGB+D 120. The\nresults show that our network outperforms state-of-the-art approaches in most\nstandard evaluation settings.\n", "date": "2023-07-22", "categories": []}
{"id": "2405.08204", "title": "A Semantic and Motion-Aware Spatiotemporal Transformer Network for\n  Action Detection", "abstract": "  This paper presents a novel spatiotemporal transformer network that\nintroduces several original components to detect actions in untrimmed videos.\nFirst, the multi-feature selective semantic attention model calculates the\ncorrelations between spatial and motion features to model spatiotemporal\ninteractions between different action semantics properly. Second, the\nmotion-aware network encodes the locations of action semantics in video frames\nutilizing the motion-aware 2D positional encoding algorithm. Such a\nmotion-aware mechanism memorizes the dynamic spatiotemporal variations in\naction frames that current methods cannot exploit. Third, the sequence-based\ntemporal attention model captures the heterogeneous temporal dependencies in\naction frames. In contrast to standard temporal attention used in natural\nlanguage processing, primarily aimed at finding similarities between linguistic\nwords, the proposed sequence-based temporal attention is designed to determine\nboth the differences and similarities between video frames that jointly define\nthe meaning of actions. The proposed approach outperforms the state-of-the-art\nsolutions on four spatiotemporal action datasets: AVA 2.2, AVA 2.1, UCF101-24,\nand EPIC-Kitchens.\n", "date": "2024-05-13", "categories": []}
{"id": "2405.08344", "title": "No Time to Waste: Squeeze Time into Channel for Mobile Video\n  Understanding", "abstract": "  Current architectures for video understanding mainly build upon 3D\nconvolutional blocks or 2D convolutions with additional operations for temporal\nmodeling. However, these methods all regard the temporal axis as a separate\ndimension of the video sequence, which requires large computation and memory\nbudgets and thus limits their usage on mobile devices. In this paper, we\npropose to squeeze the time axis of a video sequence into the channel dimension\nand present a lightweight video recognition network, term as\n\\textit{SqueezeTime}, for mobile video understanding. To enhance the temporal\nmodeling capability of the proposed network, we design a Channel-Time Learning\n(CTL) Block to capture temporal dynamics of the sequence. This module has two\ncomplementary branches, in which one branch is for temporal importance learning\nand another branch with temporal position restoring capability is to enhance\ninter-temporal object modeling ability. The proposed SqueezeTime is much\nlightweight and fast with high accuracies for mobile video understanding.\nExtensive experiments on various video recognition and action detection\nbenchmarks, i.e., Kinetics400, Kinetics600, HMDB51, AVA2.1 and THUMOS14,\ndemonstrate the superiority of our model. For example, our SqueezeTime achieves\n$+1.2\\%$ accuracy and $+80\\%$ GPU throughput gain on Kinetics400 than prior\nmethods. Codes are publicly available at\nhttps://github.com/xinghaochen/SqueezeTime and\nhttps://github.com/mindspore-lab/models/tree/master/research/huawei-noah/SqueezeTime.\n", "date": "2024-05-14", "categories": []}
{"id": "2405.08695", "title": "The impact of Compositionality in Zero-shot Multi-label action\n  recognition for Object-based tasks", "abstract": "  Addressing multi-label action recognition in videos represents a significant\nchallenge for robotic applications in dynamic environments, especially when the\nrobot is required to cooperate with humans in tasks that involve objects.\nExisting methods still struggle to recognize unseen actions or require\nextensive training data. To overcome these problems, we propose Dual-VCLIP, a\nunified approach for zero-shot multi-label action recognition. Dual-VCLIP\nenhances VCLIP, a zero-shot action recognition method, with the DualCoOp method\nfor multi-label image classification. The strength of our method is that at\ntraining time it only learns two prompts, and it is therefore much simpler than\nother methods. We validate our method on the Charades dataset that includes a\nmajority of object-based actions, demonstrating that -- despite its simplicity\n-- our method performs favorably with respect to existing methods on the\ncomplete dataset, and promising performance when tested on unseen actions. Our\ncontribution emphasizes the impact of verb-object class-splits during robots'\ntraining for new cooperative tasks, highlighting the influence on the\nperformance and giving insights into mitigating biases.\n", "date": "2024-05-14", "categories": []}
{"id": "2405.06929", "title": "PRENet: A Plane-Fit Redundancy Encoding Point Cloud Sequence Network for\n  Real-Time 3D Action Recognition", "abstract": "  Recognizing human actions from point cloud sequence has attracted tremendous\nattention from both academia and industry due to its wide applications.\nHowever, most previous studies on point cloud action recognition typically\nrequire complex networks to extract intra-frame spatial features and\ninter-frame temporal features, resulting in an excessive number of redundant\ncomputations. This leads to high latency, rendering them impractical for\nreal-world applications. To address this problem, we propose a Plane-Fit\nRedundancy Encoding point cloud sequence network named PRENet. The primary\nconcept of our approach involves the utilization of plane fitting to mitigate\nspatial redundancy within the sequence, concurrently encoding the temporal\nredundancy of the entire sequence to minimize redundant computations.\nSpecifically, our network comprises two principal modules: a Plane-Fit\nEmbedding module and a Spatio-Temporal Consistency Encoding module. The\nPlane-Fit Embedding module capitalizes on the observation that successive point\ncloud frames exhibit unique geometric features in physical space, allowing for\nthe reuse of spatially encoded data for temporal stream encoding. The\nSpatio-Temporal Consistency Encoding module amalgamates the temporal structure\nof the temporally redundant part with its corresponding spatial arrangement,\nthereby enhancing recognition accuracy. We have done numerous experiments to\nverify the effectiveness of our network. The experimental results demonstrate\nthat our method achieves almost identical recognition accuracy while being\nnearly four times faster than other state-of-the-art methods.\n", "date": "2024-05-11", "categories": []}
{"id": "2405.07723", "title": "Coarse or Fine? Recognising Action End States without Labels", "abstract": "  We focus on the problem of recognising the end state of an action in an\nimage, which is critical for understanding what action is performed and in\nwhich manner. We study this focusing on the task of predicting the coarseness\nof a cut, i.e., deciding whether an object was cut \"coarsely\" or \"finely\". No\ndataset with these annotated end states is available, so we propose an\naugmentation method to synthesise training data. We apply this method to\ncutting actions extracted from an existing action recognition dataset. Our\nmethod is object agnostic, i.e., it presupposes the location of the object but\nnot its identity. Starting from less than a hundred images of a whole object,\nwe can generate several thousands images simulating visually diverse cuts of\ndifferent coarseness. We use our synthetic data to train a model based on UNet\nand test it on real images showing coarsely/finely cut objects. Results\ndemonstrate that the model successfully recognises the end state of the cutting\naction despite the domain gap between training and testing, and that the model\ngeneralises well to unseen objects.\n", "date": "2024-05-13", "categories": []}
{"id": "2402.03019", "title": "Taylor Videos for Action Recognition", "abstract": "  Effectively extracting motions from video is a critical and long-standing\nproblem for action recognition. This problem is very challenging because\nmotions (i) do not have an explicit form, (ii) have various concepts such as\ndisplacement, velocity, and acceleration, and (iii) often contain noise caused\nby unstable pixels. Addressing these challenges, we propose the Taylor video, a\nnew video format that highlights the dominate motions (e.g., a waving hand) in\neach of its frames named the Taylor frame. Taylor video is named after Taylor\nseries, which approximates a function at a given point using important terms.\nIn the scenario of videos, we define an implicit motion-extraction function\nwhich aims to extract motions from video temporal block. In this block, using\nthe frames, the difference frames, and higher-order difference frames, we\nperform Taylor expansion to approximate this function at the starting frame. We\nshow the summation of the higher-order terms in the Taylor series gives us\ndominant motion patterns, where static objects, small and unstable motions are\nremoved. Experimentally we show that Taylor videos are effective inputs to\npopular architectures including 2D CNNs, 3D CNNs, and transformers. When used\nindividually, Taylor videos yield competitive action recognition accuracy\ncompared to RGB videos and optical flow. When fused with RGB or optical flow\nvideos, further accuracy improvement is achieved. Additionally, we apply Taylor\nvideo computation to human skeleton sequences, resulting in Taylor skeleton\nsequences that outperform the use of original skeletons for skeleton-based\naction recognition.\n", "date": "2024-02-05", "categories": []}
{"id": "2405.05354", "title": "Transfer-LMR: Heavy-Tail Driving Behavior Recognition in Diverse Traffic\n  Scenarios", "abstract": "  Recognizing driving behaviors is important for downstream tasks such as\nreasoning, planning, and navigation. Existing video recognition approaches work\nwell for common behaviors (e.g. \"drive straight\", \"brake\", \"turn left/right\").\nHowever, the performance is sub-par for underrepresented/rare behaviors\ntypically found in tail of the behavior class distribution. To address this\nshortcoming, we propose Transfer-LMR, a modular training routine for improving\nthe recognition performance across all driving behavior classes. We extensively\nevaluate our approach on METEOR and HDD datasets that contain rich yet\nheavy-tailed distribution of driving behaviors and span diverse traffic\nscenarios. The experimental results demonstrate the efficacy of our approach,\nespecially for recognizing underrepresented/rare driving behaviors.\n", "date": "2024-05-08", "categories": []}
{"id": "2405.05584", "title": "A Survey on Backbones for Deep Video Action Recognition", "abstract": "  Action recognition is a key technology in building interactive metaverses.\nWith the rapid development of deep learning, methods in action recognition have\nalso achieved great advancement. Researchers design and implement the backbones\nreferring to multiple standpoints, which leads to the diversity of methods and\nencountering new challenges. This paper reviews several action recognition\nmethods based on deep neural networks. We introduce these methods in three\nparts: 1) Two-Streams networks and their variants, which, specifically in this\npaper, use RGB video frame and optical flow modality as input; 2) 3D\nconvolutional networks, which make efforts in taking advantage of RGB modality\ndirectly while extracting different motion information is no longer necessary;\n3) Transformer-based methods, which introduce the model from natural language\nprocessing into computer vision and video understanding. We offer objective\nsights in this review and hopefully provide a reference for future research.\n", "date": "2024-05-09", "categories": []}
{"id": "2405.02538", "title": "AdaFPP: Adapt-Focused Bi-Propagating Prototype Learning for Panoramic\n  Activity Recognition", "abstract": "  Panoramic Activity Recognition (PAR) aims to identify multi-granularity\nbehaviors performed by multiple persons in panoramic scenes, including\nindividual activities, group activities, and global activities. Previous\nmethods 1) heavily rely on manually annotated detection boxes in training and\ninference, hindering further practical deployment; or 2) directly employ normal\ndetectors to detect multiple persons with varying size and spatial occlusion in\npanoramic scenes, blocking the performance gain of PAR. To this end, we\nconsider learning a detector adapting varying-size occluded persons, which is\noptimized along with the recognition module in the all-in-one framework.\nTherefore, we propose a novel Adapt-Focused bi-Propagating Prototype learning\n(AdaFPP) framework to jointly recognize individual, group, and global\nactivities in panoramic activity scenes by learning an adapt-focused detector\nand multi-granularity prototypes as the pretext tasks in an end-to-end way.\nSpecifically, to accommodate the varying sizes and spatial occlusion of\nmultiple persons in crowed panoramic scenes, we introduce a panoramic\nadapt-focuser, achieving the size-adapting detection of individuals by\ncomprehensively selecting and performing fine-grained detections on\nobject-dense sub-regions identified through original detections. In addition,\nto mitigate information loss due to inaccurate individual localizations, we\nintroduce a bi-propagation prototyper that promotes closed-loop interaction and\ninformative consistency across different granularities by facilitating\nbidirectional information propagation among the individual, group, and global\nlevels. Extensive experiments demonstrate the significant performance of AdaFPP\nand emphasize its powerful applicability for PAR.\n", "date": "2024-05-04", "categories": []}
{"id": "2405.02571", "title": "ViTALS: Vision Transformer for Action Localization in Surgical\n  Nephrectomy", "abstract": "  Surgical action localization is a challenging computer vision problem. While\nit has promising applications including automated training of surgery\nprocedures, surgical workflow optimization, etc., appropriate model design is\npivotal to accomplishing this task. Moreover, the lack of suitable medical\ndatasets adds an additional layer of complexity. To that effect, we introduce a\nnew complex dataset of nephrectomy surgeries called UroSlice. To perform the\naction localization from these videos, we propose a novel model termed as\n`ViTALS' (Vision Transformer for Action Localization in Surgical Nephrectomy).\nOur model incorporates hierarchical dilated temporal convolution layers and\ninter-layer residual connections to capture the temporal correlations at finer\nas well as coarser granularities. The proposed approach achieves\nstate-of-the-art performance on Cholec80 and UroSlice datasets (89.8% and 66.1%\naccuracy, respectively), validating its effectiveness.\n", "date": "2024-05-04", "categories": []}
{"id": "2405.03091", "title": "Research on Image Recognition Technology Based on Multimodal Deep\n  Learning", "abstract": "  This project investigates the human multi-modal behavior identification\nalgorithm utilizing deep neural networks. According to the characteristics of\ndifferent modal information, different deep neural networks are used to adapt\nto different modal video information. Through the integration of various deep\nneural networks, the algorithm successfully identifies behaviors across\nmultiple modalities. In this project, multiple cameras developed by Microsoft\nKinect were used to collect corresponding bone point data based on acquiring\nconventional images. In this way, the motion features in the image can be\nextracted. Ultimately, the behavioral characteristics discerned through both\napproaches are synthesized to facilitate the precise identification and\ncategorization of behaviors. The performance of the suggested algorithm was\nevaluated using the MSR3D data set. The findings from these experiments\nindicate that the accuracy in recognizing behaviors remains consistently high,\nsuggesting that the algorithm is reliable in various scenarios. Additionally,\nthe tests demonstrate that the algorithm substantially enhances the accuracy of\ndetecting pedestrian behaviors in video footage.\n", "date": "2024-05-06", "categories": []}
{"id": "2305.16602", "title": "Discovering Novel Actions from Open World Egocentric Videos with\n  Object-Grounded Visual Commonsense Reasoning", "abstract": "  Learning to infer labels in an open world, i.e., in an environment where the\ntarget ``labels'' are unknown, is an important characteristic for achieving\nautonomy. Foundation models, pre-trained on enormous amounts of data, have\nshown remarkable generalization skills through prompting, particularly in\nzero-shot inference. However, their performance is restricted to the\ncorrectness of the target label's search space, i.e., candidate labels provided\nin the prompt. This target search space can be unknown or exceptionally large\nin an open world, severely restricting their performance. To tackle this\nchallenging problem, we propose a two-step, neuro-symbolic framework called\nALGO - Action Learning with Grounded Object recognition that uses symbolic\nknowledge stored in large-scale knowledge bases to infer activities in\negocentric videos with limited supervision. First, we propose a neuro-symbolic\nprompting approach that uses object-centric vision-language models as a noisy\noracle to ground objects in the video through evidence-based reasoning. Second,\ndriven by prior commonsense knowledge, we discover plausible activities through\nan energy-based symbolic pattern theory framework and learn to ground\nknowledge-based action (verb) concepts in the video. Extensive experiments on\nfour publicly available datasets (EPIC-Kitchens, GTEA Gaze, GTEA Gaze Plus, and\nCharades-Ego) demonstrate its performance on open-world activity inference. We\nalso show that ALGO can be extended to zero-shot inference and demonstrate its\ncompetitive performance on the Charades-Ego dataset.\n", "date": "2023-05-26", "categories": []}
{"id": "2405.01882", "title": "Millimeter Wave Radar-based Human Activity Recognition for Healthcare\n  Monitoring Robot", "abstract": "  Healthcare monitoring is crucial, especially for the daily care of elderly\nindividuals living alone. It can detect dangerous occurrences, such as falls,\nand provide timely alerts to save lives. Non-invasive millimeter wave (mmWave)\nradar-based healthcare monitoring systems using advanced human activity\nrecognition (HAR) models have recently gained significant attention. However,\nthey encounter challenges in handling sparse point clouds, achieving real-time\ncontinuous classification, and coping with limited monitoring ranges when\nstatically mounted. To overcome these limitations, we propose RobHAR, a movable\nrobot-mounted mmWave radar system with lightweight deep neural networks for\nreal-time monitoring of human activities. Specifically, we first propose a\nsparse point cloud-based global embedding to learn the features of point clouds\nusing the light-PointNet (LPN) backbone. Then, we learn the temporal pattern\nwith a bidirectional lightweight LSTM model (BiLiLSTM). In addition, we\nimplement a transition optimization strategy, integrating the Hidden Markov\nModel (HMM) with Connectionist Temporal Classification (CTC) to improve the\naccuracy and robustness of the continuous HAR. Our experiments on three\ndatasets indicate that our method significantly outperforms the previous\nstudies in both discrete and continuous HAR tasks. Finally, we deploy our\nsystem on a movable robot-mounted edge computing platform, achieving flexible\nhealthcare monitoring in real-world scenarios.\n", "date": "2024-05-03", "categories": []}
{"id": "2405.01337", "title": "Multi-view Action Recognition via Directed Gromov-Wasserstein\n  Discrepancy", "abstract": "  Action recognition has become one of the popular research topics in computer\nvision. There are various methods based on Convolutional Networks and\nself-attention mechanisms as Transformers to solve both spatial and temporal\ndimensions problems of action recognition tasks that achieve competitive\nperformances. However, these methods lack a guarantee of the correctness of the\naction subject that the models give attention to, i.e., how to ensure an action\nrecognition model focuses on the proper action subject to make a reasonable\naction prediction. In this paper, we propose a multi-view attention consistency\nmethod that computes the similarity between two attentions from two different\nviews of the action videos using Directed Gromov-Wasserstein Discrepancy.\nFurthermore, our approach applies the idea of Neural Radiance Field to\nimplicitly render the features from novel views when training on single-view\ndatasets. Therefore, the contributions in this work are three-fold. Firstly, we\nintroduce the multi-view attention consistency to solve the problem of\nreasonable prediction in action recognition. Secondly, we define a new metric\nfor multi-view consistent attention using Directed Gromov-Wasserstein\nDiscrepancy. Thirdly, we built an action recognition model based on Video\nTransformers and Neural Radiance Fields. Compared to the recent action\nrecognition methods, the proposed approach achieves state-of-the-art results on\nthree large-scale datasets, i.e., Jester, Something-Something V2, and\nKinetics-400.\n", "date": "2024-05-02", "categories": []}
{"id": "2404.19383", "title": "Cross-Block Fine-Grained Semantic Cascade for Skeleton-Based Sports\n  Action Recognition", "abstract": "  Human action video recognition has recently attracted more attention in\napplications such as video security and sports posture correction. Popular\nsolutions, including graph convolutional networks (GCNs) that model the human\nskeleton as a spatiotemporal graph, have proven very effective. GCNs-based\nmethods with stacked blocks usually utilize top-layer semantics for\nclassification/annotation purposes. Although the global features learned\nthrough the procedure are suitable for the general classification, they have\ndifficulty capturing fine-grained action change across adjacent frames --\ndecisive factors in sports actions. In this paper, we propose a novel\n``Cross-block Fine-grained Semantic Cascade (CFSC)'' module to overcome this\nchallenge. In summary, the proposed CFSC progressively integrates shallow\nvisual knowledge into high-level blocks to allow networks to focus on action\ndetails. In particular, the CFSC module utilizes the GCN feature maps produced\nat different levels, as well as aggregated features from proceeding levels to\nconsolidate fine-grained features. In addition, a dedicated temporal\nconvolution is applied at each level to learn short-term temporal features,\nwhich will be carried over from shallow to deep layers to maximize the leverage\nof low-level details. This cross-block feature aggregation methodology, capable\nof mitigating the loss of fine-grained information, has resulted in improved\nperformance. Last, FD-7, a new action recognition dataset for fencing sports,\nwas collected and will be made publicly available. Experimental results and\nempirical analysis on public benchmarks (FSD-10) and self-collected (FD-7)\ndemonstrate the advantage of our CFSC module on learning discriminative\npatterns for action classification over others.\n", "date": "2024-04-30", "categories": []}
{"id": "2404.19542", "title": "One-Stage Open-Vocabulary Temporal Action Detection Leveraging Temporal\n  Multi-scale and Action Label Features", "abstract": "  Open-vocabulary Temporal Action Detection (Open-vocab TAD) is an advanced\nvideo analysis approach that expands Closed-vocabulary Temporal Action\nDetection (Closed-vocab TAD) capabilities. Closed-vocab TAD is typically\nconfined to localizing and classifying actions based on a predefined set of\ncategories. In contrast, Open-vocab TAD goes further and is not limited to\nthese predefined categories. This is particularly useful in real-world\nscenarios where the variety of actions in videos can be vast and not always\npredictable. The prevalent methods in Open-vocab TAD typically employ a 2-stage\napproach, which involves generating action proposals and then identifying those\nactions. However, errors made during the first stage can adversely affect the\nsubsequent action identification accuracy. Additionally, existing studies face\nchallenges in handling actions of different durations owing to the use of fixed\ntemporal processing methods. Therefore, we propose a 1-stage approach\nconsisting of two primary modules: Multi-scale Video Analysis (MVA) and\nVideo-Text Alignment (VTA). The MVA module captures actions at varying temporal\nresolutions, overcoming the challenge of detecting actions with diverse\ndurations. The VTA module leverages the synergy between visual and textual\nmodalities to precisely align video segments with corresponding action labels,\na critical step for accurate action identification in Open-vocab scenarios.\nEvaluations on widely recognized datasets THUMOS14 and ActivityNet-1.3, showed\nthat the proposed method achieved superior results compared to the other\nmethods in both Open-vocab and Closed-vocab settings. This serves as a strong\ndemonstration of the effectiveness of the proposed method in the TAD task.\n", "date": "2024-04-30", "categories": []}
{"id": "2305.08420", "title": "Exploring Few-Shot Adaptation for Activity Recognition on Diverse\n  Domains", "abstract": "  Domain adaptation is essential for activity recognition to ensure accurate\nand robust performance across diverse environments, sensor types, and data\nsources. Unsupervised domain adaptation methods have been extensively studied,\nyet, they require large-scale unlabeled data from the target domain. In this\nwork, we focus on Few-Shot Domain Adaptation for Activity Recognition\n(FSDA-AR), which leverages a very small amount of labeled target videos to\nachieve effective adaptation. This approach is appealing for applications\nbecause it only needs a few or even one labeled example per class in the target\ndomain, ideal for recognizing rare but critical activities. However, the\nexisting FSDA-AR works mostly focus on the domain adaptation on sports videos,\nwhere the domain diversity is limited. We propose a new FSDA-AR benchmark using\nfive established datasets considering the adaptation on more diverse and\nchallenging domains. Our results demonstrate that FSDA-AR performs comparably\nto unsupervised domain adaptation with significantly fewer labeled target\ndomain samples. We further propose a novel approach, RelaMiX, to better\nleverage the few labeled target domain samples as knowledge guidance. RelaMiX\nencompasses a temporal relational attention network with relation dropout,\nalongside a cross-domain information alignment mechanism. Furthermore, it\nintegrates a mechanism for mixing features within a latent space by using the\nfew-shot target domain samples. The proposed RelaMiX solution achieves\nstate-of-the-art performance on all datasets within the FSDA-AR benchmark. To\nencourage future research of few-shot domain adaptation for activity\nrecognition, our code will be publicly available at\nhttps://github.com/KPeng9510/RelaMiX.\n", "date": "2023-05-15", "categories": []}
{"id": "2404.18206", "title": "Enhancing Action Recognition from Low-Quality Skeleton Data via\n  Part-Level Knowledge Distillation", "abstract": "  Skeleton-based action recognition is vital for comprehending human-centric\nvideos and has applications in diverse domains. One of the challenges of\nskeleton-based action recognition is dealing with low-quality data, such as\nskeletons that have missing or inaccurate joints. This paper addresses the\nissue of enhancing action recognition using low-quality skeletons through a\ngeneral knowledge distillation framework. The proposed framework employs a\nteacher-student model setup, where a teacher model trained on high-quality\nskeletons guides the learning of a student model that handles low-quality\nskeletons. To bridge the gap between heterogeneous high-quality and lowquality\nskeletons, we present a novel part-based skeleton matching strategy, which\nexploits shared body parts to facilitate local action pattern learning. An\naction-specific part matrix is developed to emphasize critical parts for\ndifferent actions, enabling the student model to distill discriminative\npart-level knowledge. A novel part-level multi-sample contrastive loss achieves\nknowledge transfer from multiple high-quality skeletons to low-quality ones,\nwhich enables the proposed knowledge distillation framework to include training\nlow-quality skeletons that lack corresponding high-quality matches.\nComprehensive experiments conducted on the NTU-RGB+D, Penn Action, and SYSU 3D\nHOI datasets demonstrate the effectiveness of the proposed knowledge\ndistillation framework.\n", "date": "2024-04-28", "categories": []}
{"id": "2309.06462", "title": "Action Segmentation Using 2D Skeleton Heatmaps and Multi-Modality Fusion", "abstract": "  This paper presents a 2D skeleton-based action segmentation method with\napplications in fine-grained human activity recognition. In contrast with\nstate-of-the-art methods which directly take sequences of 3D skeleton\ncoordinates as inputs and apply Graph Convolutional Networks (GCNs) for\nspatiotemporal feature learning, our main idea is to use sequences of 2D\nskeleton heatmaps as inputs and employ Temporal Convolutional Networks (TCNs)\nto extract spatiotemporal features. Despite lacking 3D information, our\napproach yields comparable/superior performances and better robustness against\nmissing keypoints than previous methods on action segmentation datasets.\nMoreover, we improve the performances further by using both 2D skeleton\nheatmaps and RGB videos as inputs. To our best knowledge, this is the first\nwork to utilize 2D skeleton heatmap inputs and the first work to explore 2D\nskeleton+RGB fusion for action segmentation.\n", "date": "2023-09-12", "categories": []}
{"id": "2404.15719", "title": "HDBN: A Novel Hybrid Dual-branch Network for Robust Skeleton-based\n  Action Recognition", "abstract": "  Skeleton-based action recognition has gained considerable traction thanks to\nits utilization of succinct and robust skeletal representations. Nonetheless,\ncurrent methodologies often lean towards utilizing a solitary backbone to model\nskeleton modality, which can be limited by inherent flaws in the network\nbackbone. To address this and fully leverage the complementary characteristics\nof various network architectures, we propose a novel Hybrid Dual-Branch Network\n(HDBN) for robust skeleton-based action recognition, which benefits from the\ngraph convolutional network's proficiency in handling graph-structured data and\nthe powerful modeling capabilities of Transformers for global information. In\ndetail, our proposed HDBN is divided into two trunk branches: MixGCN and\nMixFormer. The two branches utilize GCNs and Transformers to model both 2D and\n3D skeletal modalities respectively. Our proposed HDBN emerged as one of the\ntop solutions in the Multi-Modal Video Reasoning and Analyzing Competition\n(MMVRAC) of 2024 ICME Grand Challenge, achieving accuracies of 47.95% and\n75.36% on two benchmarks of the UAV-Human dataset by outperforming most\nexisting methods. Our code will be publicly available at:\nhttps://github.com/liujf69/ICMEW2024-Track10.\n", "date": "2024-04-24", "categories": []}
{"id": "2404.16416", "title": "Learning Discriminative Spatio-temporal Representations for\n  Semi-supervised Action Recognition", "abstract": "  Semi-supervised action recognition aims to improve spatio-temporal reasoning\nability with a few labeled data in conjunction with a large amount of unlabeled\ndata. Albeit recent advancements, existing powerful methods are still prone to\nmaking ambiguous predictions under scarce labeled data, embodied as the\nlimitation of distinguishing different actions with similar spatio-temporal\ninformation. In this paper, we approach this problem by empowering the model\ntwo aspects of capability, namely discriminative spatial modeling and temporal\nstructure modeling for learning discriminative spatio-temporal representations.\nSpecifically, we propose an Adaptive Contrastive Learning~(ACL) strategy. It\nassesses the confidence of all unlabeled samples by the class prototypes of the\nlabeled data, and adaptively selects positive-negative samples from a\npseudo-labeled sample bank to construct contrastive learning. Additionally, we\nintroduce a Multi-scale Temporal Learning~(MTL) strategy. It could highlight\ninformative semantics from long-term clips and integrate them into the\nshort-term clip while suppressing noisy information. Afterwards, both of these\ntwo new techniques are integrated in a unified framework to encourage the model\nto make accurate predictions. Extensive experiments on UCF101, HMDB51 and\nKinetics400 show the superiority of our method over prior state-of-the-art\napproaches.\n", "date": "2024-04-25", "categories": []}
{"id": "2404.15349", "title": "A Survey on Multimodal Wearable Sensor-based Human Action Recognition", "abstract": "  The combination of increased life expectancy and falling birth rates is\nresulting in an aging population. Wearable Sensor-based Human Activity\nRecognition (WSHAR) emerges as a promising assistive technology to support the\ndaily lives of older individuals, unlocking vast potential for human-centric\napplications. However, recent surveys in WSHAR have been limited, focusing\neither solely on deep learning approaches or on a single sensor modality. In\nreal life, our human interact with the world in a multi-sensory way, where\ndiverse information sources are intricately processed and interpreted to\naccomplish a complex and unified sensing system. To give machines similar\nintelligence, multimodal machine learning, which merges data from various\nsources, has become a popular research area with recent advancements. In this\nstudy, we present a comprehensive survey from a novel perspective on how to\nleverage multimodal learning to WSHAR domain for newcomers and researchers. We\nbegin by presenting the recent sensor modalities as well as deep learning\napproaches in HAR. Subsequently, we explore the techniques used in present\nmultimodal systems for WSHAR. This includes inter-multimodal systems which\nutilize sensor modalities from both visual and non-visual systems and\nintra-multimodal systems that simply take modalities from non-visual systems.\nAfter that, we focus on current multimodal learning approaches that have\napplied to solve some of the challenges existing in WSHAR. Specifically, we\nmake extra efforts by connecting the existing multimodal literature from other\ndomains, such as computer vision and natural language processing, with current\nWSHAR area. Finally, we identify the corresponding challenges and potential\nresearch direction in current WSHAR area for further improvement.\n", "date": "2024-04-14", "categories": []}
{"id": "2404.14906", "title": "Driver Activity Classification Using Generalizable Representations from\n  Vision-Language Models", "abstract": "  Driver activity classification is crucial for ensuring road safety, with\napplications ranging from driver assistance systems to autonomous vehicle\ncontrol transitions. In this paper, we present a novel approach leveraging\ngeneralizable representations from vision-language models for driver activity\nclassification. Our method employs a Semantic Representation Late Fusion Neural\nNetwork (SRLF-Net) to process synchronized video frames from multiple\nperspectives. Each frame is encoded using a pretrained vision-language encoder,\nand the resulting embeddings are fused to generate class probability\npredictions. By leveraging contrastively-learned vision-language\nrepresentations, our approach achieves robust performance across diverse driver\nactivities. We evaluate our method on the Naturalistic Driving Action\nRecognition Dataset, demonstrating strong accuracy across many classes. Our\nresults suggest that vision-language representations offer a promising avenue\nfor driver monitoring systems, providing both accuracy and interpretability\nthrough natural language descriptors.\n", "date": "2024-04-23", "categories": []}
{"id": "2311.17241", "title": "End-to-End Temporal Action Detection with 1B Parameters Across 1000\n  Frames", "abstract": "  Recently, temporal action detection (TAD) has seen significant performance\nimprovement with end-to-end training. However, due to the memory bottleneck,\nonly models with limited scales and limited data volumes can afford end-to-end\ntraining, which inevitably restricts TAD performance. In this paper, we reduce\nthe memory consumption for end-to-end training, and manage to scale up the TAD\nbackbone to 1 billion parameters and the input video to 1,536 frames, leading\nto significant detection performance. The key to our approach lies in our\nproposed temporal-informative adapter (TIA), which is a novel lightweight\nmodule that reduces training memory. Using TIA, we free the humongous backbone\nfrom learning to adapt to the TAD task by only updating the parameters in TIA.\nTIA also leads to better TAD representation by temporally aggregating context\nfrom adjacent frames throughout the backbone. We evaluate our model across four\nrepresentative datasets. Owing to our efficient design, we are able to train\nend-to-end on VideoMAEv2-giant and achieve 75.4% mAP on THUMOS14, being the\nfirst end-to-end model to outperform the best feature-based methods. Code is\navailable at https://github.com/sming256/AdaTAD.\n", "date": "2023-11-28", "categories": []}
{"id": "2311.17948", "title": "Action-slot: Visual Action-centric Representations for Multi-label\n  Atomic Activity Recognition in Traffic Scenes", "abstract": "  In this paper, we study multi-label atomic activity recognition. Despite the\nnotable progress in action recognition, it is still challenging to recognize\natomic activities due to a deficiency in a holistic understanding of both\nmultiple road users' motions and their contextual information. In this paper,\nwe introduce Action-slot, a slot attention-based approach that learns visual\naction-centric representations, capturing both motion and contextual\ninformation. Our key idea is to design action slots that are capable of paying\nattention to regions where atomic activities occur, without the need for\nexplicit perception guidance. To further enhance slot attention, we introduce a\nbackground slot that competes with action slots, aiding the training process in\navoiding unnecessary focus on background regions devoid of activities. Yet, the\nimbalanced class distribution in the existing dataset hampers the assessment of\nrare activities. To address the limitation, we collect a synthetic dataset\ncalled TACO, which is four times larger than OATS and features a balanced\ndistribution of atomic activities. To validate the effectiveness of our method,\nwe conduct comprehensive experiments and ablation studies against various\naction recognition baselines. We also show that the performance of multi-label\natomic activity recognition on real-world datasets can be improved by\npretraining representations on TACO. We will release our source code and\ndataset. See the videos of visualization on the project page:\nhttps://hcis-lab.github.io/Action-slot/\n", "date": "2023-11-29", "categories": []}
{"id": "2404.13311", "title": "STAT: Towards Generalizable Temporal Action Localization", "abstract": "  Weakly-supervised temporal action localization (WTAL) aims to recognize and\nlocalize action instances with only video-level labels. Despite the significant\nprogress, existing methods suffer from severe performance degradation when\ntransferring to different distributions and thus may hardly adapt to real-world\nscenarios . To address this problem, we propose the Generalizable Temporal\nAction Localization task (GTAL), which focuses on improving the\ngeneralizability of action localization methods. We observed that the\nperformance decline can be primarily attributed to the lack of generalizability\nto different action scales. To address this problem, we propose STAT\n(Self-supervised Temporal Adaptive Teacher), which leverages a teacher-student\nstructure for iterative refinement. Our STAT features a refinement module and\nan alignment module. The former iteratively refines the model's output by\nleveraging contextual information and helps adapt to the target scale. The\nlatter improves the refinement process by promoting a consensus between student\nand teacher models. We conduct extensive experiments on three datasets,\nTHUMOS14, ActivityNet1.2, and HACS, and the results show that our method\nsignificantly improves the Baseline methods under the cross-distribution\nevaluation setting, even approaching the same-distribution evaluation\nperformance.\n", "date": "2024-04-20", "categories": []}
{"id": "2404.14032", "title": "1st Place Solution to the 1st SkatingVerse Challenge", "abstract": "  This paper presents the winning solution for the 1st SkatingVerse Challenge.\nWe propose a method that involves several steps. To begin, we leverage the DINO\nframework to extract the Region of Interest (ROI) and perform precise cropping\nof the raw video footage. Subsequently, we employ three distinct models, namely\nUnmasked Teacher, UniformerV2, and InfoGCN, to capture different aspects of the\ndata. By ensembling the prediction results based on logits, our solution\nattains an impressive leaderboard score of 95.73%.\n", "date": "2024-04-22", "categories": []}
{"id": "2404.11903", "title": "Simultaneous Detection and Interaction Reasoning for Object-Centric\n  Action Recognition", "abstract": "  The interactions between human and objects are important for recognizing\nobject-centric actions. Existing methods usually adopt a two-stage pipeline,\nwhere object proposals are first detected using a pretrained detector, and then\nare fed to an action recognition model for extracting video features and\nlearning the object relations for action recognition. However, since the action\nprior is unknown in the object detection stage, important objects could be\neasily overlooked, leading to inferior action recognition performance. In this\npaper, we propose an end-to-end object-centric action recognition framework\nthat simultaneously performs Detection And Interaction Reasoning in one stage.\nParticularly, after extracting video features with a base network, we create\nthree modules for concurrent object detection and interaction reasoning. First,\na Patch-based Object Decoder generates proposals from video patch tokens. Then,\nan Interactive Object Refining and Aggregation identifies important objects for\naction recognition, adjusts proposal scores based on position and appearance,\nand aggregates object-level info into a global video representation. Lastly, an\nObject Relation Modeling module encodes object relations. These three modules\ntogether with the video feature extractor can be trained jointly in an\nend-to-end fashion, thus avoiding the heavy reliance on an off-the-shelf object\ndetector, and reducing the multi-stage training burden. We conduct experiments\non two datasets, Something-Else and Ikea-Assembly, to evaluate the performance\nof our proposed approach on conventional, compositional, and few-shot action\nrecognition tasks. Through in-depth experimental analysis, we show the crucial\nrole of interactive objects in learning for action recognition, and we can\noutperform state-of-the-art methods on both datasets.\n", "date": "2024-04-18", "categories": []}
{"id": "2404.11605", "title": "VG4D: Vision-Language Model Goes 4D Video Recognition", "abstract": "  Understanding the real world through point cloud video is a crucial aspect of\nrobotics and autonomous driving systems. However, prevailing methods for 4D\npoint cloud recognition have limitations due to sensor resolution, which leads\nto a lack of detailed information. Recent advances have shown that\nVision-Language Models (VLM) pre-trained on web-scale text-image datasets can\nlearn fine-grained visual concepts that can be transferred to various\ndownstream tasks. However, effectively integrating VLM into the domain of 4D\npoint clouds remains an unresolved problem. In this work, we propose the\nVision-Language Models Goes 4D (VG4D) framework to transfer VLM knowledge from\nvisual-text pre-trained models to a 4D point cloud network. Our approach\ninvolves aligning the 4D encoder's representation with a VLM to learn a shared\nvisual and text space from training on large-scale image-text pairs. By\ntransferring the knowledge of the VLM to the 4D encoder and combining the VLM,\nour VG4D achieves improved recognition performance. To enhance the 4D encoder,\nwe modernize the classic dynamic point cloud backbone and propose an improved\nversion of PSTNet, im-PSTNet, which can efficiently model point cloud videos.\nExperiments demonstrate that our method achieves state-of-the-art performance\nfor action recognition on both the NTU RGB+D 60 dataset and the NTU RGB+D 120\ndataset. Code is available at \\url{https://github.com/Shark0-0/VG4D}.\n", "date": "2024-04-17", "categories": []}
{"id": "2312.01897", "title": "Adapting Short-Term Transformers for Action Detection in Untrimmed\n  Videos", "abstract": "  Vision Transformer (ViT) has shown high potential in video recognition, owing\nto its flexible design, adaptable self-attention mechanisms, and the efficacy\nof masked pre-training. Yet, it remains unclear how to adapt these pre-trained\nshort-term ViTs for temporal action detection (TAD) in untrimmed videos. The\nexisting works treat them as off-the-shelf feature extractors for each\nshort-trimmed snippet without capturing the fine-grained relation among\ndifferent snippets in a broader temporal context. To mitigate this issue, this\npaper focuses on designing a new mechanism for adapting these pre-trained ViT\nmodels as a unified long-form video transformer to fully unleash its modeling\npower in capturing inter-snippet relation, while still keeping low computation\noverhead and memory consumption for efficient TAD. To this end, we design\neffective cross-snippet propagation modules to gradually exchange short-term\nvideo information among different snippets from two levels. For inner-backbone\ninformation propagation, we introduce a cross-snippet propagation strategy to\nenable multi-snippet temporal feature interaction inside the backbone.For\npost-backbone information propagation, we propose temporal transformer layers\nfor further clip-level modeling. With the plain ViT-B pre-trained with\nVideoMAE, our end-to-end temporal action detector (ViT-TAD) yields a very\ncompetitive performance to previous temporal action detectors, riching up to\n69.5 average mAP on THUMOS14, 37.40 average mAP on ActivityNet-1.3 and 17.20\naverage mAP on FineAction.\n", "date": "2023-12-04", "categories": []}
{"id": "2404.07487", "title": "Fine-Grained Side Information Guided Dual-Prompts for Zero-Shot Skeleton\n  Action Recognition", "abstract": "  Skeleton-based zero-shot action recognition aims to recognize unknown human\nactions based on the learned priors of the known skeleton-based actions and a\nsemantic descriptor space shared by both known and unknown categories. However,\nprevious works focus on establishing the bridges between the known skeleton\nrepresentation space and semantic descriptions space at the coarse-grained\nlevel for recognizing unknown action categories, ignoring the fine-grained\nalignment of these two spaces, resulting in suboptimal performance in\ndistinguishing high-similarity action categories. To address these challenges,\nwe propose a novel method via Side information and dual-prompts learning for\nskeleton-based zero-shot action recognition (STAR) at the fine-grained level.\nSpecifically, 1) we decompose the skeleton into several parts based on its\ntopology structure and introduce the side information concerning multi-part\ndescriptions of human body movements for alignment between the skeleton and the\nsemantic space at the fine-grained level; 2) we design the visual-attribute and\nsemantic-part prompts to improve the intra-class compactness within the\nskeleton space and inter-class separability within the semantic space,\nrespectively, to distinguish the high-similarity actions. Extensive experiments\nshow that our method achieves state-of-the-art performance in ZSL and GZSL\nsettings on NTU RGB+D, NTU RGB+D 120, and PKU-MMD datasets.\n", "date": "2024-04-11", "categories": []}
{"id": "2404.09067", "title": "Exploring Explainability in Video Action Recognition", "abstract": "  Image Classification and Video Action Recognition are perhaps the two most\nfoundational tasks in computer vision. Consequently, explaining the inner\nworkings of trained deep neural networks is of prime importance. While numerous\nefforts focus on explaining the decisions of trained deep neural networks in\nimage classification, exploration in the domain of its temporal version, video\naction recognition, has been scant. In this work, we take a deeper look at this\nproblem. We begin by revisiting Grad-CAM, one of the popular feature\nattribution methods for Image Classification, and its extension to Video Action\nRecognition tasks and examine the method's limitations. To address these, we\nintroduce Video-TCAV, by building on TCAV for Image Classification tasks, which\naims to quantify the importance of specific concepts in the decision-making\nprocess of Video Action Recognition models. As the scalable generation of\nconcepts is still an open problem, we propose a machine-assisted approach to\ngenerate spatial and spatiotemporal concepts relevant to Video Action\nRecognition for testing Video-TCAV. We then establish the importance of\ntemporally-varying concepts by demonstrating the superiority of dynamic\nspatiotemporal concepts over trivial spatial concepts. In conclusion, we\nintroduce a framework for investigating hypotheses in action recognition and\nquantitatively testing them, thus advancing research in the explainability of\ndeep neural networks used in video action recognition.\n", "date": "2024-04-13", "categories": []}
{"id": "2404.09842", "title": "STMixer: A One-Stage Sparse Action Detector", "abstract": "  Traditional video action detectors typically adopt the two-stage pipeline,\nwhere a person detector is first employed to generate actor boxes and then 3D\nRoIAlign is used to extract actor-specific features for classification. This\ndetection paradigm requires multi-stage training and inference, and the feature\nsampling is constrained inside the box, failing to effectively leverage richer\ncontext information outside. Recently, a few query-based action detectors have\nbeen proposed to predict action instances in an end-to-end manner. However,\nthey still lack adaptability in feature sampling and decoding, thus suffering\nfrom the issues of inferior performance or slower convergence. In this paper,\nwe propose two core designs for a more flexible one-stage sparse action\ndetector. First, we present a query-based adaptive feature sampling module,\nwhich endows the detector with the flexibility of mining a group of\ndiscriminative features from the entire spatio-temporal domain. Second, we\ndevise a decoupled feature mixing module, which dynamically attends to and\nmixes video features along the spatial and temporal dimensions respectively for\nbetter feature decoding. Based on these designs, we instantiate two detection\npipelines, that is, STMixer-K for keyframe action detection and STMixer-T for\naction tubelet detection. Without bells and whistles, our STMixer detectors\nobtain state-of-the-art results on five challenging spatio-temporal action\ndetection benchmarks for keyframe action detection or action tube detection.\n", "date": "2024-04-15", "categories": []}
{"id": "2404.09951", "title": "Unifying Global and Local Scene Entities Modelling for Precise Action\n  Spotting", "abstract": "  Sports videos pose complex challenges, including cluttered backgrounds,\ncamera angle changes, small action-representing objects, and imbalanced action\nclass distribution. Existing methods for detecting actions in sports videos\nheavily rely on global features, utilizing a backbone network as a black box\nthat encompasses the entire spatial frame. However, these approaches tend to\noverlook the nuances of the scene and struggle with detecting actions that\noccupy a small portion of the frame. In particular, they face difficulties when\ndealing with action classes involving small objects, such as balls or\nyellow/red cards in soccer, which only occupy a fraction of the screen space.\nTo address these challenges, we introduce a novel approach that analyzes and\nmodels scene entities using an adaptive attention mechanism. Particularly, our\nmodel disentangles the scene content into the global environment feature and\nlocal relevant scene entities feature. To efficiently extract environmental\nfeatures while considering temporal information with less computational cost,\nwe propose the use of a 2D backbone network with a time-shift mechanism. To\naccurately capture relevant scene entities, we employ a Vision-Language model\nin conjunction with the adaptive attention mechanism. Our model has\ndemonstrated outstanding performance, securing the 1st place in the\nSoccerNet-v2 Action Spotting, FineDiving, and FineGym challenge with a\nsubstantial performance improvement of 1.6, 2.0, and 1.3 points in avg-mAP\ncompared to the runner-up methods. Furthermore, our approach offers\ninterpretability capabilities in contrast to other deep learning models, which\nare often designed as black boxes. Our code and models are released at:\nhttps://github.com/Fsoft-AIC/unifying-global-local-feature.\n", "date": "2024-04-15", "categories": []}
{"id": "2404.09964", "title": "Design and Analysis of Efficient Attention in Transformers for Social\n  Group Activity Recognition", "abstract": "  Social group activity recognition is a challenging task extended from group\nactivity recognition, where social groups must be recognized with their\nactivities and group members. Existing methods tackle this task by leveraging\nregion features of individuals following existing group activity recognition\nmethods. However, the effectiveness of region features is susceptible to person\nlocalization and variable semantics of individual actions. To overcome these\nissues, we propose leveraging attention modules in transformers to generate\nsocial group features. In this method, multiple embeddings are used to\naggregate features for a social group, each of which is assigned to a group\nmember without duplication. Due to this non-duplicated assignment, the number\nof embeddings must be significant to avoid missing group members and thus\nrenders attention in transformers ineffective. To find optimal attention\ndesigns with a large number of embeddings, we explore several design choices of\nqueries for feature aggregation and self-attention modules in transformer\ndecoders. Extensive experimental results show that the proposed method achieves\nstate-of-the-art performance and verify that the proposed attention designs are\nhighly effective on social group activity recognition.\n", "date": "2024-04-15", "categories": []}
{"id": "2404.05426", "title": "Test-Time Zero-Shot Temporal Action Localization", "abstract": "  Zero-Shot Temporal Action Localization (ZS-TAL) seeks to identify and locate\nactions in untrimmed videos unseen during training. Existing ZS-TAL methods\ninvolve fine-tuning a model on a large amount of annotated training data. While\neffective, training-based ZS-TAL approaches assume the availability of labeled\ndata for supervised learning, which can be impractical in some applications.\nFurthermore, the training process naturally induces a domain bias into the\nlearned model, which may adversely affect the model's generalization ability to\narbitrary videos. These considerations prompt us to approach the ZS-TAL problem\nfrom a radically novel perspective, relaxing the requirement for training data.\nTo this aim, we introduce a novel method that performs Test-Time adaptation for\nTemporal Action Localization (T3AL). In a nutshell, T3AL adapts a pre-trained\nVision and Language Model (VLM). T3AL operates in three steps. First, a\nvideo-level pseudo-label of the action category is computed by aggregating\ninformation from the entire video. Then, action localization is performed\nadopting a novel procedure inspired by self-supervised learning. Finally,\nframe-level textual descriptions extracted with a state-of-the-art captioning\nmodel are employed for refining the action region proposals. We validate the\neffectiveness of T3AL by conducting experiments on the THUMOS14 and the\nActivityNet-v1.3 datasets. Our results demonstrate that T3AL significantly\noutperforms zero-shot baselines based on state-of-the-art VLMs, confirming the\nbenefit of a test-time adaptation approach.\n", "date": "2024-04-08", "categories": []}
{"id": "2404.07245", "title": "Generative Resident Separation and Multi-label Classification for\n  Multi-person Activity Recognition", "abstract": "  This paper presents two models to address the problem of multi-person\nactivity recognition using ambient sensors in a home. The first model, Seq2Res,\nuses a sequence generation approach to separate sensor events from different\nresidents. The second model, BiGRU+Q2L, uses a Query2Label multi-label\nclassifier to predict multiple activities simultaneously. Performances of these\nmodels are compared to a state-of-the-art model in different experimental\nscenarios, using a state-of-the-art dataset of two residents in a home\ninstrumented with ambient sensors. These results lead to a discussion on the\nadvantages and drawbacks of resident separation and multi-label classification\nfor multi-person activity recognition.\n", "date": "2024-04-10", "categories": []}
{"id": "2404.07645", "title": "Simba: Mamba augmented U-ShiftGCN for Skeletal Action Recognition in\n  Videos", "abstract": "  Skeleton Action Recognition (SAR) involves identifying human actions using\nskeletal joint coordinates and their interconnections. While plain Transformers\nhave been attempted for this task, they still fall short compared to the\ncurrent leading methods, which are rooted in Graph Convolutional Networks\n(GCNs) due to the absence of structural priors. Recently, a novel selective\nstate space model, Mamba, has surfaced as a compelling alternative to the\nattention mechanism in Transformers, offering efficient modeling of long\nsequences. In this work, to the utmost extent of our awareness, we present the\nfirst SAR framework incorporating Mamba. Each fundamental block of our model\nadopts a novel U-ShiftGCN architecture with Mamba as its core component. The\nencoder segment of the U-ShiftGCN is devised to extract spatial features from\nthe skeletal data using downsampling vanilla Shift S-GCN blocks. These spatial\nfeatures then undergo intermediate temporal modeling facilitated by the Mamba\nblock before progressing to the encoder section, which comprises vanilla\nupsampling Shift S-GCN blocks. Additionally, a Shift T-GCN (ShiftTCN) temporal\nmodeling unit is employed before the exit of each fundamental block to refine\ntemporal representations. This particular integration of downsampling spatial,\nintermediate temporal, upsampling spatial, and ultimate temporal subunits\nyields promising results for skeleton action recognition. We dub the resulting\nmodel \\textbf{Simba}, which attains state-of-the-art performance across three\nwell-known benchmark skeleton action recognition datasets: NTU RGB+D, NTU RGB+D\n120, and Northwestern-UCLA. Interestingly, U-ShiftGCN (Simba without\nIntermediate Mamba Block) by itself is capable of performing reasonably well\nand surpasses our baseline.\n", "date": "2024-04-11", "categories": []}
{"id": "2404.07962", "title": "Live and Learn: Continual Action Clustering with Incremental Views", "abstract": "  Multi-view action clustering leverages the complementary information from\ndifferent camera views to enhance the clustering performance. Although existing\napproaches have achieved significant progress, they assume all camera views are\navailable in advance, which is impractical when the camera view is incremental\nover time. Besides, learning the invariant information among multiple camera\nviews is still a challenging issue, especially in continual learning scenario.\nAiming at these problems, we propose a novel continual action clustering (CAC)\nmethod, which is capable of learning action categories in a continual learning\nmanner. To be specific, we first devise a category memory library, which\ncaptures and stores the learned categories from historical views. Then, as a\nnew camera view arrives, we only need to maintain a consensus partition matrix,\nwhich can be updated by leveraging the incoming new camera view rather than\nkeeping all of them. Finally, a three-step alternate optimization is proposed,\nin which the category memory library and consensus partition matrix are\noptimized. The empirical experimental results on 6 realistic multi-view action\ncollections demonstrate the excellent clustering performance and time/space\nefficiency of the CAC compared with 15 state-of-the-art baselines.\n", "date": "2024-03-23", "categories": []}
{"id": "2208.11650", "title": "Lane Change Classification and Prediction with Action Recognition\n  Networks", "abstract": "  Anticipating lane change intentions of surrounding vehicles is crucial for\nefficient and safe driving decision making in an autonomous driving system.\nPrevious works often adopt physical variables such as driving speed,\nacceleration and so forth for lane change classification. However, physical\nvariables do not contain semantic information. Although 3D CNNs have been\ndeveloping rapidly, the number of methods utilising action recognition models\nand appearance feature for lane change recognition is low, and they all require\nadditional information to pre-process data. In this work, we propose an\nend-to-end framework including two action recognition methods for lane change\nrecognition, using video data collected by cameras. Our method achieves the\nbest lane change classification results using only the RGB video data of the\nPREVENTION dataset. Class activation maps demonstrate that action recognition\nmodels can efficiently extract lane change motions. A method to better extract\nmotion clues is also proposed in this paper.\n", "date": "2022-08-24", "categories": []}
{"id": "2404.06894", "title": "O-TALC: Steps Towards Combating Oversegmentation within Online Action\n  Segmentation", "abstract": "  Online temporal action segmentation shows a strong potential to facilitate\nmany HRI tasks where extended human action sequences must be tracked and\nunderstood in real time. Traditional action segmentation approaches, however,\noperate in an offline two stage approach, relying on computationally expensive\nvideo wide features for segmentation, rendering them unsuitable for online HRI\napplications. In order to facilitate online action segmentation on a stream of\nincoming video data, we introduce two methods for improved training and\ninference of backbone action recognition models, allowing them to be deployed\ndirectly for online frame level classification. Firstly, we introduce surround\ndense sampling whilst training to facilitate training vs. inference clip\nmatching and improve segment boundary predictions. Secondly, we introduce an\nOnline Temporally Aware Label Cleaning (O-TALC) strategy to explicitly reduce\noversegmentation during online inference. As our methods are backbone\ninvariant, they can be deployed with computationally efficient spatio-temporal\naction recognition models capable of operating in real time with a small\nsegmentation latency. We show our method outperforms similar online action\nsegmentation work as well as matches the performance of many offline models\nwith access to full temporal resolution when operating on challenging\nfine-grained datasets.\n", "date": "2024-04-10", "categories": []}
{"id": "2307.02730", "title": "Fine-grained Action Analysis: A Multi-modality and Multi-task Dataset of\n  Figure Skating", "abstract": "  The fine-grained action analysis of the existing action datasets is\nchallenged by insufficient action categories, low fine granularities, limited\nmodalities, and tasks. In this paper, we propose a Multi-modality and\nMulti-task dataset of Figure Skating (MMFS) which was collected from the World\nFigure Skating Championships. MMFS, which possesses action recognition and\naction quality assessment, captures RGB, skeleton, and is collected the score\nof actions from 11671 clips with 256 categories including spatial and temporal\nlabels. The key contributions of our dataset fall into three aspects as\nfollows. (1) Independently spatial and temporal categories are first proposed\nto further explore fine-grained action recognition and quality assessment. (2)\nMMFS first introduces the skeleton modality for complex fine-grained action\nquality assessment. (3) Our multi-modality and multi-task dataset encourage\nmore action analysis models. To benchmark our dataset, we adopt RGB-based and\nskeleton-based baseline methods for action recognition and action quality\nassessment.\n", "date": "2023-07-06", "categories": []}
{"id": "2404.05559", "title": "TIM: A Time Interval Machine for Audio-Visual Action Recognition", "abstract": "  Diverse actions give rise to rich audio-visual signals in long videos. Recent\nworks showcase that the two modalities of audio and video exhibit different\ntemporal extents of events and distinct labels. We address the interplay\nbetween the two modalities in long videos by explicitly modelling the temporal\nextents of audio and visual events. We propose the Time Interval Machine (TIM)\nwhere a modality-specific time interval poses as a query to a transformer\nencoder that ingests a long video input. The encoder then attends to the\nspecified interval, as well as the surrounding context in both modalities, in\norder to recognise the ongoing action.\n  We test TIM on three long audio-visual video datasets: EPIC-KITCHENS,\nPerception Test, and AVE, reporting state-of-the-art (SOTA) for recognition. On\nEPIC-KITCHENS, we beat previous SOTA that utilises LLMs and significantly\nlarger pre-training by 2.9% top-1 action recognition accuracy. Additionally, we\nshow that TIM can be adapted for action detection, using dense multi-scale\ninterval queries, outperforming SOTA on EPIC-KITCHENS-100 for most metrics, and\nshowing strong performance on the Perception Test. Our ablations show the\ncritical role of integrating the two modalities and modelling their time\nintervals in achieving this performance. Code and models at:\nhttps://github.com/JacobChalk/TIM\n", "date": "2024-04-08", "categories": []}
{"id": "2404.05849", "title": "Localizing Moments of Actions in Untrimmed Videos of Infants with Autism\n  Spectrum Disorder", "abstract": "  Autism Spectrum Disorder (ASD) presents significant challenges in early\ndiagnosis and intervention, impacting children and their families. With\nprevalence rates rising, there is a critical need for accessible and efficient\nscreening tools. Leveraging machine learning (ML) techniques, in particular\nTemporal Action Localization (TAL), holds promise for automating ASD screening.\nThis paper introduces a self-attention based TAL model designed to identify\nASD-related behaviors in infant videos. Unlike existing methods, our approach\nsimplifies complex modeling and emphasizes efficiency, which is essential for\npractical deployment in real-world scenarios. Importantly, this work\nunderscores the importance of developing computer vision methods capable of\noperating in naturilistic environments with little equipment control,\naddressing key challenges in ASD screening. This study is the first to conduct\nend-to-end temporal action localization in untrimmed videos of infants with\nASD, offering promising avenues for early intervention and support. We report\nbaseline results of behavior detection using our TAL model. We achieve 70%\naccuracy for look face, 79% accuracy for look object, 72% for smile and 65% for\nvocalization.\n", "date": "2024-04-08", "categories": []}
{"id": "2404.06243", "title": "ActNetFormer: Transformer-ResNet Hybrid Method for Semi-Supervised\n  Action Recognition in Videos", "abstract": "  Human action or activity recognition in videos is a fundamental task in\ncomputer vision with applications in surveillance and monitoring, self-driving\ncars, sports analytics, human-robot interaction and many more. Traditional\nsupervised methods require large annotated datasets for training, which are\nexpensive and time-consuming to acquire. This work proposes a novel approach\nusing Cross-Architecture Pseudo-Labeling with contrastive learning for\nsemi-supervised action recognition. Our framework leverages both labeled and\nunlabelled data to robustly learn action representations in videos, combining\npseudo-labeling with contrastive learning for effective learning from both\ntypes of samples. We introduce a novel cross-architecture approach where 3D\nConvolutional Neural Networks (3D CNNs) and video transformers (VIT) are\nutilised to capture different aspects of action representations; hence we call\nit ActNetFormer. The 3D CNNs excel at capturing spatial features and local\ndependencies in the temporal domain, while VIT excels at capturing long-range\ndependencies across frames. By integrating these complementary architectures\nwithin the ActNetFormer framework, our approach can effectively capture both\nlocal and global contextual information of an action. This comprehensive\nrepresentation learning enables the model to achieve better performance in\nsemi-supervised action recognition tasks by leveraging the strengths of each of\nthese architectures. Experimental results on standard action recognition\ndatasets demonstrate that our approach performs better than the existing\nmethods, achieving state-of-the-art performance with only a fraction of labeled\ndata. The official website of this work is available at:\nhttps://github.com/rana2149/ActNetFormer.\n", "date": "2024-04-09", "categories": []}
{"id": "2308.16018", "title": "SiT-MLP: A Simple MLP with Point-wise Topology Feature Learning for\n  Skeleton-based Action Recognition", "abstract": "  Graph convolution networks (GCNs) have achieved remarkable performance in\nskeleton-based action recognition. However, previous GCN-based methods rely on\nelaborate human priors excessively and construct complex feature aggregation\nmechanisms, which limits the generalizability and effectiveness of networks. To\nsolve these problems, we propose a novel Spatial Topology Gating Unit (STGU),\nan MLP-based variant without extra priors, to capture the co-occurrence\ntopology features that encode the spatial dependency across all joints. In\nSTGU, to learn the point-wise topology features, a new gate-based feature\ninteraction mechanism is introduced to activate the features point-to-point by\nthe attention map generated from the input sample. Based on the STGU, we\npropose the first MLP-based model, SiT-MLP, for skeleton-based action\nrecognition in this work. Compared with previous methods on three large-scale\ndatasets, SiT-MLP achieves competitive performance. In addition, SiT-MLP\nreduces the parameters significantly with favorable results. The code will be\navailable at https://github.com/BUPTSJZhang/SiT?MLP.\n", "date": "2023-08-30", "categories": []}
{"id": "2404.01518", "title": "Temporally Consistent Unbalanced Optimal Transport for Unsupervised\n  Action Segmentation", "abstract": "  We propose a novel approach to the action segmentation task for long,\nuntrimmed videos, based on solving an optimal transport problem. By encoding a\ntemporal consistency prior into a Gromov-Wasserstein problem, we are able to\ndecode a temporally consistent segmentation from a noisy affinity/matching cost\nmatrix between video frames and action classes. Unlike previous approaches, our\nmethod does not require knowing the action order for a video to attain temporal\nconsistency. Furthermore, our resulting (fused) Gromov-Wasserstein problem can\nbe efficiently solved on GPUs using a few iterations of projected mirror\ndescent. We demonstrate the effectiveness of our method in an unsupervised\nlearning setting, where our method is used to generate pseudo-labels for\nself-training. We evaluate our segmentation approach and unsupervised learning\npipeline on the Breakfast, 50-Salads, YouTube Instructions and Desktop Assembly\ndatasets, yielding state-of-the-art results for the unsupervised video action\nsegmentation task.\n", "date": "2024-04-01", "categories": []}
{"id": "2404.02405", "title": "TE-TAD: Towards Full End-to-End Temporal Action Detection via\n  Time-Aligned Coordinate Expression", "abstract": "  In this paper, we investigate that the normalized coordinate expression is a\nkey factor as reliance on hand-crafted components in query-based detectors for\ntemporal action detection (TAD). Despite significant advancements towards an\nend-to-end framework in object detection, query-based detectors have been\nlimited in achieving full end-to-end modeling in TAD. To address this issue, we\npropose \\modelname{}, a full end-to-end temporal action detection transformer\nthat integrates time-aligned coordinate expression. We reformulate coordinate\nexpression utilizing actual timeline values, ensuring length-invariant\nrepresentations from the extremely diverse video duration environment.\nFurthermore, our proposed adaptive query selection dynamically adjusts the\nnumber of queries based on video length, providing a suitable solution for\nvarying video durations compared to a fixed query set. Our approach not only\nsimplifies the TAD process by eliminating the need for hand-crafted components\nbut also significantly improves the performance of query-based detectors. Our\nTE-TAD outperforms the previous query-based detectors and achieves competitive\nperformance compared to state-of-the-art methods on popular benchmark datasets.\nCode is available at: https://github.com/Dotori-HJ/TE-TAD\n", "date": "2024-04-03", "categories": []}
{"id": "2304.00553", "title": "From Isolated Islands to Pangea: Unifying Semantic Space for Human\n  Action Understanding", "abstract": "  Action understanding has attracted long-term attention. It can be formed as\nthe mapping from the physical space to the semantic space. Typically,\nresearchers built datasets according to idiosyncratic choices to define classes\nand push the envelope of benchmarks respectively. Datasets are incompatible\nwith each other like \"Isolated Islands\" due to semantic gaps and various class\ngranularities, e.g., do housework in dataset A and wash plate in dataset B. We\nargue that we need a more principled semantic space to concentrate the\ncommunity efforts and use all datasets together to pursue generalizable action\nlearning. To this end, we design a structured action semantic space given verb\ntaxonomy hierarchy and covering massive actions. By aligning the classes of\nprevious datasets to our semantic space, we gather (image/video/skeleton/MoCap)\ndatasets into a unified database in a unified label system, i.e., bridging\n\"isolated islands\" into a \"Pangea\". Accordingly, we propose a novel model\nmapping from the physical space to semantic space to fully use Pangea. In\nextensive experiments, our new system shows significant superiority, especially\nin transfer learning. Our code and data will be made public at\nhttps://mvig-rhos.com/pangea.\n", "date": "2023-04-02", "categories": []}
{"id": "2312.07169", "title": "Semi-supervised Active Learning for Video Action Detection", "abstract": "  In this work, we focus on label efficient learning for video action\ndetection. We develop a novel semi-supervised active learning approach which\nutilizes both labeled as well as unlabeled data along with informative sample\nselection for action detection. Video action detection requires spatio-temporal\nlocalization along with classification, which poses several challenges for both\nactive learning informative sample selection as well as semi-supervised\nlearning pseudo label generation. First, we propose NoiseAug, a simple\naugmentation strategy which effectively selects informative samples for video\naction detection. Next, we propose fft-attention, a novel technique based on\nhigh-pass filtering which enables effective utilization of pseudo label for SSL\nin video action detection by emphasizing on relevant activity region within a\nvideo. We evaluate the proposed approach on three different benchmark datasets,\nUCF-101-24, JHMDB-21, and Youtube-VOS. First, we demonstrate its effectiveness\non video action detection where the proposed approach outperforms prior works\nin semi-supervised and weakly-supervised learning along with several baseline\napproaches in both UCF101-24 and JHMDB-21. Next, we also show its effectiveness\non Youtube-VOS for video object segmentation demonstrating its generalization\ncapability for other dense prediction tasks in videos. The code and models is\npublicly available at:\n\\url{https://github.com/AKASH2907/semi-sup-active-learning}.\n", "date": "2023-12-12", "categories": []}
{"id": "2404.02624", "title": "Multi-Scale Spatial-Temporal Self-Attention Graph Convolutional Networks\n  for Skeleton-based Action Recognition", "abstract": "  Skeleton-based gesture recognition methods have achieved high success using\nGraph Convolutional Network (GCN). In addition, context-dependent adaptive\ntopology as a neighborhood vertex information and attention mechanism leverages\na model to better represent actions. In this paper, we propose self-attention\nGCN hybrid model, Multi-Scale Spatial-Temporal self-attention (MSST)-GCN to\neffectively improve modeling ability to achieve state-of-the-art results on\nseveral datasets. We utilize spatial self-attention module with adaptive\ntopology to understand intra-frame interactions within a frame among different\nbody parts, and temporal self-attention module to examine correlations between\nframes of a node. These two are followed by multi-scale convolution network\nwith dilations, which not only captures the long-range temporal dependencies of\njoints but also the long-range spatial dependencies (i.e., long-distance\ndependencies) of node temporal behaviors. They are combined into high-level\nspatial-temporal representations and output the predicted action with the\nsoftmax classifier.\n", "date": "2024-04-03", "categories": []}
{"id": "2404.01591", "title": "Language Model Guided Interpretable Video Action Reasoning", "abstract": "  While neural networks have excelled in video action recognition tasks, their\nblack-box nature often obscures the understanding of their decision-making\nprocesses. Recent approaches used inherently interpretable models to analyze\nvideo actions in a manner akin to human reasoning. These models, however,\nusually fall short in performance compared to their black-box counterparts. In\nthis work, we present a new framework named Language-guided Interpretable\nAction Recognition framework (LaIAR). LaIAR leverages knowledge from language\nmodels to enhance both the recognition capabilities and the interpretability of\nvideo models. In essence, we redefine the problem of understanding video model\ndecisions as a task of aligning video and language models. Using the logical\nreasoning captured by the language model, we steer the training of the video\nmodel. This integrated approach not only improves the video model's\nadaptability to different domains but also boosts its overall performance.\nExtensive experiments on two complex video action datasets, Charades & CAD-120,\nvalidates the improved performance and interpretability of our LaIAR framework.\nThe code of LaIAR is available at https://github.com/NingWang2049/LaIAR.\n", "date": "2024-04-02", "categories": []}
{"id": "2404.01891", "title": "ASTRA: An Action Spotting TRAnsformer for Soccer Videos", "abstract": "  In this paper, we introduce ASTRA, a Transformer-based model designed for the\ntask of Action Spotting in soccer matches. ASTRA addresses several challenges\ninherent in the task and dataset, including the requirement for precise action\nlocalization, the presence of a long-tail data distribution, non-visibility in\ncertain actions, and inherent label noise. To do so, ASTRA incorporates (a) a\nTransformer encoder-decoder architecture to achieve the desired output temporal\nresolution and to produce precise predictions, (b) a balanced mixup strategy to\nhandle the long-tail distribution of the data, (c) an uncertainty-aware\ndisplacement head to capture the label variability, and (d) input audio signal\nto enhance detection of non-visible actions. Results demonstrate the\neffectiveness of ASTRA, achieving a tight Average-mAP of 66.82 on the test set.\nMoreover, in the SoccerNet 2023 Action Spotting challenge, we secure the 3rd\nposition with an Average-mAP of 70.21 on the challenge set.\n", "date": "2024-04-02", "categories": []}
{"id": "2404.01958", "title": "MESEN: Exploit Multimodal Data to Design Unimodal Human Activity\n  Recognition with Few Labels", "abstract": "  Human activity recognition (HAR) will be an essential function of various\nemerging applications. However, HAR typically encounters challenges related to\nmodality limitations and label scarcity, leading to an application gap between\ncurrent solutions and real-world requirements. In this work, we propose MESEN,\na multimodal-empowered unimodal sensing framework, to utilize unlabeled\nmultimodal data available during the HAR model design phase for unimodal HAR\nenhancement during the deployment phase. From a study on the impact of\nsupervised multimodal fusion on unimodal feature extraction, MESEN is designed\nto feature a multi-task mechanism during the multimodal-aided pre-training\nstage. With the proposed mechanism integrating cross-modal feature contrastive\nlearning and multimodal pseudo-classification aligning, MESEN exploits\nunlabeled multimodal data to extract effective unimodal features for each\nmodality. Subsequently, MESEN can adapt to downstream unimodal HAR with only a\nfew labeled samples. Extensive experiments on eight public multimodal datasets\ndemonstrate that MESEN achieves significant performance improvements over\nstate-of-the-art baselines in enhancing unimodal HAR by exploiting multimodal\ndata.\n", "date": "2024-04-02", "categories": []}
{"id": "2402.02210", "title": "Wavelet-Decoupling Contrastive Enhancement Network for Fine-Grained\n  Skeleton-Based Action Recognition", "abstract": "  Skeleton-based action recognition has attracted much attention, benefiting\nfrom its succinctness and robustness. However, the minimal inter-class\nvariation in similar action sequences often leads to confusion. The inherent\nspatiotemporal coupling characteristics make it challenging to mine the subtle\ndifferences in joint motion trajectories, which is critical for distinguishing\nconfusing fine-grained actions. To alleviate this problem, we propose a\nWavelet-Attention Decoupling (WAD) module that utilizes discrete wavelet\ntransform to effectively disentangle salient and subtle motion features in the\ntime-frequency domain. Then, the decoupling attention adaptively recalibrates\ntheir temporal responses. To further amplify the discrepancies in these subtle\nmotion features, we propose a Fine-grained Contrastive Enhancement (FCE) module\nto enhance attention towards trajectory features by contrastive learning.\nExtensive experiments are conducted on the coarse-grained dataset NTU RGB+D and\nthe fine-grained dataset FineGYM. Our methods perform competitively compared to\nstate-of-the-art methods and can discriminate confusing fine-grained actions\nwell.\n", "date": "2024-02-03", "categories": []}
{"id": "2404.00653", "title": "Dual DETRs for Multi-Label Temporal Action Detection", "abstract": "  Temporal Action Detection (TAD) aims to identify the action boundaries and\nthe corresponding category within untrimmed videos. Inspired by the success of\nDETR in object detection, several methods have adapted the query-based\nframework to the TAD task. However, these approaches primarily followed DETR to\npredict actions at the instance level (i.e., identify each action by its center\npoint), leading to sub-optimal boundary localization. To address this issue, we\npropose a new Dual-level query-based TAD framework, namely DualDETR, to detect\nactions from both instance-level and boundary-level. Decoding at different\nlevels requires semantics of different granularity, therefore we introduce a\ntwo-branch decoding structure. This structure builds distinctive decoding\nprocesses for different levels, facilitating explicit capture of temporal cues\nand semantics at each level. On top of the two-branch design, we present a\njoint query initialization strategy to align queries from both levels.\nSpecifically, we leverage encoder proposals to match queries from each level in\na one-to-one manner. Then, the matched queries are initialized using position\nand content prior from the matched action proposal. The aligned dual-level\nqueries can refine the matched proposal with complementary cues during\nsubsequent decoding. We evaluate DualDETR on three challenging multi-label TAD\nbenchmarks. The experimental results demonstrate the superior performance of\nDualDETR to the existing state-of-the-art methods, achieving a substantial\nimprovement under det-mAP and delivering impressive results under seg-mAP.\n", "date": "2024-03-31", "categories": []}
{"id": "2404.01051", "title": "Action Detection via an Image Diffusion Process", "abstract": "  Action detection aims to localize the starting and ending points of action\ninstances in untrimmed videos, and predict the classes of those instances. In\nthis paper, we make the observation that the outputs of the action detection\ntask can be formulated as images. Thus, from a novel perspective, we tackle\naction detection via a three-image generation process to generate starting\npoint, ending point and action-class predictions as images via our proposed\nAction Detection Image Diffusion (ADI-Diff) framework. Furthermore, since our\nimages differ from natural images and exhibit special properties, we further\nexplore a Discrete Action-Detection Diffusion Process and a Row-Column\nTransformer design to better handle their processing. Our ADI-Diff framework\nachieves state-of-the-art results on two widely-used datasets.\n", "date": "2024-04-01", "categories": []}
{"id": "2303.06346", "title": "3DInAction: Understanding Human Actions in 3D Point Clouds", "abstract": "  We propose a novel method for 3D point cloud action recognition.\nUnderstanding human actions in RGB videos has been widely studied in recent\nyears, however, its 3D point cloud counterpart remains under-explored. This is\nmostly due to the inherent limitation of the point cloud data modality -- lack\nof structure, permutation invariance, and varying number of points -- which\nmakes it difficult to learn a spatio-temporal representation. To address this\nlimitation, we propose the 3DinAction pipeline that first estimates patches\nmoving in time (t-patches) as a key building block, alongside a hierarchical\narchitecture that learns an informative spatio-temporal representation. We show\nthat our method achieves improved performance on existing datasets, including\nDFAUST and IKEA ASM. Code is publicly available at\nhttps://github.com/sitzikbs/3dincaction.\n", "date": "2023-03-11", "categories": []}
{"id": "2304.02560", "title": "VicTR: Video-conditioned Text Representations for Activity Recognition", "abstract": "  Vision-Language models (VLMs) have excelled in the image-domain -- especially\nin zero-shot settings -- thanks to the availability of vast pretraining data\n(i.e., paired image-text samples). However for videos, such paired data is not\nas abundant. Therefore, video-VLMs are usually designed by adapting pretrained\nimage-VLMs to the video-domain, instead of training from scratch. All such\nrecipes rely on augmenting visual embeddings with temporal information (i.e.,\nimage $\\rightarrow$ video), often keeping text embeddings unchanged or even\nbeing discarded. In this paper, we argue the contrary, that better video-VLMs\ncan be designed by focusing more on augmenting text, rather than visual\ninformation. More specifically, we introduce Video-conditioned Text\nRepresentations (VicTR): a form of text embeddings optimized w.r.t. visual\nembeddings, creating a more-flexible contrastive latent space. Our model can\nfurther make use of freely-available semantic information, in the form of\nvisually-grounded auxiliary text (e.g. object or scene information). We\nevaluate our model on few-shot, zero-shot (HMDB-51, UCF-101), short-form\n(Kinetics-400) and long-form (Charades) activity recognition benchmarks,\nshowing strong performance among video-VLMs.\n", "date": "2023-04-05", "categories": []}
{"id": "2403.20254", "title": "Benchmarking the Robustness of Temporal Action Detection Models Against\n  Temporal Corruptions", "abstract": "  Temporal action detection (TAD) aims to locate action positions and recognize\naction categories in long-term untrimmed videos. Although many methods have\nachieved promising results, their robustness has not been thoroughly studied.\nIn practice, we observe that temporal information in videos can be occasionally\ncorrupted, such as missing or blurred frames. Interestingly, existing methods\noften incur a significant performance drop even if only one frame is affected.\nTo formally evaluate the robustness, we establish two temporal corruption\nrobustness benchmarks, namely THUMOS14-C and ActivityNet-v1.3-C. In this paper,\nwe extensively analyze the robustness of seven leading TAD methods and obtain\nsome interesting findings: 1) Existing methods are particularly vulnerable to\ntemporal corruptions, and end-to-end methods are often more susceptible than\nthose with a pre-trained feature extractor; 2) Vulnerability mainly comes from\nlocalization error rather than classification error; 3) When corruptions occur\nin the middle of an action instance, TAD models tend to yield the largest\nperformance drop. Besides building a benchmark, we further develop a simple but\neffective robust training method to defend against temporal corruptions,\nthrough the FrameDrop augmentation and Temporal-Robust Consistency loss.\nRemarkably, our approach not only improves robustness but also yields promising\nimprovements on clean data. We believe that this study will serve as a\nbenchmark for future research in robust video analysis. Source code and models\nare available at https://github.com/Alvin-Zeng/temporal-robustness-benchmark.\n", "date": "2024-03-29", "categories": []}
{"id": "2312.00096", "title": "OST: Refining Text Knowledge with Optimal Spatio-Temporal Descriptor for\n  General Video Recognition", "abstract": "  Due to the resource-intensive nature of training vision-language models on\nexpansive video data, a majority of studies have centered on adapting\npre-trained image-language models to the video domain. Dominant pipelines\npropose to tackle the visual discrepancies with additional temporal learners\nwhile overlooking the substantial discrepancy for web-scaled descriptive\nnarratives and concise action category names, leading to less distinct semantic\nspace and potential performance limitations. In this work, we prioritize the\nrefinement of text knowledge to facilitate generalizable video recognition. To\naddress the limitations of the less distinct semantic space of category names,\nwe prompt a large language model (LLM) to augment action class names into\nSpatio-Temporal Descriptors thus bridging the textual discrepancy and serving\nas a knowledge base for general recognition. Moreover, to assign the best\ndescriptors with different video instances, we propose Optimal Descriptor\nSolver, forming the video recognition problem as solving the optimal matching\nflow across frame-level representations and descriptors. Comprehensive\nevaluations in zero-shot, few-shot, and fully supervised video recognition\nhighlight the effectiveness of our approach. Our best model achieves a\nstate-of-the-art zero-shot accuracy of 75.1% on Kinetics-600.\n", "date": "2023-11-30", "categories": []}
{"id": "2403.19225", "title": "Efficient and Effective Weakly-Supervised Action Segmentation via\n  Action-Transition-Aware Boundary Alignment", "abstract": "  Weakly-supervised action segmentation is a task of learning to partition a\nlong video into several action segments, where training videos are only\naccompanied by transcripts (ordered list of actions). Most of existing methods\nneed to infer pseudo segmentation for training by serial alignment between all\nframes and the transcript, which is time-consuming and hard to be parallelized\nwhile training. In this work, we aim to escape from this inefficient alignment\nwith massive but redundant frames, and instead to directly localize a few\naction transitions for pseudo segmentation generation, where a transition\nrefers to the change from an action segment to its next adjacent one in the\ntranscript. As the true transitions are submerged in noisy boundaries due to\nintra-segment visual variation, we propose a novel Action-Transition-Aware\nBoundary Alignment (ATBA) framework to efficiently and effectively filter out\nnoisy boundaries and detect transitions. In addition, to boost the semantic\nlearning in the case that noise is inevitably present in the pseudo\nsegmentation, we also introduce video-level losses to utilize the trusted\nvideo-level supervision. Extensive experiments show the effectiveness of our\napproach on both performance and training speed.\n", "date": "2024-03-28", "categories": []}
{"id": "2403.19316", "title": "Hypergraph-based Multi-View Action Recognition using Event Cameras", "abstract": "  Action recognition from video data forms a cornerstone with wide-ranging\napplications. Single-view action recognition faces limitations due to its\nreliance on a single viewpoint. In contrast, multi-view approaches capture\ncomplementary information from various viewpoints for improved accuracy.\nRecently, event cameras have emerged as innovative bio-inspired sensors,\nleading to advancements in event-based action recognition. However, existing\nworks predominantly focus on single-view scenarios, leaving a gap in multi-view\nevent data exploitation, particularly in challenges like information deficit\nand semantic misalignment. To bridge this gap, we introduce HyperMV, a\nmulti-view event-based action recognition framework. HyperMV converts discrete\nevent data into frame-like representations and extracts view-related features\nusing a shared convolutional network. By treating segments as vertices and\nconstructing hyperedges using rule-based and KNN-based strategies, a multi-view\nhypergraph neural network that captures relationships across viewpoint and\ntemporal features is established. The vertex attention hypergraph propagation\nis also introduced for enhanced feature fusion. To prompt research in this\narea, we present the largest multi-view event-based action dataset\n$\\text{THU}^{\\text{MV-EACT}}\\text{-50}$, comprising 50 actions from 6\nviewpoints, which surpasses existing datasets by over tenfold. Experimental\nresults show that HyperMV significantly outperforms baselines in both\ncross-subject and cross-view scenarios, and also exceeds the state-of-the-arts\nin frame-based multi-view action recognition.\n", "date": "2024-03-28", "categories": []}
{"id": "2207.12730", "title": "P2ANet: A Dataset and Benchmark for Dense Action Detection from Table\n  Tennis Match Broadcasting Videos", "abstract": "  While deep learning has been widely used for video analytics, such as video\nclassification and action detection, dense action detection with fast-moving\nsubjects from sports videos is still challenging. In this work, we release yet\nanother sports video benchmark \\TheName{} for \\emph{\\underline{P}}ing\n\\emph{\\underline{P}}ong-\\emph{\\underline{A}}ction detection, which consists of\n2,721 video clips collected from the broadcasting videos of professional table\ntennis matches in World Table Tennis Championships and Olympiads. We work with\na crew of table tennis professionals and referees on a specially designed\nannotation toolbox to obtain fine-grained action labels (in 14 classes) for\nevery ping-pong action that appeared in the dataset, and formulate two sets of\naction detection problems -- \\emph{action localization} and \\emph{action\nrecognition}. We evaluate a number of commonly-seen action recognition (e.g.,\nTSM, TSN, Video SwinTransformer, and Slowfast) and action localization models\n(e.g., BSN, BSN++, BMN, TCANet), using \\TheName{} for both problems, under\nvarious settings. These models can only achieve 48\\% area under the AR-AN curve\nfor localization and 82\\% top-one accuracy for recognition since the ping-pong\nactions are dense with fast-moving subjects but broadcasting videos are with\nonly 25 FPS. The results confirm that \\TheName{} is still a challenging task\nand can be used as a special benchmark for dense action detection from videos.\n", "date": "2022-07-26", "categories": []}
{"id": "2402.04599", "title": "Meet JEANIE: a Similarity Measure for 3D Skeleton Sequences via\n  Temporal-Viewpoint Alignment", "abstract": "  Video sequences exhibit significant nuisance variations (undesired effects)\nof speed of actions, temporal locations, and subjects' poses, leading to\ntemporal-viewpoint misalignment when comparing two sets of frames or evaluating\nthe similarity of two sequences. Thus, we propose Joint tEmporal and cAmera\nviewpoiNt alIgnmEnt (JEANIE) for sequence pairs. In particular, we focus on 3D\nskeleton sequences whose camera and subjects' poses can be easily manipulated\nin 3D. We evaluate JEANIE on skeletal Few-shot Action Recognition (FSAR), where\nmatching well temporal blocks (temporal chunks that make up a sequence) of\nsupport-query sequence pairs (by factoring out nuisance variations) is\nessential due to limited samples of novel classes. Given a query sequence, we\ncreate its several views by simulating several camera locations. For a support\nsequence, we match it with view-simulated query sequences, as in the popular\nDynamic Time Warping (DTW). Specifically, each support temporal block can be\nmatched to the query temporal block with the same or adjacent (next) temporal\nindex, and adjacent camera views to achieve joint local temporal-viewpoint\nwarping. JEANIE selects the smallest distance among matching paths with\ndifferent temporal-viewpoint warping patterns, an advantage over DTW which only\nperforms temporal alignment. We also propose an unsupervised FSAR akin to\nclustering of sequences with JEANIE as a distance measure. JEANIE achieves\nstate-of-the-art results on NTU-60, NTU-120, Kinetics-skeleton and UWA3D\nMultiview Activity II on supervised and unsupervised FSAR, and their\nmeta-learning inspired fusion.\n", "date": "2024-02-07", "categories": []}
{"id": "2403.15444", "title": "A Survey of IMU Based Cross-Modal Transfer Learning in Human Activity\n  Recognition", "abstract": "  Despite living in a multi-sensory world, most AI models are limited to\ntextual and visual understanding of human motion and behavior. In fact, full\nsituational awareness of human motion could best be understood through a\ncombination of sensors. In this survey we investigate how knowledge can be\ntransferred and utilized amongst modalities for Human Activity/Action\nRecognition (HAR), i.e. cross-modality transfer learning. We motivate the\nimportance and potential of IMU data and its applicability in cross-modality\nlearning as well as the importance of studying the HAR problem. We categorize\nHAR related tasks by time and abstractness and then compare various types of\nmultimodal HAR datasets. We also distinguish and expound on many related but\ninconsistently used terms in the literature, such as transfer learning, domain\nadaptation, representation learning, sensor fusion, and multimodal learning,\nand describe how cross-modal learning fits with all these concepts. We then\nreview the literature in IMU-based cross-modal transfer for HAR. The two main\napproaches for cross-modal transfer are instance-based transfer, where\ninstances of one modality are mapped to another (e.g. knowledge is transferred\nin the input space), or feature-based transfer, where the model relates the\nmodalities in an intermediate latent space (e.g. knowledge is transferred in\nthe feature space). Finally, we discuss future research directions and\napplications in cross-modal HAR.\n", "date": "2024-03-17", "categories": []}
{"id": "2403.16128", "title": "Enhancing Video Transformers for Action Understanding with VLM-aided\n  Training", "abstract": "  Owing to their ability to extract relevant spatio-temporal video embeddings,\nVision Transformers (ViTs) are currently the best performing models in video\naction understanding. However, their generalization over domains or datasets is\nsomewhat limited. In contrast, Visual Language Models (VLMs) have demonstrated\nexceptional generalization performance, but are currently unable to process\nvideos. Consequently, they cannot extract spatio-temporal patterns that are\ncrucial for action understanding. In this paper, we propose the Four-tiered\nPrompts (FTP) framework that takes advantage of the complementary strengths of\nViTs and VLMs. We retain ViTs' strong spatio-temporal representation ability\nbut improve the visual encodings to be more comprehensive and general by\naligning them with VLM outputs. The FTP framework adds four feature processors\nthat focus on specific aspects of human action in videos: action category,\naction components, action description, and context information. The VLMs are\nonly employed during training, and inference incurs a minimal computation cost.\nOur approach consistently yields state-of-the-art performance. For instance, we\nachieve remarkable top-1 accuracy of 93.8% on Kinetics-400 and 83.4% on\nSomething-Something V2, surpassing VideoMAEv2 by 2.8% and 2.6%, respectively.\n", "date": "2024-03-24", "categories": []}
{"id": "2403.14682", "title": "Deep Generative Domain Adaptation with Temporal Relation Knowledge for\n  Cross-User Activity Recognition", "abstract": "  In human activity recognition (HAR), the assumption that training and testing\ndata are independent and identically distributed (i.i.d.) often fails,\nparticularly in cross-user scenarios where data distributions vary\nsignificantly. This discrepancy highlights the limitations of conventional\ndomain adaptation methods in HAR, which typically overlook the inherent\ntemporal relations in time-series data. To bridge this gap, our study\nintroduces a Conditional Variational Autoencoder with Universal Sequence\nMapping (CVAE-USM) approach, which addresses the unique challenges of\ntime-series domain adaptation in HAR by relaxing the i.i.d. assumption and\nleveraging temporal relations to align data distributions effectively across\ndifferent users. This method combines the strengths of Variational Autoencoder\n(VAE) and Universal Sequence Mapping (USM) to capture and utilize common\ntemporal patterns between users for improved activity recognition. Our results,\nevaluated on two public HAR datasets (OPPT and PAMAP2), demonstrate that\nCVAE-USM outperforms existing state-of-the-art methods, offering a more\naccurate and generalizable solution for cross-user activity recognition.\n", "date": "2024-03-12", "categories": []}
{"id": "2311.15619", "title": "Align before Adapt: Leveraging Entity-to-Region Alignments for\n  Generalizable Video Action Recognition", "abstract": "  Large-scale visual-language pre-trained models have achieved significant\nsuccess in various video tasks. However, most existing methods follow an \"adapt\nthen align\" paradigm, which adapts pre-trained image encoders to model\nvideo-level representations and utilizes one-hot or text embedding of the\naction labels for supervision. This paradigm overlooks the challenge of mapping\nfrom static images to complicated activity concepts. In this paper, we propose\na novel \"Align before Adapt\" (ALT) paradigm. Prior to adapting to video\nrepresentation learning, we exploit the entity-to-region alignments for each\nframe. The alignments are fulfilled by matching the region-aware image\nembeddings to an offline-constructed text corpus. With the aligned entities, we\nfeed their text embeddings to a transformer-based video adapter as the queries,\nwhich can help extract the semantics of the most important entities from a\nvideo to a vector. This paradigm reuses the visual-language alignment of VLP\nduring adaptation and tries to explain an action by the underlying entities.\nThis helps understand actions by bridging the gap with complex activity\nsemantics, particularly when facing unfamiliar or unseen categories. ALT\ndemonstrates competitive performance while maintaining remarkably low\ncomputational costs. In fully supervised experiments, it achieves 88.1% top-1\naccuracy on Kinetics-400 with only 4947 GFLOPs. Moreover, ALT outperforms the\nprevious state-of-the-art methods in both zero-shot and few-shot experiments,\nemphasizing its superior generalizability across various learning scenarios.\n", "date": "2023-11-27", "categories": []}
{"id": "2403.14113", "title": "Spatio-Temporal Proximity-Aware Dual-Path Model for Panoramic Activity\n  Recognition", "abstract": "  Panoramic Activity Recognition (PAR) seeks to identify diverse human\nactivities across different scales, from individual actions to social group and\nglobal activities in crowded panoramic scenes. PAR presents two major\nchallenges: 1) recognizing the nuanced interactions among numerous individuals\nand 2) understanding multi-granular human activities. To address these, we\npropose Social Proximity-aware Dual-Path Network (SPDP-Net) based on two key\ndesign principles. First, while previous works often focus on spatial distance\namong individuals within an image, we argue to consider the spatio-temporal\nproximity. It is crucial for individual relation encoding to correctly\nunderstand social dynamics. Secondly, deviating from existing hierarchical\napproaches (individual-to-social-to-global activity), we introduce a dual-path\narchitecture for multi-granular activity recognition. This architecture\ncomprises individual-to-global and individual-to-social paths, mutually\nreinforcing each other's task with global-local context through multiple\nlayers. Through extensive experiments, we validate the effectiveness of the\nspatio-temporal proximity among individuals and the dual-path architecture in\nPAR. Furthermore, SPDP-Net achieves new state-of-the-art performance with\n46.5\\% of overall F1 score on JRDB-PAR dataset.\n", "date": "2024-03-21", "categories": []}
{"id": "2308.04834", "title": "View while Moving: Efficient Video Recognition in Long-untrimmed Videos", "abstract": "  Recent adaptive methods for efficient video recognition mostly follow the\ntwo-stage paradigm of \"preview-then-recognition\" and have achieved great\nsuccess on multiple video benchmarks. However, this two-stage paradigm involves\ntwo visits of raw frames from coarse-grained to fine-grained during inference\n(cannot be parallelized), and the captured spatiotemporal features cannot be\nreused in the second stage (due to varying granularity), being not friendly to\nefficiency and computation optimization. To this end, inspired by human\ncognition, we propose a novel recognition paradigm of \"View while Moving\" for\nefficient long-untrimmed video recognition. In contrast to the two-stage\nparadigm, our paradigm only needs to access the raw frame once. The two phases\nof coarse-grained sampling and fine-grained recognition are combined into\nunified spatiotemporal modeling, showing great performance. Moreover, we\ninvestigate the properties of semantic units in video and propose a\nhierarchical mechanism to efficiently capture and reason about the unit-level\nand video-level temporal semantics in long-untrimmed videos respectively.\nExtensive experiments on both long-untrimmed and short-trimmed videos\ndemonstrate that our approach outperforms state-of-the-art methods in terms of\naccuracy as well as efficiency, yielding new efficiency and accuracy trade-offs\nfor video spatiotemporal modeling.\n", "date": "2023-08-09", "categories": []}
{"id": "2403.11616", "title": "Multi-View Video-Based Learning: Leveraging Weak Labels for Frame-Level\n  Perception", "abstract": "  For training a video-based action recognition model that accepts multi-view\nvideo, annotating frame-level labels is tedious and difficult. However, it is\nrelatively easy to annotate sequence-level labels. This kind of coarse\nannotations are called as weak labels. However, training a multi-view\nvideo-based action recognition model with weak labels for frame-level\nperception is challenging. In this paper, we propose a novel learning\nframework, where the weak labels are first used to train a multi-view\nvideo-based base model, which is subsequently used for downstream frame-level\nperception tasks. The base model is trained to obtain individual latent\nembeddings for each view in the multi-view input. For training the model using\nthe weak labels, we propose a novel latent loss function. We also propose a\nmodel that uses the view-specific latent embeddings for downstream frame-level\naction recognition and detection tasks. The proposed framework is evaluated\nusing the MM Office dataset by comparing several baseline algorithms. The\nresults show that the proposed base model is effectively trained using weak\nlabels and the latent embeddings help the downstream models improve accuracy.\n", "date": "2024-03-18", "categories": []}
{"id": "2403.12450", "title": "Intention Action Anticipation Model with Guide-Feedback Loop Mechanism", "abstract": "  Anticipating human intention from videos has broad applications, such as\nautomatic driving, robot assistive technology, and virtual reality. This study\naddresses the problem of intention action anticipation using egocentric video\nsequences to estimate actions that indicate human intention. We propose a\nHierarchical Complete-Recent (HCR) information fusion model that makes full use\nof the features of the entire video sequence (i.e., complete features) and the\nfeatures of the video tail sequence (i.e., recent features). The HCR model has\ntwo primary mechanisms. The Guide-Feedback Loop (GFL) mechanism is proposed to\nmodel the relation between one recent feature and one complete feature. Based\non GFL, the MultiComplete-Recent Feature Aggregation (MCRFA) module is proposed\nto model the relation of one recent feature with multiscale complete features.\nBased on GFL and MCRFA, the HCR model can hierarchically explore the rich\ninterrelationships between multiscale complete features and multiscale recent\nfeatures. Through comparative and ablation experiments, we validate the\neffectiveness of our model on two well-known public datasets: EPIC-Kitchens and\nEGTEA Gaze+.\n", "date": "2024-03-19", "categories": []}
{"id": "2403.12534", "title": "ExACT: Language-guided Conceptual Reasoning and Uncertainty Estimation\n  for Event-based Action Recognition and More", "abstract": "  Event cameras have recently been shown beneficial for practical vision tasks,\nsuch as action recognition, thanks to their high temporal resolution, power\nefficiency, and reduced privacy concerns. However, current research is hindered\nby 1) the difficulty in processing events because of their prolonged duration\nand dynamic actions with complex and ambiguous semantics and 2) the redundant\naction depiction of the event frame representation with fixed stacks. We find\nlanguage naturally conveys abundant semantic information, rendering it\nstunningly superior in reducing semantic uncertainty. In light of this, we\npropose ExACT, a novel approach that, for the first time, tackles event-based\naction recognition from a cross-modal conceptualizing perspective. Our ExACT\nbrings two technical contributions. Firstly, we propose an adaptive\nfine-grained event (AFE) representation to adaptively filter out the repeated\nevents for the stationary objects while preserving dynamic ones. This subtly\nenhances the performance of ExACT without extra computational cost. Then, we\npropose a conceptual reasoning-based uncertainty estimation module, which\nsimulates the recognition process to enrich the semantic representation. In\nparticular, conceptual reasoning builds the temporal relation based on the\naction semantics, and uncertainty estimation tackles the semantic uncertainty\nof actions based on the distributional representation. Experiments show that\nour ExACT achieves superior recognition accuracy of 94.83%(+2.23%),\n90.10%(+37.47%) and 67.24% on PAF, HARDVS and our SeAct datasets respectively.\n", "date": "2024-03-19", "categories": []}
{"id": "2403.12710", "title": "Selective, Interpretable, and Motion Consistent Privacy Attribute\n  Obfuscation for Action Recognition", "abstract": "  Concerns for the privacy of individuals captured in public imagery have led\nto privacy-preserving action recognition. Existing approaches often suffer from\nissues arising through obfuscation being applied globally and a lack of\ninterpretability. Global obfuscation hides privacy sensitive regions, but also\ncontextual regions important for action recognition. Lack of interpretability\nerodes trust in these new technologies. We highlight the limitations of current\nparadigms and propose a solution: Human selected privacy templates that yield\ninterpretability by design, an obfuscation scheme that selectively hides\nattributes and also induces temporal consistency, which is important in action\nrecognition. Our approach is architecture agnostic and directly modifies input\nimagery, while existing approaches generally require architecture training. Our\napproach offers more flexibility, as no retraining is required, and outperforms\nalternatives on three widely used datasets.\n", "date": "2024-03-19", "categories": []}
{"id": "2403.01281", "title": "Fast Low-parameter Video Activity Localization in Collaborative Learning\n  Environments", "abstract": "  Research on video activity detection has primarily focused on identifying\nwell-defined human activities in short video segments. The majority of the\nresearch on video activity recognition is focused on the development of large\nparameter systems that require training on large video datasets. This paper\ndevelops a low-parameter, modular system with rapid inferencing capabilities\nthat can be trained entirely on limited datasets without requiring transfer\nlearning from large-parameter systems. The system can accurately detect and\nassociate specific activities with the students who perform the activities in\nreal-life classroom videos. Additionally, the paper develops an interactive\nweb-based application to visualize human activity maps over long real-life\nclassroom videos.\n", "date": "2024-03-02", "categories": []}
{"id": "2403.11189", "title": "Boosting Semi-Supervised Temporal Action Localization by Learning from\n  Non-Target Classes", "abstract": "  The crux of semi-supervised temporal action localization (SS-TAL) lies in\nexcavating valuable information from abundant unlabeled videos. However,\ncurrent approaches predominantly focus on building models that are robust to\nthe error-prone target class (i.e, the predicted class with the highest\nconfidence) while ignoring informative semantics within non-target classes.\nThis paper approaches SS-TAL from a novel perspective by advocating for\nlearning from non-target classes, transcending the conventional focus solely on\nthe target class. The proposed approach involves partitioning the label space\nof the predicted class distribution into distinct subspaces: target class,\npositive classes, negative classes, and ambiguous classes, aiming to mine both\npositive and negative semantics that are absent in the target class, while\nexcluding ambiguous classes. To this end, we first devise innovative strategies\nto adaptively select high-quality positive and negative classes from the label\nspace, by modeling both the confidence and rank of a class in relation to those\nof the target class. Then, we introduce novel positive and negative losses\ndesigned to guide the learning process, pushing predictions closer to positive\nclasses and away from negative classes. Finally, the positive and negative\nprocesses are integrated into a hybrid positive-negative learning framework,\nfacilitating the utilization of non-target classes in both labeled and\nunlabeled videos. Experimental results on THUMOS14 and ActivityNet v1.3\ndemonstrate the superiority of the proposed method over prior state-of-the-art\napproaches.\n", "date": "2024-03-17", "categories": []}
{"id": "2403.10082", "title": "CrossGLG: LLM Guides One-shot Skeleton-based 3D Action Recognition in a\n  Cross-level Manner", "abstract": "  Most existing one-shot skeleton-based action recognition focuses on raw\nlow-level information (e.g., joint location), and may suffer from local\ninformation loss and low generalization ability. To alleviate these, we propose\nto leverage text description generated from large language models (LLM) that\ncontain high-level human knowledge, to guide feature learning, in a\nglobal-local-global way. Particularly, during training, we design $2$ prompts\nto gain global and local text descriptions of each action from an LLM. We first\nutilize the global text description to guide the skeleton encoder focus on\ninformative joints (i.e.,global-to-local). Then we build non-local interaction\nbetween local text and joint features, to form the final global representation\n(i.e., local-to-global). To mitigate the asymmetry issue between the training\nand inference phases, we further design a dual-branch architecture that allows\nthe model to perform novel class inference without any text input, also making\nthe additional inference cost neglectable compared with the base skeleton\nencoder. Extensive experiments on three different benchmarks show that CrossGLG\nconsistently outperforms the existing SOTA methods with large margins, and the\ninference cost (model size) is only $2.8$\\% than the previous SOTA. CrossGLG\ncan also serve as a plug-and-play module that can substantially enhance the\nperformance of different SOTA skeleton encoders with a neglectable cost during\ninference. The source code will be released soon.\n", "date": "2024-03-15", "categories": []}
{"id": "2403.09056", "title": "Leveraging Foundation Model Automatic Data Augmentation Strategies and\n  Skeletal Points for Hands Action Recognition in Industrial Assembly Lines", "abstract": "  On modern industrial assembly lines, many intelligent algorithms have been\ndeveloped to replace or supervise workers. However, we found that there were\nbottlenecks in both training datasets and real-time performance when deploying\nalgorithms on actual assembly line. Therefore, we developed a promising\nstrategy for expanding industrial datasets, which utilized large models with\nstrong generalization abilities to achieve efficient, high-quality, and\nlarge-scale dataset expansion, solving the problem of insufficient and\nlow-quality industrial datasets. We also applied this strategy to video action\nrecognition. We proposed a method of converting hand action recognition\nproblems into hand skeletal trajectory classification problems, which solved\nthe real-time performance problem of industrial algorithms. In the \"hand\nmovements during wire insertion\" scenarios on the actual assembly line, the\naccuracy of hand action recognition reached 98.8\\%. We conducted detailed\nexperimental analysis to demonstrate the effectiveness and superiority of the\nmethod, and deployed the entire process on Midea's actual assembly line.\n", "date": "2024-03-14", "categories": []}
{"id": "2403.07354", "title": "BID: Boundary-Interior Decoding for Unsupervised Temporal Action\n  Localization Pre-Trainin", "abstract": "  Skeleton-based motion representations are robust for action localization and\nunderstanding for their invariance to perspective, lighting, and occlusion,\ncompared with images. Yet, they are often ambiguous and incomplete when taken\nout of context, even for human annotators. As infants discern gestures before\nassociating them with words, actions can be conceptualized before being\ngrounded with labels. Therefore, we propose the first unsupervised pre-training\nframework, Boundary-Interior Decoding (BID), that partitions a skeleton-based\nmotion sequence into discovered semantically meaningful pre-action segments. By\nfine-tuning our pre-training network with a small number of annotated data, we\nshow results out-performing SOTA methods by a large margin.\n", "date": "2024-03-12", "categories": []}
{"id": "2403.02753", "title": "Learning Group Activity Features Through Person Attribute Prediction", "abstract": "  This paper proposes Group Activity Feature (GAF) learning in which features\nof multi-person activity are learned as a compact latent vector. Unlike prior\nwork in which the manual annotation of group activities is required for\nsupervised learning, our method learns the GAF through person attribute\nprediction without group activity annotations. By learning the whole network in\nan end-to-end manner so that the GAF is required for predicting the person\nattributes of people in a group, the GAF is trained as the features of\nmulti-person activity. As a person attribute, we propose to use a person's\naction class and appearance features because the former is easy to annotate due\nto its simpleness, and the latter requires no manual annotation. In addition,\nwe introduce a location-guided attribute prediction to disentangle the complex\nGAF for extracting the features of each target person properly. Various\nexperimental results validate that our method outperforms SOTA methods\nquantitatively and qualitatively on two public datasets. Visualization of our\nGAF also demonstrates that our method learns the GAF representing fined-grained\ngroup activity classes. Code: https://github.com/chihina/GAFL-CVPR2024.\n", "date": "2024-03-05", "categories": []}
{"id": "2403.05557", "title": "Re-thinking Human Activity Recognition with Hierarchy-aware Label\n  Relationship Modeling", "abstract": "  Human Activity Recognition (HAR) has been studied for decades, from data\ncollection, learning models, to post-processing and result interpretations.\nHowever, the inherent hierarchy in the activities remains relatively\nunder-explored, despite its significant impact on model performance and\ninterpretation. In this paper, we propose H-HAR, by rethinking the HAR tasks\nfrom a fresh perspective by delving into their intricate global label\nrelationships. Rather than building multiple classifiers separately for\nmulti-layered activities, we explore the efficacy of a flat model enhanced with\ngraph-based label relationship modeling. Being hierarchy-aware, the graph-based\nlabel modeling enhances the fundamental HAR model, by incorporating intricate\nlabel relationships into the model. We validate the proposal with a multi-label\nclassifier on complex human activity data. The results highlight the advantages\nof the proposal, which can be vertically integrated into advanced HAR models to\nfurther enhance their performances.\n", "date": "2024-02-11", "categories": []}
{"id": "2403.06102", "title": "Coherent Temporal Synthesis for Incremental Action Segmentation", "abstract": "  Data replay is a successful incremental learning technique for images. It\nprevents catastrophic forgetting by keeping a reservoir of previous data,\noriginal or synthesized, to ensure the model retains past knowledge while\nadapting to novel concepts. However, its application in the video domain is\nrudimentary, as it simply stores frame exemplars for action recognition. This\npaper presents the first exploration of video data replay techniques for\nincremental action segmentation, focusing on action temporal modeling. We\npropose a Temporally Coherent Action (TCA) model, which represents actions\nusing a generative model instead of storing individual frames. The integration\nof a conditioning variable that captures temporal coherence allows our model to\nunderstand the evolution of action features over time. Therefore, action\nsegments generated by TCA for replay are diverse and temporally coherent. In a\n10-task incremental setup on the Breakfast dataset, our approach achieves\nsignificant increases in accuracy for up to 22% compared to the baselines.\n", "date": "2024-03-10", "categories": []}
{"id": "2403.06577", "title": "Transformer-based Fusion of 2D-pose and Spatio-temporal Embeddings for\n  Distracted Driver Action Recognition", "abstract": "  Classification and localization of driving actions over time is important for\nadvanced driver-assistance systems and naturalistic driving studies. Temporal\nlocalization is challenging because it requires robustness, reliability, and\naccuracy. In this study, we aim to improve the temporal localization and\nclassification accuracy performance by adapting video action recognition and 2D\nhuman-pose estimation networks to one model. Therefore, we design a\ntransformer-based fusion architecture to effectively combine 2D-pose features\nand spatio-temporal features. The model uses 2D-pose features as the positional\nembedding of the transformer architecture and spatio-temporal features as the\nmain input to the encoder of the transformer. The proposed solution is generic\nand independent of the camera numbers and positions, giving frame-based class\nprobabilities as output. Finally, the post-processing step combines information\nfrom different camera views to obtain final predictions and eliminate false\npositives. The model performs well on the A2 test set of the 2023 NVIDIA AI\nCity Challenge for naturalistic driving action recognition, achieving the\noverlap score of the organizer-defined distracted driver behaviour metric of\n0.5079.\n", "date": "2024-03-11", "categories": []}
{"id": "2403.06616", "title": "Density-Guided Label Smoothing for Temporal Localization of Driving\n  Actions", "abstract": "  Temporal localization of driving actions plays a crucial role in advanced\ndriver-assistance systems and naturalistic driving studies. However, this is a\nchallenging task due to strict requirements for robustness, reliability and\naccurate localization. In this work, we focus on improving the overall\nperformance by efficiently utilizing video action recognition networks and\nadapting these to the problem of action localization. To this end, we first\ndevelop a density-guided label smoothing technique based on label probability\ndistributions to facilitate better learning from boundary video-segments that\ntypically include multiple labels. Second, we design a post-processing step to\nefficiently fuse information from video-segments and multiple camera views into\nscene-level predictions, which facilitates elimination of false positives. Our\nmethodology yields a competitive performance on the A2 test set of the\nnaturalistic driving action recognition track of the 2022 NVIDIA AI City\nChallenge with an F1 score of 0.271.\n", "date": "2024-03-11", "categories": []}
{"id": "2403.06810", "title": "Deep Learning Approaches for Human Action Recognition in Video Data", "abstract": "  Human action recognition in videos is a critical task with significant\nimplications for numerous applications, including surveillance, sports\nanalytics, and healthcare. The challenge lies in creating models that are both\nprecise in their recognition capabilities and efficient enough for practical\nuse. This study conducts an in-depth analysis of various deep learning models\nto address this challenge. Utilizing a subset of the UCF101 Videos dataset, we\nfocus on Convolutional Neural Networks (CNNs), Recurrent Neural Networks\n(RNNs), and Two-Stream ConvNets. The research reveals that while CNNs\neffectively capture spatial features and RNNs encode temporal sequences,\nTwo-Stream ConvNets exhibit superior performance by integrating spatial and\ntemporal dimensions. These insights are distilled from the evaluation metrics\nof accuracy, precision, recall, and F1-score. The results of this study\nunderscore the potential of composite models in achieving robust human action\nrecognition and suggest avenues for future research in optimizing these models\nfor real-world deployment.\n", "date": "2024-03-11", "categories": []}
{"id": "2403.06978", "title": "Attention Prompt Tuning: Parameter-efficient Adaptation of Pre-trained\n  Models for Spatiotemporal Modeling", "abstract": "  In this paper, we introduce Attention Prompt Tuning (APT) - a computationally\nefficient variant of prompt tuning for video-based applications such as action\nrecognition. Prompt tuning approaches involve injecting a set of learnable\nprompts along with data tokens during fine-tuning while keeping the backbone\nfrozen. This approach greatly reduces the number of learnable parameters\ncompared to full tuning. For image-based downstream tasks, normally a couple of\nlearnable prompts achieve results close to those of full tuning. However,\nvideos, which contain more complex spatiotemporal information, require hundreds\nof tunable prompts to achieve reasonably good results. This reduces the\nparameter efficiency observed in images and significantly increases latency and\nthe number of floating-point operations (FLOPs) during inference. To tackle\nthese issues, we directly inject the prompts into the keys and values of the\nnon-local attention mechanism within the transformer block. Additionally, we\nintroduce a novel prompt reparameterization technique to make APT more robust\nagainst hyperparameter selection. The proposed APT approach greatly reduces the\nnumber of FLOPs and latency while achieving a significant performance boost\nover the existing parameter-efficient tuning methods on UCF101, HMDB51, and\nSSv2 datasets for action recognition. The code and pre-trained models are\navailable at https://github.com/wgcban/apt\n", "date": "2024-03-11", "categories": []}
{"id": "2208.04206", "title": "Vision-Based Activity Recognition in Children with Autism-Related\n  Behaviors", "abstract": "  Advances in machine learning and contactless sensors have enabled the\nunderstanding complex human behaviors in a healthcare setting. In particular,\nseveral deep learning systems have been introduced to enable comprehensive\nanalysis of neuro-developmental conditions such as Autism Spectrum Disorder\n(ASD). This condition affects children from their early developmental stages\nonwards, and diagnosis relies entirely on observing the child's behavior and\ndetecting behavioral cues. However, the diagnosis process is time-consuming as\nit requires long-term behavior observation, and the scarce availability of\nspecialists. We demonstrate the effect of a region-based computer vision system\nto help clinicians and parents analyze a child's behavior. For this purpose, we\nadopt and enhance a dataset for analyzing autism-related actions using videos\nof children captured in uncontrolled environments (e.g. videos collected with\nconsumer-grade cameras, in varied environments). The data is pre-processed by\ndetecting the target child in the video to reduce the impact of background\nnoise. Motivated by the effectiveness of temporal convolutional models, we\npropose both light-weight and conventional models capable of extracting action\nfeatures from video frames and classifying autism-related behaviors by\nanalyzing the relationships between frames in a video. Through extensive\nevaluations on the feature extraction and learning strategies, we demonstrate\nthat the best performance is achieved with an Inflated 3D Convnet and\nMulti-Stage Temporal Convolutional Networks, achieving a 0.83 Weighted F1-score\nfor classification of the three autism-related actions, outperforming existing\nmethods. We also propose a light-weight solution by employing the ESNet\nbackbone within the same system, achieving competitive results of 0.71 Weighted\nF1-score, and enabling potential deployment on embedded systems.\n", "date": "2022-08-08", "categories": []}
{"id": "2403.02727", "title": "HARGPT: Are LLMs Zero-Shot Human Activity Recognizers?", "abstract": "  There is an ongoing debate regarding the potential of Large Language Models\n(LLMs) as foundational models seamlessly integrated with Cyber-Physical Systems\n(CPS) for interpreting the physical world. In this paper, we carry out a case\nstudy to answer the following question: Are LLMs capable of zero-shot human\nactivity recognition (HAR). Our study, HARGPT, presents an affirmative answer\nby demonstrating that LLMs can comprehend raw IMU data and perform HAR tasks in\na zero-shot manner, with only appropriate prompts. HARGPT inputs raw IMU data\ninto LLMs and utilizes the role-play and think step-by-step strategies for\nprompting. We benchmark HARGPT on GPT4 using two public datasets of different\ninter-class similarities and compare various baselines both based on\ntraditional machine learning and state-of-the-art deep classification models.\nRemarkably, LLMs successfully recognize human activities from raw IMU data and\nconsistently outperform all the baselines on both datasets. Our findings\nindicate that by effective prompting, LLMs can interpret raw IMU data based on\ntheir knowledge base, possessing a promising potential to analyze raw sensor\ndata of the physical world effectively.\n", "date": "2024-03-05", "categories": []}
{"id": "2305.11468", "title": "Overcoming Topology Agnosticism: Enhancing Skeleton-Based Action\n  Recognition through Redefined Skeletal Topology Awareness", "abstract": "  Graph Convolutional Networks (GCNs) have long defined the state-of-the-art in\nskeleton-based action recognition, leveraging their ability to unravel the\ncomplex dynamics of human joint topology through the graph's adjacency matrix.\nHowever, an inherent flaw has come to light in these cutting-edge models: they\ntend to optimize the adjacency matrix jointly with the model weights. This\nprocess, while seemingly efficient, causes a gradual decay of bone connectivity\ndata, culminating in a model indifferent to the very topology it sought to map.\nAs a remedy, we propose a threefold strategy: (1) We forge an innovative\npathway that encodes bone connectivity by harnessing the power of graph\ndistances. This approach preserves the vital topological nuances often lost in\nconventional GCNs. (2) We highlight an oft-overlooked feature - the temporal\nmean of a skeletal sequence, which, despite its modest guise, carries highly\naction-specific information. (3) Our investigation revealed strong variations\nin joint-to-joint relationships across different actions. This finding exposes\nthe limitations of a single adjacency matrix in capturing the variations of\nrelational configurations emblematic of human movement, which we remedy by\nproposing an efficient refinement to Graph Convolutions (GC) - the BlockGC.\nThis evolution slashes parameters by a substantial margin (above 40%), while\nelevating performance beyond original GCNs. Our full model, the BlockGCN,\nestablishes new standards in skeleton-based action recognition for small model\nsizes. Its high accuracy, notably on the large-scale NTU RGB+D 120 dataset,\nstand as compelling proof of the efficacy of BlockGCN.\n", "date": "2023-05-19", "categories": []}
{"id": "2402.19258", "title": "MaskFi: Unsupervised Learning of WiFi and Vision Representations for\n  Multimodal Human Activity Recognition", "abstract": "  Human activity recognition (HAR) has been playing an increasingly important\nrole in various domains such as healthcare, security monitoring, and metaverse\ngaming. Though numerous HAR methods based on computer vision have been\ndeveloped to show prominent performance, they still suffer from poor robustness\nin adverse visual conditions in particular low illumination, which motivates\nWiFi-based HAR to serve as a good complementary modality. Existing solutions\nusing WiFi and vision modalities rely on massive labeled data that are very\ncumbersome to collect. In this paper, we propose a novel unsupervised\nmultimodal HAR solution, MaskFi, that leverages only unlabeled video and WiFi\nactivity data for model training. We propose a new algorithm, masked\nWiFi-vision modeling (MI2M), that enables the model to learn cross-modal and\nsingle-modal features by predicting the masked sections in representation\nlearning. Benefiting from our unsupervised learning procedure, the network\nrequires only a small amount of annotated data for finetuning and can adapt to\nthe new environment with better performance. We conduct extensive experiments\non two WiFi-vision datasets collected in-house, and our method achieves human\nactivity recognition and human identification in terms of both robustness and\naccuracy.\n", "date": "2024-02-29", "categories": []}
{"id": "2402.15952", "title": "ViSTec: Video Modeling for Sports Technique Recognition and Tactical\n  Analysis", "abstract": "  The immense popularity of racket sports has fueled substantial demand in\ntactical analysis with broadcast videos. However, existing manual methods\nrequire laborious annotation, and recent attempts leveraging video perception\nmodels are limited to low-level annotations like ball trajectories, overlooking\ntactics that necessitate an understanding of stroke techniques.\nState-of-the-art action segmentation models also struggle with technique\nrecognition due to frequent occlusions and motion-induced blurring in racket\nsports videos. To address these challenges, We propose ViSTec, a Video-based\nSports Technique recognition model inspired by human cognition that synergizes\nsparse visual data with rich contextual insights. Our approach integrates a\ngraph to explicitly model strategic knowledge in stroke sequences and enhance\ntechnique recognition with contextual inductive bias. A two-stage action\nperception model is jointly trained to align with the contextual knowledge in\nthe graph. Experiments demonstrate that our method outperforms existing models\nby a significant margin. Case studies with experts from the Chinese national\ntable tennis team validate our model's capacity to automate analysis for\ntechnical actions and tactical strategies. More details are available at:\nhttps://ViSTec2024.github.io/.\n", "date": "2024-02-25", "categories": []}
{"id": "2308.14500", "title": "LAC: Latent Action Composition for Skeleton-based Action Segmentation", "abstract": "  Skeleton-based action segmentation requires recognizing composable actions in\nuntrimmed videos. Current approaches decouple this problem by first extracting\nlocal visual features from skeleton sequences and then processing them by a\ntemporal model to classify frame-wise actions. However, their performances\nremain limited as the visual features cannot sufficiently express composable\nactions. In this context, we propose Latent Action Composition (LAC), a novel\nself-supervised framework aiming at learning from synthesized composable\nmotions for skeleton-based action segmentation. LAC is composed of a novel\ngeneration module towards synthesizing new sequences. Specifically, we design a\nlinear latent space in the generator to represent primitive motion. New\ncomposed motions can be synthesized by simply performing arithmetic operations\non latent representations of multiple input skeleton sequences. LAC leverages\nsuch synthesized sequences, which have large diversity and complexity, for\nlearning visual representations of skeletons in both sequence and frame spaces\nvia contrastive learning. The resulting visual encoder has a high expressive\npower and can be effectively transferred onto action segmentation tasks by\nend-to-end fine-tuning without the need for additional temporal models. We\nconduct a study focusing on transfer-learning and we show that representations\nlearned from pre-trained LAC outperform the state-of-the-art by a large margin\non TSU, Charades, PKU-MMD datasets.\n", "date": "2023-08-28", "categories": []}
{"id": "2402.05958", "title": "A comparative study on wearables and single-camera video for upper-limb\n  out-of-thelab activity recognition with different deep learning architectures", "abstract": "  The use of a wide range of computer vision solutions, and more recently\nhigh-end Inertial Measurement Units (IMU) have become increasingly popular for\nassessing human physical activity in clinical and research settings.\nNevertheless, to increase the feasibility of patient tracking in out-of-the-lab\nsettings, it is necessary to use a reduced number of devices for movement\nacquisition. Promising solutions in this context are IMU-based wearables and\nsingle camera systems. Additionally, the development of machine learning\nsystems able to recognize and digest clinically relevant data in-the-wild is\nneeded, and therefore determining the ideal input to those is crucial.\n", "date": "2024-02-04", "categories": []}
{"id": "2306.13149", "title": "\"Filling the Blanks'': Identifying Micro-activities that Compose Complex\n  Human Activities of Daily Living", "abstract": "  Complex activities of daily living (ADLs) often consist of multiple\nmicro-activities. When performed sequentially, these micro-activities help the\nuser accomplish the broad macro-activity. Naturally, a deeper understanding of\nthese micro-activities can help develop more sophisticated human activity\nrecognition (HAR) models and add explainability to their inferred conclusions.\nPrevious research has attempted to achieve this by utilizing fine-grained\nannotated data that provided the required supervision and rules for associating\nthe micro-activities to identify the macro-activity. However, this\n``bottom-up'' approach is unrealistic as getting such high-quality,\nfine-grained annotated sensor datasets is challenging, costly, and\ntime-consuming. Understanding this, in this paper, we develop AmicroN, which\nadapts a ``top-down'' approach by exploiting coarse-grained annotated data to\nexpand the macro-activities into their constituent micro-activities without any\nexternal supervision. In the backend, AmicroN uses \\textit{unsupervised}\nchange-point detection to search for the micro-activity boundaries across a\ncomplex ADL. Then, it applies a \\textit{generalized zero-shot} approach to\ncharacterize it. We evaluate AmicroN on two real-life publicly available\ndatasets and observe that AmicroN can identify the micro-activities with micro\nF\\textsubscript{1}-score $>0.75$ for both datasets. Additionally, we also\nperform an initial proof-of-concept on leveraging the state-of-the-art (SOTA)\nlarge language models (LLMs) with attribute embeddings predicted by AmicroN to\nenhance further the explainability surrounding the detection of\nmicro-activities.\n", "date": "2023-06-22", "categories": []}
{"id": "2310.18562", "title": "Optimization-Free Test-Time Adaptation for Cross-Person Activity\n  Recognition", "abstract": "  Human Activity Recognition (HAR) models often suffer from performance\ndegradation in real-world applications due to distribution shifts in activity\npatterns across individuals. Test-Time Adaptation (TTA) is an emerging learning\nparadigm that aims to utilize the test stream to adjust predictions in\nreal-time inference, which has not been explored in HAR before. However, the\nhigh computational cost of optimization-based TTA algorithms makes it\nintractable to run on resource-constrained edge devices. In this paper, we\npropose an Optimization-Free Test-Time Adaptation (OFTTA) framework for\nsensor-based HAR. OFTTA adjusts the feature extractor and linear classifier\nsimultaneously in an optimization-free manner. For the feature extractor, we\npropose Exponential DecayTest-time Normalization (EDTN) to replace the\nconventional batch normalization (CBN) layers. EDTN combines CBN and Test-time\nbatch Normalization (TBN) to extract reliable features against domain shifts\nwith TBN's influence decreasing exponentially in deeper layers. For the\nclassifier, we adjust the prediction by computing the distance between the\nfeature and the prototype, which is calculated by a maintained support set. In\naddition, the update of the support set is based on the pseudo label, which can\nbenefit from reliable features extracted by EDTN. Extensive experiments on\nthree public cross-person HAR datasets and two different TTA settings\ndemonstrate that OFTTA outperforms the state-of-the-art TTA approaches in both\nclassification performance and computational efficiency. Finally, we verify the\nsuperiority of our proposed OFTTA on edge devices, indicating possible\ndeployment in real applications. Our code is available at\nhttps://github.com/Claydon-Wang/OFTTA.\n", "date": "2023-10-28", "categories": []}
{"id": "2209.05653", "title": "Semantic2Graph: Graph-based Multi-modal Feature Fusion for Action\n  Segmentation in Videos", "abstract": "  Video action segmentation have been widely applied in many fields. Most\nprevious studies employed video-based vision models for this purpose. However,\nthey often rely on a large receptive field, LSTM or Transformer methods to\ncapture long-term dependencies within videos, leading to significant\ncomputational resource requirements. To address this challenge, graph-based\nmodel was proposed. However, previous graph-based models are less accurate.\nHence, this study introduces a graph-structured approach named Semantic2Graph,\nto model long-term dependencies in videos, thereby reducing computational costs\nand raise the accuracy. We construct a graph structure of video at the\nframe-level. Temporal edges are utilized to model the temporal relations and\naction order within videos. Additionally, we have designed positive and\nnegative semantic edges, accompanied by corresponding edge weights, to capture\nboth long-term and short-term semantic relationships in video actions. Node\nattributes encompass a rich set of multi-modal features extracted from video\ncontent, graph structures, and label text, encompassing visual, structural, and\nsemantic cues. To synthesize this multi-modal information effectively, we\nemploy a graph neural network (GNN) model to fuse multi-modal features for node\naction label classification. Experimental results demonstrate that\nSemantic2Graph outperforms state-of-the-art methods in terms of performance,\nparticularly on benchmark datasets such as GTEA and 50Salads. Multiple ablation\nexperiments further validate the effectiveness of semantic features in\nenhancing model performance. Notably, the inclusion of semantic edges in\nSemantic2Graph allows for the cost-effective capture of long-term dependencies,\naffirming its utility in addressing the challenges posed by computational\nresource constraints in video-based vision models.\n", "date": "2022-09-13", "categories": []}
{"id": "2307.07286", "title": "One-Shot Action Recognition via Multi-Scale Spatial-Temporal Skeleton\n  Matching", "abstract": "  One-shot skeleton action recognition, which aims to learn a skeleton action\nrecognition model with a single training sample, has attracted increasing\ninterest due to the challenge of collecting and annotating large-scale skeleton\naction data. However, most existing studies match skeleton sequences by\ncomparing their feature vectors directly which neglects spatial structures and\ntemporal orders of skeleton data. This paper presents a novel one-shot skeleton\naction recognition technique that handles skeleton action recognition via\nmulti-scale spatial-temporal feature matching. We represent skeleton data at\nmultiple spatial and temporal scales and achieve optimal feature matching from\ntwo perspectives. The first is multi-scale matching which captures the\nscale-wise semantic relevance of skeleton data at multiple spatial and temporal\nscales simultaneously. The second is cross-scale matching which handles\ndifferent motion magnitudes and speeds by capturing sample-wise relevance\nacross multiple scales. Extensive experiments over three large-scale datasets\n(NTU RGB+D, NTU RGB+D 120, and PKU-MMD) show that our method achieves superior\none-shot skeleton action recognition, and it outperforms the state-of-the-art\nconsistently by large margins.\n", "date": "2023-07-14", "categories": []}
{"id": "2402.02545", "title": "Classification of Tennis Actions Using Deep Learning", "abstract": "  Recent advances of deep learning makes it possible to identify specific\nevents in videos with greater precision. This has great relevance in sports\nlike tennis in order to e.g., automatically collect game statistics, or replay\nactions of specific interest for game strategy or player improvements. In this\npaper, we investigate the potential and the challenges of using deep learning\nto classify tennis actions. Three models of different size, all based on the\ndeep learning architecture SlowFast were trained and evaluated on the academic\ntennis dataset THETIS. The best models achieve a generalization accuracy of 74\n%, demonstrating a good performance for tennis action classification. We\nprovide an error analysis for the best model and pinpoint directions for\nimprovement of tennis datasets in general. We discuss the limitations of the\ndata set, general limitations of current publicly available tennis data-sets,\nand future steps needed to make progress.\n", "date": "2024-02-04", "categories": []}
{"id": "2402.03241", "title": "FROSTER: Frozen CLIP Is A Strong Teacher for Open-Vocabulary Action\n  Recognition", "abstract": "  In this paper, we introduce FROSTER, an effective framework for\nopen-vocabulary action recognition. The CLIP model has achieved remarkable\nsuccess in a range of image-based tasks, benefiting from its strong\ngeneralization capability stemming from pretaining on massive image-text pairs.\nHowever, applying CLIP directly to the open-vocabulary action recognition task\nis challenging due to the absence of temporal information in CLIP's\npretraining. Further, fine-tuning CLIP on action recognition datasets may lead\nto overfitting and hinder its generalizability, resulting in unsatisfactory\nresults when dealing with unseen actions.\n  To address these issues, FROSTER employs a residual feature distillation\napproach to ensure that CLIP retains its generalization capability while\neffectively adapting to the action recognition task. Specifically, the residual\nfeature distillation treats the frozen CLIP model as a teacher to maintain the\ngeneralizability exhibited by the original CLIP and supervises the feature\nlearning for the extraction of video-specific features to bridge the gap\nbetween images and videos. Meanwhile, it uses a residual sub-network for\nfeature distillation to reach a balance between the two distinct objectives of\nlearning generalizable and video-specific features.\n  We extensively evaluate FROSTER on open-vocabulary action recognition\nbenchmarks under both base-to-novel and cross-dataset settings. FROSTER\nconsistently achieves state-of-the-art performance on all datasets across the\nboard. Project page: https://visual-ai.github.io/froster.\n", "date": "2024-02-05", "categories": []}
{"id": "2206.02997", "title": "TadML: A fast temporal action detection with Mechanics-MLP", "abstract": "  Temporal Action Detection(TAD) is a crucial but challenging task in video\nunderstanding.It is aimed at detecting both the type and start-end frame for\neach action instance in a long, untrimmed video.Most current models adopt both\nRGB and Optical-Flow streams for the TAD task. Thus, original RGB frames must\nbe converted manually into Optical-Flow frames with additional computation and\ntime cost, which is an obstacle to achieve real-time processing. At present,\nmany models adopt two-stage strategies, which would slow the inference speed\ndown and complicatedly tuning on proposals generating.By comparison, we propose\na one-stage anchor-free temporal localization method with RGB stream only, in\nwhich a novel Newtonian Mechanics-MLP architecture is established. It has\ncomparable accuracy with all existing state-of-the-art models, while surpasses\nthe inference speed of these methods by a large margin. The typical inference\nspeed in this paper is astounding 4.44 video per second on THUMOS14. In\napplications, because there is no need to convert optical flow, the inference\nspeed will be faster.It also proves that MLP has great potential in downstream\ntasks such as TAD. The source code is available at\nhttps://github.com/BonedDeng/TadML\n", "date": "2022-06-07", "categories": []}
{"id": "2402.00293", "title": "FineBio: A Fine-Grained Video Dataset of Biological Experiments with\n  Hierarchical Annotation", "abstract": "  In the development of science, accurate and reproducible documentation of the\nexperimental process is crucial. Automatic recognition of the actions in\nexperiments from videos would help experimenters by complementing the recording\nof experiments. Towards this goal, we propose FineBio, a new fine-grained video\ndataset of people performing biological experiments. The dataset consists of\nmulti-view videos of 32 participants performing mock biological experiments\nwith a total duration of 14.5 hours. One experiment forms a hierarchical\nstructure, where a protocol consists of several steps, each further decomposed\ninto a set of atomic operations. The uniqueness of biological experiments is\nthat while they require strict adherence to steps described in each protocol,\nthere is freedom in the order of atomic operations. We provide hierarchical\nannotation on protocols, steps, atomic operations, object locations, and their\nmanipulation states, providing new challenges for structured activity\nunderstanding and hand-object interaction recognition. To find out challenges\non activity understanding in biological experiments, we introduce baseline\nmodels and results on four different tasks, including (i) step segmentation,\n(ii) atomic operation detection (iii) object detection, and (iv)\nmanipulated/affected object detection. Dataset and code are available from\nhttps://github.com/aistairc/FineBio.\n", "date": "2024-02-01", "categories": []}
{"id": "2401.17086", "title": "Active Generation Network of Human Skeleton for Action Recognition", "abstract": "  Data generation is a data augmentation technique for enhancing the\ngeneralization ability for skeleton-based human action recognition. Most\nexisting data generation methods face challenges to ensure the temporal\nconsistency of the dynamic information for action. In addition, the data\ngenerated by these methods lack diversity when only a few training samples are\navailable. To solve those problems, We propose a novel active generative\nnetwork (AGN), which can adaptively learn various action categories by motion\nstyle transfer to generate new actions when the data for a particular action is\nonly a single sample or few samples. The AGN consists of an action generation\nnetwork and an uncertainty metric network. The former, with ST-GCN as the\nBackbone, can implicitly learn the morphological features of the target action\nwhile preserving the category features of the source action. The latter guides\ngenerating actions. Specifically, an action recognition model generates\nprediction vectors for each action, which is then scored using an uncertainty\nmetric. Finally, UMN provides the uncertainty sampling basis for the generated\nactions.\n", "date": "2024-01-30", "categories": []}
{"id": "2308.12608", "title": "HR-Pro: Point-supervised Temporal Action Localization via Hierarchical\n  Reliability Propagation", "abstract": "  Point-supervised Temporal Action Localization (PSTAL) is an emerging research\ndirection for label-efficient learning. However, current methods mainly focus\non optimizing the network either at the snippet-level or the instance-level,\nneglecting the inherent reliability of point annotations at both levels. In\nthis paper, we propose a Hierarchical Reliability Propagation (HR-Pro)\nframework, which consists of two reliability-aware stages: Snippet-level\nDiscrimination Learning and Instance-level Completeness Learning, both stages\nexplore the efficient propagation of high-confidence cues in point annotations.\nFor snippet-level learning, we introduce an online-updated memory to store\nreliable snippet prototypes for each class. We then employ a Reliability-aware\nAttention Block to capture both intra-video and inter-video dependencies of\nsnippets, resulting in more discriminative and robust snippet representation.\nFor instance-level learning, we propose a point-based proposal generation\napproach as a means of connecting snippets and instances, which produces\nhigh-confidence proposals for further optimization at the instance level.\nThrough multi-level reliability-aware learning, we obtain more reliable\nconfidence scores and more accurate temporal boundaries of predicted proposals.\nOur HR-Pro achieves state-of-the-art performance on multiple challenging\nbenchmarks, including an impressive average mAP of 60.3% on THUMOS14. Notably,\nour HR-Pro largely surpasses all previous point-supervised methods, and even\noutperforms several competitive fully supervised methods. Code will be\navailable at https://github.com/pipixin321/HR-Pro.\n", "date": "2023-08-24", "categories": []}
{"id": "2401.16280", "title": "Cutup and Detect: Human Fall Detection on Cutup Untrimmed Videos Using a\n  Large Foundational Video Understanding Model", "abstract": "  This work explores the performance of a large video understanding foundation\nmodel on the downstream task of human fall detection on untrimmed video and\nleverages a pretrained vision transformer for multi-class action detection,\nwith classes: \"Fall\", \"Lying\" and \"Other/Activities of daily living (ADL)\". A\nmethod for temporal action localization that relies on a simple cutup of\nuntrimmed videos is demonstrated. The methodology includes a preprocessing\npipeline that converts datasets with timestamp action annotations into labeled\ndatasets of short action clips. Simple and effective clip-sampling strategies\nare introduced. The effectiveness of the proposed method has been empirically\nevaluated on the publicly available High-Quality Fall Simulation Dataset\n(HQFSD). The experimental results validate the performance of the proposed\npipeline. The results are promising for real-time application, and the falls\nare detected on video level with a state-of-the-art 0.96 F1 score on the HQFSD\ndataset under the given experimental settings. The source code will be made\navailable on GitHub.\n", "date": "2024-01-29", "categories": []}
{"id": "2212.04582", "title": "Towards Holistic Surgical Scene Understanding", "abstract": "  Most benchmarks for studying surgical interventions focus on a specific\nchallenge instead of leveraging the intrinsic complementarity among different\ntasks. In this work, we present a new experimental framework towards holistic\nsurgical scene understanding. First, we introduce the Phase, Step, Instrument,\nand Atomic Visual Action recognition (PSI-AVA) Dataset. PSI-AVA includes\nannotations for both long-term (Phase and Step recognition) and short-term\nreasoning (Instrument detection and novel Atomic Action recognition) in\nrobot-assisted radical prostatectomy videos. Second, we present Transformers\nfor Action, Phase, Instrument, and steps Recognition (TAPIR) as a strong\nbaseline for surgical scene understanding. TAPIR leverages our dataset's\nmulti-level annotations as it benefits from the learned representation on the\ninstrument detection task to improve its classification capacity. Our\nexperimental results in both PSI-AVA and other publicly available databases\ndemonstrate the adequacy of our framework to spur future research on holistic\nsurgical scene understanding.\n", "date": "2022-12-08", "categories": []}
{"id": "2311.12401", "title": "CASR: Refining Action Segmentation via Marginalizing Frame-levle Causal\n  Relationships", "abstract": "  Integrating deep learning and causal discovery has increased the\ninterpretability of Temporal Action Segmentation (TAS) tasks. However,\nframe-level causal relationships exist many complicated noises outside the\nsegment-level, making it infeasible to directly express macro action semantics.\nThus, we propose Causal Abstraction Segmentation Refiner (CASR), which can\nrefine TAS results from various models by enhancing video causality in\nmarginalizing frame-level casual relationships. Specifically, we define the\nequivalent frame-level casual model and segment-level causal model, so that the\ncausal adjacency matrix constructed from marginalized frame-level causal\nrelationships has the ability to represent the segmnet-level causal\nrelationships. CASR works out by reducing the difference in the causal\nadjacency matrix between we constructed and pre-segmentation results of\nbackbone models. In addition, we propose a novel evaluation metric Causal Edit\nDistance (CED) to evaluate the causal interpretability. Extensive experimental\nresults on mainstream datasets indicate that CASR significantly surpasses\nexisting various methods in action segmentation performance, as well as in\ncausal explainability and generalization.\n", "date": "2023-11-21", "categories": []}
{"id": "2401.14675", "title": "Multi-model learning by sequential reading of untrimmed videos for\n  action recognition", "abstract": "  We propose a new method for learning videos by aggregating multiple models by\nsequentially extracting video clips from untrimmed video. The proposed method\nreduces the correlation between clips by feeding clips to multiple models in\nturn and synchronizes these models through federated learning. Experimental\nresults show that the proposed method improves the performance compared to the\nno synchronization.\n", "date": "2024-01-26", "categories": []}
{"id": "2401.14838", "title": "Multi-modality action recognition based on dual feature shift in vehicle\n  cabin monitoring", "abstract": "  Driver Action Recognition (DAR) is crucial in vehicle cabin monitoring\nsystems. In real-world applications, it is common for vehicle cabins to be\nequipped with cameras featuring different modalities. However, multi-modality\nfusion strategies for the DAR task within car cabins have rarely been studied.\nIn this paper, we propose a novel yet efficient multi-modality driver action\nrecognition method based on dual feature shift, named DFS. DFS first integrates\ncomplementary features across modalities by performing modality feature\ninteraction. Meanwhile, DFS achieves the neighbour feature propagation within\nsingle modalities, by feature shifting among temporal frames. To learn common\npatterns and improve model efficiency, DFS shares feature extracting stages\namong multiple modalities. Extensive experiments have been carried out to\nverify the effectiveness of the proposed DFS model on the Drive\\&Act dataset.\nThe results demonstrate that DFS achieves good performance and improves the\nefficiency of multi-modality driver action recognition.\n", "date": "2024-01-26", "categories": []}
{"id": "2401.14034", "title": "Unsupervised Spatial-Temporal Feature Enrichment and Fidelity\n  Preservation Network for Skeleton based Action Recognition", "abstract": "  Unsupervised skeleton based action recognition has achieved remarkable\nprogress recently. Existing unsupervised learning methods suffer from severe\noverfitting problem, and thus small networks are used, significantly reducing\nthe representation capability. To address this problem, the overfitting\nmechanism behind the unsupervised learning for skeleton based action\nrecognition is first investigated. It is observed that the skeleton is already\na relatively high-level and low-dimension feature, but not in the same manifold\nas the features for action recognition. Simply applying the existing\nunsupervised learning method may tend to produce features that discriminate the\ndifferent samples instead of action classes, resulting in the overfitting\nproblem. To solve this problem, this paper presents an Unsupervised\nspatial-temporal Feature Enrichment and Fidelity Preservation framework\n(U-FEFP) to generate rich distributed features that contain all the information\nof the skeleton sequence. A spatial-temporal feature transformation subnetwork\nis developed using spatial-temporal graph convolutional network and graph\nconvolutional gate recurrent unit network as the basic feature extraction\nnetwork. The unsupervised Bootstrap Your Own Latent based learning is used to\ngenerate rich distributed features and the unsupervised pretext task based\nlearning is used to preserve the information of the skeleton sequence. The two\nunsupervised learning ways are collaborated as U-FEFP to produce robust and\ndiscriminative representations. Experimental results on three widely used\nbenchmarks, namely NTU-RGB+D-60, NTU-RGB+D-120 and PKU-MMD dataset, demonstrate\nthat the proposed U-FEFP achieves the best performance compared with the\nstate-of-the-art unsupervised learning methods. t-SNE illustrations further\nvalidate that U-FEFP can learn more discriminative features for unsupervised\nskeleton based action recognition.\n", "date": "2024-01-25", "categories": []}
{"id": "2401.14115", "title": "MIFI: MultI-camera Feature Integration for Roust 3D Distracted Driver\n  Activity Recognition", "abstract": "  Distracted driver activity recognition plays a critical role in risk\naversion-particularly beneficial in intelligent transportation systems.\nHowever, most existing methods make use of only the video from a single view\nand the difficulty-inconsistent issue is neglected. Different from them, in\nthis work, we propose a novel MultI-camera Feature Integration (MIFI) approach\nfor 3D distracted driver activity recognition by jointly modeling the data from\ndifferent camera views and explicitly re-weighting examples based on their\ndegree of difficulty. Our contributions are two-fold: (1) We propose a simple\nbut effective multi-camera feature integration framework and provide three\ntypes of feature fusion techniques. (2) To address the difficulty-inconsistent\nproblem in distracted driver activity recognition, a periodic learning method,\nnamed example re-weighting that can jointly learn the easy and hard samples, is\npresented. The experimental results on the 3MDAD dataset demonstrate that the\nproposed MIFI can consistently boost performance compared to single-view\nmodels.\n", "date": "2024-01-25", "categories": []}
{"id": "2401.13414", "title": "GTAutoAct: An Automatic Datasets Generation Framework Based on Game\n  Engine Redevelopment for Action Recognition", "abstract": "  Current datasets for action recognition tasks face limitations stemming from\ntraditional collection and generation methods, including the constrained range\nof action classes, absence of multi-viewpoint recordings, limited diversity,\npoor video quality, and labor-intensive manually collection. To address these\nchallenges, we introduce GTAutoAct, a innovative dataset generation framework\nleveraging game engine technology to facilitate advancements in action\nrecognition. GTAutoAct excels in automatically creating large-scale,\nwell-annotated datasets with extensive action classes and superior video\nquality. Our framework's distinctive contributions encompass: (1) it\ninnovatively transforms readily available coordinate-based 3D human motion into\nrotation-orientated representation with enhanced suitability in multiple\nviewpoints; (2) it employs dynamic segmentation and interpolation of rotation\nsequences to create smooth and realistic animations of action; (3) it offers\nextensively customizable animation scenes; (4) it implements an autonomous\nvideo capture and processing pipeline, featuring a randomly navigating camera,\nwith auto-trimming and labeling functionalities. Experimental results\nunderscore the framework's robustness and highlights its potential to\nsignificantly improve action recognition model training.\n", "date": "2024-01-24", "categories": []}
{"id": "2310.07189", "title": "SpikePoint: An Efficient Point-based Spiking Neural Network for Event\n  Cameras Action Recognition", "abstract": "  Event cameras are bio-inspired sensors that respond to local changes in light\nintensity and feature low latency, high energy efficiency, and high dynamic\nrange. Meanwhile, Spiking Neural Networks (SNNs) have gained significant\nattention due to their remarkable efficiency and fault tolerance. By\nsynergistically harnessing the energy efficiency inherent in event cameras and\nthe spike-based processing capabilities of SNNs, their integration could enable\nultra-low-power application scenarios, such as action recognition tasks.\nHowever, existing approaches often entail converting asynchronous events into\nconventional frames, leading to additional data mapping efforts and a loss of\nsparsity, contradicting the design concept of SNNs and event cameras. To\naddress this challenge, we propose SpikePoint, a novel end-to-end point-based\nSNN architecture. SpikePoint excels at processing sparse event cloud data,\neffectively extracting both global and local features through a singular-stage\nstructure. Leveraging the surrogate training method, SpikePoint achieves high\naccuracy with few parameters and maintains low power consumption, specifically\nemploying the identity mapping feature extractor on diverse datasets.\nSpikePoint achieves state-of-the-art (SOTA) performance on four event-based\naction recognition datasets using only 16 timesteps, surpassing other SNN\nmethods. Moreover, it also achieves SOTA performance across all methods on\nthree datasets, utilizing approximately 0.3\\% of the parameters and 0.5\\% of\npower consumption employed by artificial neural networks (ANNs). These results\nemphasize the significance of Point Cloud and pave the way for many\nultra-low-power event-based data processing applications.\n", "date": "2023-10-11", "categories": []}
{"id": "2401.11649", "title": "M2-CLIP: A Multimodal, Multi-task Adapting Framework for Video Action\n  Recognition", "abstract": "  Recently, the rise of large-scale vision-language pretrained models like\nCLIP, coupled with the technology of Parameter-Efficient FineTuning (PEFT), has\ncaptured substantial attraction in video action recognition. Nevertheless,\nprevailing approaches tend to prioritize strong supervised performance at the\nexpense of compromising the models' generalization capabilities during\ntransfer. In this paper, we introduce a novel Multimodal, Multi-task CLIP\nadapting framework named \\name to address these challenges, preserving both\nhigh supervised performance and robust transferability. Firstly, to enhance the\nindividual modality architectures, we introduce multimodal adapters to both the\nvisual and text branches. Specifically, we design a novel visual TED-Adapter,\nthat performs global Temporal Enhancement and local temporal Difference\nmodeling to improve the temporal representation capabilities of the visual\nencoder. Moreover, we adopt text encoder adapters to strengthen the learning of\nsemantic label information. Secondly, we design a multi-task decoder with a\nrich set of supervisory signals to adeptly satisfy the need for strong\nsupervised performance and generalization within a multimodal framework.\nExperimental results validate the efficacy of our approach, demonstrating\nexceptional performance in supervised learning while maintaining strong\ngeneralization in zero-shot scenarios.\n", "date": "2024-01-22", "categories": []}
{"id": "2401.11654", "title": "ActionHub: A Large-scale Action Video Description Dataset for Zero-shot\n  Action Recognition", "abstract": "  Zero-shot action recognition (ZSAR) aims to learn an alignment model between\nvideos and class descriptions of seen actions that is transferable to unseen\nactions. The text queries (class descriptions) used in existing ZSAR works,\nhowever, are often short action names that fail to capture the rich semantics\nin the videos, leading to misalignment. With the intuition that video content\ndescriptions (e.g., video captions) can provide rich contextual information of\nvisual concepts in videos, we propose to utilize human annotated video\ndescriptions to enrich the semantics of the class descriptions of each action.\nHowever, all existing action video description datasets are limited in terms of\nthe number of actions, the semantics of video descriptions, etc. To this end,\nwe collect a large-scale action video descriptions dataset named ActionHub,\nwhich covers a total of 1,211 common actions and provides 3.6 million action\nvideo descriptions. With the proposed ActionHub dataset, we further propose a\nnovel Cross-modality and Cross-action Modeling (CoCo) framework for ZSAR, which\nconsists of a Dual Cross-modality Alignment module and a Cross-action\nInvariance Mining module. Specifically, the Dual Cross-modality Alignment\nmodule utilizes both action labels and video descriptions from ActionHub to\nobtain rich class semantic features for feature alignment. The Cross-action\nInvariance Mining module exploits a cycle-reconstruction process between the\nclass semantic feature spaces of seen actions and unseen actions, aiming to\nguide the model to learn cross-action invariant representations. Extensive\nexperimental results demonstrate that our CoCo framework significantly\noutperforms the state-of-the-art on three popular ZSAR benchmarks (i.e.,\nKinetics-ZSAR, UCF101 and HMDB51) under two different learning protocols in\nZSAR. We will release our code, models, and the proposed ActionHub dataset.\n", "date": "2024-01-22", "categories": []}
{"id": "2208.09424", "title": "Hierarchical Compositional Representations for Few-shot Action\n  Recognition", "abstract": "  Recently action recognition has received more and more attention for its\ncomprehensive and practical applications in intelligent surveillance and\nhuman-computer interaction. However, few-shot action recognition has not been\nwell explored and remains challenging because of data scarcity. In this paper,\nwe propose a novel hierarchical compositional representations (HCR) learning\napproach for few-shot action recognition. Specifically, we divide a complicated\naction into several sub-actions by carefully designed hierarchical clustering\nand further decompose the sub-actions into more fine-grained spatially\nattentional sub-actions (SAS-actions). Although there exist large differences\nbetween base classes and novel classes, they can share similar patterns in\nsub-actions or SAS-actions. Furthermore, we adopt the Earth Mover's Distance in\nthe transportation problem to measure the similarity between video samples in\nterms of sub-action representations. It computes the optimal matching flows\nbetween sub-actions as distance metric, which is favorable for comparing\nfine-grained patterns. Extensive experiments show our method achieves the\nstate-of-the-art results on HMDB51, UCF101 and Kinetics datasets.\n", "date": "2022-08-19", "categories": []}
{"id": "2312.08010", "title": "EZ-CLIP: Efficient Zeroshot Video Action Recognition", "abstract": "  Recent advancements in large-scale pre-training of visual-language models on\npaired image-text data have demonstrated impressive generalization capabilities\nfor zero-shot tasks. Building on this success, efforts have been made to adapt\nthese image-based visual-language models, such as CLIP, for videos extending\ntheir zero-shot capabilities to the video domain. While these adaptations have\nshown promising results, they come at a significant computational cost and\nstruggle with effectively modeling the crucial temporal aspects inherent to the\nvideo domain. In this study, we present EZ-CLIP, a simple and efficient\nadaptation of CLIP that addresses these challenges. EZ-CLIP leverages temporal\nvisual prompting for seamless temporal adaptation, requiring no fundamental\nalterations to the core CLIP architecture while preserving its remarkable\ngeneralization abilities. Moreover, we introduce a novel learning objective\nthat guides the temporal visual prompts to focus on capturing motion, thereby\nenhancing its learning capabilities from video data. We conducted extensive\nexperiments on five different benchmark datasets, thoroughly evaluating EZ-CLIP\nfor zero-shot learning and base-to-novel video action recognition, and also\ndemonstrating its potential for few-shot generalization.Impressively, with a\nmere 5.2 million learnable parameters (as opposed to the 71.1 million in the\nprior best model), EZ-CLIP can be efficiently trained on a single GPU,\noutperforming existing approaches in several evaluations.\n", "date": "2023-12-13", "categories": []}
{"id": "2401.09606", "title": "Robustness Evaluation of Machine Learning Models for Robot Arm Action\n  Recognition in Noisy Environments", "abstract": "  In the realm of robot action recognition, identifying distinct but spatially\nproximate arm movements using vision systems in noisy environments poses a\nsignificant challenge. This paper studies robot arm action recognition in noisy\nenvironments using machine learning techniques. Specifically, a vision system\nis used to track the robot's movements followed by a deep learning model to\nextract the arm's key points. Through a comparative analysis of machine\nlearning methods, the effectiveness and robustness of this model are assessed\nin noisy environments. A case study was conducted using the Tic-Tac-Toe game in\na 3-by-3 grid environment, where the focus is to accurately identify the\nactions of the arms in selecting specific locations within this constrained\nenvironment. Experimental results show that our approach can achieve precise\nkey point detection and action classification despite the addition of noise and\nuncertainties to the dataset.\n", "date": "2024-01-17", "categories": []}
{"id": "2401.08345", "title": "Multi-view Distillation based on Multi-modal Fusion for Few-shot Action\n  Recognition(CLIP-$\\mathrm{M^2}$DF)", "abstract": "  In recent years, few-shot action recognition has attracted increasing\nattention. It generally adopts the paradigm of meta-learning. In this field,\novercoming the overlapping distribution of classes and outliers is still a\nchallenging problem based on limited samples. We believe the combination of\nMulti-modal and Multi-view can improve this issue depending on information\ncomplementarity. Therefore, we propose a method of Multi-view Distillation\nbased on Multi-modal Fusion. Firstly, a Probability Prompt Selector for the\nquery is constructed to generate probability prompt embedding based on the\ncomparison score between the prompt embeddings of the support and the visual\nembedding of the query. Secondly, we establish a Multi-view. In each view, we\nfuse the prompt embedding as consistent information with visual and the global\nor local temporal context to overcome the overlapping distribution of classes\nand outliers. Thirdly, we perform the distance fusion for the Multi-view and\nthe mutual distillation of matching ability from one to another, enabling the\nmodel to be more robust to the distribution bias. Our code is available at the\nURL: \\url{https://github.com/cofly2014/MDMF}.\n", "date": "2024-01-16", "categories": []}
{"id": "2307.08994", "title": "Human Action Recognition in Still Images Using ConViT", "abstract": "  Understanding the relationship between different parts of an image is crucial\nin a variety of applications, including object recognition, scene\nunderstanding, and image classification. Despite the fact that Convolutional\nNeural Networks (CNNs) have demonstrated impressive results in classifying and\ndetecting objects, they lack the capability to extract the relationship between\ndifferent parts of an image, which is a crucial factor in Human Action\nRecognition (HAR). To address this problem, this paper proposes a new module\nthat functions like a convolutional layer that uses Vision Transformer (ViT).\nIn the proposed model, the Vision Transformer can complement a convolutional\nneural network in a variety of tasks by helping it to effectively extract the\nrelationship among various parts of an image. It is shown that the proposed\nmodel, compared to a simple CNN, can extract meaningful parts of an image and\nsuppress the misleading parts. The proposed model has been evaluated on the\nStanford40 and PASCAL VOC 2012 action datasets and has achieved 95.5% mean\nAverage Precision (mAP) and 91.5% mAP results, respectively, which are\npromising compared to other state-of-the-art methods.\n", "date": "2023-07-18", "categories": []}
{"id": "2309.12009", "title": "Elevating Skeleton-Based Action Recognition with Efficient\n  Multi-Modality Self-Supervision", "abstract": "  Self-supervised representation learning for human action recognition has\ndeveloped rapidly in recent years. Most of the existing works are based on\nskeleton data while using a multi-modality setup. These works overlooked the\ndifferences in performance among modalities, which led to the propagation of\nerroneous knowledge between modalities while only three fundamental modalities,\ni.e., joints, bones, and motions are used, hence no additional modalities are\nexplored.\n  In this work, we first propose an Implicit Knowledge Exchange Module (IKEM)\nwhich alleviates the propagation of erroneous knowledge between low-performance\nmodalities. Then, we further propose three new modalities to enrich the\ncomplementary information between modalities. Finally, to maintain efficiency\nwhen introducing new modalities, we propose a novel teacher-student framework\nto distill the knowledge from the secondary modalities into the mandatory\nmodalities considering the relationship constrained by anchors, positives, and\nnegatives, named relational cross-modality knowledge distillation. The\nexperimental results demonstrate the effectiveness of our approach, unlocking\nthe efficient use of skeleton-based multi-modality data. Source code will be\nmade publicly available at https://github.com/desehuileng0o0/IKEM.\n", "date": "2023-09-21", "categories": []}
{"id": "2401.05477", "title": "Standardizing Your Training Process for Human Activity Recognition\n  Models: A Comprehensive Review in the Tunable Factors", "abstract": "  In recent years, deep learning has emerged as a potent tool across a\nmultitude of domains, leading to a surge in research pertaining to its\napplication in the wearable human activity recognition (WHAR) domain. Despite\nthe rapid development, concerns have been raised about the lack of\nstandardization and consistency in the procedures used for experimental model\ntraining, which may affect the reproducibility and reliability of research\nresults. In this paper, we provide an exhaustive review of contemporary deep\nlearning research in the field of WHAR and collate information pertaining to\nthe training procedure employed in various studies. Our findings suggest that a\nmajor trend is the lack of detail provided by model training protocols.\nBesides, to gain a clearer understanding of the impact of missing descriptions,\nwe utilize a control variables approach to assess the impact of key tunable\ncomponents (e.g., optimization techniques and early stopping criteria) on the\ninter-subject generalization capabilities of HAR models. With insights from the\nanalyses, we define a novel integrated training procedure tailored to the WHAR\nmodel. Empirical results derived using five well-known \\ac{whar} benchmark\ndatasets and three classical HAR model architectures demonstrate the\neffectiveness of our proposed methodology: in particular, there is a\nsignificant improvement in macro F1 leave one subject out cross-validation\nperformance.\n", "date": "2024-01-10", "categories": []}
{"id": "2307.10763", "title": "Actor-agnostic Multi-label Action Recognition with Multi-modal Query", "abstract": "  Existing action recognition methods are typically actor-specific due to the\nintrinsic topological and apparent differences among the actors. This requires\nactor-specific pose estimation (e.g., humans vs. animals), leading to\ncumbersome model design complexity and high maintenance costs. Moreover, they\noften focus on learning the visual modality alone and single-label\nclassification whilst neglecting other available information sources (e.g.,\nclass name text) and the concurrent occurrence of multiple actions. To overcome\nthese limitations, we propose a new approach called 'actor-agnostic multi-modal\nmulti-label action recognition,' which offers a unified solution for various\ntypes of actors, including humans and animals. We further formulate a novel\nMulti-modal Semantic Query Network (MSQNet) model in a transformer-based object\ndetection framework (e.g., DETR), characterized by leveraging visual and\ntextual modalities to represent the action classes better. The elimination of\nactor-specific model designs is a key advantage, as it removes the need for\nactor pose estimation altogether. Extensive experiments on five publicly\navailable benchmarks show that our MSQNet consistently outperforms the prior\narts of actor-specific alternatives on human and animal single- and multi-label\naction recognition tasks by up to 50%. Code is made available at\nhttps://github.com/mondalanindya/MSQNet.\n", "date": "2023-07-20", "categories": []}
{"id": "2401.04975", "title": "HaltingVT: Adaptive Token Halting Transformer for Efficient Video\n  Recognition", "abstract": "  Action recognition in videos poses a challenge due to its high computational\ncost, especially for Joint Space-Time video transformers (Joint VT). Despite\ntheir effectiveness, the excessive number of tokens in such architectures\nsignificantly limits their efficiency. In this paper, we propose HaltingVT, an\nefficient video transformer adaptively removing redundant video patch tokens,\nwhich is primarily composed of a Joint VT and a Glimpser module. Specifically,\nHaltingVT applies data-adaptive token reduction at each layer, resulting in a\nsignificant reduction in the overall computational cost. Besides, the Glimpser\nmodule quickly removes redundant tokens in shallow transformer layers, which\nmay even be misleading for video recognition tasks based on our observations.\nTo further encourage HaltingVT to focus on the key motion-related information\nin videos, we design an effective Motion Loss during training. HaltingVT\nacquires video analysis capabilities and token halting compression strategies\nsimultaneously in a unified training process, without requiring additional\ntraining procedures or sub-networks. On the Mini-Kinetics dataset, we achieved\n75.0% top-1 ACC with 24.2 GFLOPs, as well as 67.2% top-1 ACC with an extremely\nlow 9.9 GFLOPs. The code is available at\nhttps://github.com/dun-research/HaltingVT.\n", "date": "2024-01-10", "categories": []}
{"id": "2401.04150", "title": "Two-stream joint matching method based on contrastive learning for\n  few-shot action recognition", "abstract": "  Although few-shot action recognition based on metric learning paradigm has\nachieved significant success, it fails to address the following issues: (1)\ninadequate action relation modeling and underutilization of multi-modal\ninformation; (2) challenges in handling video matching problems with different\nlengths and speeds, and video matching problems with misalignment of video\nsub-actions. To address these issues, we propose a Two-Stream Joint Matching\nmethod based on contrastive learning (TSJM), which consists of two modules:\nMulti-modal Contrastive Learning Module (MCL) and Joint Matching Module (JMM).\nThe objective of the MCL is to extensively investigate the inter-modal mutual\ninformation relationships, thereby thoroughly extracting modal information to\nenhance the modeling of action relationships. The JMM aims to simultaneously\naddress the aforementioned video matching problems. The effectiveness of the\nproposed method is evaluated on two widely used few shot action recognition\ndatasets, namely, SSv2 and Kinetics. Comprehensive ablation experiments are\nalso conducted to substantiate the efficacy of our proposed approach.\n", "date": "2024-01-08", "categories": []}
{"id": "2307.07469", "title": "Interactive Spatiotemporal Token Attention Network for Skeleton-based\n  General Interactive Action Recognition", "abstract": "  Recognizing interactive action plays an important role in human-robot\ninteraction and collaboration. Previous methods use late fusion and\nco-attention mechanism to capture interactive relations, which have limited\nlearning capability or inefficiency to adapt to more interacting entities. With\nassumption that priors of each entity are already known, they also lack\nevaluations on a more general setting addressing the diversity of subjects. To\naddress these problems, we propose an Interactive Spatiotemporal Token\nAttention Network (ISTA-Net), which simultaneously model spatial, temporal, and\ninteractive relations. Specifically, our network contains a tokenizer to\npartition Interactive Spatiotemporal Tokens (ISTs), which is a unified way to\nrepresent motions of multiple diverse entities. By extending the entity\ndimension, ISTs provide better interactive representations. To jointly learn\nalong three dimensions in ISTs, multi-head self-attention blocks integrated\nwith 3D convolutions are designed to capture inter-token correlations. When\nmodeling correlations, a strict entity ordering is usually irrelevant for\nrecognizing interactive actions. To this end, Entity Rearrangement is proposed\nto eliminate the orderliness in ISTs for interchangeable entities. Extensive\nexperiments on four datasets verify the effectiveness of ISTA-Net by\noutperforming state-of-the-art methods. Our code is publicly available at\nhttps://github.com/Necolizer/ISTA-Net\n", "date": "2023-07-14", "categories": []}
{"id": "2401.03262", "title": "Group Activity Recognition using Unreliable Tracked Pose", "abstract": "  Group activity recognition in video is a complex task due to the need for a\nmodel to recognise the actions of all individuals in the video and their\ncomplex interactions. Recent studies propose that optimal performance is\nachieved by individually tracking each person and subsequently inputting the\nsequence of poses or cropped images/optical flow into a model. This helps the\nmodel to recognise what actions each person is performing before they are\nmerged to arrive at the group action class. However, all previous models are\nhighly reliant on high quality tracking and have only been evaluated using\nground truth tracking information. In practice it is almost impossible to\nachieve highly reliable tracking information for all individuals in a group\nactivity video. We introduce an innovative deep learning-based group activity\nrecognition approach called Rendered Pose based Group Activity Recognition\nSystem (RePGARS) which is designed to be tolerant of unreliable tracking and\npose information. Experimental results confirm that RePGARS outperforms all\nexisting group activity recognition algorithms tested which do not use ground\ntruth detection and tracking information.\n", "date": "2024-01-06", "categories": []}
{"id": "2211.09310", "title": "Language-Assisted Deep Learning for Autistic Behaviors Recognition", "abstract": "  Correctly recognizing the behaviors of children with Autism Spectrum Disorder\n(ASD) is of vital importance for the diagnosis of Autism and timely early\nintervention. However, the observation and recording during the treatment from\nthe parents of autistic children may not be accurate and objective. In such\ncases, automatic recognition systems based on computer vision and machine\nlearning (in particular deep learning) technology can alleviate this issue to a\nlarge extent. Existing human action recognition models can now achieve\npersuasive performance on challenging activity datasets, e.g. daily activity,\nand sports activity. However, problem behaviors in children with ASD are very\ndifferent from these general activities, and recognizing these problem\nbehaviors via computer vision is less studied. In this paper, we first evaluate\na strong baseline for action recognition, i.e. Video Swin Transformer, on two\nautism behaviors datasets (SSBD and ESBD) and show that it can achieve high\naccuracy and outperform the previous methods by a large margin, demonstrating\nthe feasibility of vision-based problem behaviors recognition. Moreover, we\npropose language-assisted training to further enhance the action recognition\nperformance. Specifically, we develop a two-branch multimodal deep learning\nframework by incorporating the \"freely available\" language description for each\ntype of problem behavior. Experimental results demonstrate that incorporating\nadditional language supervision can bring an obvious performance boost for the\nautism problem behaviors recognition task as compared to using the video\ninformation only (i.e. 3.49% improvement on ESBD and 1.46% on SSBD).\n", "date": "2022-11-17", "categories": []}
{"id": "2401.02138", "title": "Explore Human Parsing Modality for Action Recognition", "abstract": "  Multimodal-based action recognition methods have achieved high success using\npose and RGB modality. However, skeletons sequences lack appearance depiction\nand RGB images suffer irrelevant noise due to modality limitations. To address\nthis, we introduce human parsing feature map as a novel modality, since it can\nselectively retain effective semantic features of the body parts, while\nfiltering out most irrelevant noise. We propose a new dual-branch framework\ncalled Ensemble Human Parsing and Pose Network (EPP-Net), which is the first to\nleverage both skeletons and human parsing modalities for action recognition.\nThe first human pose branch feeds robust skeletons in graph convolutional\nnetwork to model pose features, while the second human parsing branch also\nleverages depictive parsing feature maps to model parsing festures via\nconvolutional backbones. The two high-level features will be effectively\ncombined through a late fusion strategy for better action recognition.\nExtensive experiments on NTU RGB+D and NTU RGB+D 120 benchmarks consistently\nverify the effectiveness of our proposed EPP-Net, which outperforms the\nexisting action recognition methods. Our code is available at:\nhttps://github.com/liujf69/EPP-Net-Action.\n", "date": "2024-01-04", "categories": []}
{"id": "2311.11467", "title": "SkateboardAI: The Coolest Video Action Recognition for Skateboarding", "abstract": "  Impressed by the coolest skateboarding sports program from 2021 Tokyo Olympic\nGames, we are the first to curate the original real-world video datasets\n\"SkateboardAI\" in the wild, even self-design and implement diverse uni-modal\nand multi-modal video action recognition approaches to recognize different\ntricks accurately. For uni-modal methods, we separately apply (1) CNN and LSTM;\n(2) CNN and BiLSTM; (3) CNN and BiLSTM with effective attention mechanisms; (4)\nTransformer-based action recognition pipeline. Transferred to the multi-modal\nconditions, we investigated the two-stream Inflated-3D architecture on\n\"SkateboardAI\" datasets to compare its performance with uni-modal cases. In\nsum, our objective is developing an excellent AI sport referee for the coolest\nskateboarding competitions.\n", "date": "2023-08-02", "categories": []}
{"id": "2401.00921", "title": "Skeleton2vec: A Self-supervised Learning Framework with Contextualized\n  Target Representations for Skeleton Sequence", "abstract": "  Self-supervised pre-training paradigms have been extensively explored in the\nfield of skeleton-based action recognition. In particular, methods based on\nmasked prediction have pushed the performance of pre-training to a new height.\nHowever, these methods take low-level features, such as raw joint coordinates\nor temporal motion, as prediction targets for the masked regions, which is\nsuboptimal. In this paper, we show that using high-level contextualized\nfeatures as prediction targets can achieve superior performance. Specifically,\nwe propose Skeleton2vec, a simple and efficient self-supervised 3D action\nrepresentation learning framework, which utilizes a transformer-based teacher\nencoder taking unmasked training samples as input to create latent\ncontextualized representations as prediction targets. Benefiting from the\nself-attention mechanism, the latent representations generated by the teacher\nencoder can incorporate the global context of the entire training samples,\nleading to a richer training task. Additionally, considering the high temporal\ncorrelations in skeleton sequences, we propose a motion-aware tube masking\nstrategy which divides the skeleton sequence into several tubes and performs\npersistent masking within each tube based on motion priors, thus forcing the\nmodel to build long-range spatio-temporal connections and focus on\naction-semantic richer regions. Extensive experiments on NTU-60, NTU-120, and\nPKU-MMD datasets demonstrate that our proposed Skeleton2vec outperforms\nprevious methods and achieves state-of-the-art results.\n", "date": "2024-01-01", "categories": []}
{"id": "2303.12332", "title": "Weakly-Supervised Temporal Action Localization by Inferring Salient\n  Snippet-Feature", "abstract": "  Weakly-supervised temporal action localization aims to locate action regions\nand identify action categories in untrimmed videos simultaneously by taking\nonly video-level labels as the supervision. Pseudo label generation is a\npromising strategy to solve the challenging problem, but the current methods\nignore the natural temporal structure of the video that can provide rich\ninformation to assist such a generation process. In this paper, we propose a\nnovel weakly-supervised temporal action localization method by inferring\nsalient snippet-feature. First, we design a saliency inference module that\nexploits the variation relationship between temporal neighbor snippets to\ndiscover salient snippet-features, which can reflect the significant dynamic\nchange in the video. Secondly, we introduce a boundary refinement module that\nenhances salient snippet-features through the information interaction unit.\nThen, a discrimination enhancement module is introduced to enhance the\ndiscriminative nature of snippet-features. Finally, we adopt the refined\nsnippet-features to produce high-fidelity pseudo labels, which could be used to\nsupervise the training of the action localization network. Extensive\nexperiments on two publicly available datasets, i.e., THUMOS14 and ActivityNet\nv1.3, demonstrate our proposed method achieves significant improvements\ncompared to the state-of-the-art methods.\n", "date": "2023-03-22", "categories": []}
{"id": "2305.04186", "title": "Video-Specific Query-Key Attention Modeling for Weakly-Supervised\n  Temporal Action Localization", "abstract": "  Weakly-supervised temporal action localization aims to identify and localize\nthe action instances in the untrimmed videos with only video-level action\nlabels. When humans watch videos, we can adapt our abstract-level knowledge\nabout actions in different video scenarios and detect whether some actions are\noccurring. In this paper, we mimic how humans do and bring a new perspective\nfor locating and identifying multiple actions in a video. We propose a network\nnamed VQK-Net with a video-specific query-key attention modeling that learns a\nunique query for each action category of each input video. The learned queries\nnot only contain the actions' knowledge features at the abstract level but also\nhave the ability to fit this knowledge into the target video scenario, and they\nwill be used to detect the presence of the corresponding action along the\ntemporal dimension. To better learn these action category queries, we exploit\nnot only the features of the current input video but also the correlation\nbetween different videos through a novel video-specific action category query\nlearner worked with a query similarity loss. Finally, we conduct extensive\nexperiments on three commonly used datasets (THUMOS14, ActivityNet1.2, and\nActivityNet1.3) and achieve state-of-the-art performance.\n", "date": "2023-05-07", "categories": []}
{"id": "2312.14577", "title": "PoseViNet: Distracted Driver Action Recognition Framework Using\n  Multi-View Pose Estimation and Vision Transformer", "abstract": "  Driver distraction is a principal cause of traffic accidents. In a study\nconducted by the National Highway Traffic Safety Administration, engaging in\nactivities such as interacting with in-car menus, consuming food or beverages,\nor engaging in telephonic conversations while operating a vehicle can be\nsignificant sources of driver distraction. From this viewpoint, this paper\nintroduces a novel method for detection of driver distraction using multi-view\ndriver action images. The proposed method is a vision transformer-based\nframework with pose estimation and action inference, namely PoseViNet. The\nmotivation for adding posture information is to enable the transformer to focus\nmore on key features. As a result, the framework is more adept at identifying\ncritical actions. The proposed framework is compared with various\nstate-of-the-art models using SFD3 dataset representing 10 behaviors of\ndrivers. It is found from the comparison that the PoseViNet outperforms these\nmodels. The proposed framework is also evaluated with the SynDD1 dataset\nrepresenting 16 behaviors of driver. As a result, the PoseViNet achieves 97.55%\nvalidation accuracy and 90.92% testing accuracy with the challenging dataset.\n", "date": "2023-12-22", "categories": []}
{"id": "2312.13746", "title": "Video Recognition in Portrait Mode", "abstract": "  The creation of new datasets often presents new challenges for video\nrecognition and can inspire novel ideas while addressing these challenges.\nWhile existing datasets mainly comprise landscape mode videos, our paper seeks\nto introduce portrait mode videos to the research community and highlight the\nunique challenges associated with this video format. With the growing\npopularity of smartphones and social media applications, recognizing portrait\nmode videos is becoming increasingly important. To this end, we have developed\nthe first dataset dedicated to portrait mode video recognition, namely\nPortraitMode-400. The taxonomy of PortraitMode-400 was constructed in a\ndata-driven manner, comprising 400 fine-grained categories, and rigorous\nquality assurance was implemented to ensure the accuracy of human annotations.\nIn addition to the new dataset, we conducted a comprehensive analysis of the\nimpact of video format (portrait mode versus landscape mode) on recognition\naccuracy and spatial bias due to the different formats. Furthermore, we\ndesigned extensive experiments to explore key aspects of portrait mode video\nrecognition, including the choice of data augmentation, evaluation procedure,\nthe importance of temporal information, and the role of audio modality.\nBuilding on the insights from our experimental results and the introduction of\nPortraitMode-400, our paper aims to inspire further research efforts in this\nemerging research area.\n", "date": "2023-12-21", "categories": []}
{"id": "2312.14138", "title": "Revisiting Foreground and Background Separation in Weakly-supervised\n  Temporal Action Localization: A Clustering-based Approach", "abstract": "  Weakly-supervised temporal action localization aims to localize action\ninstances in videos with only video-level action labels. Existing methods\nmainly embrace a localization-by-classification pipeline that optimizes the\nsnippet-level prediction with a video classification loss. However, this\nformulation suffers from the discrepancy between classification and detection,\nresulting in inaccurate separation of foreground and background (F\\&B)\nsnippets. To alleviate this problem, we propose to explore the underlying\nstructure among the snippets by resorting to unsupervised snippet clustering,\nrather than heavily relying on the video classification loss. Specifically, we\npropose a novel clustering-based F\\&B separation algorithm. It comprises two\ncore components: a snippet clustering component that groups the snippets into\nmultiple latent clusters and a cluster classification component that further\nclassifies the cluster as foreground or background. As there are no\nground-truth labels to train these two components, we introduce a unified\nself-labeling mechanism based on optimal transport to produce high-quality\npseudo-labels that match several plausible prior distributions. This ensures\nthat the cluster assignments of the snippets can be accurately associated with\ntheir F\\&B labels, thereby boosting the F\\&B separation. We evaluate our method\non three benchmarks: THUMOS14, ActivityNet v1.2 and v1.3. Our method achieves\npromising performance on all three benchmarks while being significantly more\nlightweight than previous methods. Code is available at\nhttps://github.com/Qinying-Liu/CASE\n", "date": "2023-12-21", "categories": []}
{"id": "2312.12250", "title": "ST(OR)2: Spatio-Temporal Object Level Reasoning for Activity Recognition\n  in the Operating Room", "abstract": "  Surgical robotics holds much promise for improving patient safety and\nclinician experience in the Operating Room (OR). However, it also comes with\nnew challenges, requiring strong team coordination and effective OR management.\nAutomatic detection of surgical activities is a key requirement for developing\nAI-based intelligent tools to tackle these challenges. The current\nstate-of-the-art surgical activity recognition methods however operate on\nimage-based representations and depend on large-scale labeled datasets whose\ncollection is time-consuming and resource-expensive. This work proposes a new\nsample-efficient and object-based approach for surgical activity recognition in\nthe OR. Our method focuses on the geometric arrangements between clinicians and\nsurgical devices, thus utilizing the significant object interaction dynamics in\nthe OR. We conduct experiments in a low-data regime study for long video\nactivity recognition. We also benchmark our method againstother object-centric\napproaches on clip-level action classification and show superior performance.\n", "date": "2023-12-19", "categories": []}
{"id": "2310.10059", "title": "Flow Dynamics Correction for Action Recognition", "abstract": "  Various research studies indicate that action recognition performance highly\ndepends on the types of motions being extracted and how accurate the human\nactions are represented. In this paper, we investigate different optical flow,\nand features extracted from these optical flow that capturing both short-term\nand long-term motion dynamics. We perform power normalization on the magnitude\ncomponent of optical flow for flow dynamics correction to boost subtle or\ndampen sudden motions. We show that existing action recognition models which\nrely on optical flow are able to get performance boosted with our corrected\noptical flow. To further improve performance, we integrate our corrected flow\ndynamics into popular models through a simple hallucination step by selecting\nonly the best performing optical flow features, and we show that by\n'translating' the CNN feature maps into these optical flow features with\ndifferent scales of motions leads to the new state-of-the-art performance on\nseveral benchmarks including HMDB-51, YUP++, fine-grained action recognition on\nMPII Cooking Activities, and large-scale Charades.\n", "date": "2023-10-16", "categories": []}
{"id": "2312.08644", "title": "Generative Model-based Feature Knowledge Distillation for Action\n  Recognition", "abstract": "  Knowledge distillation (KD), a technique widely employed in computer vision,\nhas emerged as a de facto standard for improving the performance of small\nneural networks. However, prevailing KD-based approaches in video tasks\nprimarily focus on designing loss functions and fusing cross-modal information.\nThis overlooks the spatial-temporal feature semantics, resulting in limited\nadvancements in model compression. Addressing this gap, our paper introduces an\ninnovative knowledge distillation framework, with the generative model for\ntraining a lightweight student model. In particular, the framework is organized\ninto two steps: the initial phase is Feature Representation, wherein a\ngenerative model-based attention module is trained to represent feature\nsemantics; Subsequently, the Generative-based Feature Distillation phase\nencompasses both Generative Distillation and Attention Distillation, with the\nobjective of transferring attention-based feature semantics with the generative\nmodel. The efficacy of our approach is demonstrated through comprehensive\nexperiments on diverse popular datasets, proving considerable enhancements in\nvideo action recognition task. Moreover, the effectiveness of our proposed\nframework is validated in the context of more intricate video action detection\ntask. Our code is available at https://github.com/aaai-24/Generative-based-KD.\n", "date": "2023-12-14", "categories": []}
{"id": "2308.11488", "title": "Opening the Vocabulary of Egocentric Actions", "abstract": "  Human actions in egocentric videos are often hand-object interactions\ncomposed from a verb (performed by the hand) applied to an object. Despite\ntheir extensive scaling up, egocentric datasets still face two limitations -\nsparsity of action compositions and a closed set of interacting objects. This\npaper proposes a novel open vocabulary action recognition task. Given a set of\nverbs and objects observed during training, the goal is to generalize the verbs\nto an open vocabulary of actions with seen and novel objects. To this end, we\ndecouple the verb and object predictions via an object-agnostic verb encoder\nand a prompt-based object encoder. The prompting leverages CLIP representations\nto predict an open vocabulary of interacting objects. We create open vocabulary\nbenchmarks on the EPIC-KITCHENS-100 and Assembly101 datasets; whereas\nclosed-action methods fail to generalize, our proposed method is effective. In\naddition, our object encoder significantly outperforms existing open-vocabulary\nvisual recognition methods in recognizing novel interacting objects.\n", "date": "2023-08-22", "categories": []}
{"id": "2308.12609", "title": "Cross-Video Contextual Knowledge Exploration and Exploitation for\n  Ambiguity Reduction in Weakly Supervised Temporal Action Localization", "abstract": "  Weakly supervised temporal action localization (WSTAL) aims to localize\nactions in untrimmed videos using video-level labels. Despite recent advances,\nexisting approaches mainly follow a localization-by-classification pipeline,\ngenerally processing each segment individually, thereby exploiting only limited\ncontextual information. As a result, the model will lack a comprehensive\nunderstanding (e.g. appearance and temporal structure) of various action\npatterns, leading to ambiguity in classification learning and temporal\nlocalization. Our work addresses this from a novel perspective, by exploring\nand exploiting the cross-video contextual knowledge within the dataset to\nrecover the dataset-level semantic structure of action instances via weak\nlabels only, thereby indirectly improving the holistic understanding of\nfine-grained action patterns and alleviating the aforementioned ambiguities.\nSpecifically, an end-to-end framework is proposed, including a Robust\nMemory-Guided Contrastive Learning (RMGCL) module and a Global Knowledge\nSummarization and Aggregation (GKSA) module. First, the RMGCL module explores\nthe contrast and consistency of cross-video action features, assisting in\nlearning more structured and compact embedding space, thus reducing ambiguity\nin classification learning. Further, the GKSA module is used to efficiently\nsummarize and propagate the cross-video representative action knowledge in a\nlearnable manner to promote holistic action patterns understanding, which in\nturn allows the generation of high-confidence pseudo-labels for self-learning,\nthus alleviating ambiguity in temporal localization. Extensive experiments on\nTHUMOS14, ActivityNet1.3, and FineAction demonstrate that our method\noutperforms the state-of-the-art methods, and can be easily plugged into other\nWSTAL methods.\n", "date": "2023-08-24", "categories": []}
{"id": "2312.05352", "title": "A Review of Machine Learning Methods Applied to Video Analysis Systems", "abstract": "  The paper provides a survey of the development of machine-learning techniques\nfor video analysis. The survey provides a summary of the most popular deep\nlearning methods used for human activity recognition. We discuss how popular\narchitectures perform on standard datasets and highlight the differences from\nreal-life datasets dominated by multiple activities performed by multiple\nparticipants over long periods. For real-life datasets, we describe the use of\nlow-parameter models (with 200X or 1,000X fewer parameters) that are trained to\ndetect a single activity after the relevant objects have been successfully\ndetected. Our survey then turns to a summary of machine learning methods that\nare specifically developed for working with a small number of labeled video\nsamples. Our goal here is to describe modern techniques that are specifically\ndesigned so as to minimize the amount of ground truth that is needed for\ntraining and testing video analysis systems. We provide summaries of the\ndevelopment of self-supervised learning, semi-supervised learning, active\nlearning, and zero-shot learning for applications in video analysis. For each\nmethod, we provide representative examples.\n", "date": "2023-12-08", "categories": []}
{"id": "2312.05719", "title": "DVANet: Disentangling View and Action Features for Multi-View Action\n  Recognition", "abstract": "  In this work, we present a novel approach to multi-view action recognition\nwhere we guide learned action representations to be separated from\nview-relevant information in a video. When trying to classify action instances\ncaptured from multiple viewpoints, there is a higher degree of difficulty due\nto the difference in background, occlusion, and visibility of the captured\naction from different camera angles. To tackle the various problems introduced\nin multi-view action recognition, we propose a novel configuration of learnable\ntransformer decoder queries, in conjunction with two supervised contrastive\nlosses, to enforce the learning of action features that are robust to shifts in\nviewpoints. Our disentangled feature learning occurs in two stages: the\ntransformer decoder uses separate queries to separately learn action and view\ninformation, which are then further disentangled using our two contrastive\nlosses. We show that our model and method of training significantly outperforms\nall other uni-modal models on four multi-view action recognition datasets: NTU\nRGB+D, NTU RGB+D 120, PKU-MMD, and N-UCLA. Compared to previous RGB works, we\nsee maximal improvements of 1.5\\%, 4.8\\%, 2.2\\%, and 4.8\\% on each dataset,\nrespectively.\n", "date": "2023-12-10", "categories": []}
{"id": "2312.05830", "title": "A Decoupled Spatio-Temporal Framework for Skeleton-based Action\n  Segmentation", "abstract": "  Effectively modeling discriminative spatio-temporal information is essential\nfor segmenting activities in long action sequences. However, we observe that\nexisting methods are limited in weak spatio-temporal modeling capability due to\ntwo forms of decoupled modeling: (i) cascaded interaction couples spatial and\ntemporal modeling, which over-smooths motion modeling over the long sequence,\nand (ii) joint-shared temporal modeling adopts shared weights to model each\njoint, ignoring the distinct motion patterns of different joints. We propose a\nDecoupled Spatio-Temporal Framework (DeST) to address the above issues.\nFirstly, we decouple the cascaded spatio-temporal interaction to avoid stacking\nmultiple spatio-temporal blocks, while achieving sufficient spatio-temporal\ninteraction. Specifically, DeST performs once unified spatial modeling and\ndivides the spatial features into different groups of subfeatures, which then\nadaptively interact with temporal features from different layers. Since the\ndifferent sub-features contain distinct spatial semantics, the model could\nlearn the optimal interaction pattern at each layer. Meanwhile, inspired by the\nfact that different joints move at different speeds, we propose joint-decoupled\ntemporal modeling, which employs independent trainable weights to capture\ndistinctive temporal features of each joint. On four large-scale benchmarks of\ndifferent scenes, DeST significantly outperforms current state-of-the-art\nmethods with less computational complexity.\n", "date": "2023-12-10", "categories": []}
{"id": "2312.06330", "title": "Navigating Open Set Scenarios for Skeleton-based Action Recognition", "abstract": "  In real-world scenarios, human actions often fall outside the distribution of\ntraining data, making it crucial for models to recognize known actions and\nreject unknown ones. However, using pure skeleton data in such open-set\nconditions poses challenges due to the lack of visual background cues and the\ndistinct sparse structure of body pose sequences. In this paper, we tackle the\nunexplored Open-Set Skeleton-based Action Recognition (OS-SAR) task and\nformalize the benchmark on three skeleton-based datasets. We assess the\nperformance of seven established open-set approaches on our task and identify\ntheir limits and critical generalization issues when dealing with skeleton\ninformation. To address these challenges, we propose a distance-based\ncross-modality ensemble method that leverages the cross-modal alignment of\nskeleton joints, bones, and velocities to achieve superior open-set recognition\nperformance. We refer to the key idea as CrossMax - an approach that utilizes a\nnovel cross-modality mean max discrepancy suppression mechanism to align latent\nspaces during training and a cross-modality distance-based logits refinement\nmethod during testing. CrossMax outperforms existing approaches and\nconsistently yields state-of-the-art results across all datasets and backbones.\nThe benchmark, code, and models will be released at\nhttps://github.com/KPeng9510/OS-SAR.\n", "date": "2023-12-11", "categories": []}
{"id": "2312.06598", "title": "Early Action Recognition with Action Prototypes", "abstract": "  Early action recognition is an important and challenging problem that enables\nthe recognition of an action from a partially observed video stream where the\nactivity is potentially unfinished or even not started. In this work, we\npropose a novel model that learns a prototypical representation of the full\naction for each class and uses it to regularize the architecture and the visual\nrepresentations of the partial observations. Our model is very simple in design\nand also efficient. We decompose the video into short clips, where a visual\nencoder extracts features from each clip independently. Later, a decoder\naggregates together in an online fashion features from all the clips for the\nfinal class prediction. During training, for each partial observation, the\nmodel is jointly trained to both predict the label as well as the action\nprototypical representation which acts as a regularizer. We evaluate our method\non multiple challenging real-world datasets and outperform the current\nstate-of-the-art by a significant margin. For example, on early recognition\nobserving only the first 10% of each video, our method improves the SOTA by\n+2.23 Top-1 accuracy on Something-Something-v2, +3.55 on UCF-101, +3.68 on\nSSsub21, and +5.03 on EPIC-Kitchens-55, where prior work used either\nmulti-modal inputs (e.g. optical-flow) or batched inference. Finally, we also\npresent exhaustive ablation studies to motivate the design choices we made, as\nwell as gather insights regarding what our model is learning semantically.\n", "date": "2023-12-11", "categories": []}
{"id": "2312.04266", "title": "Activity Grammars for Temporal Action Segmentation", "abstract": "  Sequence prediction on temporal data requires the ability to understand\ncompositional structures of multi-level semantics beyond individual and\ncontextual properties. The task of temporal action segmentation, which aims at\ntranslating an untrimmed activity video into a sequence of action segments,\nremains challenging for this reason. This paper addresses the problem by\nintroducing an effective activity grammar to guide neural predictions for\ntemporal action segmentation. We propose a novel grammar induction algorithm\nthat extracts a powerful context-free grammar from action sequence data. We\nalso develop an efficient generalized parser that transforms frame-level\nprobability distributions into a reliable sequence of actions according to the\ninduced grammar with recursive rules. Our approach can be combined with any\nneural network for temporal action segmentation to enhance the sequence\nprediction and discover its compositional structure. Experimental results\ndemonstrate that our method significantly improves temporal action segmentation\nin terms of both performance and interpretability on two standard benchmarks,\nBreakfast and 50 Salads.\n", "date": "2023-12-07", "categories": []}
{"id": "2312.03288", "title": "STEP CATFormer: Spatial-Temporal Effective Body-Part Cross Attention\n  Transformer for Skeleton-based Action Recognition", "abstract": "  Graph convolutional networks (GCNs) have been widely used and achieved\nremarkable results in skeleton-based action recognition. We think the key to\nskeleton-based action recognition is a skeleton hanging in frames, so we focus\non how the Graph Convolutional Convolution networks learn different topologies\nand effectively aggregate joint features in the global temporal and local\ntemporal. In this work, we propose three Channel-wise Tolopogy Graph\nConvolution based on Channel-wise Topology Refinement Graph Convolution\n(CTR-GCN). Combining CTR-GCN with two joint cross-attention modules can capture\nthe upper-lower body part and hand-foot relationship skeleton features. After\nthat, to capture features of human skeletons changing in frames we design the\nTemporal Attention Transformers to extract skeletons effectively. The Temporal\nAttention Transformers can learn the temporal features of human skeleton\nsequences. Finally, we fuse the temporal features output scale with MLP and\nclassification. We develop a powerful graph convolutional network named Spatial\nTemporal Effective Body-part Cross Attention Transformer which notably\nhigh-performance on the NTU RGB+D, NTU RGB+D 120 datasets. Our code and models\nare available at https://github.com/maclong01/STEP-CATFormer\n", "date": "2023-12-06", "categories": []}
{"id": "2312.02226", "title": "Generating Action-conditioned Prompts for Open-vocabulary Video Action\n  Recognition", "abstract": "  Exploring open-vocabulary video action recognition is a promising venture,\nwhich aims to recognize previously unseen actions within any arbitrary set of\ncategories. Existing methods typically adapt pretrained image-text models to\nthe video domain, capitalizing on their inherent strengths in generalization. A\ncommon thread among such methods is the augmentation of visual embeddings with\ntemporal information to improve the recognition of seen actions. Yet, they\ncompromise with standard less-informative action descriptions, thus faltering\nwhen confronted with novel actions. Drawing inspiration from human cognitive\nprocesses, we argue that augmenting text embeddings with human prior knowledge\nis pivotal for open-vocabulary video action recognition. To realize this, we\ninnovatively blend video models with Large Language Models (LLMs) to devise\nAction-conditioned Prompts. Specifically, we harness the knowledge in LLMs to\nproduce a set of descriptive sentences that contain distinctive features for\nidentifying given actions. Building upon this foundation, we further introduce\na multi-modal action knowledge alignment mechanism to align concepts in video\nand textual knowledge encapsulated within the prompts. Extensive experiments on\nvarious video benchmarks, including zero-shot, few-shot, and base-to-novel\ngeneralization settings, demonstrate that our method not only sets new SOTA\nperformance but also possesses excellent interpretability.\n", "date": "2023-12-04", "categories": []}
{"id": "2312.01083", "title": "Consistency Prototype Module and Motion Compensation for Few-Shot Action\n  Recognition (CLIP-CP$\\mathbf{M^2}$C)", "abstract": "  Recently, few-shot action recognition has significantly progressed by\nlearning the feature discriminability and designing suitable comparison\nmethods. Still, there are the following restrictions. (a) Previous works are\nmainly based on visual mono-modal. Although some multi-modal works use labels\nas supplementary to construct prototypes of support videos, they can not use\nthis information for query videos. The labels are not used efficiently. (b)\nMost of the works ignore the motion feature of video, although the motion\nfeatures are essential for distinguishing. We proposed a Consistency Prototype\nand Motion Compensation Network(CLIP-CP$M^2$C) to address these issues.\nFirstly, we use the CLIP for multi-modal few-shot action recognition with the\ntext-image comparison for domain adaption. Secondly, in order to make the\namount of information between the prototype and the query more similar, we\npropose a novel method to compensate for the text(prompt) information of query\nvideos when text(prompt) does not exist, which depends on a Consistency Loss.\nThirdly, we use the differential features of the adjacent frames in two\ndirections as the motion features, which explicitly embeds the network with\nmotion dynamics. We also apply the Consistency Loss to the motion features.\nExtensive experiments on standard benchmark datasets demonstrate that the\nproposed method can compete with state-of-the-art results. Our code is\navailable at the URL: https://github.com/xxx/xxx.git.\n", "date": "2023-12-02", "categories": []}
{"id": "2307.01985", "title": "Task-Specific Alignment and Multiple Level Transformer for Few-Shot\n  Action Recognition", "abstract": "  In the research field of few-shot learning, the main difference between\nimage-based and video-based is the additional temporal dimension. In recent\nyears, some works have used the Transformer to deal with frames, then get the\nattention feature and the enhanced prototype, and the results are competitive.\nHowever, some video frames may relate little to the action, and only using\nsingle frame-level or segment-level features may not mine enough information.\nWe address these problems sequentially through an end-to-end method named\n\"Task-Specific Alignment and Multiple-level Transformer Network (TSA-MLT)\". The\nfirst module (TSA) aims at filtering the action-irrelevant frames for action\nduration alignment. Affine Transformation for frame sequence in the time\ndimension is used for linear sampling. The second module (MLT) focuses on the\nMultiple-level feature of the support prototype and query sample to mine more\ninformation for the alignment, which operates on different level features. We\nadopt a fusion loss according to a fusion distance that fuses the L2 sequence\ndistance, which focuses on temporal order alignment, and the Optimal Transport\ndistance, which focuses on measuring the gap between the appearance and\nsemantics of the videos. Extensive experiments show our method achieves\nstate-of-the-art results on the HMDB51 and UCF101 datasets and a competitive\nresult on the benchmark of Kinetics and something 2-something V2 datasets. Our\ncode is available at the URL: https://github.com/cofly2014/tsa-mlt.git\n", "date": "2023-07-05", "categories": []}
{"id": "2312.00055", "title": "LEAP: LLM-Generation of Egocentric Action Programs", "abstract": "  We introduce LEAP (illustrated in Figure 1), a novel method for generating\nvideo-grounded action programs through use of a Large Language Model (LLM).\nThese action programs represent the motoric, perceptual, and structural aspects\nof action, and consist of sub-actions, pre- and post-conditions, and control\nflows. LEAP's action programs are centered on egocentric video and employ\nrecent developments in LLMs both as a source for program knowledge and as an\naggregator and assessor of multimodal video information. We apply LEAP over a\nmajority (87\\%) of the training set of the EPIC Kitchens dataset, and release\nthe resulting action programs as a publicly available dataset here\n(https://drive.google.com/drive/folders/1Cpkw_TI1IIxXdzor0pOXG3rWJWuKU5Ex?usp=drive_link).\nWe employ LEAP as a secondary source of supervision, using its action programs\nin a loss term applied to action recognition and anticipation networks. We\ndemonstrate sizable improvements in performance in both tasks due to training\nwith the LEAP dataset. Our method achieves 1st place on the EPIC Kitchens\nAction Recognition leaderboard as of November 17 among the networks restricted\nto RGB-input (see Supplementary Materials).\n", "date": "2023-11-29", "categories": []}
{"id": "2312.00348", "title": "Student Activity Recognition in Classroom Environments using Transfer\n  Learning", "abstract": "  The recent advances in artificial intelligence and deep learning facilitate\nautomation in various applications including home automation, smart\nsurveillance systems, and healthcare among others. Human Activity Recognition\nis one of its emerging applications, which can be implemented in a classroom\nenvironment to enhance safety, efficiency, and overall educational quality.\nThis paper proposes a system for detecting and recognizing the activities of\nstudents in a classroom environment. The dataset has been structured and\nrecorded by the authors since a standard dataset for this task was not\navailable at the time of this study. Transfer learning, a widely adopted method\nwithin the field of deep learning, has proven to be helpful in complex tasks\nlike image and video processing. Pretrained models including VGG-16, ResNet-50,\nInceptionV3, and Xception are used for feature extraction and classification\ntasks. Xception achieved an accuracy of 93%, on the novel classroom dataset,\noutperforming the other three models in consideration. The system proposed in\nthis study aims to introduce a safer and more productive learning environment\nfor students and educators.\n", "date": "2023-12-01", "categories": []}
{"id": "2311.17975", "title": "GeoDeformer: Geometric Deformable Transformer for Action Recognition", "abstract": "  Vision transformers have recently emerged as an effective alternative to\nconvolutional networks for action recognition. However, vision transformers\nstill struggle with geometric variations prevalent in video data. This paper\nproposes a novel approach, GeoDeformer, designed to capture the variations\ninherent in action video by integrating geometric comprehension directly into\nthe ViT architecture. Specifically, at the core of GeoDeformer is the Geometric\nDeformation Predictor, a module designed to identify and quantify potential\nspatial and temporal geometric deformations within the given video. Spatial\ndeformations adjust the geometry within individual frames, while temporal\ndeformations capture the cross-frame geometric dynamics, reflecting motion and\ntemporal progression. To demonstrate the effectiveness of our approach, we\nincorporate it into the established MViTv2 framework, replacing the standard\nself-attention blocks with GeoDeformer blocks. Our experiments at UCF101,\nHMDB51, and Mini-K200 achieve significant increases in both Top-1 and Top-5\naccuracy, establishing new state-of-the-art results with only a marginal\nincrease in computational cost. Additionally, visualizations affirm that\nGeoDeformer effectively manifests explicit geometric deformations and minimizes\ngeometric variations. Codes and checkpoints will be released.\n", "date": "2023-11-29", "categories": []}
{"id": "2311.18840", "title": "Just Add $\\pi$! Pose Induced Video Transformers for Understanding\n  Activities of Daily Living", "abstract": "  Video transformers have become the de facto standard for human action\nrecognition, yet their exclusive reliance on the RGB modality still limits\ntheir adoption in certain domains. One such domain is Activities of Daily\nLiving (ADL), where RGB alone is not sufficient to distinguish between visually\nsimilar actions, or actions observed from multiple viewpoints. To facilitate\nthe adoption of video transformers for ADL, we hypothesize that the\naugmentation of RGB with human pose information, known for its sensitivity to\nfine-grained motion and multiple viewpoints, is essential. Consequently, we\nintroduce the first Pose Induced Video Transformer: PI-ViT (or $\\pi$-ViT), a\nnovel approach that augments the RGB representations learned by video\ntransformers with 2D and 3D pose information. The key elements of $\\pi$-ViT are\ntwo plug-in modules, 2D Skeleton Induction Module and 3D Skeleton Induction\nModule, that are responsible for inducing 2D and 3D pose information into the\nRGB representations. These modules operate by performing pose-aware auxiliary\ntasks, a design choice that allows $\\pi$-ViT to discard the modules during\ninference. Notably, $\\pi$-ViT achieves the state-of-the-art performance on\nthree prominent ADL datasets, encompassing both real-world and large-scale\nRGB-D datasets, without requiring poses or additional computational overhead at\ninference.\n", "date": "2023-11-30", "categories": []}
{"id": "2311.16446", "title": "Centre Stage: Centricity-based Audio-Visual Temporal Action Detection", "abstract": "  Previous one-stage action detection approaches have modelled temporal\ndependencies using only the visual modality. In this paper, we explore\ndifferent strategies to incorporate the audio modality, using multi-scale\ncross-attention to fuse the two modalities. We also demonstrate the correlation\nbetween the distance from the timestep to the action centre and the accuracy of\nthe predicted boundaries. Thus, we propose a novel network head to estimate the\ncloseness of timesteps to the action centre, which we call the centricity\nscore. This leads to increased confidence for proposals that exhibit more\nprecise boundaries. Our method can be integrated with other one-stage\nanchor-free architectures and we demonstrate this on three recent baselines on\nthe EPIC-Kitchens-100 action detection benchmark where we achieve\nstate-of-the-art performance. Detailed ablation studies showcase the benefits\nof fusing audio and our proposed centricity scores. Code and models for our\nproposed method are publicly available at\nhttps://github.com/hanielwang/Audio-Visual-TAD.git\n", "date": "2023-11-28", "categories": []}
{"id": "2311.15916", "title": "ADM-Loc: Actionness Distribution Modeling for Point-supervised Temporal\n  Action Localization", "abstract": "  This paper addresses the challenge of point-supervised temporal action\ndetection, in which only one frame per action instance is annotated in the\ntraining set. Self-training aims to provide supplementary supervision for the\ntraining process by generating pseudo-labels (action proposals) from a base\nmodel. However, most current methods generate action proposals by applying\nmanually designed thresholds to action classification probabilities and\ntreating adjacent snippets as independent entities. As a result, these methods\nstruggle to generate complete action proposals, exhibit sensitivity to\nfluctuations in action classification scores, and generate redundant and\noverlapping action proposals. This paper proposes a novel framework termed\nADM-Loc, which stands for Actionness Distribution Modeling for point-supervised\naction Localization. ADM-Loc generates action proposals by fitting a composite\ndistribution, comprising both Gaussian and uniform distributions, to the action\nclassification signals. This fitting process is tailored to each action class\npresent in the video and is applied separately for each action instance,\nensuring the distinctiveness of their distributions. ADM-Loc significantly\nenhances the alignment between the generated action proposals and ground-truth\naction instances and offers high-quality pseudo-labels for self-training.\nMoreover, to model action boundary snippets, it enforces consistency in action\nclassification scores during training by employing Gaussian kernels, supervised\nwith the proposed loss functions. ADM-Loc outperforms the state-of-the-art\npoint-supervised methods on THUMOS14 and ActivityNet-v1.2 datasets.\n", "date": "2023-11-27", "categories": []}
{"id": "2208.01910", "title": "Multimodal Generation of Novel Action Appearances for Synthetic-to-Real\n  Recognition of Activities of Daily Living", "abstract": "  Domain shifts, such as appearance changes, are a key challenge in real-world\napplications of activity recognition models, which range from assistive\nrobotics and smart homes to driver observation in intelligent vehicles. For\nexample, while simulations are an excellent way of economical data collection,\na Synthetic-to-Real domain shift leads to a > 60% drop in accuracy when\nrecognizing activities of Daily Living (ADLs). We tackle this challenge and\nintroduce an activity domain generation framework which creates novel ADL\nappearances (novel domains) from different existing activity modalities (source\ndomains) inferred from video training data. Our framework computes human poses,\nheatmaps of body joints, and optical flow maps and uses them alongside the\noriginal RGB videos to learn the essence of source domains in order to generate\ncompletely new ADL domains. The model is optimized by maximizing the distance\nbetween the existing source appearances and the generated novel appearances\nwhile ensuring that the semantics of an activity is preserved through an\nadditional classification loss. While source data multimodality is an important\nconcept in this design, our setup does not rely on multi-sensor setups, (i.e.,\nall source modalities are inferred from a single video only.) The newly created\nactivity domains are then integrated in the training of the ADL classification\nnetworks, resulting in models far less susceptible to changes in data\ndistributions. Extensive experiments on the Synthetic-to-Real benchmark\nSims4Action demonstrate the potential of the domain generation paradigm for\ncross-domain ADL recognition, setting new state-of-the-art results. Our code is\npublicly available at https://github.com/Zrrr1997/syn2real_DG\n", "date": "2022-08-03", "categories": []}
{"id": "2311.14281", "title": "Multi-modal Instance Refinement for Cross-domain Action Recognition", "abstract": "  Unsupervised cross-domain action recognition aims at adapting the model\ntrained on an existing labeled source domain to a new unlabeled target domain.\nMost existing methods solve the task by directly aligning the feature\ndistributions of source and target domains. However, this would cause negative\ntransfer during domain adaptation due to some negative training samples in both\ndomains. In the source domain, some training samples are of low-relevance to\ntarget domain due to the difference in viewpoints, action styles, etc. In the\ntarget domain, there are some ambiguous training samples that can be easily\nclassified as another type of action under the case of source domain. The\nproblem of negative transfer has been explored in cross-domain object\ndetection, while it remains under-explored in cross-domain action recognition.\nTherefore, we propose a Multi-modal Instance Refinement (MMIR) method to\nalleviate the negative transfer based on reinforcement learning. Specifically,\na reinforcement learning agent is trained in both domains for every modality to\nrefine the training data by selecting out negative samples from each domain.\nOur method finally outperforms several other state-of-the-art baselines in\ncross-domain action recognition on the benchmark EPIC-Kitchens dataset, which\ndemonstrates the advantage of MMIR in reducing negative transfer.\n", "date": "2023-11-24", "categories": []}
{"id": "2311.12300", "title": "Challenges in Video-Based Infant Action Recognition: A Critical\n  Examination of the State of the Art", "abstract": "  Automated human action recognition, a burgeoning field within computer\nvision, boasts diverse applications spanning surveillance, security,\nhuman-computer interaction, tele-health, and sports analysis. Precise action\nrecognition in infants serves a multitude of pivotal purposes, encompassing\nsafety monitoring, developmental milestone tracking, early intervention for\ndevelopmental delays, fostering parent-infant bonds, advancing computer-aided\ndiagnostics, and contributing to the scientific comprehension of child\ndevelopment. This paper delves into the intricacies of infant action\nrecognition, a domain that has remained relatively uncharted despite the\naccomplishments in adult action recognition. In this study, we introduce a\ngroundbreaking dataset called ``InfActPrimitive'', encompassing five\nsignificant infant milestone action categories, and we incorporate specialized\npreprocessing for infant data. We conducted an extensive comparative analysis\nemploying cutting-edge skeleton-based action recognition models using this\ndataset. Our findings reveal that, although the PoseC3D model achieves the\nhighest accuracy at approximately 71%, the remaining models struggle to\naccurately capture the dynamics of infant actions. This highlights a\nsubstantial knowledge gap between infant and adult action recognition domains\nand the urgent need for data-efficient pipeline models.\n", "date": "2023-11-21", "categories": []}
{"id": "2311.12344", "title": "Modality Mixer Exploiting Complementary Information for Multi-modal\n  Action Recognition", "abstract": "  Due to the distinctive characteristics of sensors, each modality exhibits\nunique physical properties. For this reason, in the context of multi-modal\naction recognition, it is important to consider not only the overall action\ncontent but also the complementary nature of different modalities. In this\npaper, we propose a novel network, named Modality Mixer (M-Mixer) network,\nwhich effectively leverages and incorporates the complementary information\nacross modalities with the temporal context of actions for action recognition.\nA key component of our proposed M-Mixer is the Multi-modal Contextualization\nUnit (MCU), a simple yet effective recurrent unit. Our MCU is responsible for\ntemporally encoding a sequence of one modality (e.g., RGB) with action content\nfeatures of other modalities (e.g., depth and infrared modalities). This\nprocess encourages M-Mixer network to exploit global action content and also to\nsupplement complementary information of other modalities. Furthermore, to\nextract appropriate complementary information regarding to the given modality\nsettings, we introduce a new module, named Complementary Feature Extraction\nModule (CFEM). CFEM incorporates sepearte learnable query embeddings for each\nmodality, which guide CFEM to extract complementary information and global\naction content from the other modalities. As a result, our proposed method\noutperforms state-of-the-art methods on NTU RGB+D 60, NTU RGB+D 120, and\nNW-UCLA datasets. Moreover, through comprehensive ablation studies, we further\nvalidate the effectiveness of our proposed method.\n", "date": "2023-11-21", "categories": []}
{"id": "2311.12674", "title": "Contrastive Left-Right Wearable Sensors (IMUs) Consistency Matching for\n  HAR", "abstract": "  Machine learning algorithms are improving rapidly, but annotating training\ndata remains a bottleneck for many applications. In this paper, we show how\nreal data can be used for self-supervised learning without any transformations\nby taking advantage of the symmetry present in the activities. Our approach\ninvolves contrastive matching of two different sensors (left and right wrist or\nleg-worn IMUs) to make representations of co-occurring sensor data more similar\nand those of non-co-occurring sensor data more different. We test our approach\non the Opportunity and MM-Fit datasets. In MM-Fit we show significant\nimprovement over the baseline supervised and self-supervised method SimCLR,\nwhile for Opportunity there is significant improvement over the supervised\nbaseline and slight improvement when compared to SimCLR. Moreover, our method\nimproves supervised baselines even when using only a small amount of the data\nfor training. Future work should explore under which conditions our method is\nbeneficial for human activity recognition systems and other related\napplications.\n", "date": "2023-11-21", "categories": []}
{"id": "2304.06866", "title": "PMI Sampler: Patch Similarity Guided Frame Selection for Aerial Action\n  Recognition", "abstract": "  We present a new algorithm for selection of informative frames in video\naction recognition. Our approach is designed for aerial videos captured using a\nmoving camera where human actors occupy a small spatial resolution of video\nframes. Our algorithm utilizes the motion bias within aerial videos, which\nenables the selection of motion-salient frames. We introduce the concept of\npatch mutual information (PMI) score to quantify the motion bias between\nadjacent frames, by measuring the similarity of patches. We use this score to\nassess the amount of discriminative motion information contained in one frame\nrelative to another. We present an adaptive frame selection strategy using\nshifted leaky ReLu and cumulative distribution function, which ensures that the\nsampled frames comprehensively cover all the essential segments with high\nmotion salience. Our approach can be integrated with any action recognition\nmodel to enhance its accuracy. In practice, our method achieves a relative\nimprovement of 2.2 - 13.8% in top-1 accuracy on UAV-Human, 6.8% on NEC Drone,\nand 9.0% on Diving48 datasets.\n", "date": "2023-04-14", "categories": []}
{"id": "2303.02575", "title": "MITFAS: Mutual Information based Temporal Feature Alignment and Sampling\n  for Aerial Video Action Recognition", "abstract": "  We present a novel approach for action recognition in UAV videos. Our\nformulation is designed to handle occlusion and viewpoint changes caused by the\nmovement of a UAV. We use the concept of mutual information to compute and\nalign the regions corresponding to human action or motion in the temporal\ndomain. This enables our recognition model to learn from the key features\nassociated with the motion. We also propose a novel frame sampling method that\nuses joint mutual information to acquire the most informative frame sequence in\nUAV videos. We have integrated our approach with X3D and evaluated the\nperformance on multiple datasets. In practice, we achieve 18.9% improvement in\nTop-1 accuracy over current state-of-the-art methods on UAV-Human(Li et al.,\n2021), 7.3% improvement on Drone-Action(Perera et al., 2019), and 7.16%\nimprovement on NEC Drones(Choi et al., 2020).\n", "date": "2023-03-05", "categories": []}
{"id": "2211.05483", "title": "Unsupervised Deep Learning-based clustering for Human Activity\n  Recognition", "abstract": "  One of the main problems in applying deep learning techniques to recognize\nactivities of daily living (ADLs) based on inertial sensors is the lack of\nappropriately large labelled datasets to train deep learning-based models. A\nlarge amount of data would be available due to the wide spread of mobile\ndevices equipped with inertial sensors that can collect data to recognize human\nactivities. Unfortunately, this data is not labelled. The paper proposes DISC\n(Deep Inertial Sensory Clustering), a DL-based clustering architecture that\nautomatically labels multi-dimensional inertial signals. In particular, the\narchitecture combines a recurrent AutoEncoder and a clustering criterion to\npredict unlabelled human activities-related signals. The proposed architecture\nis evaluated on three publicly available HAR datasets and compared with four\nwell-known end-to-end deep clustering approaches. The experiments demonstrate\nthe effectiveness of DISC on both clustering accuracy and normalized mutual\ninformation metrics.\n", "date": "2022-11-10", "categories": []}
{"id": "2311.06218", "title": "Semantic-aware Video Representation for Few-shot Action Recognition", "abstract": "  Recent work on action recognition leverages 3D features and textual\ninformation to achieve state-of-the-art performance. However, most of the\ncurrent few-shot action recognition methods still rely on 2D frame-level\nrepresentations, often require additional components to model temporal\nrelations, and employ complex distance functions to achieve accurate alignment\nof these representations. In addition, existing methods struggle to effectively\nintegrate textual semantics, some resorting to concatenation or addition of\ntextual and visual features, and some using text merely as an additional\nsupervision without truly achieving feature fusion and information transfer\nfrom different modalities. In this work, we propose a simple yet effective\nSemantic-Aware Few-Shot Action Recognition (SAFSAR) model to address these\nissues. We show that directly leveraging a 3D feature extractor combined with\nan effective feature-fusion scheme, and a simple cosine similarity for\nclassification can yield better performance without the need of extra\ncomponents for temporal modeling or complex distance functions. We introduce an\ninnovative scheme to encode the textual semantics into the video representation\nwhich adaptively fuses features from text and video, and encourages the visual\nencoder to extract more semantically consistent features. In this scheme,\nSAFSAR achieves alignment and fusion in a compact way. Experiments on five\nchallenging few-shot action recognition benchmarks under various settings\ndemonstrate that the proposed SAFSAR model significantly improves the\nstate-of-the-art performance.\n", "date": "2023-11-10", "categories": []}
{"id": "2311.06231", "title": "Learning Human Action Recognition Representations Without Real Humans", "abstract": "  Pre-training on massive video datasets has become essential to achieve high\naction recognition performance on smaller downstream datasets. However, most\nlarge-scale video datasets contain images of people and hence are accompanied\nwith issues related to privacy, ethics, and data protection, often preventing\nthem from being publicly shared for reproducible research. Existing work has\nattempted to alleviate these problems by blurring faces, downsampling videos,\nor training on synthetic data. On the other hand, analysis on the\ntransferability of privacy-preserving pre-trained models to downstream tasks\nhas been limited. In this work, we study this problem by first asking the\nquestion: can we pre-train models for human action recognition with data that\ndoes not include real humans? To this end, we present, for the first time, a\nbenchmark that leverages real-world videos with humans removed and synthetic\ndata containing virtual humans to pre-train a model. We then evaluate the\ntransferability of the representation learned on this data to a diverse set of\ndownstream action recognition benchmarks. Furthermore, we propose a novel\npre-training strategy, called Privacy-Preserving MAE-Align, to effectively\ncombine synthetic data and human-removed real data. Our approach outperforms\nprevious baselines by up to 5% and closes the performance gap between human and\nno-human action recognition representations on downstream tasks, for both\nlinear probing and fine-tuning. Our benchmark, code, and models are available\nat https://github.com/howardzh01/PPMA .\n", "date": "2023-11-10", "categories": []}
{"id": "2311.04236", "title": "Distributed Agent-Based Collaborative Learning in Cross-Individual\n  Wearable Sensor-Based Human Activity Recognition", "abstract": "  The rapid growth of wearable sensor technologies holds substantial promise\nfor the field of personalized and context-aware Human Activity Recognition.\nGiven the inherently decentralized nature of data sources within this domain,\nthe utilization of multi-agent systems with their inherent decentralization\ncapabilities presents an opportunity to facilitate the development of scalable,\nadaptable, and privacy-conscious methodologies. This paper introduces a\ncollaborative distributed learning approach rooted in multi-agent principles,\nwherein individual users of sensor-equipped devices function as agents within a\ndistributed network, collectively contributing to the comprehensive process of\nlearning and classifying human activities. In this proposed methodology, not\nonly is the privacy of activity monitoring data upheld for each individual,\neliminating the need for an external server to oversee the learning process,\nbut the system also exhibits the potential to surmount the limitations of\nconventional centralized models and adapt to the unique attributes of each\nuser. The proposed approach has been empirically tested on two publicly\naccessible human activity recognition datasets, specifically PAMAP2 and HARTH,\nacross varying settings. The provided empirical results conclusively highlight\nthe efficacy of inter-individual collaborative learning when contrasted with\ncentralized configurations, both in terms of local and global generalization.\n", "date": "2023-11-06", "categories": []}
{"id": "2311.00729", "title": "ZEETAD: Adapting Pretrained Vision-Language Model for Zero-Shot\n  End-to-End Temporal Action Detection", "abstract": "  Temporal action detection (TAD) involves the localization and classification\nof action instances within untrimmed videos. While standard TAD follows fully\nsupervised learning with closed-set setting on large training data, recent\nzero-shot TAD methods showcase the promising open-set setting by leveraging\nlarge-scale contrastive visual-language (ViL) pretrained models. However,\nexisting zero-shot TAD methods have limitations on how to properly construct\nthe strong relationship between two interdependent tasks of localization and\nclassification and adapt ViL model to video understanding. In this work, we\npresent ZEETAD, featuring two modules: dual-localization and zero-shot proposal\nclassification. The former is a Transformer-based module that detects action\nevents while selectively collecting crucial semantic embeddings for later\nrecognition. The latter one, CLIP-based module, generates semantic embeddings\nfrom text and frame inputs for each temporal unit. Additionally, we enhance\ndiscriminative capability on unseen classes by minimally updating the frozen\nCLIP encoder with lightweight adapters. Extensive experiments on THUMOS14 and\nActivityNet-1.3 datasets demonstrate our approach's superior performance in\nzero-shot TAD and effective knowledge transfer from ViL models to unseen action\ncategories.\n", "date": "2023-11-01", "categories": []}
{"id": "2311.03390", "title": "FPGA-QHAR: Throughput-Optimized for Quantized Human Action Recognition\n  on The Edge", "abstract": "  Accelerating Human Action Recognition (HAR) efficiently for real-time\nsurveillance and robotic systems on edge chips remains a challenging research\nfield, given its high computational and memory requirements. This paper\nproposed an integrated end-to-end HAR scalable HW/SW accelerator co-design\nbased on an enhanced 8-bit quantized Two-Stream SimpleNet-PyTorch CNN\narchitecture. Our network accelerator was trained on UCF101 and UCF24 datasets\nand implemented on edge SoC-FPGA. Our development uses partially streaming\ndataflow architecture to achieve higher throughput versus network design and\nresource utilization trade-off. We also fused all convolutional, batch-norm,\nand ReLU operations into a single homogeneous layer and utilized the\nLucas-Kanade motion flow method to enable a high parallelism accelerator design\nand optimized on-chip engine computing.Furthermore, our proposed methodology\nachieved nearly 81% prediction accuracy with an approximately 24 FPS real-time\ninference throughput at 187MHz on ZCU104, which is 1.7x - 1.9x higher than the\nprior research. Lastly, the designed framework was benchmarked against several\nhardware chips for higher throughput and performance measurements and is now\navailable as an open-source project on GitHub for training and implementation\non edge platforms.\n", "date": "2023-11-04", "categories": []}
{"id": "2311.03106", "title": "Unified Multi-modal Unsupervised Representation Learning for\n  Skeleton-based Action Understanding", "abstract": "  Unsupervised pre-training has shown great success in skeleton-based action\nunderstanding recently. Existing works typically train separate\nmodality-specific models, then integrate the multi-modal information for action\nunderstanding by a late-fusion strategy. Although these approaches have\nachieved significant performance, they suffer from the complex yet redundant\nmulti-stream model designs, each of which is also limited to the fixed input\nskeleton modality. To alleviate these issues, in this paper, we propose a\nUnified Multimodal Unsupervised Representation Learning framework, called\nUmURL, which exploits an efficient early-fusion strategy to jointly encode the\nmulti-modal features in a single-stream manner. Specifically, instead of\ndesigning separate modality-specific optimization processes for uni-modal\nunsupervised learning, we feed different modality inputs into the same stream\nwith an early-fusion strategy to learn their multi-modal features for reducing\nmodel complexity. To ensure that the fused multi-modal features do not exhibit\nmodality bias, i.e., being dominated by a certain modality input, we further\npropose both intra- and inter-modal consistency learning to guarantee that the\nmulti-modal features contain the complete semantics of each modal via feature\ndecomposition and distinct alignment. In this manner, our framework is able to\nlearn the unified representations of uni-modal or multi-modal skeleton input,\nwhich is flexible to different kinds of modality input for robust action\nunderstanding in practical cases. Extensive experiments conducted on three\nlarge-scale datasets, i.e., NTU-60, NTU-120, and PKU-MMD II, demonstrate that\nUmURL is highly efficient, possessing the approximate complexity with the\nuni-modal methods, while achieving new state-of-the-art performance across\nvarious downstream task scenarios in skeleton-based action representation\nlearning.\n", "date": "2023-11-06", "categories": []}
{"id": "2309.13425", "title": "MiliPoint: A Point Cloud Dataset for mmWave Radar", "abstract": "  Millimetre-wave (mmWave) radar has emerged as an attractive and\ncost-effective alternative for human activity sensing compared to traditional\ncamera-based systems. mmWave radars are also non-intrusive, providing better\nprotection for user privacy. However, as a Radio Frequency (RF) based\ntechnology, mmWave radars rely on capturing reflected signals from objects,\nmaking them more prone to noise compared to cameras. This raises an intriguing\nquestion for the deep learning community: Can we develop more effective point\nset-based deep learning methods for such attractive sensors?\n  To answer this question, our work, termed MiliPoint, delves into this idea by\nproviding a large-scale, open dataset for the community to explore how mmWave\nradars can be utilised for human activity recognition. Moreover, MiliPoint\nstands out as it is larger in size than existing datasets, has more diverse\nhuman actions represented, and encompasses all three key tasks in human\nactivity recognition. We have also established a range of point-based deep\nneural networks such as DGCNN, PointNet++ and PointTransformer, on MiliPoint,\nwhich can serve to set the ground baseline for further development.\n", "date": "2023-09-23", "categories": []}
{"id": "2311.01283", "title": "Distilling Knowledge from CNN-Transformer Models for Enhanced Human\n  Action Recognition", "abstract": "  This paper presents a study on improving human action recognition through the\nutilization of knowledge distillation, and the combination of CNN and ViT\nmodels. The research aims to enhance the performance and efficiency of smaller\nstudent models by transferring knowledge from larger teacher models. The\nproposed method employs a Transformer vision network as the student model,\nwhile a convolutional network serves as the teacher model. The teacher model\nextracts local image features, whereas the student model focuses on global\nfeatures using an attention mechanism. The Vision Transformer (ViT)\narchitecture is introduced as a robust framework for capturing global\ndependencies in images. Additionally, advanced variants of ViT, namely PVT,\nConvit, MVIT, Swin Transformer, and Twins, are discussed, highlighting their\ncontributions to computer vision tasks. The ConvNeXt model is introduced as a\nteacher model, known for its efficiency and effectiveness in computer vision.\nThe paper presents performance results for human action recognition on the\nStanford 40 dataset, comparing the accuracy and mAP of student models trained\nwith and without knowledge distillation. The findings illustrate that the\nsuggested approach significantly improves the accuracy and mAP when compared to\ntraining networks under regular settings. These findings emphasize the\npotential of combining local and global features in action recognition tasks.\n", "date": "2023-11-02", "categories": []}
{"id": "2308.03950", "title": "Zero-shot Skeleton-based Action Recognition via Mutual Information\n  Estimation and Maximization", "abstract": "  Zero-shot skeleton-based action recognition aims to recognize actions of\nunseen categories after training on data of seen categories. The key is to\nbuild the connection between visual and semantic space from seen to unseen\nclasses. Previous studies have primarily focused on encoding sequences into a\nsingular feature vector, with subsequent mapping the features to an identical\nanchor point within the embedded space. Their performance is hindered by 1) the\nignorance of the global visual/semantic distribution alignment, which results\nin a limitation to capture the true interdependence between the two spaces. 2)\nthe negligence of temporal information since the frame-wise features with rich\naction clues are directly pooled into a single feature vector. We propose a new\nzero-shot skeleton-based action recognition method via mutual information (MI)\nestimation and maximization. Specifically, 1) we maximize the MI between visual\nand semantic space for distribution alignment; 2) we leverage the temporal\ninformation for estimating the MI by encouraging MI to increase as more frames\nare observed. Extensive experiments on three large-scale skeleton action\ndatasets confirm the effectiveness of our method. Code:\nhttps://github.com/YujieOuO/SMIE.\n", "date": "2023-08-07", "categories": []}
{"id": "2310.17493", "title": "A Hybrid Graph Network for Complex Activity Detection in Video", "abstract": "  Interpretation and understanding of video presents a challenging computer\nvision task in numerous fields - e.g. autonomous driving and sports analytics.\nExisting approaches to interpreting the actions taking place within a video\nclip are based upon Temporal Action Localisation (TAL), which typically\nidentifies short-term actions. The emerging field of Complex Activity Detection\n(CompAD) extends this analysis to long-term activities, with a deeper\nunderstanding obtained by modelling the internal structure of a complex\nactivity taking place within the video. We address the CompAD problem using a\nhybrid graph neural network which combines attention applied to a graph\nencoding the local (short-term) dynamic scene with a temporal graph modelling\nthe overall long-duration activity. Our approach is as follows: i) Firstly, we\npropose a novel feature extraction technique which, for each video snippet,\ngenerates spatiotemporal `tubes' for the active elements (`agents') in the\n(local) scene by detecting individual objects, tracking them and then\nextracting 3D features from all the agent tubes as well as the overall scene.\nii) Next, we construct a local scene graph where each node (representing either\nan agent tube or the scene) is connected to all other nodes. Attention is then\napplied to this graph to obtain an overall representation of the local dynamic\nscene. iii) Finally, all local scene graph representations are interconnected\nvia a temporal graph, to estimate the complex activity class together with its\nstart and end time. The proposed framework outperforms all previous\nstate-of-the-art methods on all three datasets including ActivityNet-1.3,\nThumos-14, and ROAD.\n", "date": "2023-10-26", "categories": []}
{"id": "2307.06947", "title": "Video-FocalNets: Spatio-Temporal Focal Modulation for Video Action\n  Recognition", "abstract": "  Recent video recognition models utilize Transformer models for long-range\nspatio-temporal context modeling. Video transformer designs are based on\nself-attention that can model global context at a high computational cost. In\ncomparison, convolutional designs for videos offer an efficient alternative but\nlack long-range dependency modeling. Towards achieving the best of both\ndesigns, this work proposes Video-FocalNet, an effective and efficient\narchitecture for video recognition that models both local and global contexts.\nVideo-FocalNet is based on a spatio-temporal focal modulation architecture that\nreverses the interaction and aggregation steps of self-attention for better\nefficiency. Further, the aggregation step and the interaction step are both\nimplemented using efficient convolution and element-wise multiplication\noperations that are computationally less expensive than their self-attention\ncounterparts on video representations. We extensively explore the design space\nof focal modulation-based spatio-temporal context modeling and demonstrate our\nparallel spatial and temporal encoding design to be the optimal choice.\nVideo-FocalNets perform favorably well against the state-of-the-art\ntransformer-based models for video recognition on five large-scale datasets\n(Kinetics-400, Kinetics-600, SS-v2, Diving-48, and ActivityNet-1.3) at a lower\ncomputational cost. Our code/models are released at\nhttps://github.com/TalalWasim/Video-FocalNets.\n", "date": "2023-07-13", "categories": []}
{"id": "2305.02324", "title": "Cross-Stream Contrastive Learning for Self-Supervised Skeleton-Based\n  Action Recognition", "abstract": "  Self-supervised skeleton-based action recognition enjoys a rapid growth along\nwith the development of contrastive learning. The existing methods rely on\nimposing invariance to augmentations of 3D skeleton within a single data\nstream, which merely leverages the easy positive pairs and limits the ability\nto explore the complicated movement patterns. In this paper, we advocate that\nthe defect of single-stream contrast and the lack of necessary feature\ntransformation are responsible for easy positives, and therefore propose a\nCross-Stream Contrastive Learning framework for skeleton-based action\nRepresentation learning (CSCLR). Specifically, the proposed CSCLR not only\nutilizes intra-stream contrast pairs, but introduces inter-stream contrast\npairs as hard samples to formulate a better representation learning. Besides,\nto further exploit the potential of positive pairs and increase the robustness\nof self-supervised representation learning, we propose a Positive Feature\nTransformation (PFT) strategy which adopts feature-level manipulation to\nincrease the variance of positive pairs. To validate the effectiveness of our\nmethod, we conduct extensive experiments on three benchmark datasets NTU-RGB+D\n60, NTU-RGB+D 120 and PKU-MMD. Experimental results show that our proposed\nCSCLR exceeds the state-of-the-art methods on a diverse range of evaluation\nprotocols.\n", "date": "2023-05-03", "categories": []}
{"id": "2309.01270", "title": "COMEDIAN: Self-Supervised Learning and Knowledge Distillation for Action\n  Spotting using Transformers", "abstract": "  We present COMEDIAN, a novel pipeline to initialize spatiotemporal\ntransformers for action spotting, which involves self-supervised learning and\nknowledge distillation. Action spotting is a timestamp-level temporal action\ndetection task. Our pipeline consists of three steps, with two initialization\nstages. First, we perform self-supervised initialization of a spatial\ntransformer using short videos as input. Additionally, we initialize a temporal\ntransformer that enhances the spatial transformer's outputs with global context\nthrough knowledge distillation from a pre-computed feature bank aligned with\neach short video segment. In the final step, we fine-tune the transformers to\nthe action spotting task. The experiments, conducted on the SoccerNet-v2\ndataset, demonstrate state-of-the-art performance and validate the\neffectiveness of COMEDIAN's pretraining paradigm. Our results highlight several\nadvantages of our pretraining pipeline, including improved performance and\nfaster convergence compared to non-pretrained models.\n", "date": "2023-09-03", "categories": []}
{"id": "2310.17323", "title": "IndustReal: A Dataset for Procedure Step Recognition Handling Execution\n  Errors in Egocentric Videos in an Industrial-Like Setting", "abstract": "  Although action recognition for procedural tasks has received notable\nattention, it has a fundamental flaw in that no measure of success for actions\nis provided. This limits the applicability of such systems especially within\nthe industrial domain, since the outcome of procedural actions is often\nsignificantly more important than the mere execution. To address this\nlimitation, we define the novel task of procedure step recognition (PSR),\nfocusing on recognizing the correct completion and order of procedural steps.\nAlongside the new task, we also present the multi-modal IndustReal dataset.\nUnlike currently available datasets, IndustReal contains procedural errors\n(such as omissions) as well as execution errors. A significant part of these\nerrors are exclusively present in the validation and test sets, making\nIndustReal suitable to evaluate robustness of algorithms to new, unseen\nmistakes. Additionally, to encourage reproducibility and allow for scalable\napproaches trained on synthetic data, the 3D models of all parts are publicly\navailable. Annotations and benchmark performance are provided for action\nrecognition and assembly state detection, as well as the new PSR task.\nIndustReal, along with the code and model weights, is available at:\nhttps://github.com/TimSchoonbeek/IndustReal .\n", "date": "2023-10-26", "categories": []}
{"id": "2310.17421", "title": "Distribution of Action Movements (DAM): A Descriptor for Human Action\n  Recognition", "abstract": "  Human action recognition from skeletal data is an important and active area\nof research in which the state of the art has not yet achieved near-perfect\naccuracy on many well-known datasets. In this paper, we introduce the\nDistribution of Action Movements Descriptor, a novel action descriptor based on\nthe distribution of the directions of the motions of the joints between frames,\nover the set of all possible motions in the dataset. The descriptor is computed\nas a normalized histogram over a set of representative directions of the\njoints, which are in turn obtained via clustering. While the descriptor is\nglobal in the sense that it represents the overall distribution of movement\ndirections of an action, it is able to partially retain its temporal structure\nby applying a windowing scheme.\n  The descriptor, together with a standard classifier, outperforms several\nstate-of-the-art techniques on many well-known datasets.\n", "date": "2023-10-26", "categories": []}
{"id": "2310.15324", "title": "Videoprompter: an ensemble of foundational models for zero-shot video\n  understanding", "abstract": "  Vision-language models (VLMs) classify the query video by calculating a\nsimilarity score between the visual features and text-based class label\nrepresentations. Recently, large language models (LLMs) have been used to\nenrich the text-based class labels by enhancing the descriptiveness of the\nclass names. However, these improvements are restricted to the text-based\nclassifier only, and the query visual features are not considered. In this\npaper, we propose a framework which combines pre-trained discriminative VLMs\nwith pre-trained generative video-to-text and text-to-text models. We introduce\ntwo key modifications to the standard zero-shot setting. First, we propose\nlanguage-guided visual feature enhancement and employ a video-to-text model to\nconvert the query video to its descriptive form. The resulting descriptions\ncontain vital visual cues of the query video, such as what objects are present\nand their spatio-temporal interactions. These descriptive cues provide\nadditional semantic knowledge to VLMs to enhance their zeroshot performance.\nSecond, we propose video-specific prompts to LLMs to generate more meaningful\ndescriptions to enrich class label representations. Specifically, we introduce\nprompt techniques to create a Tree Hierarchy of Categories for class names,\noffering a higher-level action context for additional visual cues, We\ndemonstrate the effectiveness of our approach in video understanding across\nthree different zero-shot settings: 1) video action recognition, 2)\nvideo-to-text and textto-video retrieval, and 3) time-sensitive video tasks.\nConsistent improvements across multiple benchmarks and with various VLMs\ndemonstrate the effectiveness of our proposed framework. Our code will be made\npublicly available.\n", "date": "2023-10-23", "categories": []}
{"id": "2310.15568", "title": "I$^2$MD: 3D Action Representation Learning with Inter- and Intra-modal\n  Mutual Distillation", "abstract": "  Recent progresses on self-supervised 3D human action representation learning\nare largely attributed to contrastive learning. However, in conventional\ncontrastive frameworks, the rich complementarity between different skeleton\nmodalities remains under-explored. Moreover, optimized with distinguishing\nself-augmented samples, models struggle with numerous similar positive\ninstances in the case of limited action categories. In this work, we tackle the\naforementioned problems by introducing a general Inter- and Intra-modal Mutual\nDistillation (I$^2$MD) framework. In I$^2$MD, we first re-formulate the\ncross-modal interaction as a Cross-modal Mutual Distillation (CMD) process.\nDifferent from existing distillation solutions that transfer the knowledge of a\npre-trained and fixed teacher to the student, in CMD, the knowledge is\ncontinuously updated and bidirectionally distilled between modalities during\npre-training. To alleviate the interference of similar samples and exploit\ntheir underlying contexts, we further design the Intra-modal Mutual\nDistillation (IMD) strategy, In IMD, the Dynamic Neighbors Aggregation (DNA)\nmechanism is first introduced, where an additional cluster-level discrimination\nbranch is instantiated in each modality. It adaptively aggregates\nhighly-correlated neighboring features, forming local cluster-level\ncontrasting. Mutual distillation is then performed between the two branches for\ncross-level knowledge exchange. Extensive experiments on three datasets show\nthat our approach sets a series of new records.\n", "date": "2023-10-24", "categories": []}
{"id": "2210.10352", "title": "Temporal Action Segmentation: An Analysis of Modern Techniques", "abstract": "  Temporal action segmentation (TAS) in videos aims at densely identifying\nvideo frames in minutes-long videos with multiple action classes. As a\nlong-range video understanding task, researchers have developed an extended\ncollection of methods and examined their performance using various benchmarks.\nDespite the rapid growth of TAS techniques in recent years, no systematic\nsurvey has been conducted in these sectors. This survey analyzes and summarizes\nthe most significant contributions and trends. In particular, we first examine\nthe task definition, common benchmarks, types of supervision, and prevalent\nevaluation measures. In addition, we systematically investigate two essential\ntechniques of this topic, i.e., frame representation and temporal modeling,\nwhich have been studied extensively in the literature. We then conduct a\nthorough review of existing TAS works categorized by their levels of\nsupervision and conclude our survey by identifying and emphasizing several\nresearch gaps. In addition, we have curated a list of TAS resources, which is\navailable at https://github.com/nus-cvml/awesome-temporal-action-segmentation.\n", "date": "2022-10-19", "categories": []}
{"id": "2310.14390", "title": "Cross-Domain HAR: Few Shot Transfer Learning for Human Activity\n  Recognition", "abstract": "  The ubiquitous availability of smartphones and smartwatches with integrated\ninertial measurement units (IMUs) enables straightforward capturing of human\nactivities. For specific applications of sensor based human activity\nrecognition (HAR), however, logistical challenges and burgeoning costs render\nespecially the ground truth annotation of such data a difficult endeavor,\nresulting in limited scale and diversity of datasets. Transfer learning, i.e.,\nleveraging publicly available labeled datasets to first learn useful\nrepresentations that can then be fine-tuned using limited amounts of labeled\ndata from a target domain, can alleviate some of the performance issues of\ncontemporary HAR systems. Yet they can fail when the differences between source\nand target conditions are too large and/ or only few samples from a target\napplication domain are available, each of which are typical challenges in\nreal-world human activity recognition scenarios. In this paper, we present an\napproach for economic use of publicly available labeled HAR datasets for\neffective transfer learning. We introduce a novel transfer learning framework,\nCross-Domain HAR, which follows the teacher-student self-training paradigm to\nmore effectively recognize activities with very limited label information. It\nbridges conceptual gaps between source and target domains, including sensor\nlocations and type of activities. Through our extensive experimental evaluation\non a range of benchmark datasets, we demonstrate the effectiveness of our\napproach for practically relevant few shot activity recognition scenarios. We\nalso present a detailed analysis into how the individual components of our\nframework affect downstream performance.\n", "date": "2023-10-22", "categories": []}
{"id": "2310.14416", "title": "ConViViT -- A Deep Neural Network Combining Convolutions and Factorized\n  Self-Attention for Human Activity Recognition", "abstract": "  The Transformer architecture has gained significant popularity in computer\nvision tasks due to its capacity to generalize and capture long-range\ndependencies. This characteristic makes it well-suited for generating\nspatiotemporal tokens from videos. On the other hand, convolutions serve as the\nfundamental backbone for processing images and videos, as they efficiently\naggregate information within small local neighborhoods to create spatial tokens\nthat describe the spatial dimension of a video. While both CNN-based\narchitectures and pure transformer architectures are extensively studied and\nutilized by researchers, the effective combination of these two backbones has\nnot received comparable attention in the field of activity recognition. In this\nresearch, we propose a novel approach that leverages the strengths of both CNNs\nand Transformers in an hybrid architecture for performing activity recognition\nusing RGB videos. Specifically, we suggest employing a CNN network to enhance\nthe video representation by generating a 128-channel video that effectively\nseparates the human performing the activity from the background. Subsequently,\nthe output of the CNN module is fed into a transformer to extract\nspatiotemporal tokens, which are then used for classification purposes. Our\narchitecture has achieved new SOTA results with 90.05 \\%, 99.6\\%, and 95.09\\%\non HMDB51, UCF101, and ETRI-Activity3D respectively.\n", "date": "2023-10-22", "categories": []}
{"id": "2310.14556", "title": "S3Aug: Segmentation, Sampling, and Shift for Action Recognition", "abstract": "  Action recognition is a well-established area of research in computer vision.\nIn this paper, we propose S3Aug, a video data augmenatation for action\nrecognition. Unlike conventional video data augmentation methods that involve\ncutting and pasting regions from two videos, the proposed method generates new\nvideos from a single training video through segmentation and label-to-image\ntransformation. Furthermore, the proposed method modifies certain categories of\nlabel images by sampling to generate a variety of videos, and shifts\nintermediate features to enhance the temporal coherency between frames of the\ngenerate videos. Experimental results on the UCF101, HMDB51, and Mimetics\ndatasets demonstrate the effectiveness of the proposed method, paricularlly for\nout-of-context videos of the Mimetics dataset.\n", "date": "2023-10-23", "categories": []}
{"id": "2301.03462", "title": "Unleashing the Power of Shared Label Structures for Human Activity\n  Recognition", "abstract": "  Current human activity recognition (HAR) techniques regard activity labels as\ninteger class IDs without explicitly modeling the semantics of class labels. We\nobserve that different activity names often have shared structures. For\nexample, \"open door\" and \"open fridge\" both have \"open\" as the action; \"kicking\nsoccer ball\" and \"playing tennis ball\" both have \"ball\" as the object. Such\nshared structures in label names can be translated to the similarity in sensory\ndata and modeling common structures would help uncover knowledge across\ndifferent activities, especially for activities with limited samples. In this\npaper, we propose SHARE, a HAR framework that takes into account shared\nstructures of label names for different activities. To exploit the shared\nstructures, SHARE comprises an encoder for extracting features from input\nsensory time series and a decoder for generating label names as a token\nsequence. We also propose three label augmentation techniques to help the model\nmore effectively capture semantic structures across activities, including a\nbasic token-level augmentation, and two enhanced embedding-level and\nsequence-level augmentations utilizing the capabilities of pre-trained models.\nSHARE outperforms state-of-the-art HAR models in extensive experiments on seven\nHAR benchmark datasets. We also evaluate in few-shot learning and label\nimbalance settings and observe even more significant performance gap.\n", "date": "2023-01-01", "categories": []}
{"id": "2310.13039", "title": "Human Pose-based Estimation, Tracking and Action Recognition with Deep\n  Learning: A Survey", "abstract": "  Human pose analysis has garnered significant attention within both the\nresearch community and practical applications, owing to its expanding array of\nuses, including gaming, video surveillance, sports performance analysis, and\nhuman-computer interactions, among others. The advent of deep learning has\nsignificantly improved the accuracy of pose capture, making pose-based\napplications increasingly practical. This paper presents a comprehensive survey\nof pose-based applications utilizing deep learning, encompassing pose\nestimation, pose tracking, and action recognition.Pose estimation involves the\ndetermination of human joint positions from images or image sequences. Pose\ntracking is an emerging research direction aimed at generating consistent human\npose trajectories over time. Action recognition, on the other hand, targets the\nidentification of action types using pose estimation or tracking data. These\nthree tasks are intricately interconnected, with the latter often reliant on\nthe former. In this survey, we comprehensively review related works, spanning\nfrom single-person pose estimation to multi-person pose estimation, from 2D\npose estimation to 3D pose estimation, from single image to video, from mining\ntemporal context gradually to pose tracking, and lastly from tracking to\npose-based action recognition. As a survey centered on the application of deep\nlearning to pose analysis, we explicitly discuss both the strengths and\nlimitations of existing techniques. Notably, we emphasize methodologies for\nintegrating these three tasks into a unified framework within video sequences.\nAdditionally, we explore the challenges involved and outline potential\ndirections for future research.\n", "date": "2023-10-19", "categories": []}
{"id": "2310.12085", "title": "On the Benefit of Generative Foundation Models for Human Activity\n  Recognition", "abstract": "  In human activity recognition (HAR), the limited availability of annotated\ndata presents a significant challenge. Drawing inspiration from the latest\nadvancements in generative AI, including Large Language Models (LLMs) and\nmotion synthesis models, we believe that generative AI can address this data\nscarcity by autonomously generating virtual IMU data from text descriptions.\nBeyond this, we spotlight several promising research pathways that could\nbenefit from generative AI for the community, including the generating\nbenchmark datasets, the development of foundational models specific to HAR, the\nexploration of hierarchical structures within HAR, breaking down complex\nactivities, and applications in health sensing and activity summarization.\n", "date": "2023-10-18", "categories": []}
{"id": "2207.05515", "title": "Compound Prototype Matching for Few-shot Action Recognition", "abstract": "  Few-shot action recognition aims to recognize novel action classes using only\na small number of labeled training samples. In this work, we propose a novel\napproach that first summarizes each video into compound prototypes consisting\nof a group of global prototypes and a group of focused prototypes, and then\ncompares video similarity based on the prototypes. Each global prototype is\nencouraged to summarize a specific aspect from the entire video, for example,\nthe start/evolution of the action. Since no clear annotation is provided for\nthe global prototypes, we use a group of focused prototypes to focus on certain\ntimestamps in the video. We compare video similarity by matching the compound\nprototypes between the support and query videos. The global prototypes are\ndirectly matched to compare videos from the same perspective, for example, to\ncompare whether two actions start similarly. For the focused prototypes, since\nactions have various temporal variations in the videos, we apply bipartite\nmatching to allow the comparison of actions with different temporal positions\nand shifts. Experiments demonstrate that our proposed method achieves\nstate-of-the-art results on multiple benchmarks.\n", "date": "2022-07-12", "categories": []}
{"id": "2304.08799", "title": "Self-Supervised 3D Action Representation Learning with Skeleton Cloud\n  Colorization", "abstract": "  3D Skeleton-based human action recognition has attracted increasing attention\nin recent years. Most of the existing work focuses on supervised learning which\nrequires a large number of labeled action sequences that are often expensive\nand time-consuming to annotate. In this paper, we address self-supervised 3D\naction representation learning for skeleton-based action recognition. We\ninvestigate self-supervised representation learning and design a novel skeleton\ncloud colorization technique that is capable of learning spatial and temporal\nskeleton representations from unlabeled skeleton sequence data. We represent a\nskeleton action sequence as a 3D skeleton cloud and colorize each point in the\ncloud according to its temporal and spatial orders in the original\n(unannotated) skeleton sequence. Leveraging the colorized skeleton point cloud,\nwe design an auto-encoder framework that can learn spatial-temporal features\nfrom the artificial color labels of skeleton joints effectively. Specifically,\nwe design a two-steam pretraining network that leverages fine-grained and\ncoarse-grained colorization to learn multi-scale spatial-temporal features. In\naddition, we design a Masked Skeleton Cloud Repainting task that can pretrain\nthe designed auto-encoder framework to learn informative representations. We\nevaluate our skeleton cloud colorization approach with linear classifiers\ntrained under different configurations, including unsupervised,\nsemi-supervised, fully-supervised, and transfer learning settings. Extensive\nexperiments on NTU RGB+D, NTU RGB+D 120, PKU-MMD, NW-UCLA, and UWA3D datasets\nshow that the proposed method outperforms existing unsupervised and\nsemi-supervised 3D action recognition methods by large margins and achieves\ncompetitive performance in supervised 3D action recognition as well.\n", "date": "2023-04-18", "categories": []}
{"id": "2307.06527", "title": "Free-Form Composition Networks for Egocentric Action Recognition", "abstract": "  Egocentric action recognition is gaining significant attention in the field\nof human action recognition. In this paper, we address data scarcity issue in\negocentric action recognition from a compositional generalization perspective.\nTo tackle this problem, we propose a free-form composition network (FFCN) that\ncan simultaneously learn disentangled verb, preposition, and noun\nrepresentations, and then use them to compose new samples in the feature space\nfor rare classes of action videos. First, we use a graph to capture the\nspatial-temporal relations among different hand/object instances in each action\nvideo. We thus decompose each action into a set of verb and preposition\nspatial-temporal representations using the edge features in the graph. The\ntemporal decomposition extracts verb and preposition representations from\ndifferent video frames, while the spatial decomposition adaptively learns verb\nand preposition representations from action-related instances in each frame.\nWith these spatial-temporal representations of verbs and prepositions, we can\ncompose new samples for those rare classes in a free-form manner, which is not\nrestricted to a rigid form of a verb and a noun. The proposed FFCN can directly\ngenerate new training data samples for rare classes, hence significantly\nimprove action recognition performance. We evaluated our method on three\npopular egocentric action recognition datasets, Something-Something V2, H2O,\nand EPIC-KITCHENS-100, and the experimental results demonstrate the\neffectiveness of the proposed method for handling data scarcity problems,\nincluding long-tailed and few-shot egocentric action recognition.\n", "date": "2023-07-13", "categories": []}
{"id": "2310.10125", "title": "Few-shot Action Recognition with Captioning Foundation Models", "abstract": "  Transferring vision-language knowledge from pretrained multimodal foundation\nmodels to various downstream tasks is a promising direction. However, most\ncurrent few-shot action recognition methods are still limited to a single\nvisual modality input due to the high cost of annotating additional textual\ndescriptions. In this paper, we develop an effective plug-and-play framework\ncalled CapFSAR to exploit the knowledge of multimodal models without manually\nannotating text. To be specific, we first utilize a captioning foundation model\n(i.e., BLIP) to extract visual features and automatically generate associated\ncaptions for input videos. Then, we apply a text encoder to the synthetic\ncaptions to obtain representative text embeddings. Finally, a visual-text\naggregation module based on Transformer is further designed to incorporate\ncross-modal spatio-temporal complementary information for reliable few-shot\nmatching. In this way, CapFSAR can benefit from powerful multimodal knowledge\nof pretrained foundation models, yielding more comprehensive classification in\nthe low-shot regime. Extensive experiments on multiple standard few-shot\nbenchmarks demonstrate that the proposed CapFSAR performs favorably against\nexisting methods and achieves state-of-the-art performance. The code will be\nmade publicly available.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.10131", "title": "3DYoga90: A Hierarchical Video Dataset for Yoga Pose Understanding", "abstract": "  The increasing popularity of exercises including yoga and Pilates has created\na greater demand for professional exercise video datasets in the realm of\nartificial intelligence. In this study, we developed 3DYoga901, which is\norganized within a three-level label hierarchy. We have expanded the number of\nposes from an existing state-of-the-art dataset, increasing it from 82 to 90\nposes. Our dataset includes meticulously curated RGB yoga pose videos and 3D\nskeleton sequences. This dataset was created by a dedicated team of six\nindividuals, including yoga instructors. It stands out as one of the most\ncomprehensive open datasets, featuring the largest collection of RGB videos and\n3D skeleton sequences among publicly available resources. This contribution has\nthe potential to significantly advance the field of yoga action recognition and\npose assessment. Additionally, we conducted experiments to evaluate the\npracticality of our proposed dataset. We employed three different model\nvariants for benchmarking purposes.\n", "date": "2023-10-16", "categories": []}
{"id": "2310.09114", "title": "Timestamp-supervised Wearable-based Activity Segmentation and\n  Recognition with Contrastive Learning and Order-Preserving Optimal Transport", "abstract": "  Human activity recognition (HAR) with wearables is one of the serviceable\ntechnologies in ubiquitous and mobile computing applications. The\nsliding-window scheme is widely adopted while suffering from the multi-class\nwindows problem. As a result, there is a growing focus on joint segmentation\nand recognition with deep-learning methods, aiming at simultaneously dealing\nwith HAR and time-series segmentation issues. However, obtaining the full\nactivity annotations of wearable data sequences is resource-intensive or\ntime-consuming, while unsupervised methods yield poor performance. To address\nthese challenges, we propose a novel method for joint activity segmentation and\nrecognition with timestamp supervision, in which only a single annotated sample\nis needed in each activity segment. However, the limited information of sparse\nannotations exacerbates the gap between recognition and segmentation tasks,\nleading to sub-optimal model performance. Therefore, the prototypes are\nestimated by class-activation maps to form a sample-to-prototype contrast\nmodule for well-structured embeddings. Moreover, with the optimal transport\ntheory, our approach generates the sample-level pseudo-labels that take\nadvantage of unlabeled data between timestamp annotations for further\nperformance improvement. Comprehensive experiments on four public HAR datasets\ndemonstrate that our model trained with timestamp supervision is superior to\nthe state-of-the-art weakly-supervised methods and achieves comparable\nperformance to the fully-supervised approaches.\n", "date": "2023-10-13", "categories": []}
{"id": "2210.09222", "title": "MMTSA: Multimodal Temporal Segment Attention Network for Efficient Human\n  Activity Recognition", "abstract": "  Multimodal sensors provide complementary information to develop accurate\nmachine-learning methods for human activity recognition (HAR), but introduce\nsignificantly higher computational load, which reduces efficiency. This paper\nproposes an efficient multimodal neural architecture for HAR using an RGB\ncamera and inertial measurement units (IMUs) called Multimodal Temporal Segment\nAttention Network (MMTSA). MMTSA first transforms IMU sensor data into a\ntemporal and structure-preserving gray-scale image using the Gramian Angular\nField (GAF), representing the inherent properties of human activities. MMTSA\nthen applies a multimodal sparse sampling method to reduce data redundancy.\nLastly, MMTSA adopts an inter-segment attention module for efficient multimodal\nfusion. Using three well-established public datasets, we evaluated MMTSA's\neffectiveness and efficiency in HAR. Results show that our method achieves\nsuperior performance improvements 11.13% of cross-subject F1-score on the MMAct\ndataset than the previous state-of-the-art (SOTA) methods. The ablation study\nand analysis suggest that MMTSA's effectiveness in fusing multimodal data for\naccurate HAR. The efficiency evaluation on an edge device showed that MMTSA\nachieved significantly better accuracy, lower computational load, and lower\ninference latency than SOTA methods.\n", "date": "2022-10-14", "categories": []}
{"id": "2302.09018", "title": "Self-supervised Action Representation Learning from Partial\n  Spatio-Temporal Skeleton Sequences", "abstract": "  Self-supervised learning has demonstrated remarkable capability in\nrepresentation learning for skeleton-based action recognition. Existing methods\nmainly focus on applying global data augmentation to generate different views\nof the skeleton sequence for contrastive learning. However, due to the rich\naction clues in the skeleton sequences, existing methods may only take a global\nperspective to learn to discriminate different skeletons without thoroughly\nleveraging the local relationship between different skeleton joints and video\nframes, which is essential for real-world applications. In this work, we\npropose a Partial Spatio-Temporal Learning (PSTL) framework to exploit the\nlocal relationship from a partial skeleton sequences built by a unique\nspatio-temporal masking strategy. Specifically, we construct a\nnegative-sample-free triplet steam structure that is composed of an anchor\nstream without any masking, a spatial masking stream with Central Spatial\nMasking (CSM), and a temporal masking stream with Motion Attention Temporal\nMasking (MATM). The feature cross-correlation matrix is measured between the\nanchor stream and the other two masking streams, respectively. (1) Central\nSpatial Masking discards selected joints from the feature calculation process,\nwhere the joints with a higher degree of centrality have a higher possibility\nof being selected. (2) Motion Attention Temporal Masking leverages the motion\nof action and remove frames that move faster with a higher possibility. Our\nmethod achieves SOTA performance on NTU-60, NTU-120 and PKU-MMD under various\ndownstream tasks. A practical evaluation is performed where some skeleton\njoints are lost in downstream tasks. In contrast to previous methods that\nsuffer from large performance drops, our PSTL can still achieve remarkable\nresults, validating the robustness of our method. Code:\nhttps://github.com/YujieOuO/PSTL.git.\n", "date": "2023-02-17", "categories": []}
{"id": "2310.08451", "title": "Proving the Potential of Skeleton Based Action Recognition to Automate\n  the Analysis of Manual Processes", "abstract": "  In manufacturing sectors such as textiles and electronics, manual processes\nare a fundamental part of production. The analysis and monitoring of the\nprocesses is necessary for efficient production design. Traditional methods for\nanalyzing manual processes are complex, expensive, and inflexible. Compared to\nestablished approaches such as Methods-Time-Measurement (MTM), machine learning\n(ML) methods promise: Higher flexibility, self-sufficient & permanent use,\nlower costs. In this work, based on a video stream, the current motion class in\na manual assembly process is detected. With information on the current motion,\nKey-Performance-Indicators (KPIs) can be derived easily. A skeleton-based\naction recognition approach is taken, as this field recently shows major\nsuccess in machine vision tasks. For skeleton-based action recognition in\nmanual assembly, no sufficient pre-work could be found. Therefore, a ML\npipeline is developed, to enable extensive research on different (pre-)\nprocessing methods and neural nets. Suitable well generalizing approaches are\nfound, proving the potential of ML to enhance analyzation of manual processes.\nModels detect the current motion, performed by an operator in manual assembly,\nbut the results can be transferred to all kinds of manual processes.\n", "date": "2023-10-12", "categories": []}
{"id": "2209.13808", "title": "Streaming Video Temporal Action Segmentation In Real Time", "abstract": "  Temporal action segmentation (TAS) is a critical step toward long-term video\nunderstanding. Recent studies follow a pattern that builds models based on\nfeatures instead of raw video picture information. However, we claim those\nmodels are trained complicatedly and limit application scenarios. It is hard\nfor them to segment human actions of video in real time because they must work\nafter the full video features are extracted. As the real-time action\nsegmentation task is different from TAS task, we define it as streaming video\nreal-time temporal action segmentation (SVTAS) task. In this paper, we propose\na real-time end-to-end multi-modality model for SVTAS task. More specifically,\nunder the circumstances that we cannot get any future information, we segment\nthe current human action of streaming video chunk in real time. Furthermore,\nthe model we propose combines the last steaming video chunk feature extracted\nby language model with the current image feature extracted by image model to\nimprove the quantity of real-time temporal action segmentation. To the best of\nour knowledge, it is the first multi-modality real-time temporal action\nsegmentation model. Under the same evaluation criteria as full video temporal\naction segmentation, our model segments human action in real time with less\nthan 40% of state-of-the-art model computation and achieves 90% of the accuracy\nof the full video state-of-the-art model.\n", "date": "2022-09-28", "categories": []}
{"id": "2303.09756", "title": "Video Action Recognition with Attentive Semantic Units", "abstract": "  Visual-Language Models (VLMs) have significantly advanced action video\nrecognition. Supervised by the semantics of action labels, recent works adapt\nthe visual branch of VLMs to learn video representations. Despite the\neffectiveness proved by these works, we believe that the potential of VLMs has\nyet to be fully harnessed. In light of this, we exploit the semantic units (SU)\nhiding behind the action labels and leverage their correlations with\nfine-grained items in frames for more accurate action recognition. SUs are\nentities extracted from the language descriptions of the entire action set,\nincluding body parts, objects, scenes, and motions. To further enhance the\nalignments between visual contents and the SUs, we introduce a multi-region\nmodule (MRA) to the visual branch of the VLM. The MRA allows the perception of\nregion-aware visual features beyond the original global feature. Our method\nadaptively attends to and selects relevant SUs with visual features of frames.\nWith a cross-modal decoder, the selected SUs serve to decode spatiotemporal\nvideo representations. In summary, the SUs as the medium can boost\ndiscriminative ability and transferability. Specifically, in fully-supervised\nlearning, our method achieved 87.8% top-1 accuracy on Kinetics-400. In K=2\nfew-shot experiments, our method surpassed the previous state-of-the-art by\n+7.1% and +15.0% on HMDB-51 and UCF-101, respectively.\n", "date": "2023-03-17", "categories": []}
{"id": "2310.05938", "title": "Component attention network for multimodal dance improvisation\n  recognition", "abstract": "  Dance improvisation is an active research topic in the arts. Motion analysis\nof improvised dance can be challenging due to its unique dynamics. Data-driven\ndance motion analysis, including recognition and generation, is often limited\nto skeletal data. However, data of other modalities, such as audio, can be\nrecorded and benefit downstream tasks. This paper explores the application and\nperformance of multimodal fusion methods for human motion recognition in the\ncontext of dance improvisation. We propose an attention-based model, component\nattention network (CANet), for multimodal fusion on three levels: 1) feature\nfusion with CANet, 2) model fusion with CANet and graph convolutional network\n(GCN), and 3) late fusion with a voting strategy. We conduct thorough\nexperiments to analyze the impact of each modality in different fusion methods\nand distinguish critical temporal or component features. We show that our\nproposed model outperforms the two baseline methods, demonstrating its\npotential for analyzing improvisation in dance.\n", "date": "2023-08-24", "categories": []}
{"id": "2308.14900", "title": "BIT: Bi-Level Temporal Modeling for Efficient Supervised Action\n  Segmentation", "abstract": "  We address the task of supervised action segmentation which aims to partition\na video into non-overlapping segments, each representing a different action.\nRecent works apply transformers to perform temporal modeling at the\nframe-level, which suffer from high computational cost and cannot well capture\naction dependencies over long temporal horizons. To address these issues, we\npropose an efficient BI-level Temporal modeling (BIT) framework that learns\nexplicit action tokens to represent action segments, in parallel performs\ntemporal modeling on frame and action levels, while maintaining a low\ncomputational cost. Our model contains (i) a frame branch that uses convolution\nto learn frame-level relationships, (ii) an action branch that uses transformer\nto learn action-level dependencies with a small set of action tokens and (iii)\ncross-attentions to allow communication between the two branches. We apply and\nextend a set-prediction objective to allow each action token to represent one\nor multiple action segments, thus can avoid learning a large number of tokens\nover long videos with many segments. Thanks to the design of our action branch,\nwe can also seamlessly leverage textual transcripts of videos (when available)\nto help action segmentation by using them to initialize the action tokens. We\nevaluate our model on four video datasets (two egocentric and two third-person)\nfor action segmentation with and without transcripts, showing that BIT\nsignificantly improves the state-of-the-art accuracy with much lower\ncomputational cost (30 times faster) compared to existing transformer-based\nmethods.\n", "date": "2023-08-28", "categories": []}
{"id": "2310.05511", "title": "Proposal-based Temporal Action Localization with Point-level Supervision", "abstract": "  Point-level supervised temporal action localization (PTAL) aims at\nrecognizing and localizing actions in untrimmed videos where only a single\npoint (frame) within every action instance is annotated in training data.\nWithout temporal annotations, most previous works adopt the multiple instance\nlearning (MIL) framework, where the input video is segmented into\nnon-overlapped short snippets, and action classification is performed\nindependently on every short snippet. We argue that the MIL framework is\nsuboptimal for PTAL because it operates on separated short snippets that\ncontain limited temporal information. Therefore, the classifier only focuses on\nseveral easy-to-distinguish snippets instead of discovering the whole action\ninstance without missing any relevant snippets. To alleviate this problem, we\npropose a novel method that localizes actions by generating and evaluating\naction proposals of flexible duration that involve more comprehensive temporal\ninformation. Moreover, we introduce an efficient clustering algorithm to\nefficiently generate dense pseudo labels that provide stronger supervision, and\na fine-grained contrastive loss to further refine the quality of pseudo labels.\nExperiments show that our proposed method achieves competitive or superior\nperformance to the state-of-the-art methods and some fully-supervised methods\non four benchmarks: ActivityNet 1.3, THUMOS 14, GTEA, and BEOID datasets.\n", "date": "2023-10-09", "categories": []}
{"id": "2310.03760", "title": "Investigating Deep Neural Network Architecture and Feature Extraction\n  Designs for Sensor-based Human Activity Recognition", "abstract": "  The extensive ubiquitous availability of sensors in smart devices and the\nInternet of Things (IoT) has opened up the possibilities for implementing\nsensor-based activity recognition. As opposed to traditional sensor time-series\nprocessing and hand-engineered feature extraction, in light of deep learning's\nproven effectiveness across various domains, numerous deep methods have been\nexplored to tackle the challenges in activity recognition, outperforming the\ntraditional signal processing and traditional machine learning approaches. In\nthis work, by performing extensive experimental studies on two human activity\nrecognition datasets, we investigate the performance of common deep learning\nand machine learning approaches as well as different training mechanisms (such\nas contrastive learning), and various feature representations extracted from\nthe sensor time-series data and measure their effectiveness for the human\nactivity recognition task.\n", "date": "2023-09-26", "categories": []}
{"id": "2310.03108", "title": "Reinforcement Learning-based Mixture of Vision Transformers for Video\n  Violence Recognition", "abstract": "  Video violence recognition based on deep learning concerns accurate yet\nscalable human violence recognition. Currently, most state-of-the-art video\nviolence recognition studies use CNN-based models to represent and categorize\nvideos. However, recent studies suggest that pre-trained transformers are more\naccurate than CNN-based models on various video analysis benchmarks. Yet these\nmodels are not thoroughly evaluated for video violence recognition. This paper\nintroduces a novel transformer-based Mixture of Experts (MoE) video violence\nrecognition system. Through an intelligent combination of large vision\ntransformers and efficient transformer architectures, the proposed system not\nonly takes advantage of the vision transformer architecture but also reduces\nthe cost of utilizing large vision transformers. The proposed architecture\nmaximizes violence recognition system accuracy while actively reducing\ncomputational costs through a reinforcement learning-based router. The\nempirical results show the proposed MoE architecture's superiority over\nCNN-based models by achieving 92.4% accuracy on the RWF dataset.\n", "date": "2023-10-04", "categories": []}
{"id": "2310.03288", "title": "PoseAction: Action Recognition for Patients in the Ward using Deep\n  Learning Approaches", "abstract": "  Real-time intelligent detection and prediction of subjects' behavior\nparticularly their movements or actions is critical in the ward. This approach\noffers the advantage of reducing in-hospital care costs and improving the\nefficiency of healthcare workers, which is especially true for scenarios at\nnight or during peak admission periods. Therefore, in this work, we propose\nusing computer vision (CV) and deep learning (DL) methods for detecting\nsubjects and recognizing their actions. We utilize OpenPose as an accurate\nsubject detector for recognizing the positions of human subjects in the video\nstream. Additionally, we employ AlphAction's Asynchronous Interaction\nAggregation (AIA) network to predict the actions of detected subjects. This\nintegrated model, referred to as PoseAction, is proposed. At the same time, the\nproposed model is further trained to predict 12 common actions in ward areas,\nsuch as staggering, chest pain, and falling down, using medical-related video\nclips from the NTU RGB+D and NTU RGB+D 120 datasets. The results demonstrate\nthat PoseAction achieves the highest classification mAP of 98.72% (IoU@0.5).\nAdditionally, this study develops an online real-time mode for action\nrecognition, which strongly supports the clinical translation of PoseAction.\nFurthermore, using OpenPose's function for recognizing face key points, we also\nimplement face blurring, which is a practical solution to address the privacy\nprotection concerns of patients and healthcare workers. Nevertheless, the\ntraining data for PoseAction is currently limited, particularly in terms of\nlabel diversity. Consequently, the subsequent step involves utilizing a more\ndiverse dataset (including general actions) to train the model's parameters for\nimproved generalization.\n", "date": "2023-10-05", "categories": []}
{"id": "2310.03377", "title": "ACT-Net: Anchor-context Action Detection in Surgery Videos", "abstract": "  Recognition and localization of surgical detailed actions is an essential\ncomponent of developing a context-aware decision support system. However, most\nexisting detection algorithms fail to provide high-accuracy action classes even\nhaving their locations, as they do not consider the surgery procedure's\nregularity in the whole video. This limitation hinders their application.\nMoreover, implementing the predictions in clinical applications seriously needs\nto convey model confidence to earn entrustment, which is unexplored in surgical\naction prediction. In this paper, to accurately detect fine-grained actions\nthat happen at every moment, we propose an anchor-context action detection\nnetwork (ACTNet), including an anchor-context detection (ACD) module and a\nclass conditional diffusion (CCD) module, to answer the following questions: 1)\nwhere the actions happen; 2) what actions are; 3) how confidence predictions\nare. Specifically, the proposed ACD module spatially and temporally highlights\nthe regions interacting with the extracted anchor in surgery video, which\noutputs action location and its class distribution based on anchor-context\ninteractions. Considering the full distribution of action classes in videos,\nthe CCD module adopts a denoising diffusion-based generative model conditioned\non our ACD estimator to further reconstruct accurately the action predictions.\nMoreover, we utilize the stochastic nature of the diffusion model outputs to\naccess model confidence for each prediction. Our method reports the\nstate-of-the-art performance, with improvements of 4.0% mAP against baseline on\nthe surgical video dataset.\n", "date": "2023-10-05", "categories": []}
{"id": "2310.03456", "title": "Multi-Resolution Audio-Visual Feature Fusion for Temporal Action\n  Localization", "abstract": "  Temporal Action Localization (TAL) aims to identify actions' start, end, and\nclass labels in untrimmed videos. While recent advancements using transformer\nnetworks and Feature Pyramid Networks (FPN) have enhanced visual feature\nrecognition in TAL tasks, less progress has been made in the integration of\naudio features into such frameworks. This paper introduces the Multi-Resolution\nAudio-Visual Feature Fusion (MRAV-FF), an innovative method to merge\naudio-visual data across different temporal resolutions. Central to our\napproach is a hierarchical gated cross-attention mechanism, which discerningly\nweighs the importance of audio information at diverse temporal scales. Such a\ntechnique not only refines the precision of regression boundaries but also\nbolsters classification confidence. Importantly, MRAV-FF is versatile, making\nit compatible with existing FPN TAL architectures and offering a significant\nenhancement in performance when audio data is available.\n", "date": "2023-10-05", "categories": []}
{"id": "2310.02887", "title": "A Grammatical Compositional Model for Video Action Detection", "abstract": "  Analysis of human actions in videos demands understanding complex human\ndynamics, as well as the interaction between actors and context. However, these\ninteraction relationships usually exhibit large intra-class variations from\ndiverse human poses or object manipulations, and fine-grained inter-class\ndifferences between similar actions. Thus the performance of existing methods\nis severely limited. Motivated by the observation that interactive actions can\nbe decomposed into actor dynamics and participating objects or humans, we\npropose to investigate the composite property of them. In this paper, we\npresent a novel Grammatical Compositional Model (GCM) for action detection\nbased on typical And-Or graphs. Our model exploits the intrinsic structures and\nlatent relationships of actions in a hierarchical manner to harness both the\ncompositionality of grammar models and the capability of expressing rich\nfeatures of DNNs. The proposed model can be readily embodied into a neural\nnetwork module for efficient optimization in an end-to-end manner. Extensive\nexperiments are conducted on the AVA dataset and the Something-Else task to\ndemonstrate the superiority of our model, meanwhile the interpretability is\nenhanced through an inference parsing procedure.\n", "date": "2023-10-04", "categories": []}
{"id": "2310.00670", "title": "A Hierarchical Graph-based Approach for Recognition and Description\n  Generation of Bimanual Actions in Videos", "abstract": "  Nuanced understanding and the generation of detailed descriptive content for\n(bimanual) manipulation actions in videos is important for disciplines such as\nrobotics, human-computer interaction, and video content analysis. This study\ndescribes a novel method, integrating graph based modeling with layered\nhierarchical attention mechanisms, resulting in higher precision and better\ncomprehensiveness of video descriptions. To achieve this, we encode, first, the\nspatio-temporal inter dependencies between objects and actions with scene\ngraphs and we combine this, in a second step, with a novel 3-level architecture\ncreating a hierarchical attention mechanism using Graph Attention Networks\n(GATs). The 3-level GAT architecture allows recognizing local, but also global\ncontextual elements. This way several descriptions with different semantic\ncomplexity can be generated in parallel for the same video clip, enhancing the\ndiscriminative accuracy of action recognition and action description. The\nperformance of our approach is empirically tested using several 2D and 3D\ndatasets. By comparing our method to the state of the art we consistently\nobtain better performance concerning accuracy, precision, and contextual\nrelevance when evaluating action recognition as well as description generation.\nIn a large set of ablation experiments we also assess the role of the different\ncomponents of our model. With our multi-level approach the system obtains\ndifferent semantic description depths, often observed in descriptions made by\ndifferent people, too. Furthermore, better insight into bimanual hand-object\ninteractions as achieved by our model may portend advancements in the field of\nrobotics, enabling the emulation of intricate human actions with heightened\nprecision.\n", "date": "2023-10-01", "categories": []}
{"id": "2310.00831", "title": "Action Recognition Utilizing YGAR Dataset", "abstract": "  The scarcity of high quality actions video data is a bottleneck in the\nresearch and application of action recognition. Although significant effort has\nbeen made in this area, there still exist gaps in the range of available data\ntypes a more flexible and comprehensive data set could help bridge. In this\npaper, we present a new 3D actions data simulation engine and generate 3 sets\nof sample data to demonstrate its current functionalities. With the new data\ngeneration process, we demonstrate its applications to image classifications,\naction recognitions and potential to evolve into a system that would allow the\nexploration of much more complex action recognition tasks. In order to show off\nthese capabilities, we also train and test a list of commonly used models for\nimage recognition to demonstrate the potential applications and capabilities of\nthe data sets and their generation process.\n", "date": "2023-10-02", "categories": []}
{"id": "2211.12883", "title": "Mitigating and Evaluating Static Bias of Action Representations in the\n  Background and the Foreground", "abstract": "  In video action recognition, shortcut static features can interfere with the\nlearning of motion features, resulting in poor out-of-distribution (OOD)\ngeneralization. The video background is clearly a source of static bias, but\nthe video foreground, such as the clothing of the actor, can also provide\nstatic bias. In this paper, we empirically verify the existence of foreground\nstatic bias by creating test videos with conflicting signals from the static\nand moving portions of the video. To tackle this issue, we propose a simple yet\neffective technique, StillMix, to learn robust action representations.\nSpecifically, StillMix identifies bias-inducing video frames using a 2D\nreference network and mixes them with videos for training, serving as effective\nbias suppression even when we cannot explicitly extract the source of bias\nwithin each video frame or enumerate types of bias. Finally, to precisely\nevaluate static bias, we synthesize two new benchmarks, SCUBA for static cues\nin the background, and SCUFO for static cues in the foreground. With extensive\nexperiments, we demonstrate that StillMix mitigates both types of static bias\nand improves video representations for downstream applications. Code is\navailable at https://github.com/lihaoxin05/StillMix.\n", "date": "2022-11-23", "categories": []}
{"id": "2309.15635", "title": "Position and Orientation-Aware One-Shot Learning for Medical Action\n  Recognition from Signal Data", "abstract": "  In this work, we propose a position and orientation-aware one-shot learning\nframework for medical action recognition from signal data. The proposed\nframework comprises two stages and each stage includes signal-level image\ngeneration (SIG), cross-attention (CsA), dynamic time warping (DTW) modules and\nthe information fusion between the proposed privacy-preserved position and\norientation features. The proposed SIG method aims to transform the raw\nskeleton data into privacy-preserved features for training. The CsA module is\ndeveloped to guide the network in reducing medical action recognition bias and\nmore focusing on important human body parts for each specific action, aimed at\naddressing similar medical action related issues. Moreover, the DTW module is\nemployed to minimize temporal mismatching between instances and further improve\nmodel performance. Furthermore, the proposed privacy-preserved\norientation-level features are utilized to assist the position-level features\nin both of the two stages for enhancing medical action recognition performance.\nExtensive experimental results on the widely-used and well-known NTU RGB+D 60,\nNTU RGB+D 120, and PKU-MMD datasets all demonstrate the effectiveness of the\nproposed method, which outperforms the other state-of-the-art methods with\ngeneral dataset partitioning by 2.7%, 6.2% and 4.1%, respectively.\n", "date": "2023-09-27", "categories": []}
{"id": "2308.09946", "title": "Weakly-Supervised Action Localization by Hierarchically-structured\n  Latent Attention Modeling", "abstract": "  Weakly-supervised action localization aims to recognize and localize action\ninstancese in untrimmed videos with only video-level labels. Most existing\nmodels rely on multiple instance learning(MIL), where the predictions of\nunlabeled instances are supervised by classifying labeled bags. The MIL-based\nmethods are relatively well studied with cogent performance achieved on\nclassification but not on localization. Generally, they locate temporal regions\nby the video-level classification but overlook the temporal variations of\nfeature semantics. To address this problem, we propose a novel attention-based\nhierarchically-structured latent model to learn the temporal variations of\nfeature semantics. Specifically, our model entails two components, the first is\nan unsupervised change-points detection module that detects change-points by\nlearning the latent representations of video features in a temporal hierarchy\nbased on their rates of change, and the second is an attention-based\nclassification model that selects the change-points of the foreground as the\nboundaries. To evaluate the effectiveness of our model, we conduct extensive\nexperiments on two benchmark datasets, THUMOS-14 and ActivityNet-v1.3. The\nexperiments show that our method outperforms current state-of-the-art methods,\nand even achieves comparable performance with fully-supervised methods.\n", "date": "2023-08-19", "categories": []}
{"id": "2211.09558", "title": "ReLER@ZJU Submission to the Ego4D Moment Queries Challenge 2022", "abstract": "  In this report, we present the ReLER@ZJU1 submission to the Ego4D Moment\nQueries Challenge in ECCV 2022. In this task, the goal is to retrieve and\nlocalize all instances of possible activities in egocentric videos. Ego4D\ndataset is challenging for the temporal action localization task as the\ntemporal duration of the videos is quite long and each video contains multiple\naction instances with fine-grained action classes. To address these problems,\nwe utilize a multi-scale transformer to classify different action categories\nand predict the boundary of each instance. Moreover, in order to better capture\nthe long-term temporal dependencies in the long videos, we propose a\nsegment-level recurrence mechanism. Compared with directly feeding all video\nfeatures to the transformer encoder, the proposed segment-level recurrence\nmechanism alleviates the optimization difficulties and achieves better\nperformance. The final submission achieved Recall@1,tIoU=0.5 score of 37.24,\naverage mAP score of 17.67 and took 3-rd place on the leaderboard.\n", "date": "2022-11-17", "categories": []}
{"id": "2308.05430", "title": "Ensemble Modeling for Multimodal Visual Action Recognition", "abstract": "  In this work, we propose an ensemble modeling approach for multimodal action\nrecognition. We independently train individual modality models using a variant\nof focal loss tailored to handle the long-tailed distribution of the MECCANO\n[21] dataset. Based on the underlying principle of focal loss, which captures\nthe relationship between tail (scarce) classes and their prediction\ndifficulties, we propose an exponentially decaying variant of focal loss for\nour current task. It initially emphasizes learning from the hard misclassified\nexamples and gradually adapts to the entire range of examples in the dataset.\nThis annealing process encourages the model to strike a balance between\nfocusing on the sparse set of hard samples, while still leveraging the\ninformation provided by the easier ones. Additionally, we opt for the late\nfusion strategy to combine the resultant probability distributions from RGB and\nDepth modalities for final action prediction. Experimental evaluations on the\nMECCANO dataset demonstrate the effectiveness of our approach.\n", "date": "2023-08-10", "categories": []}
{"id": "2308.11358", "title": "How Much Temporal Long-Term Context is Needed for Action Segmentation?", "abstract": "  Modeling long-term context in videos is crucial for many fine-grained tasks\nincluding temporal action segmentation. An interesting question that is still\nopen is how much long-term temporal context is needed for optimal performance.\nWhile transformers can model the long-term context of a video, this becomes\ncomputationally prohibitive for long videos. Recent works on temporal action\nsegmentation thus combine temporal convolutional networks with self-attentions\nthat are computed only for a local temporal window. While these approaches show\ngood results, their performance is limited by their inability to capture the\nfull context of a video. In this work, we try to answer how much long-term\ntemporal context is required for temporal action segmentation by introducing a\ntransformer-based model that leverages sparse attention to capture the full\ncontext of a video. We compare our model with the current state of the art on\nthree datasets for temporal action segmentation, namely 50Salads, Breakfast,\nand Assembly101. Our experiments show that modeling the full context of a video\nis necessary to obtain the best performance for temporal action segmentation.\n", "date": "2023-08-22", "categories": []}
{"id": "2309.13962", "title": "Egocentric RGB+Depth Action Recognition in Industry-Like Settings", "abstract": "  Action recognition from an egocentric viewpoint is a crucial perception task\nin robotics and enables a wide range of human-robot interactions. While most\ncomputer vision approaches prioritize the RGB camera, the Depth modality -\nwhich can further amplify the subtleties of actions from an egocentric\nperspective - remains underexplored. Our work focuses on recognizing actions\nfrom egocentric RGB and Depth modalities in an industry-like environment. To\nstudy this problem, we consider the recent MECCANO dataset, which provides a\nwide range of assembling actions. Our framework is based on the 3D Video SWIN\nTransformer to encode both RGB and Depth modalities effectively. To address the\ninherent skewness in real-world multimodal action occurrences, we propose a\ntraining strategy using an exponentially decaying variant of the focal loss\nmodulating factor. Additionally, to leverage the information in both RGB and\nDepth modalities, we opt for late fusion to combine the predictions from each\nmodality. We thoroughly evaluate our method on the action recognition task of\nthe MECCANO dataset, and it significantly outperforms the prior work. Notably,\nour method also secured first place at the multimodal action recognition\nchallenge at ICIAP 2023.\n", "date": "2023-09-25", "categories": []}
{"id": "2302.09187", "title": "Video Action Recognition Collaborative Learning with Dynamics via\n  PSO-ConvNet Transformer", "abstract": "  Recognizing human actions in video sequences, known as Human Action\nRecognition (HAR), is a challenging task in pattern recognition. While\nConvolutional Neural Networks (ConvNets) have shown remarkable success in image\nrecognition, they are not always directly applicable to HAR, as temporal\nfeatures are critical for accurate classification. In this paper, we propose a\nnovel dynamic PSO-ConvNet model for learning actions in videos, building on our\nrecent work in image recognition. Our approach leverages a framework where the\nweight vector of each neural network represents the position of a particle in\nphase space, and particles share their current weight vectors and gradient\nestimates of the Loss function. To extend our approach to video, we integrate\nConvNets with state-of-the-art temporal methods such as Transformer and\nRecurrent Neural Networks. Our experimental results on the UCF-101 dataset\ndemonstrate substantial improvements of up to 9% in accuracy, which confirms\nthe effectiveness of our proposed method. In addition, we conducted experiments\non larger and more variety of datasets including Kinetics-400 and HMDB-51 and\nobtained preference for Collaborative Learning in comparison with\nNon-Collaborative Learning (Individual Learning). Overall, our dynamic\nPSO-ConvNet model provides a promising direction for improving HAR by better\ncapturing the spatio-temporal dynamics of human actions in videos. The code is\navailable at\nhttps://github.com/leonlha/Video-Action-Recognition-Collaborative-Learning-with-Dynamics-via-PSO-ConvNet-Transformer.\n", "date": "2023-02-17", "categories": []}
{"id": "2309.11718", "title": "CPR-Coach: Recognizing Composite Error Actions based on Single-class\n  Training", "abstract": "  The fine-grained medical action analysis task has received considerable\nattention from pattern recognition communities recently, but it faces the\nproblems of data and algorithm shortage. Cardiopulmonary Resuscitation (CPR) is\nan essential skill in emergency treatment. Currently, the assessment of CPR\nskills mainly depends on dummies and trainers, leading to high training costs\nand low efficiency. For the first time, this paper constructs a vision-based\nsystem to complete error action recognition and skill assessment in CPR.\nSpecifically, we define 13 types of single-error actions and 74 types of\ncomposite error actions during external cardiac compression and then develop a\nvideo dataset named CPR-Coach. By taking the CPR-Coach as a benchmark, this\npaper thoroughly investigates and compares the performance of existing action\nrecognition models based on different data modalities. To solve the unavoidable\nSingle-class Training & Multi-class Testing problem, we propose a\nhumancognition-inspired framework named ImagineNet to improve the model's\nmultierror recognition performance under restricted supervision. Extensive\nexperiments verify the effectiveness of the framework. We hope this work could\nadvance research toward fine-grained medical action analysis and skill\nassessment. The CPR-Coach dataset and the code of ImagineNet are publicly\navailable on Github.\n", "date": "2023-09-21", "categories": []}
{"id": "2304.04688", "title": "Interaction-Aware Prompting for Zero-Shot Spatio-Temporal Action\n  Detection", "abstract": "  The goal of spatial-temporal action detection is to determine the time and\nplace where each person's action occurs in a video and classify the\ncorresponding action category. Most of the existing methods adopt\nfully-supervised learning, which requires a large amount of training data,\nmaking it very difficult to achieve zero-shot learning. In this paper, we\npropose to utilize a pre-trained visual-language model to extract the\nrepresentative image and text features, and model the relationship between\nthese features through different interaction modules to obtain the interaction\nfeature. In addition, we use this feature to prompt each label to obtain more\nappropriate text features. Finally, we calculate the similarity between the\ninteraction feature and the text feature for each label to determine the action\ncategory. Our experiments on J-HMDB and UCF101-24 datasets demonstrate that the\nproposed interaction module and prompting make the visual-language features\nbetter aligned, thus achieving excellent accuracy for zero-shot spatio-temporal\naction detection. The code will be available at\nhttps://github.com/webber2933/iCLIP.\n", "date": "2023-04-10", "categories": []}
{"id": "2309.11445", "title": "SkeleTR: Towrads Skeleton-based Action Recognition in the Wild", "abstract": "  We present SkeleTR, a new framework for skeleton-based action recognition. In\ncontrast to prior work, which focuses mainly on controlled environments, we\ntarget more general scenarios that typically involve a variable number of\npeople and various forms of interaction between people. SkeleTR works with a\ntwo-stage paradigm. It first models the intra-person skeleton dynamics for each\nskeleton sequence with graph convolutions, and then uses stacked Transformer\nencoders to capture person interactions that are important for action\nrecognition in general scenarios. To mitigate the negative impact of inaccurate\nskeleton associations, SkeleTR takes relative short skeleton sequences as input\nand increases the number of sequences. As a unified solution, SkeleTR can be\ndirectly applied to multiple skeleton-based action tasks, including video-level\naction classification, instance-level action detection, and group-level\nactivity recognition. It also enables transfer learning and joint training\nacross different action tasks and datasets, which result in performance\nimprovement. When evaluated on various skeleton-based action recognition\nbenchmarks, SkeleTR achieves the state-of-the-art performance.\n", "date": "2023-09-20", "categories": []}
{"id": "2309.10001", "title": "CaSAR: Contact-aware Skeletal Action Recognition", "abstract": "  Skeletal Action recognition from an egocentric view is important for\napplications such as interfaces in AR/VR glasses and human-robot interaction,\nwhere the device has limited resources. Most of the existing skeletal action\nrecognition approaches use 3D coordinates of hand joints and 8-corner\nrectangular bounding boxes of objects as inputs, but they do not capture how\nthe hands and objects interact with each other within the spatial context. In\nthis paper, we present a new framework called Contact-aware Skeletal Action\nRecognition (CaSAR). It uses novel representations of hand-object interaction\nthat encompass spatial information: 1) contact points where the hand joints\nmeet the objects, 2) distant points where the hand joints are far away from the\nobject and nearly not involved in the current action. Our framework is able to\nlearn how the hands touch or stay away from the objects for each frame of the\naction sequence, and use this information to predict the action class. We\ndemonstrate that our approach achieves the state-of-the-art accuracy of 91.3%\nand 98.4% on two public datasets, H2O and FPHA, respectively.\n", "date": "2023-09-17", "categories": []}
{"id": "2309.10765", "title": "MAGIC-TBR: Multiview Attention Fusion for Transformer-based Bodily\n  Behavior Recognition in Group Settings", "abstract": "  Bodily behavioral language is an important social cue, and its automated\nanalysis helps in enhancing the understanding of artificial intelligence\nsystems. Furthermore, behavioral language cues are essential for active\nengagement in social agent-based user interactions. Despite the progress made\nin computer vision for tasks like head and body pose estimation, there is still\na need to explore the detection of finer behaviors such as gesturing, grooming,\nor fumbling. This paper proposes a multiview attention fusion method named\nMAGIC-TBR that combines features extracted from videos and their corresponding\nDiscrete Cosine Transform coefficients via a transformer-based approach. The\nexperiments are conducted on the BBSI dataset and the results demonstrate the\neffectiveness of the proposed feature fusion with multiview attention. The code\nis available at: https://github.com/surbhimadan92/MAGIC-TBR\n", "date": "2023-09-19", "categories": []}
{"id": "2309.09060", "title": "Sub-action Prototype Learning for Point-level Weakly-supervised Temporal\n  Action Localization", "abstract": "  Point-level weakly-supervised temporal action localization (PWTAL) aims to\nlocalize actions with only a single timestamp annotation for each action\ninstance. Existing methods tend to mine dense pseudo labels to alleviate the\nlabel sparsity, but overlook the potential sub-action temporal structures,\nresulting in inferior performance. To tackle this problem, we propose a novel\nsub-action prototype learning framework (SPL-Loc) which comprises Sub-action\nPrototype Clustering (SPC) and Ordered Prototype Alignment (OPA). SPC\nadaptively extracts representative sub-action prototypes which are capable to\nperceive the temporal scale and spatial content variation of action instances.\nOPA selects relevant prototypes to provide completeness clue for pseudo label\ngeneration by applying a temporal alignment loss. As a result, pseudo labels\nare derived from alignment results to improve action boundary prediction.\nExtensive experiments on three popular benchmarks demonstrate that the proposed\nSPL-Loc significantly outperforms existing SOTA PWTAL methods.\n", "date": "2023-09-16", "categories": []}
{"id": "2309.09592", "title": "Multi-Semantic Fusion Model for Generalized Zero-Shot Skeleton-Based\n  Action Recognition", "abstract": "  Generalized zero-shot skeleton-based action recognition (GZSSAR) is a new\nchallenging problem in computer vision community, which requires models to\nrecognize actions without any training samples. Previous studies only utilize\nthe action labels of verb phrases as the semantic prototypes for learning the\nmapping from skeleton-based actions to a shared semantic space. However, the\nlimited semantic information of action labels restricts the generalization\nability of skeleton features for recognizing unseen actions. In order to solve\nthis dilemma, we propose a multi-semantic fusion (MSF) model for improving the\nperformance of GZSSAR, where two kinds of class-level textual descriptions\n(i.e., action descriptions and motion descriptions), are collected as auxiliary\nsemantic information to enhance the learning efficacy of generalizable skeleton\nfeatures. Specially, a pre-trained language encoder takes the action\ndescriptions, motion descriptions and original class labels as inputs to obtain\nrich semantic features for each action class, while a skeleton encoder is\nimplemented to extract skeleton features. Then, a variational autoencoder (VAE)\nbased generative module is performed to learn a cross-modal alignment between\nskeleton and semantic features. Finally, a classification module is built to\nrecognize the action categories of input samples, where a seen-unseen\nclassification gate is adopted to predict whether the sample comes from seen\naction classes or not in GZSSAR. The superior performance in comparisons with\nprevious models validates the effectiveness of the proposed MSF model on\nGZSSAR.\n", "date": "2023-09-18", "categories": []}
{"id": "2309.03989", "title": "CDFSL-V: Cross-Domain Few-Shot Learning for Videos", "abstract": "  Few-shot video action recognition is an effective approach to recognizing new\ncategories with only a few labeled examples, thereby reducing the challenges\nassociated with collecting and annotating large-scale video datasets. Existing\nmethods in video action recognition rely on large labeled datasets from the\nsame domain. However, this setup is not realistic as novel categories may come\nfrom different data domains that may have different spatial and temporal\ncharacteristics. This dissimilarity between the source and target domains can\npose a significant challenge, rendering traditional few-shot action recognition\ntechniques ineffective. To address this issue, in this work, we propose a novel\ncross-domain few-shot video action recognition method that leverages\nself-supervised learning and curriculum learning to balance the information\nfrom the source and target domains. To be particular, our method employs a\nmasked autoencoder-based self-supervised training objective to learn from both\nsource and target data in a self-supervised manner. Then a progressive\ncurriculum balances learning the discriminative information from the source\ndataset with the generic information learned from the target domain. Initially,\nour curriculum utilizes supervised learning to learn class discriminative\nfeatures from the source data. As the training progresses, we transition to\nlearning target-domain-specific features. We propose a progressive curriculum\nto encourage the emergence of rich features in the target domain based on class\ndiscriminative supervised features in the source domain. We evaluate our method\non several challenging benchmark datasets and demonstrate that our approach\noutperforms existing cross-domain few-shot learning techniques. Our code is\navailable at https://github.com/Sarinda251/CDFSL-V\n", "date": "2023-09-07", "categories": []}
{"id": "2309.08087", "title": "hear-your-action: human action recognition by ultrasound active sensing", "abstract": "  Action recognition is a key technology for many industrial applications.\nMethods using visual information such as images are very popular. However,\nprivacy issues prevent widespread usage due to the inclusion of private\ninformation, such as visible faces and scene backgrounds, which are not\nnecessary for recognizing user action. In this paper, we propose a\nprivacy-preserving action recognition by ultrasound active sensing. As action\nrecognition from ultrasound active sensing in a non-invasive manner is not well\ninvestigated, we create a new dataset for action recognition and conduct a\ncomparison of features for classification. We calculated feature values by\nfocusing on the temporal variation of the amplitude of ultrasound reflected\nwaves and performed classification using a support vector machine and VGG for\neight fundamental action classes. We confirmed that our method achieved an\naccuracy of 97.9% when trained and evaluated on the same person and in the same\nenvironment. Additionally, our method achieved an accuracy of 89.5% even when\ntrained and evaluated on different people. We also report the analyses of\naccuracies in various conditions and limitations.\n", "date": "2023-09-15", "categories": []}
{"id": "2309.07911", "title": "Disentangling Spatial and Temporal Learning for Efficient Image-to-Video\n  Transfer Learning", "abstract": "  Recently, large-scale pre-trained language-image models like CLIP have shown\nextraordinary capabilities for understanding spatial contents, but naively\ntransferring such models to video recognition still suffers from unsatisfactory\ntemporal modeling capabilities. Existing methods insert tunable structures into\nor in parallel with the pre-trained model, which either requires\nback-propagation through the whole pre-trained model and is thus\nresource-demanding, or is limited by the temporal reasoning capability of the\npre-trained structure. In this work, we present DiST, which disentangles the\nlearning of spatial and temporal aspects of videos. Specifically, DiST uses a\ndual-encoder structure, where a pre-trained foundation model acts as the\nspatial encoder, and a lightweight network is introduced as the temporal\nencoder. An integration branch is inserted between the encoders to fuse\nspatio-temporal information. The disentangled spatial and temporal learning in\nDiST is highly efficient because it avoids the back-propagation of massive\npre-trained parameters. Meanwhile, we empirically show that disentangled\nlearning with an extra network for integration benefits both spatial and\ntemporal understanding. Extensive experiments on five benchmarks show that DiST\ndelivers better performance than existing state-of-the-art methods by\nconvincing gaps. When pre-training on the large-scale Kinetics-710, we achieve\n89.7% on Kinetics-400 with a frozen ViT-L model, which verifies the scalability\nof DiST. Codes and models can be found in\nhttps://github.com/alibaba-mmai-research/DiST.\n", "date": "2023-09-14", "categories": []}
{"id": "2305.15701", "title": "Action Sensitivity Learning for Temporal Action Localization", "abstract": "  Temporal action localization (TAL), which involves recognizing and locating\naction instances, is a challenging task in video understanding. Most existing\napproaches directly predict action classes and regress offsets to boundaries,\nwhile overlooking the discrepant importance of each frame. In this paper, we\npropose an Action Sensitivity Learning framework (ASL) to tackle this task,\nwhich aims to assess the value of each frame and then leverage the generated\naction sensitivity to recalibrate the training procedure. We first introduce a\nlightweight Action Sensitivity Evaluator to learn the action sensitivity at the\nclass level and instance level, respectively. The outputs of the two branches\nare combined to reweight the gradient of the two sub-tasks. Moreover, based on\nthe action sensitivity of each frame, we design an Action Sensitive Contrastive\nLoss to enhance features, where the action-aware frames are sampled as positive\npairs to push away the action-irrelevant frames. The extensive studies on\nvarious action localization benchmarks (i.e., MultiThumos, Charades,\nEgo4D-Moment Queries v1.0, Epic-Kitchens 100, Thumos14 and ActivityNet1.3) show\nthat ASL surpasses the state-of-the-art in terms of average-mAP under multiple\ntypes of scenarios, e.g., single-labeled, densely-labeled and egocentric.\n", "date": "2023-05-25", "categories": []}
{"id": "2309.06951", "title": "TransNet: A Transfer Learning-Based Network for Human Action Recognition", "abstract": "  Human action recognition (HAR) is a high-level and significant research area\nin computer vision due to its ubiquitous applications. The main limitations of\nthe current HAR models are their complex structures and lengthy training time.\nIn this paper, we propose a simple yet versatile and effective end-to-end deep\nlearning architecture, coined as TransNet, for HAR. TransNet decomposes the\ncomplex 3D-CNNs into 2D- and 1D-CNNs, where the 2D- and 1D-CNN components\nextract spatial features and temporal patterns in videos, respectively.\nBenefiting from its concise architecture, TransNet is ideally compatible with\nany pretrained state-of-the-art 2D-CNN models in other fields, being\ntransferred to serve the HAR task. In other words, it naturally leverages the\npower and success of transfer learning for HAR, bringing huge advantages in\nterms of efficiency and effectiveness. Extensive experimental results and the\ncomparison with the state-of-the-art models demonstrate the superior\nperformance of the proposed TransNet in HAR in terms of flexibility, model\ncomplexity, training speed and classification accuracy.\n", "date": "2023-09-13", "categories": []}
{"id": "2309.05834", "title": "SCD-Net: Spatiotemporal Clues Disentanglement Network for\n  Self-supervised Skeleton-based Action Recognition", "abstract": "  Contrastive learning has achieved great success in skeleton-based action\nrecognition. However, most existing approaches encode the skeleton sequences as\nentangled spatiotemporal representations and confine the contrasts to the same\nlevel of representation. Instead, this paper introduces a novel contrastive\nlearning framework, namely Spatiotemporal Clues Disentanglement Network\n(SCD-Net). Specifically, we integrate the decoupling module with a feature\nextractor to derive explicit clues from spatial and temporal domains\nrespectively. As for the training of SCD-Net, with a constructed global anchor,\nwe encourage the interaction between the anchor and extracted clues. Further,\nwe propose a new masking strategy with structural constraints to strengthen the\ncontextual associations, leveraging the latest development from masked image\nmodelling into the proposed SCD-Net. We conduct extensive evaluations on the\nNTU-RGB+D (60&120) and PKU-MMD (I&II) datasets, covering various downstream\ntasks such as action recognition, action retrieval, transfer learning, and\nsemi-supervised learning. The experimental results demonstrate the\neffectiveness of our method, which outperforms the existing state-of-the-art\n(SOTA) approaches significantly.\n", "date": "2023-09-11", "categories": []}
{"id": "2309.06130", "title": "JOADAA: joint online action detection and action anticipation", "abstract": "  Action anticipation involves forecasting future actions by connecting past\nevents to future ones. However, this reasoning ignores the real-life hierarchy\nof events which is considered to be composed of three main parts: past,\npresent, and future. We argue that considering these three main parts and their\ndependencies could improve performance. On the other hand, online action\ndetection is the task of predicting actions in a streaming manner. In this\ncase, one has access only to the past and present information. Therefore, in\nonline action detection (OAD) the existing approaches miss semantics or future\ninformation which limits their performance. To sum up, for both of these tasks,\nthe complete set of knowledge (past-present-future) is missing, which makes it\nchallenging to infer action dependencies, therefore having low performances. To\naddress this limitation, we propose to fuse both tasks into a single uniform\narchitecture. By combining action anticipation and online action detection, our\napproach can cover the missing dependencies of future information in online\naction detection. This method referred to as JOADAA, presents a uniform model\nthat jointly performs action anticipation and online action detection. We\nvalidate our proposed model on three challenging datasets: THUMOS'14, which is\na sparsely annotated dataset with one action per time step, CHARADES, and\nMulti-THUMOS, two densely annotated datasets with more complex scenarios.\nJOADAA achieves SOTA results on these benchmarks for both tasks.\n", "date": "2023-09-12", "categories": []}
{"id": "2309.06276", "title": "OTAS: Unsupervised Boundary Detection for Object-Centric Temporal Action\n  Segmentation", "abstract": "  Temporal action segmentation is typically achieved by discovering the\ndramatic variances in global visual descriptors. In this paper, we explore the\nmerits of local features by proposing the unsupervised framework of\nObject-centric Temporal Action Segmentation (OTAS). Broadly speaking, OTAS\nconsists of self-supervised global and local feature extraction modules as well\nas a boundary selection module that fuses the features and detects salient\nboundaries for action segmentation. As a second contribution, we discuss the\npros and cons of existing frame-level and boundary-level evaluation metrics.\nThrough extensive experiments, we find OTAS is superior to the previous\nstate-of-the-art method by $41\\%$ on average in terms of our recommended F1\nscore. Surprisingly, OTAS even outperforms the ground-truth human annotations\nin the user study. Moreover, OTAS is efficient enough to allow real-time\ninference.\n", "date": "2023-09-12", "categories": []}
{"id": "2308.12006", "title": "Multi-stage Factorized Spatio-Temporal Representation for RGB-D Action\n  and Gesture Recognition", "abstract": "  RGB-D action and gesture recognition remain an interesting topic in\nhuman-centered scene understanding, primarily due to the multiple granularities\nand large variation in human motion. Although many RGB-D based action and\ngesture recognition approaches have demonstrated remarkable results by\nutilizing highly integrated spatio-temporal representations across multiple\nmodalities (i.e., RGB and depth data), they still encounter several challenges.\nFirstly, vanilla 3D convolution makes it hard to capture fine-grained motion\ndifferences between local clips under different modalities. Secondly, the\nintricate nature of highly integrated spatio-temporal modeling can lead to\noptimization difficulties. Thirdly, duplicate and unnecessary information can\nadd complexity and complicate entangled spatio-temporal modeling. To address\nthe above issues, we propose an innovative heuristic architecture called\nMulti-stage Factorized Spatio-Temporal (MFST) for RGB-D action and gesture\nrecognition. The proposed MFST model comprises a 3D Central Difference\nConvolution Stem (CDC-Stem) module and multiple factorized spatio-temporal\nstages. The CDC-Stem enriches fine-grained temporal perception, and the\nmultiple hierarchical spatio-temporal stages construct dimension-independent\nhigher-order semantic primitives. Specifically, the CDC-Stem module captures\nbottom-level spatio-temporal features and passes them successively to the\nfollowing spatio-temporal factored stages to capture the hierarchical spatial\nand temporal features through the Multi- Scale Convolution and Transformer\n(MSC-Trans) hybrid block and Weight-shared Multi-Scale Transformer (WMS-Trans)\nblock. The seamless integration of these innovative designs results in a robust\nspatio-temporal representation that outperforms state-of-the-art approaches on\nRGB-D action and gesture recognition datasets.\n", "date": "2023-08-23", "categories": []}
{"id": "2309.05032", "title": "Unified Contrastive Fusion Transformer for Multimodal Human Action\n  Recognition", "abstract": "  Various types of sensors have been considered to develop human action\nrecognition (HAR) models. Robust HAR performance can be achieved by fusing\nmultimodal data acquired by different sensors. In this paper, we introduce a\nnew multimodal fusion architecture, referred to as Unified Contrastive Fusion\nTransformer (UCFFormer) designed to integrate data with diverse distributions\nto enhance HAR performance. Based on the embedding features extracted from each\nmodality, UCFFormer employs the Unified Transformer to capture the\ninter-dependency among embeddings in both time and modality domains. We present\nthe Factorized Time-Modality Attention to perform self-attention efficiently\nfor the Unified Transformer. UCFFormer also incorporates contrastive learning\nto reduce the discrepancy in feature distributions across various modalities,\nthus generating semantically aligned features for information fusion.\nPerformance evaluation conducted on two popular datasets, UTD-MHAD and NTU\nRGB+D, demonstrates that UCFFormer achieves state-of-the-art performance,\noutperforming competing methods by considerable margins.\n", "date": "2023-09-10", "categories": []}
{"id": "2309.05590", "title": "Temporal Action Localization with Enhanced Instant Discriminability", "abstract": "  Temporal action detection (TAD) aims to detect all action boundaries and\ntheir corresponding categories in an untrimmed video. The unclear boundaries of\nactions in videos often result in imprecise predictions of action boundaries by\nexisting methods. To resolve this issue, we propose a one-stage framework named\nTriDet. First, we propose a Trident-head to model the action boundary via an\nestimated relative probability distribution around the boundary. Then, we\nanalyze the rank-loss problem (i.e. instant discriminability deterioration) in\ntransformer-based methods and propose an efficient scalable-granularity\nperception (SGP) layer to mitigate this issue. To further push the limit of\ninstant discriminability in the video backbone, we leverage the strong\nrepresentation capability of pretrained large models and investigate their\nperformance on TAD. Last, considering the adequate spatial-temporal context for\nclassification, we design a decoupled feature pyramid network with separate\nfeature pyramids to incorporate rich spatial context from the large model for\nlocalization. Experimental results demonstrate the robustness of TriDet and its\nstate-of-the-art performance on multiple TAD datasets, including hierarchical\n(multilabel) TAD datasets.\n", "date": "2023-09-11", "categories": []}
{"id": "2208.05318", "title": "Generative Action Description Prompts for Skeleton-based Action\n  Recognition", "abstract": "  Skeleton-based action recognition has recently received considerable\nattention. Current approaches to skeleton-based action recognition are\ntypically formulated as one-hot classification tasks and do not fully exploit\nthe semantic relations between actions. For example, \"make victory sign\" and\n\"thumb up\" are two actions of hand gestures, whose major difference lies in the\nmovement of hands. This information is agnostic from the categorical one-hot\nencoding of action classes but could be unveiled from the action description.\nTherefore, utilizing action description in training could potentially benefit\nrepresentation learning. In this work, we propose a Generative\nAction-description Prompts (GAP) approach for skeleton-based action\nrecognition. More specifically, we employ a pre-trained large-scale language\nmodel as the knowledge engine to automatically generate text descriptions for\nbody parts movements of actions, and propose a multi-modal training scheme by\nutilizing the text encoder to generate feature vectors for different body parts\nand supervise the skeleton encoder for action representation learning.\nExperiments show that our proposed GAP method achieves noticeable improvements\nover various baseline models without extra computation cost at inference. GAP\nachieves new state-of-the-arts on popular skeleton-based action recognition\nbenchmarks, including NTU RGB+D, NTU RGB+D 120 and NW-UCLA. The source code is\navailable at https://github.com/MartinXM/GAP.\n", "date": "2022-08-10", "categories": []}
{"id": "2209.02986", "title": "Shifting Perspective to See Difference: A Novel Multi-View Method for\n  Skeleton based Action Recognition", "abstract": "  Skeleton-based human action recognition is a longstanding challenge due to\nits complex dynamics. Some fine-grain details of the dynamics play a vital role\nin classification. The existing work largely focuses on designing incremental\nneural networks with more complicated adjacent matrices to capture the details\nof joints relationships. However, they still have difficulties distinguishing\nactions that have broadly similar motion patterns but belong to different\ncategories. Interestingly, we found that the subtle differences in motion\npatterns can be significantly amplified and become easy for audience to\ndistinct through specified view directions, where this property haven't been\nfully explored before. Drastically different from previous work, we boost the\nperformance by proposing a conceptually simple yet effective Multi-view\nstrategy that recognizes actions from a collection of dynamic view features.\nSpecifically, we design a novel Skeleton-Anchor Proposal (SAP) module which\ncontains a Multi-head structure to learn a set of views. For feature learning\nof different views, we introduce a novel Angle Representation to transform the\nactions under different views and feed the transformations into the baseline\nmodel. Our module can work seamlessly with the existing action classification\nmodel. Incorporated with baseline models, our SAP module exhibits clear\nperformance gains on many challenging benchmarks. Moreover, comprehensive\nexperiments show that our model consistently beats down the state-of-the-art\nand remains effective and robust especially when dealing with corrupted data.\nRelated code will be available on https://github.com/ideal-idea/SAP .\n", "date": "2022-09-07", "categories": []}
{"id": "2309.00696", "title": "AAN: Attributes-Aware Network for Temporal Action Detection", "abstract": "  The challenge of long-term video understanding remains constrained by the\nefficient extraction of object semantics and the modelling of their\nrelationships for downstream tasks. Although the CLIP visual features exhibit\ndiscriminative properties for various vision tasks, particularly in object\nencoding, they are suboptimal for long-term video understanding. To address\nthis issue, we present the Attributes-Aware Network (AAN), which consists of\ntwo key components: the Attributes Extractor and a Graph Reasoning block. These\ncomponents facilitate the extraction of object-centric attributes and the\nmodelling of their relationships within the video. By leveraging CLIP features,\nAAN outperforms state-of-the-art approaches on two popular action detection\ndatasets: Charades and Toyota Smarthome Untrimmed datasets.\n", "date": "2023-09-01", "categories": []}
{"id": "2309.01262", "title": "Multimodal Contrastive Learning with Hard Negative Sampling for Human\n  Activity Recognition", "abstract": "  Human Activity Recognition (HAR) systems have been extensively studied by the\nvision and ubiquitous computing communities due to their practical applications\nin daily life, such as smart homes, surveillance, and health monitoring.\n  Typically, this process is supervised in nature and the development of such\nsystems requires access to large quantities of annotated data.\n  However, the higher costs and challenges associated with obtaining good\nquality annotations have rendered the application of self-supervised methods an\nattractive option and contrastive learning comprises one such method.\n  However, a major component of successful contrastive learning is the\nselection of good positive and negative samples.\n  Although positive samples are directly obtainable, sampling good negative\nsamples remain a challenge.\n  As human activities can be recorded by several modalities like camera and IMU\nsensors, we propose a hard negative sampling method for multimodal HAR with a\nhard negative sampling loss for skeleton and IMU data pairs.\n  We exploit hard negatives that have different labels from the anchor but are\nprojected nearby in the latent space using an adjustable concentration\nparameter.\n  Through extensive experiments on two benchmark datasets: UTD-MHAD and MMAct,\nwe demonstrate the robustness of our approach forlearning strong feature\nrepresentation for HAR tasks, and on the limited data setting.\n  We further show that our model outperforms all other state-of-the-art methods\nfor UTD-MHAD dataset, and self-supervised methods for MMAct: Cross session,\neven when uni-modal data are used during downstream activity recognition.\n", "date": "2023-09-03", "categories": []}
{"id": "2309.01265", "title": "SOAR: Scene-debiasing Open-set Action Recognition", "abstract": "  Deep learning models have a risk of utilizing spurious clues to make\npredictions, such as recognizing actions based on the background scene. This\nissue can severely degrade the open-set action recognition performance when the\ntesting samples have different scene distributions from the training samples.\nTo mitigate this problem, we propose a novel method, called Scene-debiasing\nOpen-set Action Recognition (SOAR), which features an adversarial scene\nreconstruction module and an adaptive adversarial scene classification module.\nThe former prevents the decoder from reconstructing the video background given\nvideo features, and thus helps reduce the background information in feature\nlearning. The latter aims to confuse scene type classification given video\nfeatures, with a specific emphasis on the action foreground, and helps to learn\nscene-invariant information. In addition, we design an experiment to quantify\nthe scene bias. The results indicate that the current open-set action\nrecognizers are biased toward the scene, and our proposed SOAR method better\nmitigates such bias. Furthermore, our extensive experiments demonstrate that\nour method outperforms state-of-the-art methods, and the ablation studies\nconfirm the effectiveness of our proposed modules.\n", "date": "2023-09-03", "categories": []}
{"id": "2309.02423", "title": "EgoPCA: A New Framework for Egocentric Hand-Object Interaction\n  Understanding", "abstract": "  With the surge in attention to Egocentric Hand-Object Interaction (Ego-HOI),\nlarge-scale datasets such as Ego4D and EPIC-KITCHENS have been proposed.\nHowever, most current research is built on resources derived from third-person\nvideo action recognition. This inherent domain gap between first- and\nthird-person action videos, which have not been adequately addressed before,\nmakes current Ego-HOI suboptimal. This paper rethinks and proposes a new\nframework as an infrastructure to advance Ego-HOI recognition by Probing,\nCuration and Adaption (EgoPCA). We contribute comprehensive pre-train sets,\nbalanced test sets and a new baseline, which are complete with a\ntraining-finetuning strategy. With our new framework, we not only achieve\nstate-of-the-art performance on Ego-HOI benchmarks but also build several new\nand effective mechanisms and settings to advance further research. We believe\nour data and the findings will pave a new way for Ego-HOI understanding. Code\nand data are available at https://mvig-rhos.com/ego_pca\n", "date": "2023-09-05", "categories": []}
{"id": "2308.16325", "title": "Two-Stage Violence Detection Using ViTPose and Classification Models at\n  Smart Airports", "abstract": "  This study introduces an innovative violence detection framework tailored to\nthe unique requirements of smart airports, where prompt responses to violent\nsituations are crucial. The proposed framework harnesses the power of ViTPose\nfor human pose estimation. It employs a CNN - BiLSTM network to analyse spatial\nand temporal information within keypoints sequences, enabling the accurate\nclassification of violent behaviour in real time. Seamlessly integrated within\nthe SAFE (Situational Awareness for Enhanced Security framework of SAAB, the\nsolution underwent integrated testing to ensure robust performance in real\nworld scenarios. The AIRTLab dataset, characterized by its high video quality\nand relevance to surveillance scenarios, is utilized in this study to enhance\nthe model's accuracy and mitigate false positives. As airports face increased\nfoot traffic in the post pandemic era, implementing AI driven violence\ndetection systems, such as the one proposed, is paramount for improving\nsecurity, expediting response times, and promoting data informed decision\nmaking. The implementation of this framework not only diminishes the\nprobability of violent events but also assists surveillance teams in\neffectively addressing potential threats, ultimately fostering a more secure\nand protected aviation sector. Codes are available at:\nhttps://github.com/Asami-1/GDP.\n", "date": "2023-08-30", "categories": []}
{"id": "2308.16552", "title": "Prompt-enhanced Hierarchical Transformer Elevating Cardiopulmonary\n  Resuscitation Instruction via Temporal Action Segmentation", "abstract": "  The vast majority of people who suffer unexpected cardiac arrest are\nperformed cardiopulmonary resuscitation (CPR) by passersby in a desperate\nattempt to restore life, but endeavors turn out to be fruitless on account of\ndisqualification. Fortunately, many pieces of research manifest that\ndisciplined training will help to elevate the success rate of resuscitation,\nwhich constantly desires a seamless combination of novel techniques to yield\nfurther advancement. To this end, we collect a custom CPR video dataset in\nwhich trainees make efforts to behave resuscitation on mannequins independently\nin adherence to approved guidelines, thereby devising an auxiliary toolbox to\nassist supervision and rectification of intermediate potential issues via\nmodern deep learning methodologies. Our research empirically views this problem\nas a temporal action segmentation (TAS) task in computer vision, which aims to\nsegment an untrimmed video at a frame-wise level. Here, we propose a\nPrompt-enhanced hierarchical Transformer (PhiTrans) that integrates three\nindispensable modules, including a textual prompt-based Video Features\nExtractor (VFE), a transformer-based Action Segmentation Executor (ASE), and a\nregression-based Prediction Refinement Calibrator (PRC). The backbone of the\nmodel preferentially derives from applications in three approved public\ndatasets (GTEA, 50Salads, and Breakfast) collected for TAS tasks, which\naccounts for the excavation of the segmentation pipeline on the CPR dataset. In\ngeneral, we unprecedentedly probe into a feasible pipeline that genuinely\nelevates the CPR instruction qualification via action segmentation in\nconjunction with cutting-edge deep learning techniques. Associated experiments\nadvocate our implementation with multiple metrics surpassing 91.0%.\n", "date": "2023-08-31", "categories": []}
{"id": "2308.14843", "title": "Robust Activity Recognition for Adaptive Worker-Robot Interaction using\n  Transfer Learning", "abstract": "  Human activity recognition (HAR) using machine learning has shown tremendous\npromise in detecting construction workers' activities. HAR has many\napplications in human-robot interaction research to enable robots'\nunderstanding of human counterparts' activities. However, many existing HAR\napproaches lack robustness, generalizability, and adaptability. This paper\nproposes a transfer learning methodology for activity recognition of\nconstruction workers that requires orders of magnitude less data and compute\ntime for comparable or better classification accuracy. The developed algorithm\ntransfers features from a model pre-trained by the original authors and\nfine-tunes them for the downstream task of activity recognition in\nconstruction. The model was pre-trained on Kinetics-400, a large-scale\nvideo-based human activity recognition dataset with 400 distinct classes. The\nmodel was fine-tuned and tested using videos captured from manual material\nhandling (MMH) activities found on YouTube. Results indicate that the\nfine-tuned model can recognize distinct MMH tasks in a robust and adaptive\nmanner which is crucial for the widespread deployment of collaborative robots\nin construction.\n", "date": "2023-08-28", "categories": []}
{"id": "2308.15345", "title": "IndGIC: Supervised Action Recognition under Low Illumination", "abstract": "  Technologies of human action recognition in the dark are gaining more and\nmore attention as huge demand in surveillance, motion control and\nhuman-computer interaction. However, because of limitation in image enhancement\nmethod and low-lighting video datasets, e.g. labeling cost, existing methods\nmeet some problems. Some video-based approached are effect and efficient in\nspecific datasets but cannot generalize to most cases while others methods\nusing multiple sensors rely heavily to prior knowledge to deal with noisy\nnature from video stream. In this paper, we proposes action recognition method\nusing deep multi-input network. Furthermore, we proposed a Independent Gamma\nIntensity Corretion (Ind-GIC) to enhance poor-illumination video, generating\none gamma for one frame to increase enhancement performance. To prove our\nmethod is effective, there is some evaluation and comparison between our method\nand existing methods. Experimental results show that our model achieves high\naccuracy in on ARID dataset.\n", "date": "2023-08-29", "categories": []}
{"id": "2212.02053", "title": "Day2Dark: Pseudo-Supervised Activity Recognition beyond Silent Daylight", "abstract": "  This paper strives to recognize activities in the dark, as well as in the\nday. We first establish that state-of-the-art activity recognizers are\neffective during the day, but not trustworthy in the dark. The main causes are\nthe limited availability of labeled dark videos to learn from, as well as the\ndistribution shift towards the lower color contrast at test-time. To compensate\nfor the lack of labeled dark videos, we introduce a pseudo-supervised learning\nscheme, which utilizes easy to obtain unlabeled and task-irrelevant dark videos\nto improve an activity recognizer in low light. As the lower color contrast\nresults in visual information loss, we further propose to incorporate the\ncomplementary activity information within audio, which is invariant to\nillumination. Since the usefulness of audio and visual features differs\ndepending on the amount of illumination, we introduce our `darkness-adaptive'\naudio-visual recognizer. Experiments on EPIC-Kitchens, Kinetics-Sound, and\nCharades demonstrate our proposals are superior to image enhancement, domain\nadaptation and alternative audio-visual fusion methods, and can even improve\nrobustness to local darkness caused by occlusions. Project page:\nhttps://xiaobai1217.github.io/Day2Dark/\n", "date": "2022-12-05", "categories": []}
{"id": "2303.12149", "title": "SPARTAN: Self-supervised Spatiotemporal Transformers Approach to Group\n  Activity Recognition", "abstract": "  In this paper, we propose a new, simple, and effective Self-supervised\nSpatio-temporal Transformers (SPARTAN) approach to Group Activity Recognition\n(GAR) using unlabeled video data. Given a video, we create local and global\nSpatio-temporal views with varying spatial patch sizes and frame rates. The\nproposed self-supervised objective aims to match the features of these\ncontrasting views representing the same video to be consistent with the\nvariations in spatiotemporal domains. To the best of our knowledge, the\nproposed mechanism is one of the first works to alleviate the weakly supervised\nsetting of GAR using the encoders in video transformers. Furthermore, using the\nadvantage of transformer models, our proposed approach supports long-term\nrelationship modeling along spatio-temporal dimensions. The proposed SPARTAN\napproach performs well on two group activity recognition benchmarks, including\nNBA and Volleyball datasets, by surpassing the state-of-the-art results by a\nsignificant margin in terms of MCA and MPCA metrics.\n", "date": "2023-03-06", "categories": []}
{"id": "2304.08451", "title": "Efficient Video Action Detection with Token Dropout and Context\n  Refinement", "abstract": "  Streaming video clips with large-scale video tokens impede vision\ntransformers (ViTs) for efficient recognition, especially in video action\ndetection where sufficient spatiotemporal representations are required for\nprecise actor identification. In this work, we propose an end-to-end framework\nfor efficient video action detection (EVAD) based on vanilla ViTs. Our EVAD\nconsists of two specialized designs for video action detection. First, we\npropose a spatiotemporal token dropout from a keyframe-centric perspective. In\na video clip, we maintain all tokens from its keyframe, preserve tokens\nrelevant to actor motions from other frames, and drop out the remaining tokens\nin this clip. Second, we refine scene context by leveraging remaining tokens\nfor better recognizing actor identities. The region of interest (RoI) in our\naction detector is expanded into temporal domain. The captured spatiotemporal\nactor identity representations are refined via scene context in a decoder with\nthe attention mechanism. These two designs make our EVAD efficient while\nmaintaining accuracy, which is validated on three benchmark datasets (i.e.,\nAVA, UCF101-24, JHMDB). Compared to the vanilla ViT backbone, our EVAD reduces\nthe overall GFLOPs by 43% and improves real-time inference speed by 40% with no\nperformance degradation. Moreover, even at similar computational costs, our\nEVAD can improve the performance by 1.1 mAP with higher resolution inputs. Code\nis available at https://github.com/MCG-NJU/EVAD.\n", "date": "2023-04-17", "categories": []}
{"id": "2308.13711", "title": "EventTransAct: A video transformer-based framework for Event-camera\n  based action recognition", "abstract": "  Recognizing and comprehending human actions and gestures is a crucial\nperception requirement for robots to interact with humans and carry out tasks\nin diverse domains, including service robotics, healthcare, and manufacturing.\nEvent cameras, with their ability to capture fast-moving objects at a high\ntemporal resolution, offer new opportunities compared to standard action\nrecognition in RGB videos. However, previous research on event camera action\nrecognition has primarily focused on sensor-specific network architectures and\nimage encoding, which may not be suitable for new sensors and limit the use of\nrecent advancements in transformer-based architectures. In this study, we\nemploy a computationally efficient model, namely the video transformer network\n(VTN), which initially acquires spatial embeddings per event-frame and then\nutilizes a temporal self-attention mechanism. In order to better adopt the VTN\nfor the sparse and fine-grained nature of event data, we design\nEvent-Contrastive Loss ($\\mathcal{L}_{EC}$) and event-specific augmentations.\nProposed $\\mathcal{L}_{EC}$ promotes learning fine-grained spatial cues in the\nspatial backbone of VTN by contrasting temporally misaligned frames. We\nevaluate our method on real-world action recognition of N-EPIC Kitchens\ndataset, and achieve state-of-the-art results on both protocols - testing in\nseen kitchen (\\textbf{74.9\\%} accuracy) and testing in unseen kitchens\n(\\textbf{42.43\\% and 46.66\\% Accuracy}). Our approach also takes less\ncomputation time compared to competitive prior approaches, which demonstrates\nthe potential of our framework \\textit{EventTransAct} for real-world\napplications of event-camera based action recognition. Project Page:\n\\url{https://tristandb8.github.io/EventTransAct_webpage/}\n", "date": "2023-08-25", "categories": []}
{"id": "2308.13866", "title": "Improving Video Violence Recognition with Human Interaction Learning on\n  3D Skeleton Point Clouds", "abstract": "  Deep learning has proved to be very effective in video action recognition.\nVideo violence recognition attempts to learn the human multi-dynamic behaviours\nin more complex scenarios. In this work, we develop a method for video violence\nrecognition from a new perspective of skeleton points. Unlike the previous\nworks, we first formulate 3D skeleton point clouds from human skeleton\nsequences extracted from videos and then perform interaction learning on these\n3D skeleton point clouds. Specifically, we propose two types of Skeleton Points\nInteraction Learning (SPIL) strategies: (i) Local-SPIL: by constructing a\nspecific weight distribution strategy between local regional points, Local-SPIL\naims to selectively focus on the most relevant parts of them based on their\nfeatures and spatial-temporal position information. In order to capture diverse\ntypes of relation information, a multi-head mechanism is designed to aggregate\ndifferent features from independent heads to jointly handle different types of\nrelationships between points. (ii) Global-SPIL: to better learn and refine the\nfeatures of the unordered and unstructured skeleton points, Global-SPIL employs\nthe self-attention layer that operates directly on the sampled points, which\ncan help to make the output more permutation-invariant and well-suited for our\ntask. Extensive experimental results validate the effectiveness of our approach\nand show that our model outperforms the existing networks and achieves new\nstate-of-the-art performance on video violence datasets.\n", "date": "2023-08-26", "categories": []}
{"id": "2308.13082", "title": "Benchmarking Data Efficiency and Computational Efficiency of Temporal\n  Action Localization Models", "abstract": "  In temporal action localization, given an input video, the goal is to predict\nwhich actions it contains, where they begin, and where they end. Training and\ntesting current state-of-the-art deep learning models requires access to large\namounts of data and computational power. However, gathering such data is\nchallenging and computational resources might be limited. This work explores\nand measures how current deep temporal action localization models perform in\nsettings constrained by the amount of data or computational power. We measure\ndata efficiency by training each model on a subset of the training set. We find\nthat TemporalMaxer outperforms other models in data-limited settings.\nFurthermore, we recommend TriDet when training time is limited. To test the\nefficiency of the models during inference, we pass videos of different lengths\nthrough each model. We find that TemporalMaxer requires the least computational\nresources, likely due to its simple architecture.\n", "date": "2023-08-24", "categories": []}
{"id": "2306.07703", "title": "E2E-LOAD: End-to-End Long-form Online Action Detection", "abstract": "  Recently, there has been a growing trend toward feature-based approaches for\nOnline Action Detection (OAD). However, these approaches have limitations due\nto their fixed backbone design, which ignores the potential capability of a\ntrainable backbone. In this paper, we propose the first end-to-end OAD model,\ntermed E2E-LOAD, designed to address the major challenge of OAD, namely,\nlong-term understanding and efficient online reasoning. Specifically, our\nproposed approach adopts an initial spatial model that is shared by all frames\nand maintains a long sequence cache for inference at a low computational cost.\nWe also advocate an asymmetric spatial-temporal model for long-form and\nshort-form modeling effectively. Furthermore, we propose a novel and efficient\ninference mechanism that accelerates heavy spatial-temporal exploration.\nExtensive ablation studies and experiments demonstrate the effectiveness and\nefficiency of our proposed method. Notably, we achieve 17.3 (+12.6) FPS for\nend-to-end OAD with 72.4%~(+1.2%), 90.3%~(+0.7%), and 48.1%~(+26.0%) mAP on\nTHMOUS14, TVSeries, and HDD, respectively, which is 3x faster than previous\napproaches. The source code will be made publicly available.\n", "date": "2023-06-13", "categories": []}
{"id": "2306.08713", "title": "What can a cook in Italy teach a mechanic in India? Action Recognition\n  Generalisation Over Scenarios and Locations", "abstract": "  We propose and address a new generalisation problem: can a model trained for\naction recognition successfully classify actions when they are performed within\na previously unseen scenario and in a previously unseen location? To answer\nthis question, we introduce the Action Recognition Generalisation Over\nscenarios and locations dataset (ARGO1M), which contains 1.1M video clips from\nthe large-scale Ego4D dataset, across 10 scenarios and 13 locations. We\ndemonstrate recognition models struggle to generalise over 10 proposed test\nsplits, each of an unseen scenario in an unseen location. We thus propose CIR,\na method to represent each video as a Cross-Instance Reconstruction of videos\nfrom other domains. Reconstructions are paired with text narrations to guide\nthe learning of a domain generalisable representation. We provide extensive\nanalysis and ablations on ARGO1M that show CIR outperforms prior domain\ngeneralisation works on all test splits. Code and data:\nhttps://chiaraplizz.github.io/what-can-a-cook/.\n", "date": "2023-06-14", "categories": []}
{"id": "2308.12501", "title": "DD-GCN: Directed Diffusion Graph Convolutional Network for\n  Skeleton-based Human Action Recognition", "abstract": "  Graph Convolutional Networks (GCNs) have been widely used in skeleton-based\nhuman action recognition. In GCN-based methods, the spatio-temporal graph is\nfundamental for capturing motion patterns. However, existing approaches ignore\nthe physical dependency and synchronized spatio-temporal correlations between\njoints, which limits the representation capability of GCNs. To solve these\nproblems, we construct the directed diffusion graph for action modeling and\nintroduce the activity partition strategy to optimize the weight sharing\nmechanism of graph convolution kernels. In addition, we present the\nspatio-temporal synchronization encoder to embed synchronized spatio-temporal\nsemantics. Finally, we propose Directed Diffusion Graph Convolutional Network\n(DD-GCN) for action recognition, and the experiments on three public datasets:\nNTU-RGB+D, NTU-RGB+D 120, and NW-UCLA, demonstrate the state-of-the-art\nperformance of our method.\n", "date": "2023-08-24", "categories": []}
{"id": "2308.12840", "title": "FaceTouch: Detecting hand-to-face touch with supervised contrastive\n  learning to assist in tracing infectious disease", "abstract": "  Through our respiratory system, many viruses and diseases frequently spread\nand pass from one person to another. Covid-19 served as an example of how\ncrucial it is to track down and cut back on contacts to stop its spread. There\nis a clear gap in finding automatic methods that can detect hand-to-face\ncontact in complex urban scenes or indoors. In this paper, we introduce a\ncomputer vision framework, called FaceTouch, based on deep learning. It\ncomprises deep sub-models to detect humans and analyse their actions. FaceTouch\nseeks to detect hand-to-face touches in the wild, such as through video chats,\nbus footage, or CCTV feeds. Despite partial occlusion of faces, the introduced\nsystem learns to detect face touches from the RGB representation of a given\nscene by utilising the representation of the body gestures such as arm\nmovement. This has been demonstrated to be useful in complex urban scenarios\nbeyond simply identifying hand movement and its closeness to faces. Relying on\nSupervised Contrastive Learning, the introduced model is trained on our\ncollected dataset, given the absence of other benchmark datasets. The framework\nshows a strong validation in unseen datasets which opens the door for potential\ndeployment.\n", "date": "2023-08-24", "categories": []}
{"id": "2307.08908", "title": "What Can Simple Arithmetic Operations Do for Temporal Modeling?", "abstract": "  Temporal modeling plays a crucial role in understanding video content. To\ntackle this problem, previous studies built complicated temporal relations\nthrough time sequence thanks to the development of computationally powerful\ndevices. In this work, we explore the potential of four simple arithmetic\noperations for temporal modeling. Specifically, we first capture auxiliary\ntemporal cues by computing addition, subtraction, multiplication, and division\nbetween pairs of extracted frame features. Then, we extract corresponding\nfeatures from these cues to benefit the original temporal-irrespective domain.\nWe term such a simple pipeline as an Arithmetic Temporal Module (ATM), which\noperates on the stem of a visual backbone with a plug-and-play style. We\nconduct comprehensive ablation studies on the instantiation of ATMs and\ndemonstrate that this module provides powerful temporal modeling capability at\na low computational cost. Moreover, the ATM is compatible with both CNNs- and\nViTs-based architectures. Our results show that ATM achieves superior\nperformance over several popular video benchmarks. Specifically, on\nSomething-Something V1, V2 and Kinetics-400, we reach top-1 accuracy of 65.6%,\n74.6%, and 89.4% respectively. The code is available at\nhttps://github.com/whwu95/ATM.\n", "date": "2023-07-18", "categories": []}
{"id": "2308.09139", "title": "The Unreasonable Effectiveness of Large Language-Vision Models for\n  Source-free Video Domain Adaptation", "abstract": "  Source-Free Video Unsupervised Domain Adaptation (SFVUDA) task consists in\nadapting an action recognition model, trained on a labelled source dataset, to\nan unlabelled target dataset, without accessing the actual source data. The\nprevious approaches have attempted to address SFVUDA by leveraging\nself-supervision (e.g., enforcing temporal consistency) derived from the target\ndata itself. In this work, we take an orthogonal approach by exploiting\n\"web-supervision\" from Large Language-Vision Models (LLVMs), driven by the\nrationale that LLVMs contain a rich world prior surprisingly robust to\ndomain-shift. We showcase the unreasonable effectiveness of integrating LLVMs\nfor SFVUDA by devising an intuitive and parameter-efficient method, which we\nname Domain Adaptation with Large Language-Vision models (DALL-V), that\ndistills the world prior and complementary source model information into a\nstudent network tailored for the target. Despite the simplicity, DALL-V\nachieves significant improvement over state-of-the-art SFVUDA methods.\n", "date": "2023-08-17", "categories": []}
{"id": "2308.11244", "title": "Are current long-term video understanding datasets long-term?", "abstract": "  Many real-world applications, from sport analysis to surveillance, benefit\nfrom automatic long-term action recognition. In the current deep learning\nparadigm for automatic action recognition, it is imperative that models are\ntrained and tested on datasets and tasks that evaluate if such models actually\nlearn and reason over long-term information. In this work, we propose a method\nto evaluate how suitable a video dataset is to evaluate models for long-term\naction recognition. To this end, we define a long-term action as excluding all\nthe videos that can be correctly recognized using solely short-term\ninformation. We test this definition on existing long-term classification tasks\non three popular real-world datasets, namely Breakfast, CrossTask and LVU, to\ndetermine if these datasets are truly evaluating long-term recognition. Our\nstudy reveals that these datasets can be effectively solved using shortcuts\nbased on short-term information. Following this finding, we encourage long-term\naction recognition researchers to make use of datasets that need long-term\ninformation to be solved.\n", "date": "2023-08-22", "categories": []}
{"id": "2308.11249", "title": "Video BagNet: short temporal receptive fields increase robustness in\n  long-term action recognition", "abstract": "  Previous work on long-term video action recognition relies on deep\n3D-convolutional models that have a large temporal receptive field (RF). We\nargue that these models are not always the best choice for temporal modeling in\nvideos. A large temporal receptive field allows the model to encode the exact\nsub-action order of a video, which causes a performance decrease when testing\nvideos have a different sub-action order. In this work, we investigate whether\nwe can improve the model robustness to the sub-action order by shrinking the\ntemporal receptive field of action recognition models. For this, we design\nVideo BagNet, a variant of the 3D ResNet-50 model with the temporal receptive\nfield size limited to 1, 9, 17 or 33 frames. We analyze Video BagNet on\nsynthetic and real-world video datasets and experimentally compare models with\nvarying temporal receptive fields. We find that short receptive fields are\nrobust to sub-action order changes, while larger temporal receptive fields are\nsensitive to the sub-action order.\n", "date": "2023-08-22", "categories": []}
{"id": "2303.13505", "title": "A Large-scale Study of Spatiotemporal Representation Learning with a New\n  Benchmark on Action Recognition", "abstract": "  The goal of building a benchmark (suite of datasets) is to provide a unified\nprotocol for fair evaluation and thus facilitate the evolution of a specific\narea. Nonetheless, we point out that existing protocols of action recognition\ncould yield partial evaluations due to several limitations. To comprehensively\nprobe the effectiveness of spatiotemporal representation learning, we introduce\nBEAR, a new BEnchmark on video Action Recognition. BEAR is a collection of 18\nvideo datasets grouped into 5 categories (anomaly, gesture, daily, sports, and\ninstructional), which covers a diverse set of real-world applications. With\nBEAR, we thoroughly evaluate 6 common spatiotemporal models pre-trained by both\nsupervised and self-supervised learning. We also report transfer performance\nvia standard finetuning, few-shot finetuning, and unsupervised domain\nadaptation. Our observation suggests that current state-of-the-art cannot\nsolidly guarantee high performance on datasets close to real-world\napplications, and we hope BEAR can serve as a fair and challenging evaluation\nbenchmark to gain insights on building next-generation spatiotemporal learners.\nOur dataset, code, and models are released at:\nhttps://github.com/AndongDeng/BEAR\n", "date": "2023-03-23", "categories": []}
{"id": "2308.04589", "title": "Temporal DINO: A Self-supervised Video Strategy to Enhance Action\n  Prediction", "abstract": "  The emerging field of action prediction plays a vital role in various\ncomputer vision applications such as autonomous driving, activity analysis and\nhuman-computer interaction. Despite significant advancements, accurately\npredicting future actions remains a challenging problem due to high\ndimensionality, complex dynamics and uncertainties inherent in video data.\nTraditional supervised approaches require large amounts of labelled data, which\nis expensive and time-consuming to obtain. This paper introduces a novel\nself-supervised video strategy for enhancing action prediction inspired by DINO\n(self-distillation with no labels). The Temporal-DINO approach employs two\nmodels; a 'student' processing past frames; and a 'teacher' processing both\npast and future frames, enabling a broader temporal context. During training,\nthe teacher guides the student to learn future context by only observing past\nframes. The strategy is evaluated on ROAD dataset for the action prediction\ndownstream task using 3D-ResNet, Transformer, and LSTM architectures. The\nexperimental results showcase significant improvements in prediction\nperformance across these architectures, with our method achieving an average\nenhancement of 9.9% Precision Points (PP), highlighting its effectiveness in\nenhancing the backbones' capabilities of capturing long-term dependencies.\nFurthermore, our approach demonstrates efficiency regarding the pretraining\ndataset size and the number of epochs required. This method overcomes\nlimitations present in other approaches, including considering various backbone\narchitectures, addressing multiple prediction horizons, reducing reliance on\nhand-crafted augmentations, and streamlining the pretraining process into a\nsingle stage. These findings highlight the potential of our approach in diverse\nvideo-based tasks such as activity recognition, motion planning, and scene\nunderstanding.\n", "date": "2023-08-08", "categories": []}
{"id": "2308.07558", "title": "Action Class Relation Detection and Classification Across Multiple Video\n  Datasets", "abstract": "  The Meta Video Dataset (MetaVD) provides annotated relations between action\nclasses in major datasets for human action recognition in videos. Although\nthese annotated relations enable dataset augmentation, it is only applicable to\nthose covered by MetaVD. For an external dataset to enjoy the same benefit, the\nrelations between its action classes and those in MetaVD need to be determined.\nTo address this issue, we consider two new machine learning tasks: action class\nrelation detection and classification. We propose a unified model to predict\nrelations between action classes, using language and visual information\nassociated with classes. Experimental results show that (i) pre-trained recent\nneural network models for texts and videos contribute to high predictive\nperformance, (ii) the relation prediction based on action label texts is more\naccurate than based on videos, and (iii) a blending approach that combines\npredictions by both modalities can further improve the predictive performance\nin some cases.\n", "date": "2023-08-15", "categories": []}
{"id": "2308.09897", "title": "Spatial-Temporal Alignment Network for Action Recognition", "abstract": "  This paper studies introducing viewpoint invariant feature representations in\nexisting action recognition architecture. Despite significant progress in\naction recognition, efficiently handling geometric variations in large-scale\ndatasets remains challenging. To tackle this problem, we propose a novel\nSpatial-Temporal Alignment Network (STAN), which explicitly learns geometric\ninvariant representations for action recognition. Notably, the STAN model is\nlight-weighted and generic, which could be plugged into existing action\nrecognition models (e.g., MViTv2) with a low extra computational cost. We test\nour STAN model on widely-used datasets like UCF101 and HMDB51. The experimental\nresults show that the STAN model can consistently improve the state-of-the-art\nmodels in action recognition tasks in trained-from-scratch settings.\n", "date": "2023-08-19", "categories": []}
{"id": "2308.09993", "title": "TTPOINT: A Tensorized Point Cloud Network for Lightweight Action\n  Recognition with Event Cameras", "abstract": "  Event cameras have gained popularity in computer vision due to their data\nsparsity, high dynamic range, and low latency. As a bio-inspired sensor, event\ncameras generate sparse and asynchronous data, which is inherently incompatible\nwith the traditional frame-based method. Alternatively, the point-based method\ncan avoid additional modality transformation and naturally adapt to the\nsparsity of events. Still, it typically cannot reach a comparable accuracy as\nthe frame-based method. We propose a lightweight and generalized point cloud\nnetwork called TTPOINT which achieves competitive results even compared to the\nstate-of-the-art (SOTA) frame-based method in action recognition tasks while\nonly using 1.5 % of the computational resources. The model is adept at\nabstracting local and global geometry by hierarchy structure. By leveraging\ntensor-train compressed feature extractors, TTPOINT can be designed with\nminimal parameters and computational complexity. Additionally, we developed a\nstraightforward downsampling algorithm to maintain the spatio-temporal feature.\nIn the experiment, TTPOINT emerged as the SOTA method on three datasets while\nalso attaining SOTA among point cloud methods on all five datasets. Moreover,\nby using the tensor-train decomposition method, the accuracy of the proposed\nTTPOINT is almost unaffected while compressing the parameter size by 55 % in\nall five datasets.\n", "date": "2023-08-19", "categories": []}
{"id": "2308.10570", "title": "Self-Feedback DETR for Temporal Action Detection", "abstract": "  Temporal Action Detection (TAD) is challenging but fundamental for real-world\nvideo applications. Recently, DETR-based models have been devised for TAD but\nhave not performed well yet. In this paper, we point out the problem in the\nself-attention of DETR for TAD; the attention modules focus on a few key\nelements, called temporal collapse problem. It degrades the capability of the\nencoder and decoder since their self-attention modules play no role. To solve\nthe problem, we propose a novel framework, Self-DETR, which utilizes\ncross-attention maps of the decoder to reactivate self-attention modules. We\nrecover the relationship between encoder features by simple matrix\nmultiplication of the cross-attention map and its transpose. Likewise, we also\nget the information within decoder queries. By guiding collapsed self-attention\nmaps with the guidance map calculated, we settle down the temporal collapse of\nself-attention modules in the encoder and decoder. Our extensive experiments\ndemonstrate that Self-DETR resolves the temporal collapse problem by keeping\nhigh diversity of attention over all layers.\n", "date": "2023-08-21", "categories": []}
{"id": "2212.05638", "title": "Cross-Modal Learning with 3D Deformable Attention for Action Recognition", "abstract": "  An important challenge in vision-based action recognition is the embedding of\nspatiotemporal features with two or more heterogeneous modalities into a single\nfeature. In this study, we propose a new 3D deformable transformer for action\nrecognition with adaptive spatiotemporal receptive fields and a cross-modal\nlearning scheme. The 3D deformable transformer consists of three attention\nmodules: 3D deformability, local joint stride, and temporal stride attention.\nThe two cross-modal tokens are input into the 3D deformable attention module to\ncreate a cross-attention token with a reflected spatiotemporal correlation.\nLocal joint stride attention is applied to spatially combine attention and pose\ntokens. Temporal stride attention temporally reduces the number of input tokens\nin the attention module and supports temporal expression learning without the\nsimultaneous use of all tokens. The deformable transformer iterates L-times and\ncombines the last cross-modal token for classification. The proposed 3D\ndeformable transformer was tested on the NTU60, NTU120, FineGYM, and PennAction\ndatasets, and showed results better than or similar to pre-trained\nstate-of-the-art methods even without a pre-training process. In addition, by\nvisualizing important joints and correlations during action recognition through\nspatial joint and temporal stride attention, the possibility of achieving an\nexplainable potential for action recognition is presented.\n", "date": "2022-12-12", "categories": []}
{"id": "2308.05681", "title": "Hard No-Box Adversarial Attack on Skeleton-Based Human Action\n  Recognition with Skeleton-Motion-Informed Gradient", "abstract": "  Recently, methods for skeleton-based human activity recognition have been\nshown to be vulnerable to adversarial attacks. However, these attack methods\nrequire either the full knowledge of the victim (i.e. white-box attacks),\naccess to training data (i.e. transfer-based attacks) or frequent model queries\n(i.e. black-box attacks). All their requirements are highly restrictive,\nraising the question of how detrimental the vulnerability is. In this paper, we\nshow that the vulnerability indeed exists. To this end, we consider a new\nattack task: the attacker has no access to the victim model or the training\ndata or labels, where we coin the term hard no-box attack. Specifically, we\nfirst learn a motion manifold where we define an adversarial loss to compute a\nnew gradient for the attack, named skeleton-motion-informed (SMI) gradient. Our\ngradient contains information of the motion dynamics, which is different from\nexisting gradient-based attack methods that compute the loss gradient assuming\neach dimension in the data is independent. The SMI gradient can augment many\ngradient-based attack methods, leading to a new family of no-box attack\nmethods. Extensive evaluation and comparison show that our method imposes a\nreal threat to existing classifiers. They also show that the SMI gradient\nimproves the transferability and imperceptibility of adversarial samples in\nboth no-box and transfer-based black-box settings.\n", "date": "2023-08-10", "categories": []}
{"id": "2308.09268", "title": "Progression-Guided Temporal Action Detection in Videos", "abstract": "  We present a novel framework, Action Progression Network (APN), for temporal\naction detection (TAD) in videos. The framework locates actions in videos by\ndetecting the action evolution process. To encode the action evolution, we\nquantify a complete action process into 101 ordered stages (0\\%, 1\\%, ...,\n100\\%), referred to as action progressions. We then train a neural network to\nrecognize the action progressions. The framework detects action boundaries by\ndetecting complete action processes in the videos, e.g., a video segment with\ndetected action progressions closely follow the sequence 0\\%, 1\\%, ..., 100\\%.\nThe framework offers three major advantages: (1) Our neural networks are\ntrained end-to-end, contrasting conventional methods that optimize modules\nseparately; (2) The APN is trained using action frames exclusively, enabling\nmodels to be trained on action classification datasets and robust to videos\nwith temporal background styles differing from those in training; (3) Our\nframework effectively avoids detecting incomplete actions and excels in\ndetecting long-lasting actions due to the fine-grained and explicit encoding of\nthe temporal structure of actions. Leveraging these advantages, the APN\nachieves competitive performance and significantly surpasses its counterparts\nin detecting long-lasting actions. With an IoU threshold of 0.5, the APN\nachieves a mean Average Precision (mAP) of 58.3\\% on the THUMOS14 dataset and\n98.9\\% mAP on the DFMAD70 dataset.\n", "date": "2023-08-18", "categories": []}
{"id": "2308.09327", "title": "Unlimited Knowledge Distillation for Action Recognition in the Dark", "abstract": "  Dark videos often lose essential information, which causes the knowledge\nlearned by networks is not enough to accurately recognize actions. Existing\nknowledge assembling methods require massive GPU memory to distill the\nknowledge from multiple teacher models into a student model. In action\nrecognition, this drawback becomes serious due to much computation required by\nvideo process. Constrained by limited computation source, these approaches are\ninfeasible. To address this issue, we propose an unlimited knowledge\ndistillation (UKD) in this paper. Compared with existing knowledge assembling\nmethods, our UKD can effectively assemble different knowledge without\nintroducing high GPU memory consumption. Thus, the number of teaching models\nfor distillation is unlimited. With our UKD, the network's learned knowledge\ncan be remarkably enriched. Our experiments show that the single stream network\ndistilled with our UKD even surpasses a two-stream network. Extensive\nexperiments are conducted on the ARID dataset.\n", "date": "2023-08-18", "categories": []}
{"id": "2308.09346", "title": "Boosting Few-shot Action Recognition with Graph-guided Hybrid Matching", "abstract": "  Class prototype construction and matching are core aspects of few-shot action\nrecognition. Previous methods mainly focus on designing spatiotemporal relation\nmodeling modules or complex temporal alignment algorithms. Despite the\npromising results, they ignored the value of class prototype construction and\nmatching, leading to unsatisfactory performance in recognizing similar\ncategories in every task. In this paper, we propose GgHM, a new framework with\nGraph-guided Hybrid Matching. Concretely, we learn task-oriented features by\nthe guidance of a graph neural network during class prototype construction,\noptimizing the intra- and inter-class feature correlation explicitly. Next, we\ndesign a hybrid matching strategy, combining frame-level and tuple-level\nmatching to classify videos with multivariate styles. We additionally propose a\nlearnable dense temporal modeling module to enhance the video feature temporal\nrepresentation to build a more solid foundation for the matching process. GgHM\nshows consistent improvements over other challenging baselines on several\nfew-shot datasets, demonstrating the effectiveness of our method. The code will\nbe publicly available at https://github.com/jiazheng-xing/GgHM.\n", "date": "2023-08-18", "categories": []}
{"id": "2308.09685", "title": "Audiovisual Moments in Time: A Large-Scale Annotated Dataset of\n  Audiovisual Actions", "abstract": "  We present Audiovisual Moments in Time (AVMIT), a large-scale dataset of\naudiovisual action events. In an extensive annotation task 11 participants\nlabelled a subset of 3-second audiovisual videos from the Moments in Time\ndataset (MIT). For each trial, participants assessed whether the labelled\naudiovisual action event was present and whether it was the most prominent\nfeature of the video. The dataset includes the annotation of 57,177 audiovisual\nvideos, each independently evaluated by 3 of 11 trained participants. From this\ninitial collection, we created a curated test set of 16 distinct action\nclasses, with 60 videos each (960 videos). We also offer 2 sets of pre-computed\naudiovisual feature embeddings, using VGGish/YamNet for audio data and\nVGG16/EfficientNetB0 for visual data, thereby lowering the barrier to entry for\naudiovisual DNN research. We explored the advantages of AVMIT annotations and\nfeature embeddings to improve performance on audiovisual event recognition. A\nseries of 6 Recurrent Neural Networks (RNNs) were trained on either\nAVMIT-filtered audiovisual events or modality-agnostic events from MIT, and\nthen tested on our audiovisual test set. In all RNNs, top 1 accuracy was\nincreased by 2.71-5.94\\% by training exclusively on audiovisual events, even\noutweighing a three-fold increase in training data. We anticipate that the\nnewly annotated AVMIT dataset will serve as a valuable resource for research\nand comparative experiments involving computational models and human\nparticipants, specifically when addressing research questions where audiovisual\ncorrespondence is of critical importance.\n", "date": "2023-08-18", "categories": []}
{"id": "2304.10465", "title": "Implicit Temporal Modeling with Learnable Alignment for Video\n  Recognition", "abstract": "  Contrastive language-image pretraining (CLIP) has demonstrated remarkable\nsuccess in various image tasks. However, how to extend CLIP with effective\ntemporal modeling is still an open and crucial problem. Existing factorized or\njoint spatial-temporal modeling trades off between the efficiency and\nperformance. While modeling temporal information within straight through tube\nis widely adopted in literature, we find that simple frame alignment already\nprovides enough essence without temporal attention. To this end, in this paper,\nwe proposed a novel Implicit Learnable Alignment (ILA) method, which minimizes\nthe temporal modeling effort while achieving incredibly high performance.\nSpecifically, for a frame pair, an interactive point is predicted in each\nframe, serving as a mutual information rich region. By enhancing the features\naround the interactive point, two frames are implicitly aligned. The aligned\nfeatures are then pooled into a single token, which is leveraged in the\nsubsequent spatial self-attention. Our method allows eliminating the costly or\ninsufficient temporal self-attention in video. Extensive experiments on\nbenchmarks demonstrate the superiority and generality of our module.\nParticularly, the proposed ILA achieves a top-1 accuracy of 88.7% on\nKinetics-400 with much fewer FLOPs compared with Swin-L and ViViT-H. Code is\nreleased at https://github.com/Francis-Rings/ILA .\n", "date": "2023-04-20", "categories": []}
{"id": "2308.07571", "title": "Ske2Grid: Skeleton-to-Grid Representation Learning for Action\n  Recognition", "abstract": "  This paper presents Ske2Grid, a new representation learning framework for\nimproved skeleton-based action recognition. In Ske2Grid, we define a regular\nconvolution operation upon a novel grid representation of human skeleton, which\nis a compact image-like grid patch constructed and learned through three novel\ndesigns. Specifically, we propose a graph-node index transform (GIT) to\nconstruct a regular grid patch through assigning the nodes in the skeleton\ngraph one by one to the desired grid cells. To ensure that GIT is a bijection\nand enrich the expressiveness of the grid representation, an up-sampling\ntransform (UPT) is learned to interpolate the skeleton graph nodes for filling\nthe grid patch to the full. To resolve the problem when the one-step UPT is\naggressive and further exploit the representation capability of the grid patch\nwith increasing spatial size, a progressive learning strategy (PLS) is proposed\nwhich decouples the UPT into multiple steps and aligns them to multiple paired\nGITs through a compact cascaded design learned progressively. We construct\nnetworks upon prevailing graph convolution networks and conduct experiments on\nsix mainstream skeleton-based action recognition datasets. Experiments show\nthat our Ske2Grid significantly outperforms existing GCN-based solutions under\ndifferent benchmark settings, without bells and whistles. Code and models are\navailable at https://github.com/OSVAI/Ske2Grid\n", "date": "2023-08-15", "categories": []}
{"id": "2308.07893", "title": "Memory-and-Anticipation Transformer for Online Action Understanding", "abstract": "  Most existing forecasting systems are memory-based methods, which attempt to\nmimic human forecasting ability by employing various memory mechanisms and have\nprogressed in temporal modeling for memory dependency. Nevertheless, an obvious\nweakness of this paradigm is that it can only model limited historical\ndependence and can not transcend the past. In this paper, we rethink the\ntemporal dependence of event evolution and propose a novel\nmemory-anticipation-based paradigm to model an entire temporal structure,\nincluding the past, present, and future. Based on this idea, we present\nMemory-and-Anticipation Transformer (MAT), a memory-anticipation-based\napproach, to address the online action detection and anticipation tasks. In\naddition, owing to the inherent superiority of MAT, it can process online\naction detection and anticipation tasks in a unified manner. The proposed MAT\nmodel is tested on four challenging benchmarks TVSeries, THUMOS'14, HDD, and\nEPIC-Kitchens-100, for online action detection and anticipation tasks, and it\nsignificantly outperforms all existing methods. Code is available at\nhttps://github.com/Echo0125/Memory-and-Anticipation-Transformer.\n", "date": "2023-08-15", "categories": []}
{"id": "2303.17959", "title": "Diffusion Action Segmentation", "abstract": "  Temporal action segmentation is crucial for understanding long-form videos.\nPrevious works on this task commonly adopt an iterative refinement paradigm by\nusing multi-stage models. We propose a novel framework via denoising diffusion\nmodels, which nonetheless shares the same inherent spirit of such iterative\nrefinement. In this framework, action predictions are iteratively generated\nfrom random noise with input video features as conditions. To enhance the\nmodeling of three striking characteristics of human actions, including the\nposition prior, the boundary ambiguity, and the relational dependency, we\ndevise a unified masking strategy for the conditioning inputs in our framework.\nExtensive experiments on three benchmark datasets, i.e., GTEA, 50Salads, and\nBreakfast, are performed and the proposed method achieves superior or\ncomparable results to state-of-the-art methods, showing the effectiveness of a\ngenerative approach for action segmentation.\n", "date": "2023-03-31", "categories": []}
{"id": "2308.06897", "title": "Orthogonal Temporal Interpolation for Zero-Shot Video Recognition", "abstract": "  Zero-shot video recognition (ZSVR) is a task that aims to recognize video\ncategories that have not been seen during the model training process. Recently,\nvision-language models (VLMs) pre-trained on large-scale image-text pairs have\ndemonstrated impressive transferability for ZSVR. To make VLMs applicable to\nthe video domain, existing methods often use an additional temporal learning\nmodule after the image-level encoder to learn the temporal relationships among\nvideo frames. Unfortunately, for video from unseen categories, we observe an\nabnormal phenomenon where the model that uses spatial-temporal feature performs\nmuch worse than the model that removes temporal learning module and uses only\nspatial feature. We conjecture that improper temporal modeling on video\ndisrupts the spatial feature of the video. To verify our hypothesis, we propose\nFeature Factorization to retain the orthogonal temporal feature of the video\nand use interpolation to construct refined spatial-temporal feature. The model\nusing appropriately refined spatial-temporal feature performs better than the\none using only spatial feature, which verifies the effectiveness of the\northogonal temporal feature for the ZSVR task. Therefore, an Orthogonal\nTemporal Interpolation module is designed to learn a better refined\nspatial-temporal video feature during training. Additionally, a Matching Loss\nis introduced to improve the quality of the orthogonal temporal feature. We\npropose a model called OTI for ZSVR by employing orthogonal temporal\ninterpolation and the matching loss based on VLMs. The ZSVR accuracies on\npopular video datasets (i.e., Kinetics-600, UCF101 and HMDB51) show that OTI\noutperforms the previous state-of-the-art method by a clear margin.\n", "date": "2023-08-14", "categories": []}
{"id": "2308.07092", "title": "Masked Motion Predictors are Strong 3D Action Representation Learners", "abstract": "  In 3D human action recognition, limited supervised data makes it challenging\nto fully tap into the modeling potential of powerful networks such as\ntransformers. As a result, researchers have been actively investigating\neffective self-supervised pre-training strategies. In this work, we show that\ninstead of following the prevalent pretext task to perform masked\nself-component reconstruction in human joints, explicit contextual motion\nmodeling is key to the success of learning effective feature representation for\n3D action recognition. Formally, we propose the Masked Motion Prediction (MAMP)\nframework. To be specific, the proposed MAMP takes as input the masked\nspatio-temporal skeleton sequence and predicts the corresponding temporal\nmotion of the masked human joints. Considering the high temporal redundancy of\nthe skeleton sequence, in our MAMP, the motion information also acts as an\nempirical semantic richness prior that guide the masking process, promoting\nbetter attention to semantically rich temporal regions. Extensive experiments\non NTU-60, NTU-120, and PKU-MMD datasets show that the proposed MAMP\npre-training substantially improves the performance of the adopted vanilla\ntransformer, achieving state-of-the-art results without bells and whistles. The\nsource code of our MAMP is available at https://github.com/maoyunyao/MAMP.\n", "date": "2023-08-14", "categories": []}
{"id": "2308.07119", "title": "On the Importance of Spatial Relations for Few-shot Action Recognition", "abstract": "  Deep learning has achieved great success in video recognition, yet still\nstruggles to recognize novel actions when faced with only a few examples. To\ntackle this challenge, few-shot action recognition methods have been proposed\nto transfer knowledge from a source dataset to a novel target dataset with only\none or a few labeled videos. However, existing methods mainly focus on modeling\nthe temporal relations between the query and support videos while ignoring the\nspatial relations. In this paper, we find that the spatial misalignment between\nobjects also occurs in videos, notably more common than the temporal\ninconsistency. We are thus motivated to investigate the importance of spatial\nrelations and propose a more accurate few-shot action recognition method that\nleverages both spatial and temporal information. Particularly, a novel Spatial\nAlignment Cross Transformer (SA-CT) which learns to re-adjust the spatial\nrelations and incorporates the temporal information is contributed. Experiments\nreveal that, even without using any temporal information, the performance of\nSA-CT is comparable to temporal based methods on 3/4 benchmarks. To further\nincorporate the temporal information, we propose a simple yet effective\nTemporal Mixer module. The Temporal Mixer enhances the video representation and\nimproves the performance of the full SA-CT model, achieving very competitive\nresults. In this work, we also exploit large-scale pretrained models for\nfew-shot action recognition, providing useful insights for this research\ndirection.\n", "date": "2023-08-14", "categories": []}
{"id": "2307.00464", "title": "Human-to-Human Interaction Detection", "abstract": "  A comprehensive understanding of interested human-to-human interactions in\nvideo streams, such as queuing, handshaking, fighting and chasing, is of\nimmense importance to the surveillance of public security in regions like\ncampuses, squares and parks. Different from conventional human interaction\nrecognition, which uses choreographed videos as inputs, neglects concurrent\ninteractive groups, and performs detection and recognition in separate stages,\nwe introduce a new task named human-to-human interaction detection (HID). HID\ndevotes to detecting subjects, recognizing person-wise actions, and grouping\npeople according to their interactive relations, in one model. First, based on\nthe popular AVA dataset created for action detection, we establish a new HID\nbenchmark, termed AVA-Interaction (AVA-I), by adding annotations on interactive\nrelations in a frame-by-frame manner. AVA-I consists of 85,254 frames and\n86,338 interactive groups, and each image includes up to 4 concurrent\ninteractive groups. Second, we present a novel baseline approach SaMFormer for\nHID, containing a visual feature extractor, a split stage which leverages a\nTransformer-based model to decode action instances and interactive groups, and\na merging stage which reconstructs the relationship between instances and\ngroups. All SaMFormer components are jointly trained in an end-to-end manner.\nExtensive experiments on AVA-I validate the superiority of SaMFormer over\nrepresentative methods. The dataset and code will be made public to encourage\nmore follow-up studies.\n", "date": "2023-07-02", "categories": []}
{"id": "2308.04549", "title": "Prune Spatio-temporal Tokens by Semantic-aware Temporal Accumulation", "abstract": "  Transformers have become the primary backbone of the computer vision\ncommunity due to their impressive performance. However, the unfriendly\ncomputation cost impedes their potential in the video recognition domain. To\noptimize the speed-accuracy trade-off, we propose Semantic-aware Temporal\nAccumulation score (STA) to prune spatio-temporal tokens integrally. STA score\nconsiders two critical factors: temporal redundancy and semantic importance.\nThe former depicts a specific region based on whether it is a new occurrence or\na seen entity by aggregating token-to-token similarity in consecutive frames\nwhile the latter evaluates each token based on its contribution to the overall\nprediction. As a result, tokens with higher scores of STA carry more temporal\nredundancy as well as lower semantics thus being pruned. Based on the STA\nscore, we are able to progressively prune the tokens without introducing any\nadditional parameters or requiring further re-training. We directly apply the\nSTA module to off-the-shelf ViT and VideoSwin backbones, and the empirical\nresults on Kinetics-400 and Something-Something V2 achieve over 30% computation\nreduction with a negligible ~0.2% accuracy drop. The code is released at\nhttps://github.com/Mark12Ding/STA.\n", "date": "2023-08-08", "categories": []}
{"id": "2308.04828", "title": "Seeing in Flowing: Adapting CLIP for Action Recognition with Motion\n  Prompts Learning", "abstract": "  The Contrastive Language-Image Pre-training (CLIP) has recently shown\nremarkable generalization on \"zero-shot\" training and has applied to many\ndownstream tasks. We explore the adaptation of CLIP to achieve a more efficient\nand generalized action recognition method. We propose that the key lies in\nexplicitly modeling the motion cues flowing in video frames. To that end, we\ndesign a two-stream motion modeling block to capture motion and spatial\ninformation at the same time. And then, the obtained motion cues are utilized\nto drive a dynamic prompts learner to generate motion-aware prompts, which\ncontain much semantic information concerning human actions. In addition, we\npropose a multimodal communication block to achieve a collaborative learning\nand further improve the performance. We conduct extensive experiments on\nHMDB-51, UCF-101, and Kinetics-400 datasets. Our method outperforms most\nexisting state-of-the-art methods by a significant margin on \"few-shot\" and\n\"zero-shot\" training. We also achieve competitive performance on \"closed-set\"\ntraining with extremely few trainable parameters and additional computational\ncosts.\n", "date": "2023-08-09", "categories": []}
{"id": "2308.04934", "title": "JEDI: Joint Expert Distillation in a Semi-Supervised Multi-Dataset\n  Student-Teacher Scenario for Video Action Recognition", "abstract": "  We propose JEDI, a multi-dataset semi-supervised learning method, which\nefficiently combines knowledge from multiple experts, learned on different\ndatasets, to train and improve the performance of individual, per dataset,\nstudent models. Our approach achieves this by addressing two important problems\nin current machine learning research: generalization across datasets and\nlimitations of supervised training due to scarcity of labeled data. We start\nwith an arbitrary number of experts, pretrained on their own specific dataset,\nwhich form the initial set of student models. The teachers are immediately\nderived by concatenating the feature representations from the penultimate\nlayers of the students. We then train all models in a student-teacher\nsemi-supervised learning scenario until convergence. In our efficient approach,\nstudent-teacher training is carried out jointly and end-to-end, showing that\nboth students and teachers improve their generalization capacity during\ntraining. We validate our approach on four video action recognition datasets.\nBy simultaneously considering all datasets within a unified semi-supervised\nsetting, we demonstrate significant improvements over the initial experts.\n", "date": "2023-08-09", "categories": []}
{"id": "2308.05051", "title": "PAT: Position-Aware Transformer for Dense Multi-Label Action Detection", "abstract": "  We present PAT, a transformer-based network that learns complex temporal\nco-occurrence action dependencies in a video by exploiting multi-scale temporal\nfeatures. In existing methods, the self-attention mechanism in transformers\nloses the temporal positional information, which is essential for robust action\ndetection. To address this issue, we (i) embed relative positional encoding in\nthe self-attention mechanism and (ii) exploit multi-scale temporal\nrelationships by designing a novel non hierarchical network, in contrast to the\nrecent transformer-based approaches that use a hierarchical structure. We argue\nthat joining the self-attention mechanism with multiple sub-sampling processes\nin the hierarchical approaches results in increased loss of positional\ninformation. We evaluate the performance of our proposed approach on two\nchallenging dense multi-label benchmark datasets, and show that PAT improves\nthe current state-of-the-art result by 1.1% and 0.6% mAP on the Charades and\nMultiTHUMOS datasets, respectively, thereby achieving the new state-of-the-art\nmAP at 26.5% and 44.6%, respectively. We also perform extensive ablation\nstudies to examine the impact of the different components of our proposed\nnetwork.\n", "date": "2023-08-09", "categories": []}
{"id": "2308.03908", "title": "ViLP: Knowledge Exploration using Vision, Language, and Pose Embeddings\n  for Video Action Recognition", "abstract": "  Video Action Recognition (VAR) is a challenging task due to its inherent\ncomplexities. Though different approaches have been explored in the literature,\ndesigning a unified framework to recognize a large number of human actions is\nstill a challenging problem. Recently, Multi-Modal Learning (MML) has\ndemonstrated promising results in this domain. In literature, 2D skeleton or\npose modality has often been used for this task, either independently or in\nconjunction with the visual information (RGB modality) present in videos.\nHowever, the combination of pose, visual information, and text attributes has\nnot been explored yet, though text and pose attributes independently have been\nproven to be effective in numerous computer vision tasks. In this paper, we\npresent the first pose augmented Vision-language model (VLM) for VAR. Notably,\nour scheme achieves an accuracy of 92.81% and 73.02% on two popular human video\naction recognition benchmark datasets, UCF-101 and HMDB-51, respectively, even\nwithout any video data pre-training, and an accuracy of 96.11% and 75.75% after\nkinetics pre-training.\n", "date": "2023-08-07", "categories": []}
{"id": "2308.03975", "title": "Prompted Contrast with Masked Motion Modeling: Towards Versatile 3D\n  Action Representation Learning", "abstract": "  Self-supervised learning has proved effective for skeleton-based human action\nunderstanding, which is an important yet challenging topic. Previous works\nmainly rely on contrastive learning or masked motion modeling paradigm to model\nthe skeleton relations. However, the sequence-level and joint-level\nrepresentation learning cannot be effectively and simultaneously handled by\nthese methods. As a result, the learned representations fail to generalize to\ndifferent downstream tasks. Moreover, combining these two paradigms in a naive\nmanner leaves the synergy between them untapped and can lead to interference in\ntraining. To address these problems, we propose Prompted Contrast with Masked\nMotion Modeling, PCM$^{\\rm 3}$, for versatile 3D action representation\nlearning. Our method integrates the contrastive learning and masked prediction\ntasks in a mutually beneficial manner, which substantially boosts the\ngeneralization capacity for various downstream tasks. Specifically, masked\nprediction provides novel training views for contrastive learning, which in\nturn guides the masked prediction training with high-level semantic\ninformation. Moreover, we propose a dual-prompted multi-task pretraining\nstrategy, which further improves model representations by reducing the\ninterference caused by learning the two different pretext tasks. Extensive\nexperiments on five downstream tasks under three large-scale datasets are\nconducted, demonstrating the superior generalization capacity of PCM$^{\\rm 3}$\ncompared to the state-of-the-art works. Our project is publicly available at:\nhttps://jhang2020.github.io/Projects/PCM3/PCM3.html .\n", "date": "2023-08-08", "categories": []}
{"id": "2304.01199", "title": "On the Benefits of 3D Pose and Tracking for Human Action Recognition", "abstract": "  In this work we study the benefits of using tracking and 3D poses for action\nrecognition. To achieve this, we take the Lagrangian view on analysing actions\nover a trajectory of human motion rather than at a fixed point in space. Taking\nthis stand allows us to use the tracklets of people to predict their actions.\nIn this spirit, first we show the benefits of using 3D pose to infer actions,\nand study person-person interactions. Subsequently, we propose a Lagrangian\nAction Recognition model by fusing 3D pose and contextualized appearance over\ntracklets. To this end, our method achieves state-of-the-art performance on the\nAVA v2.2 dataset on both pose only settings and on standard benchmark settings.\nWhen reasoning about the action using only pose cues, our pose model achieves\n+10.0 mAP gain over the corresponding state-of-the-art while our fused model\nhas a gain of +2.8 mAP over the best state-of-the-art model. Code and results\nare available at: https://brjathu.github.io/LART\n", "date": "2023-04-03", "categories": []}
{"id": "2307.16415", "title": "DDG-Net: Discriminability-Driven Graph Network for Weakly-supervised\n  Temporal Action Localization", "abstract": "  Weakly-supervised temporal action localization (WTAL) is a practical yet\nchallenging task. Due to large-scale datasets, most existing methods use a\nnetwork pretrained in other datasets to extract features, which are not\nsuitable enough for WTAL. To address this problem, researchers design several\nmodules for feature enhancement, which improve the performance of the\nlocalization module, especially modeling the temporal relationship between\nsnippets. However, all of them neglect the adverse effects of ambiguous\ninformation, which would reduce the discriminability of others. Considering\nthis phenomenon, we propose Discriminability-Driven Graph Network (DDG-Net),\nwhich explicitly models ambiguous snippets and discriminative snippets with\nwell-designed connections, preventing the transmission of ambiguous information\nand enhancing the discriminability of snippet-level representations.\nAdditionally, we propose feature consistency loss to prevent the assimilation\nof features and drive the graph convolution network to generate more\ndiscriminative representations. Extensive experiments on THUMOS14 and\nActivityNet1.2 benchmarks demonstrate the effectiveness of DDG-Net,\nestablishing new state-of-the-art results on both datasets. Source code is\navailable at \\url{https://github.com/XiaojunTang22/ICCV2023-DDGNet}.\n", "date": "2023-07-31", "categories": []}
{"id": "2308.03063", "title": "M$^3$Net: Multi-view Encoding, Matching, and Fusion for Few-shot\n  Fine-grained Action Recognition", "abstract": "  Due to the scarcity of manually annotated data required for fine-grained\nvideo understanding, few-shot fine-grained (FS-FG) action recognition has\ngained significant attention, with the aim of classifying novel fine-grained\naction categories with only a few labeled instances. Despite the progress made\nin FS coarse-grained action recognition, current approaches encounter two\nchallenges when dealing with the fine-grained action categories: the inability\nto capture subtle action details and the insufficiency of learning from limited\ndata that exhibit high intra-class variance and inter-class similarity. To\naddress these limitations, we propose M$^3$Net, a matching-based framework for\nFS-FG action recognition, which incorporates \\textit{multi-view encoding},\n\\textit{multi-view matching}, and \\textit{multi-view fusion} to facilitate\nembedding encoding, similarity matching, and decision making across multiple\nviewpoints. \\textit{Multi-view encoding} captures rich contextual details from\nthe intra-frame, intra-video, and intra-episode perspectives, generating\ncustomized higher-order embeddings for fine-grained data. \\textit{Multi-view\nmatching} integrates various matching functions enabling flexible relation\nmodeling within limited samples to handle multi-scale spatio-temporal\nvariations by leveraging the instance-specific, category-specific, and\ntask-specific perspectives. \\textit{Multi-view fusion} consists of\nmatching-predictions fusion and matching-losses fusion over the above views,\nwhere the former promotes mutual complementarity and the latter enhances\nembedding generalizability by employing multi-task collaborative learning.\nExplainable visualizations and experimental results on three challenging\nbenchmarks demonstrate the superiority of M$^3$Net in capturing fine-grained\naction details and achieving state-of-the-art performance for FS-FG action\nrecognition.\n", "date": "2023-08-06", "categories": []}
{"id": "2308.03741", "title": "MAiVAR-T: Multimodal Audio-image and Video Action Recognizer using\n  Transformers", "abstract": "  In line with the human capacity to perceive the world by simultaneously\nprocessing and integrating high-dimensional inputs from multiple modalities\nlike vision and audio, we propose a novel model, MAiVAR-T (Multimodal\nAudio-Image to Video Action Recognition Transformer). This model employs an\nintuitive approach for the combination of audio-image and video modalities,\nwith a primary aim to escalate the effectiveness of multimodal human action\nrecognition (MHAR). At the core of MAiVAR-T lies the significance of distilling\nsubstantial representations from the audio modality and transmuting these into\nthe image domain. Subsequently, this audio-image depiction is fused with the\nvideo modality to formulate a unified representation. This concerted approach\nstrives to exploit the contextual richness inherent in both audio and video\nmodalities, thereby promoting action recognition. In contrast to existing\nstate-of-the-art strategies that focus solely on audio or video modalities,\nMAiVAR-T demonstrates superior performance. Our extensive empirical evaluations\nconducted on a benchmark action recognition dataset corroborate the model's\nremarkable performance. This underscores the potential enhancements derived\nfrom integrating audio and video modalities for action recognition purposes.\n", "date": "2023-08-01", "categories": []}
{"id": "2308.01526", "title": "Data Augmentation for Human Behavior Analysis in Multi-Person\n  Conversations", "abstract": "  In this paper, we present the solution of our team HFUT-VUT for the\nMultiMediate Grand Challenge 2023 at ACM Multimedia 2023. The solution covers\nthree sub-challenges: bodily behavior recognition, eye contact detection, and\nnext speaker prediction. We select Swin Transformer as the baseline and exploit\ndata augmentation strategies to address the above three tasks. Specifically, we\ncrop the raw video to remove the noise from other parts. At the same time, we\nutilize data augmentation to improve the generalization of the model. As a\nresult, our solution achieves the best results of 0.6262 for bodily behavior\nrecognition in terms of mean average precision and the accuracy of 0.7771 for\neye contact detection on the corresponding test set. In addition, our approach\nalso achieves comparable results of 0.5281 for the next speaker prediction in\nterms of unweighted average recall.\n", "date": "2023-08-03", "categories": []}
{"id": "2308.01618", "title": "A Survey on Deep Learning-based Spatio-temporal Action Detection", "abstract": "  Spatio-temporal action detection (STAD) aims to classify the actions present\nin a video and localize them in space and time. It has become a particularly\nactive area of research in computer vision because of its explosively emerging\nreal-world applications, such as autonomous driving, visual surveillance,\nentertainment, etc. Many efforts have been devoted in recent years to building\na robust and effective framework for STAD. This paper provides a comprehensive\nreview of the state-of-the-art deep learning-based methods for STAD. Firstly, a\ntaxonomy is developed to organize these methods. Next, the linking algorithms,\nwhich aim to associate the frame- or clip-level detection results together to\nform action tubes, are reviewed. Then, the commonly used benchmark datasets and\nevaluation metrics are introduced, and the performance of state-of-the-art\nmodels is compared. At last, this paper is concluded, and a set of potential\nresearch directions of STAD are discussed.\n", "date": "2023-08-03", "categories": []}
{"id": "2208.06174", "title": "Two-person Graph Convolutional Network for Skeleton-based Human\n  Interaction Recognition", "abstract": "  Graph convolutional networks (GCNs) have been the predominant methods in\nskeleton-based human action recognition, including human-human interaction\nrecognition. However, when dealing with interaction sequences, current\nGCN-based methods simply split the two-person skeleton into two discrete graphs\nand perform graph convolution separately as done for single-person action\nclassification. Such operations ignore rich interactive information and hinder\neffective spatial inter-body relationship modeling. To overcome the above\nshortcoming, we introduce a novel unified two-person graph to represent\ninter-body and intra-body correlations between joints. Experiments show\naccuracy improvements in recognizing both interactions and individual actions\nwhen utilizing the proposed two-person graph topology. In addition, We design\nseveral graph labeling strategies to supervise the model to learn discriminant\nspatial-temporal interactive features. Finally, we propose a two-person graph\nconvolutional network (2P-GCN). Our model achieves state-of-the-art results on\nfour benchmarks of three interaction datasets: SBU, interaction subsets of\nNTU-RGB+D and NTU-RGB+D 120.\n", "date": "2022-08-12", "categories": []}
{"id": "2307.16803", "title": "DPMix: Mixture of Depth and Point Cloud Video Experts for 4D Action\n  Segmentation", "abstract": "  In this technical report, we present our findings from the research conducted\non the Human-Object Interaction 4D (HOI4D) dataset for egocentric action\nsegmentation task. As a relatively novel research area, point cloud video\nmethods might not be good at temporal modeling, especially for long point cloud\nvideos (\\eg, 150 frames). In contrast, traditional video understanding methods\nhave been well developed. Their effectiveness on temporal modeling has been\nwidely verified on many large scale video datasets. Therefore, we convert point\ncloud videos into depth videos and employ traditional video modeling methods to\nimprove 4D action segmentation. By ensembling depth and point cloud video\nmethods, the accuracy is significantly improved. The proposed method, named\nMixture of Depth and Point cloud video experts (DPMix), achieved the first\nplace in the 4D Action Segmentation Track of the HOI4D Challenge 2023.\n", "date": "2023-07-31", "categories": []}
{"id": "2307.14866", "title": "Sample Less, Learn More: Efficient Action Recognition via Frame Feature\n  Restoration", "abstract": "  Training an effective video action recognition model poses significant\ncomputational challenges, particularly under limited resource budgets. Current\nmethods primarily aim to either reduce model size or utilize pre-trained\nmodels, limiting their adaptability to various backbone architectures. This\npaper investigates the issue of over-sampled frames, a prevalent problem in\nmany approaches yet it has received relatively little attention. Despite the\nuse of fewer frames being a potential solution, this approach often results in\na substantial decline in performance. To address this issue, we propose a novel\nmethod to restore the intermediate features for two sparsely sampled and\nadjacent video frames. This feature restoration technique brings a negligible\nincrease in computational requirements compared to resource-intensive image\nencoders, such as ViT. To evaluate the effectiveness of our method, we conduct\nextensive experiments on four public datasets, including Kinetics-400,\nActivityNet, UCF-101, and HMDB-51. With the integration of our method, the\nefficiency of three commonly used baselines has been improved by over 50%, with\na mere 0.5% reduction in recognition accuracy. In addition, our method also\nsurprisingly helps improve the generalization ability of the models under\nzero-shot settings.\n", "date": "2023-07-27", "categories": []}
{"id": "2302.13434", "title": "Spatial-temporal Transformer-guided Diffusion based Data Augmentation\n  for Efficient Skeleton-based Action Recognition", "abstract": "  Recently, skeleton-based human action has become a hot research topic because\nthe compact representation of human skeletons brings new blood to this research\ndomain. As a result, researchers began to notice the importance of using RGB or\nother sensors to analyze human action by extracting skeleton information.\nLeveraging the rapid development of deep learning (DL), a significant number of\nskeleton-based human action approaches have been presented with fine-designed\nDL structures recently. However, a well-trained DL model always demands\nhigh-quality and sufficient data, which is hard to obtain without costing high\nexpenses and human labor. In this paper, we introduce a novel data augmentation\nmethod for skeleton-based action recognition tasks, which can effectively\ngenerate high-quality and diverse sequential actions. In order to obtain\nnatural and realistic action sequences, we propose denoising diffusion\nprobabilistic models (DDPMs) that can generate a series of synthetic action\nsequences, and their generation process is precisely guided by a\nspatial-temporal transformer (ST-Trans). Experimental results show that our\nmethod outperforms the state-of-the-art (SOTA) motion generation approaches on\ndifferent naturality and diversity metrics. It proves that its high-quality\nsynthetic data can also be effectively deployed to existing action recognition\nmodels with significant performance improvement.\n", "date": "2023-02-26", "categories": []}
{"id": "2307.07768", "title": "SoccerKDNet: A Knowledge Distillation Framework for Action Recognition\n  in Soccer Videos", "abstract": "  Classifying player actions from soccer videos is a challenging problem, which\nhas become increasingly important in sports analytics over the years. Most\nstate-of-the-art methods employ highly complex offline networks, which makes it\ndifficult to deploy such models in resource constrained scenarios. Here, in\nthis paper we propose a novel end-to-end knowledge distillation based transfer\nlearning network pre-trained on the Kinetics400 dataset and then perform\nextensive analysis on the learned framework by introducing a unique loss\nparameterization. We also introduce a new dataset named SoccerDB1 containing\n448 videos and consisting of 4 diverse classes each of players playing soccer.\nFurthermore, we introduce an unique loss parameter that help us linearly weigh\nthe extent to which the predictions of each network are utilized. Finally, we\nalso perform a thorough performance study using various changed\nhyperparameters. We also benchmark the first classification results on the new\nSoccerDB1 dataset obtaining 67.20% validation accuracy. Apart from\noutperforming prior arts significantly, our model also generalizes to new\ndatasets easily. The dataset has been made publicly available at:\nhttps://bit.ly/soccerdb1\n", "date": "2023-07-15", "categories": []}
{"id": "2307.13541", "title": "Group Activity Recognition in Computer Vision: A Comprehensive Review,\n  Challenges, and Future Perspectives", "abstract": "  Group activity recognition is a hot topic in computer vision. Recognizing\nactivities through group relationships plays a vital role in group activity\nrecognition. It holds practical implications in various scenarios, such as\nvideo analysis, surveillance, automatic driving, and understanding social\nactivities. The model's key capabilities encompass efficiently modeling\nhierarchical relationships within a scene and accurately extracting distinctive\nspatiotemporal features from groups. Given this technology's extensive\napplicability, identifying group activities has garnered significant research\nattention. This work examines the current progress in technology for\nrecognizing group activities, with a specific focus on global interactivity and\nactivities. Firstly, we comprehensively review the pertinent literature and\nvarious group activity recognition approaches, from traditional methodologies\nto the latest methods based on spatial structure, descriptors, non-deep\nlearning, hierarchical recurrent neural networks (HRNN), relationship models,\nand attention mechanisms. Subsequently, we present the relational network and\nrelational architectures for each module. Thirdly, we investigate methods for\nrecognizing group activity and compare their performance with state-of-the-art\ntechnologies. We summarize the existing challenges and provide comprehensive\nguidance for newcomers to understand group activity recognition. Furthermore,\nwe review emerging perspectives in group activity recognition to explore new\ndirections and possibilities.\n", "date": "2023-07-25", "categories": []}
{"id": "2303.08914", "title": "MAtch, eXpand and Improve: Unsupervised Finetuning for Zero-Shot Action\n  Recognition with Language Knowledge", "abstract": "  Large scale Vision-Language (VL) models have shown tremendous success in\naligning representations between visual and text modalities. This enables\nremarkable progress in zero-shot recognition, image generation & editing, and\nmany other exciting tasks. However, VL models tend to over-represent objects\nwhile paying much less attention to verbs, and require additional tuning on\nvideo data for best zero-shot action recognition performance. While previous\nwork relied on large-scale, fully-annotated data, in this work we propose an\nunsupervised approach. We adapt a VL model for zero-shot and few-shot action\nrecognition using a collection of unlabeled videos and an unpaired action\ndictionary. Based on that, we leverage Large Language Models and VL models to\nbuild a text bag for each unlabeled video via matching, text expansion and\ncaptioning. We use those bags in a Multiple Instance Learning setup to adapt an\nimage-text backbone to video data. Although finetuned on unlabeled video data,\nour resulting models demonstrate high transferability to numerous unseen\nzero-shot downstream tasks, improving the base VL model performance by up to\n14\\%, and even comparing favorably to fully-supervised baselines in both\nzero-shot and few-shot video recognition transfer. The code will be released\nlater at \\url{https://github.com/wlin-at/MAXI}.\n", "date": "2023-03-15", "categories": []}
{"id": "2307.12837", "title": "EPIC-KITCHENS-100 Unsupervised Domain Adaptation Challenge: Mixed\n  Sequences Prediction", "abstract": "  This report presents the technical details of our approach for the\nEPIC-Kitchens-100 Unsupervised Domain Adaptation (UDA) Challenge in Action\nRecognition. Our approach is based on the idea that the order in which actions\nare performed is similar between the source and target domains. Based on this,\nwe generate a modified sequence by randomly combining actions from the source\nand target domains. As only unlabelled target data are available under the UDA\nsetting, we use a standard pseudo-labeling strategy for extracting action\nlabels for the target. We then ask the network to predict the resulting action\nsequence. This allows to integrate information from both domains during\ntraining and to achieve better transfer results on target. Additionally, to\nbetter incorporate sequence information, we use a language model to filter\nunlikely sequences. Lastly, we employed a co-occurrence matrix to eliminate\nunseen combinations of verbs and nouns. Our submission, labeled as 'sshayan',\ncan be found on the leaderboard, where it currently holds the 2nd position for\n'verb' and the 4th position for both 'noun' and 'action'.\n", "date": "2023-07-24", "categories": []}
{"id": "2307.10624", "title": "Joint Skeletal and Semantic Embedding Loss for Micro-gesture\n  Classification", "abstract": "  In this paper, we briefly introduce the solution of our team HFUT-VUT for the\nMicros-gesture Classification in the MiGA challenge at IJCAI 2023. The\nmicro-gesture classification task aims at recognizing the action category of a\ngiven video based on the skeleton data. For this task, we propose a\n3D-CNNs-based micro-gesture recognition network, which incorporates a skeletal\nand semantic embedding loss to improve action classification performance.\nFinally, we rank 1st in the Micro-gesture Classification Challenge, surpassing\nthe second-place team in terms of Top-1 accuracy by 1.10%.\n", "date": "2023-07-20", "categories": []}
{"id": "2212.04761", "title": "Leveraging Spatio-Temporal Dependency for Skeleton-Based Action\n  Recognition", "abstract": "  Skeleton-based action recognition has attracted considerable attention due to\nits compact representation of the human body's skeletal sructure. Many recent\nmethods have achieved remarkable performance using graph convolutional networks\n(GCNs) and convolutional neural networks (CNNs), which extract spatial and\ntemporal features, respectively. Although spatial and temporal dependencies in\nthe human skeleton have been explored separately, spatio-temporal dependency is\nrarely considered. In this paper, we propose the Spatio-Temporal Curve Network\n(STC-Net) to effectively leverage the spatio-temporal dependency of the human\nskeleton. Our proposed network consists of two novel elements: 1) The\nSpatio-Temporal Curve (STC) module; and 2) Dilated Kernels for Graph\nConvolution (DK-GC). The STC module dynamically adjusts the receptive field by\nidentifying meaningful node connections between every adjacent frame and\ngenerating spatio-temporal curves based on the identified node connections,\nproviding an adaptive spatio-temporal coverage. In addition, we propose DK-GC\nto consider long-range dependencies, which results in a large receptive field\nwithout any additional parameters by applying an extended kernel to the given\nadjacency matrices of the graph. Our STC-Net combines these two modules and\nachieves state-of-the-art performance on four skeleton-based action recognition\nbenchmarks.\n", "date": "2022-12-09", "categories": []}
{"id": "2303.01589", "title": "AZTR: Aerial Video Action Recognition with Auto Zoom and Temporal\n  Reasoning", "abstract": "  We propose a novel approach for aerial video action recognition. Our method\nis designed for videos captured using UAVs and can run on edge or mobile\ndevices. We present a learning-based approach that uses customized auto zoom to\nautomatically identify the human target and scale it appropriately. This makes\nit easier to extract the key features and reduces the computational overhead.\nWe also present an efficient temporal reasoning algorithm to capture the action\ninformation along the spatial and temporal domains within a controllable\ncomputational cost. Our approach has been implemented and evaluated both on the\ndesktop with high-end GPUs and on the low power Robotics RB5 Platform for\nrobots and drones. In practice, we achieve 6.1-7.4% improvement over SOTA in\nTop-1 accuracy on the RoCoG-v2 dataset, 8.3-10.4% improvement on the UAV-Human\ndataset and 3.2% improvement on the Drone Action dataset.\n", "date": "2023-03-02", "categories": []}
{"id": "2304.06403", "title": "Leveraging triplet loss for unsupervised action segmentation", "abstract": "  In this paper, we propose a novel fully unsupervised framework that learns\naction representations suitable for the action segmentation task from the\nsingle input video itself, without requiring any training data. Our method is a\ndeep metric learning approach rooted in a shallow network with a triplet loss\noperating on similarity distributions and a novel triplet selection strategy\nthat effectively models temporal and semantic priors to discover actions in the\nnew representational space. Under these circumstances, we successfully recover\ntemporal boundaries in the learned action representations with higher quality\ncompared with existing unsupervised approaches. The proposed method is\nevaluated on two widely used benchmark datasets for the action segmentation\ntask and it achieves competitive performance by applying a generic clustering\nalgorithm on the learned representations.\n", "date": "2023-04-13", "categories": []}
{"id": "2307.07483", "title": "Multimodal Distillation for Egocentric Action Recognition", "abstract": "  The focal point of egocentric video understanding is modelling hand-object\ninteractions. Standard models, e.g. CNNs or Vision Transformers, which receive\nRGB frames as input perform well. However, their performance improves further\nby employing additional input modalities that provide complementary cues, such\nas object detections, optical flow, audio, etc. The added complexity of the\nmodality-specific modules, on the other hand, makes these models impractical\nfor deployment. The goal of this work is to retain the performance of such a\nmultimodal approach, while using only the RGB frames as input at inference\ntime. We demonstrate that for egocentric action recognition on the\nEpic-Kitchens and the Something-Something datasets, students which are taught\nby multimodal teachers tend to be more accurate and better calibrated than\narchitecturally equivalent models trained on ground truth labels in a unimodal\nor multimodal fashion. We further adopt a principled multimodal knowledge\ndistillation framework, allowing us to deal with issues which occur when\napplying multimodal knowledge distillation in a naive manner. Lastly, we\ndemonstrate the achieved reduction in computational complexity, and show that\nour approach maintains higher performance with the reduction of the number of\ninput views. We release our code at\nhttps://github.com/gorjanradevski/multimodal-distillation.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.08944", "title": "Siamese Networks for Weakly Supervised Human Activity Recognition", "abstract": "  Deep learning has been successfully applied to human activity recognition.\nHowever, training deep neural networks requires explicitly labeled data which\nis difficult to acquire. In this paper, we present a model with multiple\nsiamese networks that are trained by using only the information about the\nsimilarity between pairs of data samples without knowing the explicit labels.\nThe trained model maps the activity data samples into fixed size representation\nvectors such that the distance between the vectors in the representation space\napproximates the similarity of the data samples in the input space. Thus, the\ntrained model can work as a metric for a wide range of different clustering\nalgorithms. The training process minimizes a similarity loss function that\nforces the distance metric to be small for pairs of samples from the same kind\nof activity, and large for pairs of samples from different kinds of activities.\nWe evaluate the model on three datasets to verify its effectiveness in\nsegmentation and recognition of continuous human activity sequences.\n", "date": "2023-07-18", "categories": []}
{"id": "2307.09238", "title": "Fusing Hand and Body Skeletons for Human Action Recognition in Assembly", "abstract": "  As collaborative robots (cobots) continue to gain popularity in industrial\nmanufacturing, effective human-robot collaboration becomes crucial. Cobots\nshould be able to recognize human actions to assist with assembly tasks and act\nautonomously. To achieve this, skeleton-based approaches are often used due to\ntheir ability to generalize across various people and environments. Although\nbody skeleton approaches are widely used for action recognition, they may not\nbe accurate enough for assembly actions where the worker's fingers and hands\nplay a significant role. To address this limitation, we propose a method in\nwhich less detailed body skeletons are combined with highly detailed hand\nskeletons. We investigate CNNs and transformers, the latter of which are\nparticularly adept at extracting and combining important information from both\nskeleton types using attention. This paper demonstrates the effectiveness of\nour proposed approach in enhancing action recognition in assembly scenarios.\n", "date": "2023-07-18", "categories": []}
{"id": "2210.03600", "title": "BlanketSet -- A clinical real-world in-bed action recognition and\n  qualitative semi-synchronised MoCap dataset", "abstract": "  Clinical in-bed video-based human motion analysis is a very relevant computer\nvision topic for several relevant biomedical applications. Nevertheless, the\nmain public large datasets (e.g. ImageNet or 3DPW) used for deep learning\napproaches lack annotated examples for these clinical scenarios. To address\nthis issue, we introduce BlanketSet, an RGB-IR-D action recognition dataset of\nsequences performed in a hospital bed. This dataset has the potential to help\nbridge the improvements attained in more general large datasets to these\nclinical scenarios. Information on how to access the dataset is available at\nhttps://rdm.inesctec.pt/dataset/nis-2022-004.\n", "date": "2022-10-07", "categories": []}
{"id": "2307.07977", "title": "Integrating Human Parsing and Pose Network for Human Action Recognition", "abstract": "  Human skeletons and RGB sequences are both widely-adopted input modalities\nfor human action recognition. However, skeletons lack appearance features and\ncolor data suffer large amount of irrelevant depiction. To address this, we\nintroduce human parsing feature map as a novel modality, since it can\nselectively retain spatiotemporal features of the body parts, while filtering\nout noises regarding outfits, backgrounds, etc. We propose an Integrating Human\nParsing and Pose Network (IPP-Net) for action recognition, which is the first\nto leverage both skeletons and human parsing feature maps in dual-branch\napproach. The human pose branch feeds compact skeletal representations of\ndifferent modalities in graph convolutional network to model pose features. In\nhuman parsing branch, multi-frame body-part parsing features are extracted with\nhuman detector and parser, which is later learnt using a convolutional\nbackbone. A late ensemble of two branches is adopted to get final predictions,\nconsidering both robust keypoints and rich semantic body-part features.\nExtensive experiments on NTU RGB+D and NTU RGB+D 120 benchmarks consistently\nverify the effectiveness of the proposed IPP-Net, which outperforms the\nexisting action recognition methods. Our code is publicly available at\nhttps://github.com/liujf69/IPP-Net-Parsing .\n", "date": "2023-07-16", "categories": []}
{"id": "2307.08476", "title": "SkeletonMAE: Graph-based Masked Autoencoder for Skeleton Sequence\n  Pre-training", "abstract": "  Skeleton sequence representation learning has shown great advantages for\naction recognition due to its promising ability to model human joints and\ntopology. However, the current methods usually require sufficient labeled data\nfor training computationally expensive models, which is labor-intensive and\ntime-consuming. Moreover, these methods ignore how to utilize the fine-grained\ndependencies among different skeleton joints to pre-train an efficient skeleton\nsequence learning model that can generalize well across different datasets. In\nthis paper, we propose an efficient skeleton sequence learning framework, named\nSkeleton Sequence Learning (SSL). To comprehensively capture the human pose and\nobtain discriminative skeleton sequence representation, we build an asymmetric\ngraph-based encoder-decoder pre-training architecture named SkeletonMAE, which\nembeds skeleton joint sequence into Graph Convolutional Network (GCN) and\nreconstructs the masked skeleton joints and edges based on the prior human\ntopology knowledge. Then, the pre-trained SkeletonMAE encoder is integrated\nwith the Spatial-Temporal Representation Learning (STRL) module to build the\nSSL framework. Extensive experimental results show that our SSL generalizes\nwell across different datasets and outperforms the state-of-the-art\nself-supervised skeleton-based action recognition methods on FineGym, Diving48,\nNTU 60 and NTU 120 datasets. Additionally, we obtain comparable performance to\nsome fully supervised methods. The code is avaliable at\nhttps://github.com/HongYan1123/SkeletonMAE.\n", "date": "2023-07-17", "categories": []}
{"id": "2303.14863", "title": "DiffTAD: Temporal Action Detection with Proposal Denoising Diffusion", "abstract": "  We propose a new formulation of temporal action detection (TAD) with\ndenoising diffusion, DiffTAD in short. Taking as input random temporal\nproposals, it can yield action proposals accurately given an untrimmed long\nvideo. This presents a generative modeling perspective, against previous\ndiscriminative learning manners. This capability is achieved by first diffusing\nthe ground-truth proposals to random ones (i.e., the forward/noising process)\nand then learning to reverse the noising process (i.e., the backward/denoising\nprocess). Concretely, we establish the denoising process in the Transformer\ndecoder (e.g., DETR) by introducing a temporal location query design with\nfaster convergence in training. We further propose a cross-step selective\nconditioning algorithm for inference acceleration. Extensive evaluations on\nActivityNet and THUMOS show that our DiffTAD achieves top performance compared\nto previous art alternatives. The code will be made available at\nhttps://github.com/sauradip/DiffusionTAD.\n", "date": "2023-03-27", "categories": []}
{"id": "2307.07265", "title": "AudioInceptionNeXt: TCL AI LAB Submission to EPIC-SOUND\n  Audio-Based-Interaction-Recognition Challenge 2023", "abstract": "  This report presents the technical details of our submission to the 2023\nEpic-Kitchen EPIC-SOUNDS Audio-Based Interaction Recognition Challenge. The\ntask is to learn the mapping from audio samples to their corresponding action\nlabels. To achieve this goal, we propose a simple yet effective single-stream\nCNN-based architecture called AudioInceptionNeXt that operates on the\ntime-frequency log-mel-spectrogram of the audio samples. Motivated by the\ndesign of the InceptionNeXt, we propose parallel multi-scale depthwise\nseparable convolutional kernels in the AudioInceptionNeXt block, which enable\nthe model to learn the time and frequency information more effectively. The\nlarge-scale separable kernels capture the long duration of activities and the\nglobal frequency semantic information, while the small-scale separable kernels\ncapture the short duration of activities and local details of frequency\ninformation. Our approach achieved 55.43% of top-1 accuracy on the challenge\ntest set, ranked as 1st on the public leaderboard. Codes are available\nanonymously at https://github.com/StevenLauHKHK/AudioInceptionNeXt.git.\n", "date": "2023-07-14", "categories": []}
{"id": "2307.06569", "title": "A Study on Differentiable Logic and LLMs for EPIC-KITCHENS-100\n  Unsupervised Domain Adaptation Challenge for Action Recognition 2023", "abstract": "  In this technical report, we present our findings from a study conducted on\nthe EPIC-KITCHENS-100 Unsupervised Domain Adaptation task for Action\nRecognition. Our research focuses on the innovative application of a\ndifferentiable logic loss in the training to leverage the co-occurrence\nrelations between verb and noun, as well as the pre-trained Large Language\nModels (LLMs) to generate the logic rules for the adaptation to unseen action\nlabels. Specifically, the model's predictions are treated as the truth\nassignment of a co-occurrence logic formula to compute the logic loss, which\nmeasures the consistency between the predictions and the logic constraints. By\nusing the verb-noun co-occurrence matrix generated from the dataset, we observe\na moderate improvement in model performance compared to our baseline framework.\nTo further enhance the model's adaptability to novel action labels, we\nexperiment with rules generated using GPT-3.5, which leads to a slight decrease\nin performance. These findings shed light on the potential and challenges of\nincorporating differentiable logic and LLMs for knowledge extraction in\nunsupervised domain adaptation for action recognition. Our final submission\n(entitled `NS-LLM') achieved the first place in terms of top-1 action\nrecognition accuracy.\n", "date": "2023-07-13", "categories": []}
{"id": "2307.05784", "title": "EgoAdapt: A multi-stream evaluation study of adaptation to real-world\n  egocentric user video", "abstract": "  In egocentric action recognition a single population model is typically\ntrained and subsequently embodied on a head-mounted device, such as an\naugmented reality headset. While this model remains static for new users and\nenvironments, we introduce an adaptive paradigm of two phases, where after\npretraining a population model, the model adapts on-device and online to the\nuser's experience. This setting is highly challenging due to the change from\npopulation to user domain and the distribution shifts in the user's data\nstream. Coping with the latter in-stream distribution shifts is the focus of\ncontinual learning, where progress has been rooted in controlled benchmarks but\nchallenges faced in real-world applications often remain unaddressed. We\nintroduce EgoAdapt, a benchmark for real-world egocentric action recognition\nthat facilitates our two-phased adaptive paradigm, and real-world challenges\nnaturally occur in the egocentric video streams from Ego4d, such as long-tailed\naction distributions and large-scale classification over 2740 actions. We\nintroduce an evaluation framework that directly exploits the user's data stream\nwith new metrics to measure the adaptation gain over the population model,\nonline generalization, and hindsight performance. In contrast to single-stream\nevaluation in existing works, our framework proposes a meta-evaluation that\naggregates the results from 50 independent user streams. We provide an\nextensive empirical study for finetuning and experience replay.\n", "date": "2023-07-11", "categories": []}
{"id": "2307.04516", "title": "An Examination of Wearable Sensors and Video Data Capture for Human\n  Exercise Classification", "abstract": "  Wearable sensors such as Inertial Measurement Units (IMUs) are often used to\nassess the performance of human exercise. Common approaches use handcrafted\nfeatures based on domain expertise or automatically extracted features using\ntime series analysis. Multiple sensors are required to achieve high\nclassification accuracy, which is not very practical. These sensors require\ncalibration and synchronization and may lead to discomfort over longer time\nperiods. Recent work utilizing computer vision techniques has shown similar\nperformance using video, without the need for manual feature engineering, and\navoiding some pitfalls such as sensor calibration and placement on the body. In\nthis paper, we compare the performance of IMUs to a video-based approach for\nhuman exercise classification on two real-world datasets consisting of Military\nPress and Rowing exercises. We compare the performance using a single camera\nthat captures video in the frontal view versus using 5 IMUs placed on different\nparts of the body. We observe that an approach based on a single camera can\noutperform a single IMU by 10 percentage points on average. Additionally, a\nminimum of 3 IMUs are required to outperform a single camera. We observe that\nworking with the raw data using multivariate time series classifiers\noutperforms traditional approaches based on handcrafted or automatically\nextracted features. Finally, we show that an ensemble model combining the data\nfrom a single camera with a single IMU outperforms either data modality. Our\nwork opens up new and more realistic avenues for this application, where a\nvideo captured using a readily available smartphone camera, combined with a\nsingle sensor, can be used for effective human exercise classification.\n", "date": "2023-07-10", "categories": []}
{"id": "2304.09789", "title": "Automatic Interaction and Activity Recognition from Videos of Human\n  Manual Demonstrations with Application to Anomaly Detection", "abstract": "  This paper presents a new method to describe spatio-temporal relations\nbetween objects and hands, to recognize both interactions and activities within\nvideo demonstrations of manual tasks. The approach exploits Scene Graphs to\nextract key interaction features from image sequences while simultaneously\nencoding motion patterns and context. Additionally, the method introduces\nevent-based automatic video segmentation and clustering, which allow for the\ngrouping of similar events and detect if a monitored activity is executed\ncorrectly. The effectiveness of the approach was demonstrated in two\nmulti-subject experiments, showing the ability to recognize and cluster\nhand-object and object-object interactions without prior knowledge of the\nactivity, as well as matching the same activity performed by different\nsubjects.\n", "date": "2023-04-19", "categories": []}
{"id": "2307.01467", "title": "Technical Report for Ego4D Long Term Action Anticipation Challenge 2023", "abstract": "  In this report, we describe the technical details of our approach for the\nEgo4D Long-Term Action Anticipation Challenge 2023. The aim of this task is to\npredict a sequence of future actions that will take place at an arbitrary time\nor later, given an input video. To accomplish this task, we introduce three\nimprovements to the baseline model, which consists of an encoder that generates\nclip-level features from the video, an aggregator that integrates multiple\nclip-level features, and a decoder that outputs Z future actions. 1) Model\nensemble of SlowFast and SlowFast-CLIP; 2) Label smoothing to relax order\nconstraints for future actions; 3) Constraining the prediction of the action\nclass (verb, noun) based on word co-occurrence. Our method outperformed the\nbaseline performance and recorded as second place solution on the public\nleaderboard.\n", "date": "2023-07-04", "categories": []}
{"id": "2307.02025", "title": "NMS Threshold matters for Ego4D Moment Queries -- 2nd place solution to\n  the Ego4D Moment Queries Challenge 2023", "abstract": "  This report describes our submission to the Ego4D Moment Queries Challenge\n2023. Our submission extends ActionFormer, a latest method for temporal action\nlocalization. Our extension combines an improved ground-truth assignment\nstrategy during training and a refined version of SoftNMS at inference time.\nOur solution is ranked 2nd on the public leaderboard with 26.62% average mAP\nand 45.69% Recall@1x at tIoU=0.5 on the test set, significantly outperforming\nthe strong baseline from 2023 challenge. Our code is available at\nhttps://github.com/happyharrycn/actionformer_release.\n", "date": "2023-07-05", "categories": []}
{"id": "2307.01168", "title": "Don't freeze: Finetune encoders for better Self-Supervised HAR", "abstract": "  Recently self-supervised learning has been proposed in the field of human\nactivity recognition as a solution to the labelled data availability problem.\nThe idea being that by using pretext tasks such as reconstruction or\ncontrastive predictive coding, useful representations can be learned that then\ncan be used for classification. Those approaches follow the pretrain, freeze\nand fine-tune procedure. In this paper we will show how a simple change - not\nfreezing the representation - leads to substantial performance gains across\npretext tasks. The improvement was found in all four investigated datasets and\nacross all four pretext tasks and is inversely proportional to amount of\nlabelled data. Moreover the effect is present whether the pretext task is\ncarried on the Capture24 dataset or directly in unlabelled data of the target\ndataset.\n", "date": "2023-07-03", "categories": []}
{"id": "2306.16545", "title": "Palm: Predicting Actions through Language Models @ Ego4D Long-Term\n  Action Anticipation Challenge 2023", "abstract": "  We present Palm, a solution to the Long-Term Action Anticipation (LTA) task\nutilizing vision-language and large language models. Given an input video with\nannotated action periods, the LTA task aims to predict possible future actions.\nWe hypothesize that an optimal solution should capture the interdependency\nbetween past and future actions, and be able to infer future actions based on\nthe structure and dependency encoded in the past actions. Large language models\nhave demonstrated remarkable commonsense-based reasoning ability. Inspired by\nthat, Palm chains an image captioning model and a large language model. It\npredicts future actions based on frame descriptions and action labels extracted\nfrom the input videos. Our method outperforms other participants in the EGO4D\nLTA challenge and achieves the best performance in terms of action prediction.\nOur code is available at https://github.com/DanDoge/Palm\n", "date": "2023-06-28", "categories": []}
{"id": "2306.16682", "title": "Streaming egocentric action anticipation: An evaluation scheme and\n  approach", "abstract": "  Egocentric action anticipation aims to predict the future actions the camera\nwearer will perform from the observation of the past. While predictions about\nthe future should be available before the predicted events take place, most\napproaches do not pay attention to the computational time required to make such\npredictions. As a result, current evaluation schemes assume that predictions\nare available right after the input video is observed, i.e., presuming a\nnegligible runtime, which may lead to overly optimistic evaluations. We propose\na streaming egocentric action evaluation scheme which assumes that predictions\nare performed online and made available only after the model has processed the\ncurrent input segment, which depends on its runtime. To evaluate all models\nconsidering the same prediction horizon, we hence propose that slower models\nshould base their predictions on temporal segments sampled ahead of time. Based\non the observation that model runtime can affect performance in the considered\nstreaming evaluation scenario, we further propose a lightweight action\nanticipation model based on feed-forward 3D CNNs which is optimized using\nknowledge distillation techniques with a novel past-to-future distillation\nloss. Experiments on the three popular datasets EPIC-KITCHENS-55,\nEPIC-KITCHENS-100 and EGTEA Gaze+ show that (i) the proposed evaluation scheme\ninduces a different ranking on state-of-the-art methods as compared to classic\nevaluations, (ii) lightweight approaches tend to outmatch more computationally\nexpensive ones, and (iii) the proposed model based on feed-forward 3D CNNs and\nknowledge distillation outperforms current art in the streaming egocentric\naction anticipation scenario.\n", "date": "2023-06-29", "categories": []}
{"id": "2306.15765", "title": "A Novel Two Stream Decision Level Fusion of Vision and Inertial Sensors\n  Data for Automatic Multimodal Human Activity Recognition System", "abstract": "  This paper presents a novel multimodal human activity recognition system. It\nuses a two-stream decision level fusion of vision and inertial sensors. In the\nfirst stream, raw RGB frames are passed to a part affinity field-based pose\nestimation network to detect the keypoints of the user. These keypoints are\nthen pre-processed and inputted in a sliding window fashion to a specially\ndesigned convolutional neural network for the spatial feature extraction\nfollowed by regularized LSTMs to calculate the temporal features. The outputs\nof LSTM networks are then inputted to fully connected layers for\nclassification. In the second stream, data obtained from inertial sensors are\npre-processed and inputted to regularized LSTMs for the feature extraction\nfollowed by fully connected layers for the classification. At this stage, the\nSoftMax scores of two streams are then fused using the decision level fusion\nwhich gives the final prediction. Extensive experiments are conducted to\nevaluate the performance. Four multimodal standard benchmark datasets (UP-Fall\ndetection, UTD-MHAD, Berkeley-MHAD, and C-MHAD) are used for experimentations.\nThe accuracies obtained by the proposed system are 96.9 %, 97.6 %, 98.7 %, and\n95.9 % respectively on the UP-Fall Detection, UTDMHAD, Berkeley-MHAD, and\nC-MHAD datasets. These results are far superior than the current\nstate-of-the-art methods.\n", "date": "2023-06-27", "categories": []}
{"id": "2306.16357", "title": "Theater Aid System for the Visually Impaired Through Transfer Learning\n  of Spatio-Temporal Graph Convolution Networks", "abstract": "  The aim of this research is to recognize human actions performed on stage to\naid visually impaired and blind individuals. To achieve this, we have created a\ntheatre human action recognition system that uses skeleton data captured by\ndepth image as input. We collected new samples of human actions in a theatre\nenvironment, and then tested the transfer learning technique with three\npre-trained Spatio-Temporal Graph Convolution Networks for skeleton-based human\naction recognition: the spatio-temporal graph convolution network, the\ntwo-stream adaptive graph convolution network, and the multi-scale disentangled\nunified graph convolution network. We selected the NTU-RGBD human action\nbenchmark as the source domain and used our collected dataset as the target\ndomain. We analyzed the transferability of the pre-trained models and proposed\ntwo configurations to apply and adapt the transfer learning technique to the\ndiversity between the source and target domains. The use of transfer learning\nhelped to improve the performance of the human action system within the context\nof theatre. The results indicate that Spatio-Temporal Graph Convolution\nNetworks is positively transferred, and there was an improvement in performance\ncompared to the baseline without transfer learning.\n", "date": "2023-06-28", "categories": []}
{"id": "2306.15045", "title": "Action Anticipation with Goal Consistency", "abstract": "  In this paper, we address the problem of short-term action anticipation,\ni.e., we want to predict an upcoming action one second before it happens. We\npropose to harness high-level intent information to anticipate actions that\nwill take place in the future. To this end, we incorporate an additional goal\nprediction branch into our model and propose a consistency loss function that\nencourages the anticipated actions to conform to the high-level goal pursued in\nthe video. In our experiments, we show the effectiveness of the proposed\napproach and demonstrate that our method achieves state-of-the-art results on\ntwo large-scale datasets: Assembly101 and COIN.\n", "date": "2023-06-26", "categories": []}
{"id": "2306.15321", "title": "Multi-Dimensional Refinement Graph Convolutional Network with Robust\n  Decouple Loss for Fine-Grained Skeleton-Based Action Recognition", "abstract": "  Graph convolutional networks have been widely used in skeleton-based action\nrecognition. However, existing approaches are limited in fine-grained action\nrecognition due to the similarity of inter-class data. Moreover, the noisy data\nfrom pose extraction increases the challenge of fine-grained recognition. In\nthis work, we propose a flexible attention block called Channel-Variable\nSpatial-Temporal Attention (CVSTA) to enhance the discriminative power of\nspatial-temporal joints and obtain a more compact intra-class feature\ndistribution. Based on CVSTA, we construct a Multi-Dimensional Refinement Graph\nConvolutional Network (MDR-GCN), which can improve the discrimination among\nchannel-, joint- and frame-level features for fine-grained actions.\nFurthermore, we propose a Robust Decouple Loss (RDL), which significantly\nboosts the effect of the CVSTA and reduces the impact of noise. The proposed\nmethod combining MDR-GCN with RDL outperforms the known state-of-the-art\nskeleton-based approaches on fine-grained datasets, FineGym99 and FSD-10, and\nalso on the coarse dataset NTU-RGB+D X-view version.\n", "date": "2023-06-27", "categories": []}
{"id": "2305.15283", "title": "High Speed Human Action Recognition using a Photonic Reservoir Computer", "abstract": "  The recognition of human actions in videos is one of the most active research\nfields in computer vision. The canonical approach consists in a more or less\ncomplex preprocessing stages of the raw video data, followed by a relatively\nsimple classification algorithm. Here we address recognition of human actions\nusing the reservoir computing algorithm, which allows us to focus on the\nclassifier stage. We introduce a new training method for the reservoir\ncomputer, based on \"Timesteps Of Interest\", which combines in a simple way\nshort and long time scales. We study the performance of this algorithm using\nboth numerical simulations and a photonic implementation based on a single\nnon-linear node and a delay line on the well known KTH dataset. We solve the\ntask with high accuracy and speed, to the point of allowing for processing\nmultiple video streams in real time. The present work is thus an important step\ntowards developing efficient dedicated hardware for video processing.\n", "date": "2023-05-24", "categories": []}
{"id": "2306.13735", "title": "Combining Public Human Activity Recognition Datasets to Mitigate Labeled\n  Data Scarcity", "abstract": "  The use of supervised learning for Human Activity Recognition (HAR) on mobile\ndevices leads to strong classification performances. Such an approach, however,\nrequires large amounts of labeled data, both for the initial training of the\nmodels and for their customization on specific clients (whose data often differ\ngreatly from the training data). This is actually impractical to obtain due to\nthe costs, intrusiveness, and time-consuming nature of data annotation.\nMoreover, even with the help of a significant amount of labeled data, model\ndeployment on heterogeneous clients faces difficulties in generalizing well on\nunseen data. Other domains, like Computer Vision or Natural Language\nProcessing, have proposed the notion of pre-trained models, leveraging large\ncorpora, to reduce the need for annotated data and better manage heterogeneity.\nThis promising approach has not been implemented in the HAR domain so far\nbecause of the lack of public datasets of sufficient size. In this paper, we\npropose a novel strategy to combine publicly available datasets with the goal\nof learning a generalized HAR model that can be fine-tuned using a limited\namount of labeled data on an unseen target domain. Our experimental evaluation,\nwhich includes experimenting with different state-of-the-art neural network\narchitectures, shows that combining public datasets can significantly reduce\nthe number of labeled samples required to achieve satisfactory performance on\nan unseen target domain.\n", "date": "2023-06-23", "categories": []}
{"id": "2306.13285", "title": "Learning Scene Flow With Skeleton Guidance For 3D Action Recognition", "abstract": "  Among the existing modalities for 3D action recognition, 3D flow has been\npoorly examined, although conveying rich motion information cues for human\nactions. Presumably, its susceptibility to noise renders it intractable, thus\nchallenging the learning process within deep models. This work demonstrates the\nuse of 3D flow sequence by a deep spatiotemporal model and further proposes an\nincremental two-level spatial attention mechanism, guided from skeleton domain,\nfor emphasizing motion features close to the body joint areas and according to\ntheir informativeness. Towards this end, an extended deep skeleton model is\nalso introduced to learn the most discriminant action motion dynamics, so as to\nestimate an informativeness score for each joint. Subsequently, a late fusion\nscheme is adopted between the two models for learning the high level\ncross-modal correlations. Experimental results on the currently largest and\nmost challenging dataset NTU RGB+D, demonstrate the effectiveness of the\nproposed approach, achieving state-of-the-art results.\n", "date": "2023-06-23", "categories": []}
{"id": "2209.12221", "title": "Hand Hygiene Assessment via Joint Step Segmentation and Key Action\n  Scorer", "abstract": "  Hand hygiene is a standard six-step hand-washing action proposed by the World\nHealth Organization (WHO). However, there is no good way to supervise medical\nstaff to do hand hygiene, which brings the potential risk of disease spread.\nExisting action assessment works usually make an overall quality prediction on\nan entire video. However, the internal structures of hand hygiene action are\nimportant in hand hygiene assessment. Therefore, we propose a novel\nfine-grained learning framework to perform step segmentation and key action\nscorer in a joint manner for accurate hand hygiene assessment. Existing\ntemporal segmentation methods usually employ multi-stage convolutional network\nto improve the segmentation robustness, but easily lead to over-segmentation\ndue to the lack of the long-range dependence. To address this issue, we design\na multi-stage convolution-transformer network for step segmentation. Based on\nthe observation that each hand-washing step involves several key actions which\ndetermine the hand-washing quality, we design a set of key action scorers to\nevaluate the quality of key actions in each step. In addition, there lacks a\nunified dataset in hand hygiene assessment. Therefore, under the supervision of\nmedical staff, we contribute a video dataset that contains 300 video sequences\nwith fine-grained annotations. Extensive experiments on the dataset suggest\nthat our method well assesses hand hygiene videos and achieves outstanding\nperformance.\n", "date": "2022-09-25", "categories": []}
{"id": "2210.15804", "title": "Handwashing Action Detection System for an Autonomous Social Robot", "abstract": "  Young children are at an increased risk of contracting contagious diseases\nsuch as COVID-19 due to improper hand hygiene. An autonomous social agent that\nobserves children while handwashing and encourages good hand washing practices\ncould provide an opportunity for handwashing behavior to become a habit. In\nthis article, we present a human action recognition system, which is part of\nthe vision system of a social robot platform, to assist children in developing\na correct handwashing technique. A modified convolution neural network (CNN)\narchitecture with Channel Spatial Attention Bilinear Pooling (CSAB) frame, with\na VGG-16 architecture as the backbone is trained and validated on an augmented\ndataset. The modified architecture generalizes well with an accuracy of 90% for\nthe WHO-prescribed handwashing steps even in an unseen environment. Our\nfindings indicate that the approach can recognize even subtle hand movements in\nthe video and can be used for gesture detection and classification in social\nrobotics.\n", "date": "2022-10-27", "categories": []}
{"id": "2306.11046", "title": "FSAR: Federated Skeleton-based Action Recognition with Adaptive Topology\n  Structure and Knowledge Distillation", "abstract": "  Existing skeleton-based action recognition methods typically follow a\ncentralized learning paradigm, which can pose privacy concerns when exposing\nhuman-related videos. Federated Learning (FL) has attracted much attention due\nto its outstanding advantages in privacy-preserving. However, directly applying\nFL approaches to skeleton videos suffers from unstable training. In this paper,\nwe investigate and discover that the heterogeneous human topology graph\nstructure is the crucial factor hindering training stability. To address this\nlimitation, we pioneer a novel Federated Skeleton-based Action Recognition\n(FSAR) paradigm, which enables the construction of a globally generalized model\nwithout accessing local sensitive data. Specifically, we introduce an Adaptive\nTopology Structure (ATS), separating generalization and personalization by\nlearning a domain-invariant topology shared across clients and a\ndomain-specific topology decoupled from global model aggregation.Furthermore,\nwe explore Multi-grain Knowledge Distillation (MKD) to mitigate the discrepancy\nbetween clients and server caused by distinct updating patterns through\naligning shallow block-wise motion features. Extensive experiments on multiple\ndatasets demonstrate that FSAR outperforms state-of-the-art FL-based methods\nwhile inherently protecting privacy.\n", "date": "2023-06-19", "categories": []}
{"id": "2306.11287", "title": "Spatiotemporal Pyramidal CNN with Depth-Wise Separable Convolution for\n  Eye Blinking Detection in the Wild", "abstract": "  Eye blinking detection in the wild plays an essential role in deception\ndetection, driving fatigue detection, etc. Despite the fact that numerous\nattempts have already been made, the majority of them have encountered\ndifficulties, such as the derived eye images having different resolutions as\nthe distance between the face and the camera changes; or the requirement of a\nlightweight detection model to obtain a short inference time in order to\nperform in real-time. In this research, two problems are addressed: how the eye\nblinking detection model can learn efficiently from different resolutions of\neye pictures in diverse conditions; and how to reduce the size of the detection\nmodel for faster inference time. We propose to utilize upsampling and\ndownsampling the input eye images to the same resolution as one potential\nsolution for the first problem, then find out which interpolation method can\nresult in the highest performance of the detection model. For the second\nproblem, although a recent spatiotemporal convolutional neural network used for\neye blinking detection has a strong capacity to extract both spatial and\ntemporal characteristics, it remains having a high number of network\nparameters, leading to high inference time. Therefore, using Depth-wise\nSeparable Convolution rather than conventional convolution layers inside each\nbranch is considered in this paper as a feasible solution.\n", "date": "2023-06-20", "categories": []}
{"id": "2306.11726", "title": "How can objects help action recognition?", "abstract": "  Current state-of-the-art video models process a video clip as a long sequence\nof spatio-temporal tokens. However, they do not explicitly model objects, their\ninteractions across the video, and instead process all the tokens in the video.\nIn this paper, we investigate how we can use knowledge of objects to design\nbetter video models, namely to process fewer tokens and to improve recognition\naccuracy. This is in contrast to prior works which either drop tokens at the\ncost of accuracy, or increase accuracy whilst also increasing the computation\nrequired. First, we propose an object-guided token sampling strategy that\nenables us to retain a small fraction of the input tokens with minimal impact\non accuracy. And second, we propose an object-aware attention module that\nenriches our feature representation with object information and improves\noverall accuracy. Our resulting framework achieves better performance when\nusing fewer tokens than strong baselines. In particular, we match our baseline\nwith 30%, 40%, and 60% of the input tokens on SomethingElse,\nSomething-something v2, and Epic-Kitchens, respectively. When we use our model\nto process the same number of tokens as our baseline, we improve by 0.6 to 4.2\npoints on these datasets.\n", "date": "2023-06-20", "categories": []}
{"id": "2306.09331", "title": "Seeing the Pose in the Pixels: Learning Pose-Aware Representations in\n  Vision Transformers", "abstract": "  Human perception of surroundings is often guided by the various poses present\nwithin the environment. Many computer vision tasks, such as human action\nrecognition and robot imitation learning, rely on pose-based entities like\nhuman skeletons or robotic arms. However, conventional Vision Transformer (ViT)\nmodels uniformly process all patches, neglecting valuable pose priors in input\nvideos. We argue that incorporating poses into RGB data is advantageous for\nlearning fine-grained and viewpoint-agnostic representations. Consequently, we\nintroduce two strategies for learning pose-aware representations in ViTs. The\nfirst method, called Pose-aware Attention Block (PAAB), is a plug-and-play ViT\nblock that performs localized attention on pose regions within videos. The\nsecond method, dubbed Pose-Aware Auxiliary Task (PAAT), presents an auxiliary\npose prediction task optimized jointly with the primary ViT task. Although\ntheir functionalities differ, both methods succeed in learning pose-aware\nrepresentations, enhancing performance in multiple diverse downstream tasks.\nOur experiments, conducted across seven datasets, reveal the efficacy of both\npose-aware methods on three video analysis tasks, with PAAT holding a slight\nedge over PAAB. Both PAAT and PAAB surpass their respective backbone\nTransformers by up to 9.8% in real-world action recognition and 21.8% in\nmulti-view robotic video alignment. Code is available at\nhttps://github.com/dominickrei/PoseAwareVT.\n", "date": "2023-06-15", "categories": []}
{"id": "2211.09146", "title": "A Unified Multimodal De- and Re-coupling Framework for RGB-D Motion\n  Recognition", "abstract": "  Motion recognition is a promising direction in computer vision, but the\ntraining of video classification models is much harder than images due to\ninsufficient data and considerable parameters. To get around this, some works\nstrive to explore multimodal cues from RGB-D data. Although improving motion\nrecognition to some extent, these methods still face sub-optimal situations in\nthe following aspects: (i) Data augmentation, i.e., the scale of the RGB-D\ndatasets is still limited, and few efforts have been made to explore novel data\naugmentation strategies for videos; (ii) Optimization mechanism, i.e., the\ntightly space-time-entangled network structure brings more challenges to\nspatiotemporal information modeling; And (iii) cross-modal knowledge fusion,\ni.e., the high similarity between multimodal representations caused to\ninsufficient late fusion. To alleviate these drawbacks, we propose to improve\nRGB-D-based motion recognition both from data and algorithm perspectives in\nthis paper. In more detail, firstly, we introduce a novel video data\naugmentation method dubbed ShuffleMix, which acts as a supplement to MixUp, to\nprovide additional temporal regularization for motion recognition. Secondly, a\nUnified Multimodal De-coupling and multi-stage Re-coupling framework, termed\nUMDR, is proposed for video representation learning. Finally, a novel\ncross-modal Complement Feature Catcher (CFCer) is explored to mine potential\ncommonalities features in multimodal information as the auxiliary fusion\nstream, to improve the late fusion results. The seamless combination of these\nnovel designs forms a robust spatiotemporal representation and achieves better\nperformance than state-of-the-art methods on four public motion datasets.\nSpecifically, UMDR achieves unprecedented improvements of +4.5% on the Chalearn\nIsoGD dataset. Our code is available at\nhttps://github.com/zhoubenjia/MotionRGBD-PAMI.\n", "date": "2022-11-16", "categories": []}
{"id": "2301.08874", "title": "Improving Zero-Shot Action Recognition using Human Instruction with Text\n  Description", "abstract": "  Zero-shot action recognition, which recognizes actions in videos without\nhaving received any training examples, is gaining wide attention considering it\ncan save labor costs and training time. Nevertheless, the performance of\nzero-shot learning is still unsatisfactory, which limits its practical\napplication. To solve this problem, this study proposes a framework to improve\nzero-shot action recognition using human instructions with text descriptions.\nThe proposed framework manually describes video contents, which incurs some\nlabor costs; in many situations, the labor costs are worth it. We manually\nannotate text features for each action, which can be a word, phrase, or\nsentence. Then by computing the matching degrees between the video and all text\nfeatures, we can predict the class of the video. Furthermore, the proposed\nmodel can also be combined with other models to improve its accuracy. In\naddition, our model can be continuously optimized to improve the accuracy by\nrepeating human instructions. The results with UCF101 and HMDB51 showed that\nour model achieved the best accuracy and improved the accuracies of other\nmodels.\n", "date": "2023-01-21", "categories": []}
{"id": "2301.10900", "title": "Graph Contrastive Learning for Skeleton-based Action Recognition", "abstract": "  In the field of skeleton-based action recognition, current top-performing\ngraph convolutional networks (GCNs) exploit intra-sequence context to construct\nadaptive graphs for feature aggregation. However, we argue that such context is\nstill \\textit{local} since the rich cross-sequence relations have not been\nexplicitly investigated. In this paper, we propose a graph contrastive learning\nframework for skeleton-based action recognition (\\textit{SkeletonGCL}) to\nexplore the \\textit{global} context across all sequences. In specific,\nSkeletonGCL associates graph learning across sequences by enforcing graphs to\nbe class-discriminative, \\emph{i.e.,} intra-class compact and inter-class\ndispersed, which improves the GCN capacity to distinguish various action\npatterns. Besides, two memory banks are designed to enrich cross-sequence\ncontext from two complementary levels, \\emph{i.e.,} instance and semantic\nlevels, enabling graph contrastive learning in multiple context scales.\nConsequently, SkeletonGCL establishes a new training paradigm, and it can be\nseamlessly incorporated into current GCNs. Without loss of generality, we\ncombine SkeletonGCL with three GCNs (2S-ACGN, CTR-GCN, and InfoGCN), and\nachieve consistent improvements on NTU60, NTU120, and NW-UCLA benchmarks. The\nsource code will be available at\n\\url{https://github.com/OliverHxh/SkeletonGCL}.\n", "date": "2023-01-26", "categories": []}
{"id": "2306.05772", "title": "A Boosted Model Ensembling Approach to Ball Action Spotting in Videos:\n  The Runner-Up Solution to CVPR'23 SoccerNet Challenge", "abstract": "  This technical report presents our solution to Ball Action Spotting in\nvideos. Our method reached second place in the CVPR'23 SoccerNet Challenge.\nDetails of this challenge can be found at\nhttps://www.soccer-net.org/tasks/ball-action-spotting. Our approach is\ndeveloped based on a baseline model termed E2E-Spot, which was provided by the\norganizer of this competition. We first generated several variants of the\nE2E-Spot model, resulting in a candidate model set. We then proposed a strategy\nfor selecting appropriate model members from this set and assigning an\nappropriate weight to each model. The aim of this strategy is to boost the\nperformance of the resulting model ensemble. Therefore, we call our approach\nBoosted Model Ensembling (BME). Our code is available at\nhttps://github.com/ZJLAB-AMMI/E2E-Spot-MBS.\n", "date": "2023-06-09", "categories": []}
{"id": "2306.02140", "title": "Unsupervised Human Activity Recognition through Two-stage Prompting with\n  ChatGPT", "abstract": "  Wearable sensor devices, which offer the advantage of recording daily objects\nused by a person while performing an activity, enable the feasibility of\nunsupervised Human Activity Recognition (HAR). Unfortunately, previous\nunsupervised approaches using the usage sequence of objects usually require a\nproper description of activities manually prepared by humans. Instead, we\nleverage the knowledge embedded in a Large Language Model (LLM) of ChatGPT.\nBecause the sequence of objects robustly characterizes the activity identity,\nit is possible that ChatGPT already learned the association between activities\nand objects from existing contexts. However, previous prompt engineering for\nChatGPT exhibits limited generalization ability when dealing with a list of\nwords (i.e., sequence of objects) due to the similar weighting assigned to each\nword in the list. In this study, we propose a two-stage prompt engineering,\nwhich first guides ChatGPT to generate activity descriptions associated with\nobjects while emphasizing important objects for distinguishing similar\nactivities; then outputs activity classes and explanations for enhancing the\ncontexts that are helpful for HAR. To the best of our knowledge, this is the\nfirst study that utilizes ChatGPT to recognize activities using objects in an\nunsupervised manner. We conducted our approach on three datasets and\ndemonstrated the state-of-the-art performance.\n", "date": "2023-06-03", "categories": []}
{"id": "2306.05844", "title": "How Object Information Improves Skeleton-based Human Action Recognition\n  in Assembly Tasks", "abstract": "  As the use of collaborative robots (cobots) in industrial manufacturing\ncontinues to grow, human action recognition for effective human-robot\ncollaboration becomes increasingly important. This ability is crucial for\ncobots to act autonomously and assist in assembly tasks. Recently,\nskeleton-based approaches are often used as they tend to generalize better to\ndifferent people and environments. However, when processing skeletons alone,\ninformation about the objects a human interacts with is lost. Therefore, we\npresent a novel approach of integrating object information into skeleton-based\naction recognition. We enhance two state-of-the-art methods by treating object\ncenters as further skeleton joints. Our experiments on the assembly dataset\nIKEA ASM show that our approach improves the performance of these\nstate-of-the-art methods to a large extent when combining skeleton joints with\nobjects predicted by a state-of-the-art instance segmentation model. Our\nresearch sheds light on the benefits of combining skeleton joints with object\ninformation for human action recognition in assembly tasks. We analyze the\neffect of the object detector on the combination for action classification and\ndiscuss the important factors that must be taken into account.\n", "date": "2023-06-09", "categories": []}
{"id": "2306.06078", "title": "Cheating off your neighbors: Improving activity recognition through\n  corroboration", "abstract": "  Understanding the complexity of human activities solely through an\nindividual's data can be challenging. However, in many situations, surrounding\nindividuals are likely performing similar activities, while existing human\nactivity recognition approaches focus almost exclusively on individual\nmeasurements and largely ignore the context of the activity. Consider two\nactivities: attending a small group meeting and working at an office desk. From\nsolely an individual's perspective, it can be difficult to differentiate\nbetween these activities as they may appear very similar, even though they are\nmarkedly different. Yet, by observing others nearby, it can be possible to\ndistinguish between these activities. In this paper, we propose an approach to\nenhance the prediction accuracy of an individual's activities by incorporating\ninsights from surrounding individuals. We have collected a real-world dataset\nfrom 20 participants with over 58 hours of data including activities such as\nattending lectures, having meetings, working in the office, and eating\ntogether. Compared to observing a single person in isolation, our proposed\napproach significantly improves accuracy. We regard this work as a first step\nin collaborative activity recognition, opening new possibilities for\nunderstanding human activity in group settings.\n", "date": "2023-05-27", "categories": []}
{"id": "2302.06848", "title": "YOWOv2: A Stronger yet Efficient Multi-level Detection Framework for\n  Real-time Spatio-temporal Action Detection", "abstract": "  Designing a real-time framework for the spatio-temporal action detection task\nis still a challenge. In this paper, we propose a novel real-time action\ndetection framework, YOWOv2. In this new framework, YOWOv2 takes advantage of\nboth the 3D backbone and 2D backbone for accurate action detection. A\nmulti-level detection pipeline is designed to detect action instances of\ndifferent scales. To achieve this goal, we carefully build a simple and\nefficient 2D backbone with a feature pyramid network to extract different\nlevels of classification features and regression features. For the 3D backbone,\nwe adopt the existing efficient 3D CNN to save development time. By combining\n3D backbones and 2D backbones of different sizes, we design a YOWOv2 family\nincluding YOWOv2-Tiny, YOWOv2-Medium, and YOWOv2-Large. We also introduce the\npopular dynamic label assignment strategy and anchor-free mechanism to make the\nYOWOv2 consistent with the advanced model architecture design. With our\nimprovement, YOWOv2 is significantly superior to YOWO, and can still keep\nreal-time detection. Without any bells and whistles, YOWOv2 achieves 87.0 %\nframe mAP and 52.8 % video mAP with over 20 FPS on the UCF101-24. On the AVA,\nYOWOv2 achieves 21.7 % frame mAP with over 20 FPS. Our code is available on\nhttps://github.com/yjh0410/YOWOv2.\n", "date": "2023-02-14", "categories": []}
{"id": "2306.04822", "title": "Optimizing ViViT Training: Time and Memory Reduction for Action\n  Recognition", "abstract": "  In this paper, we address the challenges posed by the substantial training\ntime and memory consumption associated with video transformers, focusing on the\nViViT (Video Vision Transformer) model, in particular the Factorised Encoder\nversion, as our baseline for action recognition tasks. The factorised encoder\nvariant follows the late-fusion approach that is adopted by many state of the\nart approaches. Despite standing out for its favorable speed/accuracy tradeoffs\namong the different variants of ViViT, its considerable training time and\nmemory requirements still pose a significant barrier to entry. Our method is\ndesigned to lower this barrier and is based on the idea of freezing the spatial\ntransformer during training. This leads to a low accuracy model if naively\ndone. But we show that by (1) appropriately initializing the temporal\ntransformer (a module responsible for processing temporal information) (2)\nintroducing a compact adapter model connecting frozen spatial representations\n((a module that selectively focuses on regions of the input image) to the\ntemporal transformer, we can enjoy the benefits of freezing the spatial\ntransformer without sacrificing accuracy. Through extensive experimentation\nover 6 benchmarks, we demonstrate that our proposed training strategy\nsignificantly reduces training costs (by $\\sim 50\\%$) and memory consumption\nwhile maintaining or slightly improving performance by up to 1.79\\% compared to\nthe baseline model. Our approach additionally unlocks the capability to utilize\nlarger image transformer models as our spatial transformer and access more\nframes with the same memory consumption.\n", "date": "2023-06-07", "categories": []}
{"id": "2306.05147", "title": "Human Action Recognition in Egocentric Perspective Using 2D Object and\n  Hands Pose", "abstract": "  Egocentric action recognition is essential for healthcare and assistive\ntechnology that relies on egocentric cameras because it allows for the\nautomatic and continuous monitoring of activities of daily living (ADLs)\nwithout requiring any conscious effort from the user. This study explores the\nfeasibility of using 2D hand and object pose information for egocentric action\nrecognition. While current literature focuses on 3D hand pose information, our\nwork shows that using 2D skeleton data is a promising approach for hand-based\naction classification, might offer privacy enhancement, and could be less\ncomputationally demanding. The study uses a state-of-the-art transformer-based\nmethod to classify sequences and achieves validation results of 94%,\noutperforming other existing solutions. The accuracy of the test subset drops\nto 76%, indicating the need for further generalization improvement. This\nresearch highlights the potential of 2D hand and object pose information for\naction recognition tasks and offers a promising alternative to 3D-based\nmethods.\n", "date": "2023-06-08", "categories": []}
{"id": "2212.00724", "title": "SWL-Adapt: An Unsupervised Domain Adaptation Model with Sample Weight\n  Learning for Cross-User Wearable Human Activity Recognition", "abstract": "  In practice, Wearable Human Activity Recognition (WHAR) models usually face\nperformance degradation on the new user due to user variance. Unsupervised\ndomain adaptation (UDA) becomes the natural solution to cross-user WHAR under\nannotation scarcity. Existing UDA models usually align samples across domains\nwithout differentiation, which ignores the difference among samples. In this\npaper, we propose an unsupervised domain adaptation model with sample weight\nlearning (SWL-Adapt) for cross-user WHAR. SWL-Adapt calculates sample weights\naccording to the classification loss and domain discrimination loss of each\nsample with a parameterized network. We introduce the meta-optimization based\nupdate rule to learn this network end-to-end, which is guided by\nmeta-classification loss on the selected pseudo-labeled target samples.\nTherefore, this network can fit a weighting function according to the\ncross-user WHAR task at hand, which is superior to existing sample\ndifferentiation rules fixed for special scenarios. Extensive experiments on\nthree public WHAR datasets demonstrate that SWL-Adapt achieves the\nstate-of-the-art performance on the cross-user WHAR task, outperforming the\nbest baseline by an average of 3.1% and 5.3% in accuracy and macro F1 score,\nrespectively.\n", "date": "2022-11-25", "categories": []}
{"id": "2306.01075", "title": "Pedestrian Crossing Action Recognition and Trajectory Prediction with 3D\n  Human Keypoints", "abstract": "  Accurate understanding and prediction of human behaviors are critical\nprerequisites for autonomous vehicles, especially in highly dynamic and\ninteractive scenarios such as intersections in dense urban areas. In this work,\nwe aim at identifying crossing pedestrians and predicting their future\ntrajectories. To achieve these goals, we not only need the context information\nof road geometry and other traffic participants but also need fine-grained\ninformation of the human pose, motion and activity, which can be inferred from\nhuman keypoints. In this paper, we propose a novel multi-task learning\nframework for pedestrian crossing action recognition and trajectory prediction,\nwhich utilizes 3D human keypoints extracted from raw sensor data to capture\nrich information on human pose and activity. Moreover, we propose to apply two\nauxiliary tasks and contrastive learning to enable auxiliary supervisions to\nimprove the learned keypoints representation, which further enhances the\nperformance of major tasks. We validate our approach on a large-scale in-house\ndataset, as well as a public benchmark dataset, and show that our approach\nachieves state-of-the-art performance on a wide range of evaluation metrics.\nThe effectiveness of each model component is validated in a detailed ablation\nstudy.\n", "date": "2023-06-01", "categories": []}
{"id": "2302.00624", "title": "Open-VCLIP: Transforming CLIP to an Open-vocabulary Video Model via\n  Interpolated Weight Optimization", "abstract": "  Contrastive Language-Image Pretraining (CLIP) has demonstrated impressive\nzero-shot learning abilities for image understanding, yet limited effort has\nbeen made to investigate CLIP for zero-shot video recognition. We introduce\nOpen-VCLIP, a simple yet effective approach that transforms CLIP into a strong\nzero-shot video classifier that can recognize unseen actions and events at test\ntime. Our framework extends CLIP with minimal modifications to model\nspatial-temporal relationships in videos, making it a specialized video\nclassifier, while striving for generalization. We formally show that training\nan Open-VCLIP is equivalent to continual learning with zero historical data. To\naddress this problem, we propose Interpolated Weight Optimization, which\nutilizes the benefit of weight interpolation in both training and test time. We\nevaluate our method on three popular and challenging action recognition\ndatasets following various zero-shot evaluation protocols and we demonstrate\nour approach outperforms state-of-the-art methods by clear margins. In\nparticular, we achieve 87.9%, 58.3%, 81.1% zero-shot accuracy on UCF, HMDB and\nKinetics-600 respectively, outperforming state-of-the-art methods by 8.3%, 7.8%\nand 12.2%. Code is released at https://github.com/wengzejia1/Open-VCLIP.\n", "date": "2023-02-01", "categories": []}
{"id": "2305.19624", "title": "A Multi-Modal Transformer Network for Action Detection", "abstract": "  This paper proposes a novel multi-modal transformer network for detecting\nactions in untrimmed videos. To enrich the action features, our transformer\nnetwork utilizes a new multi-modal attention mechanism that computes the\ncorrelations between different spatial and motion modalities combinations.\nExploring such correlations for actions has not been attempted previously. To\nuse the motion and spatial modality more effectively, we suggest an algorithm\nthat corrects the motion distortion caused by camera movement. Such motion\ndistortion, common in untrimmed videos, severely reduces the expressive power\nof motion features such as optical flow fields. Our proposed algorithm\noutperforms the state-of-the-art methods on two public benchmarks, THUMOS14 and\nActivityNet. We also conducted comparative experiments on our new instructional\nactivity dataset, including a large set of challenging classroom videos\ncaptured from elementary schools.\n", "date": "2023-05-31", "categories": []}
{"id": "2301.13090", "title": "Action Capsules: Human Skeleton Action Recognition", "abstract": "  Due to the compact and rich high-level representations offered,\nskeleton-based human action recognition has recently become a highly active\nresearch topic. Previous studies have demonstrated that investigating joint\nrelationships in spatial and temporal dimensions provides effective information\ncritical to action recognition. However, effectively encoding global\ndependencies of joints during spatio-temporal feature extraction is still\nchallenging. In this paper, we introduce Action Capsule which identifies\naction-related key joints by considering the latent correlation of joints in a\nskeleton sequence. We show that, during inference, our end-to-end network pays\nattention to a set of joints specific to each action, whose encoded\nspatio-temporal features are aggregated to recognize the action. Additionally,\nthe use of multiple stages of action capsules enhances the ability of the\nnetwork to classify similar actions. Consequently, our network outperforms the\nstate-of-the-art approaches on the N-UCLA dataset and obtains competitive\nresults on the NTURGBD dataset. This is while our approach has significantly\nlower computational requirements based on GFLOPs measurements.\n", "date": "2023-01-30", "categories": []}
{"id": "2305.17861", "title": "Proposal-Based Multiple Instance Learning for Weakly-Supervised Temporal\n  Action Localization", "abstract": "  Weakly-supervised temporal action localization aims to localize and recognize\nactions in untrimmed videos with only video-level category labels during\ntraining. Without instance-level annotations, most existing methods follow the\nSegment-based Multiple Instance Learning (S-MIL) framework, where the\npredictions of segments are supervised by the labels of videos. However, the\nobjective for acquiring segment-level scores during training is not consistent\nwith the target for acquiring proposal-level scores during testing, leading to\nsuboptimal results. To deal with this problem, we propose a novel\nProposal-based Multiple Instance Learning (P-MIL) framework that directly\nclassifies the candidate proposals in both the training and testing stages,\nwhich includes three key designs: 1) a surrounding contrastive feature\nextraction module to suppress the discriminative short proposals by considering\nthe surrounding contrastive information, 2) a proposal completeness evaluation\nmodule to inhibit the low-quality proposals with the guidance of the\ncompleteness pseudo labels, and 3) an instance-level rank consistency loss to\nachieve robust detection by leveraging the complementarity of RGB and FLOW\nmodalities. Extensive experimental results on two challenging benchmarks\nincluding THUMOS14 and ActivityNet demonstrate the superior performance of our\nmethod.\n", "date": "2023-05-29", "categories": []}
{"id": "2208.12448", "title": "CMD: Self-supervised 3D Action Representation Learning with Cross-modal\n  Mutual Distillation", "abstract": "  In 3D action recognition, there exists rich complementary information between\nskeleton modalities. Nevertheless, how to model and utilize this information\nremains a challenging problem for self-supervised 3D action representation\nlearning. In this work, we formulate the cross-modal interaction as a\nbidirectional knowledge distillation problem. Different from classic\ndistillation solutions that transfer the knowledge of a fixed and pre-trained\nteacher to the student, in this work, the knowledge is continuously updated and\nbidirectionally distilled between modalities. To this end, we propose a new\nCross-modal Mutual Distillation (CMD) framework with the following designs. On\nthe one hand, the neighboring similarity distribution is introduced to model\nthe knowledge learned in each modality, where the relational information is\nnaturally suitable for the contrastive frameworks. On the other hand,\nasymmetrical configurations are used for teacher and student to stabilize the\ndistillation process and to transfer high-confidence information between\nmodalities. By derivation, we find that the cross-modal positive mining in\nprevious works can be regarded as a degenerated version of our CMD. We perform\nextensive experiments on NTU RGB+D 60, NTU RGB+D 120, and PKU-MMD II datasets.\nOur approach outperforms existing self-supervised methods and sets a series of\nnew records. The code is available at: https://github.com/maoyunyao/CMD\n", "date": "2022-08-26", "categories": []}
{"id": "2305.15692", "title": "Deep Neural Networks in Video Human Action Recognition: A Review", "abstract": "  Currently, video behavior recognition is one of the most foundational tasks\nof computer vision. The 2D neural networks of deep learning are built for\nrecognizing pixel-level information such as images with RGB, RGB-D, or optical\nflow formats, with the current increasingly wide usage of surveillance video\nand more tasks related to human action recognition. There are increasing tasks\nrequiring temporal information for frames dependency analysis. The researchers\nhave widely studied video-based recognition rather than\nimage-based(pixel-based) only to extract more informative elements from\ngeometry tasks. Our current related research addresses multiple novel proposed\nresearch works and compares their advantages and disadvantages between the\nderived deep learning frameworks rather than machine learning frameworks. The\ncomparison happened between existing frameworks and datasets, which are video\nformat data only. Due to the specific properties of human actions and the\nincreasingly wide usage of deep neural networks, we collected all research\nworks within the last three years between 2020 to 2022. In our article, the\nperformance of deep neural networks surpassed most of the techniques in the\nfeature learning and extraction tasks, especially video action recognition.\n", "date": "2023-05-25", "categories": []}
{"id": "2305.11365", "title": "Enhancing Transformer Backbone for Egocentric Video Action Segmentation", "abstract": "  Egocentric temporal action segmentation in videos is a crucial task in\ncomputer vision with applications in various fields such as mixed reality,\nhuman behavior analysis, and robotics. Although recent research has utilized\nadvanced visual-language frameworks, transformers remain the backbone of action\nsegmentation models. Therefore, it is necessary to improve transformers to\nenhance the robustness of action segmentation models. In this work, we propose\ntwo novel ideas to enhance the state-of-the-art transformer for action\nsegmentation. First, we introduce a dual dilated attention mechanism to\nadaptively capture hierarchical representations in both local-to-global and\nglobal-to-local contexts. Second, we incorporate cross-connections between the\nencoder and decoder blocks to prevent the loss of local context by the decoder.\nWe also utilize state-of-the-art visual-language representation learning\ntechniques to extract richer and more compact features for our transformer. Our\nproposed approach outperforms other state-of-the-art methods on the Georgia\nTech Egocentric Activities (GTEA) and HOI4D Office Tools datasets, and we\nvalidate our introduced components with ablation studies. The source code and\nsupplementary materials are publicly available on\nhttps://www.sail-nu.com/dxformer.\n", "date": "2023-05-19", "categories": []}
{"id": "2303.15086", "title": "Learning Action Changes by Measuring Verb-Adverb Textual Relationships", "abstract": "  The goal of this work is to understand the way actions are performed in\nvideos. That is, given a video, we aim to predict an adverb indicating a\nmodification applied to the action (e.g. cut \"finely\"). We cast this problem as\na regression task. We measure textual relationships between verbs and adverbs\nto generate a regression target representing the action change we aim to learn.\nWe test our approach on a range of datasets and achieve state-of-the-art\nresults on both adverb prediction and antonym classification. Furthermore, we\noutperform previous work when we lift two commonly assumed conditions: the\navailability of action labels during testing and the pairing of adverbs as\nantonyms. Existing datasets for adverb recognition are either noisy, which\nmakes learning difficult, or contain actions whose appearance is not influenced\nby adverbs, which makes evaluation less reliable. To address this, we collect a\nnew high quality dataset: Adverbs in Recipes (AIR). We focus on instructional\nrecipes videos, curating a set of actions that exhibit meaningful visual\nchanges when performed differently. Videos in AIR are more tightly trimmed and\nwere manually reviewed by multiple annotators to ensure high labelling quality.\nResults show that models learn better from AIR given its cleaner videos. At the\nsame time, adverb prediction on AIR is challenging, demonstrating that there is\nconsiderable room for improvement.\n", "date": "2023-03-27", "categories": []}
{"id": "2206.03891", "title": "PrivHAR: Recognizing Human Actions From Privacy-preserving Lens", "abstract": "  The accelerated use of digital cameras prompts an increasing concern about\nprivacy and security, particularly in applications such as action recognition.\nIn this paper, we propose an optimizing framework to provide robust visual\nprivacy protection along the human action recognition pipeline. Our framework\nparameterizes the camera lens to successfully degrade the quality of the videos\nto inhibit privacy attributes and protect against adversarial attacks while\nmaintaining relevant features for activity recognition. We validate our\napproach with extensive simulations and hardware experiments.\n", "date": "2022-06-08", "categories": []}
{"id": "2301.01380", "title": "Ego-Only: Egocentric Action Detection without Exocentric Transferring", "abstract": "  We present Ego-Only, the first approach that enables state-of-the-art action\ndetection on egocentric (first-person) videos without any form of exocentric\n(third-person) transferring. Despite the content and appearance gap separating\nthe two domains, large-scale exocentric transferring has been the default\nchoice for egocentric action detection. This is because prior works found that\negocentric models are difficult to train from scratch and that transferring\nfrom exocentric representations leads to improved accuracy. However, in this\npaper, we revisit this common belief. Motivated by the large gap separating the\ntwo domains, we propose a strategy that enables effective training of\negocentric models without exocentric transferring. Our Ego-Only approach is\nsimple. It trains the video representation with a masked autoencoder finetuned\nfor temporal segmentation. The learned features are then fed to an\noff-the-shelf temporal action localization method to detect actions. We find\nthat this renders exocentric transferring unnecessary by showing remarkably\nstrong results achieved by this simple Ego-Only approach on three established\negocentric video datasets: Ego4D, EPIC-Kitchens-100, and Charades-Ego. On both\naction detection and action recognition, Ego-Only outperforms previous best\nexocentric transferring methods that use orders of magnitude more labels.\nEgo-Only sets new state-of-the-art results on these datasets and benchmarks\nwithout exocentric data.\n", "date": "2023-01-03", "categories": []}
{"id": "2305.12398", "title": "Language Knowledge-Assisted Representation Learning for Skeleton-Based\n  Action Recognition", "abstract": "  How humans understand and recognize the actions of others is a complex\nneuroscientific problem that involves a combination of cognitive mechanisms and\nneural networks. Research has shown that humans have brain areas that recognize\nactions that process top-down attentional information, such as the\ntemporoparietal association area. Also, humans have brain regions dedicated to\nunderstanding the minds of others and analyzing their intentions, such as the\nmedial prefrontal cortex of the temporal lobe. Skeleton-based action\nrecognition creates mappings for the complex connections between the human\nskeleton movement patterns and behaviors. Although existing studies encoded\nmeaningful node relationships and synthesized action representations for\nclassification with good results, few of them considered incorporating a priori\nknowledge to aid potential representation learning for better performance.\nLA-GCN proposes a graph convolution network using large-scale language models\n(LLM) knowledge assistance. First, the LLM knowledge is mapped into a priori\nglobal relationship (GPR) topology and a priori category relationship (CPR)\ntopology between nodes. The GPR guides the generation of new \"bone\"\nrepresentations, aiming to emphasize essential node information from the data\nlevel. The CPR mapping simulates category prior knowledge in human brain\nregions, encoded by the PC-AC module and used to add additional\nsupervision-forcing the model to learn class-distinguishable features. In\naddition, to improve information transfer efficiency in topology modeling, we\npropose multi-hop attention graph convolution. It aggregates each node's\nk-order neighbor simultaneously to speed up model convergence. LA-GCN reaches\nstate-of-the-art on NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets.\n", "date": "2023-05-21", "categories": []}
{"id": "2301.05575", "title": "Deep learning-based approaches for human motion decoding in smart\n  walkers for rehabilitation", "abstract": "  Gait disabilities are among the most frequent worldwide. Their treatment\nrelies on rehabilitation therapies, in which smart walkers are being introduced\nto empower the user's recovery and autonomy, while reducing the clinicians\neffort. For that, these should be able to decode human motion and needs, as\nearly as possible. Current walkers decode motion intention using information of\nwearable or embedded sensors, namely inertial units, force and hall sensors,\nand lasers, whose main limitations imply an expensive solution or hinder the\nperception of human movement. Smart walkers commonly lack a seamless\nhuman-robot interaction, which intuitively understands human motions. A\ncontactless approach is proposed in this work, addressing human motion decoding\nas an early action recognition/detection problematic, using RGB-D cameras. We\nstudied different deep learning-based algorithms, organised in three different\napproaches, to process lower body RGB-D video sequences, recorded from an\nembedded camera of a smart walker, and classify them into 4 classes (stop,\nwalk, turn right/left). A custom dataset involving 15 healthy participants\nwalking with the device was acquired and prepared, resulting in 28800 balanced\nRGB-D frames, to train and evaluate the deep networks. The best results were\nattained by a convolutional neural network with a channel attention mechanism,\nreaching accuracy values of 99.61% and above 93%, for offline early\ndetection/recognition and trial simulations, respectively. Following the\nhypothesis that human lower body features encode prominent information,\nfostering a more robust prediction towards real-time applications, the\nalgorithm focus was also evaluated using Dice metric, leading to values\nslightly higher than 30%. Promising results were attained for early action\ndetection as a human motion decoding strategy, with enhancements in the focus\nof the proposed architectures.\n", "date": "2023-01-13", "categories": []}
{"id": "2305.08877", "title": "M$^2$DAR: Multi-View Multi-Scale Driver Action Recognition with Vision\n  Transformer", "abstract": "  Ensuring traffic safety and preventing accidents is a critical goal in daily\ndriving, where the advancement of computer vision technologies can be leveraged\nto achieve this goal. In this paper, we present a multi-view, multi-scale\nframework for naturalistic driving action recognition and localization in\nuntrimmed videos, namely M$^2$DAR, with a particular focus on detecting\ndistracted driving behaviors. Our system features a weight-sharing, multi-scale\nTransformer-based action recognition network that learns robust hierarchical\nrepresentations. Furthermore, we propose a new election algorithm consisting of\naggregation, filtering, merging, and selection processes to refine the\npreliminary results from the action recognition module across multiple views.\nExtensive experiments conducted on the 7th AI City Challenge Track 3 dataset\ndemonstrate the effectiveness of our approach, where we achieved an overlap\nscore of 0.5921 on the A2 test set. Our source code is available at\n\\url{https://github.com/PurdueDigitalTwin/M2DAR}.\n", "date": "2023-05-13", "categories": []}
{"id": "2305.09539", "title": "Learning Higher-order Object Interactions for Keypoint-based Video\n  Understanding", "abstract": "  Action recognition is an important problem that requires identifying actions\nin video by learning complex interactions across scene actors and objects.\nHowever, modern deep-learning based networks often require significant\ncomputation, and may capture scene context using various modalities that\nfurther increases compute costs. Efficient methods such as those used for AR/VR\noften only use human-keypoint information but suffer from a loss of scene\ncontext that hurts accuracy. In this paper, we describe an action-localization\nmethod, KeyNet, that uses only the keypoint data for tracking and action\nrecognition. Specifically, KeyNet introduces the use of object based keypoint\ninformation to capture context in the scene. Our method illustrates how to\nbuild a structured intermediate representation that allows modeling\nhigher-order interactions in the scene from object and human keypoints without\nusing any RGB information. We find that KeyNet is able to track and classify\nhuman actions at just 5 FPS. More importantly, we demonstrate that object\nkeypoints can be modeled to recover any loss in context from using keypoint\ninformation over AVA action and Kinetics datasets.\n", "date": "2023-05-16", "categories": []}
{"id": "2305.08191", "title": "Is end-to-end learning enough for fitness activity recognition?", "abstract": "  End-to-end learning has taken hold of many computer vision tasks, in\nparticular, related to still images, with task-specific optimization yielding\nvery strong performance. Nevertheless, human-centric action recognition is\nstill largely dominated by hand-crafted pipelines, and only individual\ncomponents are replaced by neural networks that typically operate on individual\nframes. As a testbed to study the relevance of such pipelines, we present a new\nfully annotated video dataset of fitness activities. Any recognition\ncapabilities in this domain are almost exclusively a function of human poses\nand their temporal dynamics, so pose-based solutions should perform well. We\nshow that, with this labelled data, end-to-end learning on raw pixels can\ncompete with state-of-the-art action recognition pipelines based on pose\nestimation. We also show that end-to-end learning can support temporally\nfine-grained tasks such as real-time repetition counting.\n", "date": "2023-05-14", "categories": []}
{"id": "2305.07214", "title": "MMG-Ego4D: Multi-Modal Generalization in Egocentric Action Recognition", "abstract": "  In this paper, we study a novel problem in egocentric action recognition,\nwhich we term as \"Multimodal Generalization\" (MMG). MMG aims to study how\nsystems can generalize when data from certain modalities is limited or even\ncompletely missing. We thoroughly investigate MMG in the context of standard\nsupervised action recognition and the more challenging few-shot setting for\nlearning new action categories. MMG consists of two novel scenarios, designed\nto support security, and efficiency considerations in real-world applications:\n(1) missing modality generalization where some modalities that were present\nduring the train time are missing during the inference time, and (2)\ncross-modal zero-shot generalization, where the modalities present during the\ninference time and the training time are disjoint. To enable this\ninvestigation, we construct a new dataset MMG-Ego4D containing data points with\nvideo, audio, and inertial motion sensor (IMU) modalities. Our dataset is\nderived from Ego4D dataset, but processed and thoroughly re-annotated by human\nexperts to facilitate research in the MMG problem. We evaluate a diverse array\nof models on MMG-Ego4D and propose new methods with improved generalization\nability. In particular, we introduce a new fusion module with modality dropout\ntraining, contrastive-based alignment training, and a novel cross-modal\nprototypical loss for better few-shot performance. We hope this study will\nserve as a benchmark and guide future research in multimodal generalization\nproblems. The benchmark and code will be available at\nhttps://github.com/facebookresearch/MMG_Ego4D.\n", "date": "2023-05-12", "categories": []}
{"id": "2209.02399", "title": "SkeletonMAE: Spatial-Temporal Masked Autoencoders for Self-supervised\n  Skeleton Action Recognition", "abstract": "  Fully supervised skeleton-based action recognition has achieved great\nprogress with the blooming of deep learning techniques. However, these methods\nrequire sufficient labeled data which is not easy to obtain. In contrast,\nself-supervised skeleton-based action recognition has attracted more attention.\nWith utilizing the unlabeled data, more generalizable features can be learned\nto alleviate the overfitting problem and reduce the demand of massive labeled\ntraining data. Inspired by the MAE, we propose a spatial-temporal masked\nautoencoder framework for self-supervised 3D skeleton-based action recognition\n(SkeletonMAE). Following MAE's masking and reconstruction pipeline, we utilize\na skeleton-based encoder-decoder transformer architecture to reconstruct the\nmasked skeleton sequences. A novel masking strategy, named Spatial-Temporal\nMasking, is introduced in terms of both joint-level and frame-level for the\nskeleton sequence. This pre-training strategy makes the encoder output\ngeneralizable skeleton features with spatial and temporal dependencies. Given\nthe unmasked skeleton sequence, the encoder is fine-tuned for the action\nrecognition task. Extensive experiments show that our SkeletonMAE achieves\nremarkable performance and outperforms the state-of-the-art methods on both NTU\nRGB+D and NTU RGB+D 120 datasets.\n", "date": "2022-09-01", "categories": []}
{"id": "2305.00666", "title": "Part Aware Contrastive Learning for Self-Supervised Action Recognition", "abstract": "  In recent years, remarkable results have been achieved in self-supervised\naction recognition using skeleton sequences with contrastive learning. It has\nbeen observed that the semantic distinction of human action features is often\nrepresented by local body parts, such as legs or hands, which are advantageous\nfor skeleton-based action recognition. This paper proposes an attention-based\ncontrastive learning framework for skeleton representation learning, called\nSkeAttnCLR, which integrates local similarity and global features for\nskeleton-based action representations. To achieve this, a multi-head attention\nmask module is employed to learn the soft attention mask features from the\nskeletons, suppressing non-salient local features while accentuating local\nsalient features, thereby bringing similar local features closer in the feature\nspace. Additionally, ample contrastive pairs are generated by expanding\ncontrastive pairs based on salient and non-salient features with global\nfeatures, which guide the network to learn the semantic representations of the\nentire skeleton. Therefore, with the attention mask mechanism, SkeAttnCLR\nlearns local features under different data augmentation views. The experiment\nresults demonstrate that the inclusion of local feature similarity\nsignificantly enhances skeleton-based action representation. Our proposed\nSkeAttnCLR outperforms state-of-the-art methods on NTURGB+D, NTU120-RGB+D, and\nPKU-MMD datasets.\n", "date": "2023-05-01", "categories": []}
{"id": "2305.06437", "title": "Self-Supervised Video Representation Learning via Latent Time Navigation", "abstract": "  Self-supervised video representation learning aimed at maximizing similarity\nbetween different temporal segments of one video, in order to enforce feature\npersistence over time. This leads to loss of pertinent information related to\ntemporal relationships, rendering actions such as `enter' and `leave' to be\nindistinguishable. To mitigate this limitation, we propose Latent Time\nNavigation (LTN), a time-parameterized contrastive learning strategy that is\nstreamlined to capture fine-grained motions. Specifically, we maximize the\nrepresentation similarity between different video segments from one video,\nwhile maintaining their representations time-aware along a subspace of the\nlatent representation code including an orthogonal basis to represent temporal\nchanges. Our extensive experimental analysis suggests that learning video\nrepresentations by LTN consistently improves performance of action\nclassification in fine-grained and human-oriented tasks (e.g., on Toyota\nSmarthome dataset). In addition, we demonstrate that our proposed model, when\npre-trained on Kinetics-400, generalizes well onto the unseen real world video\nbenchmark datasets UCF101 and HMDB51, achieving state-of-the-art performance in\naction recognition.\n", "date": "2023-05-10", "categories": []}
{"id": "2305.06114", "title": "Few-shot Action Recognition via Intra- and Inter-Video Information\n  Maximization", "abstract": "  Current few-shot action recognition involves two primary sources of\ninformation for classification:(1) intra-video information, determined by frame\ncontent within a single video clip, and (2) inter-video information, measured\nby relationships (e.g., feature similarity) among videos. However, existing\nmethods inadequately exploit these two information sources. In terms of\nintra-video information, current sampling operations for input videos may omit\ncritical action information, reducing the utilization efficiency of video data.\nFor the inter-video information, the action misalignment among videos makes it\nchallenging to calculate precise relationships. Moreover, how to jointly\nconsider both inter- and intra-video information remains under-explored for\nfew-shot action recognition. To this end, we propose a novel framework, Video\nInformation Maximization (VIM), for few-shot video action recognition. VIM is\nequipped with an adaptive spatial-temporal video sampler and a spatiotemporal\naction alignment model to maximize intra- and inter-video information,\nrespectively. The video sampler adaptively selects important frames and\namplifies critical spatial regions for each input video based on the task at\nhand. This preserves and emphasizes informative parts of video clips while\neliminating interference at the data level. The alignment model performs\ntemporal and spatial action alignment sequentially at the feature level,\nleading to more precise measurements of inter-video similarity. Finally, These\ngoals are facilitated by incorporating additional loss terms based on mutual\ninformation measurement. Consequently, VIM acts to maximize the distinctiveness\nof video information from limited video data. Extensive experimental results on\npublic datasets for few-shot action recognition demonstrate the effectiveness\nand benefits of our framework.\n", "date": "2023-05-10", "categories": []}
{"id": "2305.05583", "title": "Group Activity Recognition via Dynamic Composition and Interaction", "abstract": "  Previous group activity recognition approaches were limited to reasoning\nusing human relations or finding important subgroups and tended to ignore\nindispensable group composition and human-object interactions. This absence\nmakes a partial interpretation of the scene and increases the interference of\nirrelevant actions on the results. Therefore, we propose our DynamicFormer with\nDynamic composition Module (DcM) and Dynamic interaction Module (DiM) to model\nrelations and locations of persons and discriminate the contribution of\nparticipants, respectively. Our findings on group composition and human-object\ninteraction inspire our core idea. Group composition tells us the location of\npeople and their relations inside the group, while interaction reflects the\nrelation between humans and objects outside the group. We utilize spatial and\ntemporal encoders in DcM to model our dynamic composition and build DiM to\nexplore interaction with a novel GCN, which has a transformer inside to\nconsider the temporal neighbors of human/object. Also, a Multi-level Dynamic\nIntegration is employed to integrate features from different levels. We conduct\nextensive experiments on two public datasets and show that our method achieves\nstate-of-the-art.\n", "date": "2023-05-09", "categories": []}
{"id": "2305.03810", "title": "Distilled Mid-Fusion Transformer Networks for Multi-Modal Human Activity\n  Recognition", "abstract": "  Human Activity Recognition is an important task in many human-computer\ncollaborative scenarios, whilst having various practical applications. Although\nuni-modal approaches have been extensively studied, they suffer from data\nquality and require modality-specific feature engineering, thus not being\nrobust and effective enough for real-world deployment. By utilizing various\nsensors, Multi-modal Human Activity Recognition could utilize the complementary\ninformation to build models that can generalize well. While deep learning\nmethods have shown promising results, their potential in extracting salient\nmulti-modal spatial-temporal features and better fusing complementary\ninformation has not been fully explored. Also, reducing the complexity of the\nmulti-modal approach for edge deployment is another problem yet to resolve. To\nresolve the issues, a knowledge distillation-based Multi-modal Mid-Fusion\napproach, DMFT, is proposed to conduct informative feature extraction and\nfusion to resolve the Multi-modal Human Activity Recognition task efficiently.\nDMFT first encodes the multi-modal input data into a unified representation.\nThen the DMFT teacher model applies an attentive multi-modal spatial-temporal\ntransformer module that extracts the salient spatial-temporal features. A\ntemporal mid-fusion module is also proposed to further fuse the temporal\nfeatures. Then the knowledge distillation method is applied to transfer the\nlearned representation from the teacher model to a simpler DMFT student model,\nwhich consists of a lite version of the multi-modal spatial-temporal\ntransformer module, to produce the results. Evaluation of DMFT was conducted on\ntwo public multi-modal human activity recognition datasets with various\nstate-of-the-art approaches. The experimental results demonstrate that the\nmodel achieves competitive performance in terms of effectiveness, scalability,\nand robustness.\n", "date": "2023-05-05", "categories": []}
{"id": "2305.03187", "title": "Generating Virtual On-body Accelerometer Data from Virtual Textual\n  Descriptions for Human Activity Recognition", "abstract": "  The development of robust, generalized models in human activity recognition\n(HAR) has been hindered by the scarcity of large-scale, labeled data sets.\nRecent work has shown that virtual IMU data extracted from videos using\ncomputer vision techniques can lead to substantial performance improvements\nwhen training HAR models combined with small portions of real IMU data.\nInspired by recent advances in motion synthesis from textual descriptions and\nconnecting Large Language Models (LLMs) to various AI models, we introduce an\nautomated pipeline that first uses ChatGPT to generate diverse textual\ndescriptions of activities. These textual descriptions are then used to\ngenerate 3D human motion sequences via a motion synthesis model, T2M-GPT, and\nlater converted to streams of virtual IMU data. We benchmarked our approach on\nthree HAR datasets (RealWorld, PAMAP2, and USC-HAD) and demonstrate that the\nuse of virtual IMU training data generated using our new approach leads to\nsignificantly improved HAR model performance compared to only using real IMU\ndata. Our approach contributes to the growing field of cross-modality transfer\nmethods and illustrate how HAR models can be improved through the generation of\nvirtual training data that do not require any manual effort.\n", "date": "2023-05-04", "categories": []}
{"id": "2305.01733", "title": "Cross-view Action Recognition via Contrastive View-invariant\n  Representation", "abstract": "  Cross view action recognition (CVAR) seeks to recognize a human action when\nobserved from a previously unseen viewpoint. This is a challenging problem\nsince the appearance of an action changes significantly with the viewpoint.\nApplications of CVAR include surveillance and monitoring of assisted living\nfacilities where is not practical or feasible to collect large amounts of\ntraining data when adding a new camera. We present a simple yet efficient CVAR\nframework to learn invariant features from either RGB videos, 3D skeleton data,\nor both. The proposed approach outperforms the current state-of-the-art\nachieving similar levels of performance across input modalities: 99.4% (RGB)\nand 99.9% (3D skeletons), 99.4% (RGB) and 99.9% (3D Skeletons), 97.3% (RGB),\nand 99.2% (3D skeletons), and 84.4%(RGB) for the N-UCLA, NTU-RGB+D 60,\nNTU-RGB+D 120, and UWA3DII datasets, respectively.\n", "date": "2023-05-02", "categories": []}
{"id": "2210.12527", "title": "Baby Physical Safety Monitoring in Smart Home Using Action Recognition\n  System", "abstract": "  Humans are able to intuitively deduce actions that took place between two\nstates in observations via deductive reasoning. This is because the brain\noperates on a bidirectional communication model, which has radically improved\nthe accuracy of recognition and prediction based on features connected to\nprevious experiences. During the past decade, deep learning models for action\nrecognition have significantly improved. However, deep neural networks struggle\nwith these tasks on a smaller dataset for specific Action Recognition (AR)\ntasks. As with most action recognition tasks, the ambiguity of accurately\ndescribing activities in spatial-temporal data is a drawback that can be\novercome by curating suitable datasets, including careful annotations and\npreprocessing of video data for analyzing various recognition tasks. In this\nstudy, we present a novel lightweight framework combining transfer learning\ntechniques with a Conv2D LSTM layer to extract features from the pre-trained\nI3D model on the Kinetics dataset for a new AR task (Smart Baby Care) that\nrequires a smaller dataset and less computational resources. Furthermore, we\ndeveloped a benchmark dataset and an automated model that uses LSTM convolution\nwith I3D (ConvLSTM-I3D) for recognizing and predicting baby activities in a\nsmart baby room. Finally, we implemented video augmentation to improve model\nperformance on the smart baby care task. Compared to other benchmark models,\nour experimental framework achieved better performance with less computational\nresources.\n", "date": "2022-10-22", "categories": []}
{"id": "2305.00607", "title": "Boosting Weakly-Supervised Temporal Action Localization with Text\n  Information", "abstract": "  Due to the lack of temporal annotation, current Weakly-supervised Temporal\nAction Localization (WTAL) methods are generally stuck into over-complete or\nincomplete localization. In this paper, we aim to leverage the text information\nto boost WTAL from two aspects, i.e., (a) the discriminative objective to\nenlarge the inter-class difference, thus reducing the over-complete; (b) the\ngenerative objective to enhance the intra-class integrity, thus finding more\ncomplete temporal boundaries. For the discriminative objective, we propose a\nText-Segment Mining (TSM) mechanism, which constructs a text description based\non the action class label, and regards the text as the query to mine all\nclass-related segments. Without the temporal annotation of actions, TSM\ncompares the text query with the entire videos across the dataset to mine the\nbest matching segments while ignoring irrelevant ones. Due to the shared\nsub-actions in different categories of videos, merely applying TSM is too\nstrict to neglect the semantic-related segments, which results in incomplete\nlocalization. We further introduce a generative objective named Video-text\nLanguage Completion (VLC), which focuses on all semantic-related segments from\nvideos to complete the text sentence. We achieve the state-of-the-art\nperformance on THUMOS14 and ActivityNet1.3. Surprisingly, we also find our\nproposed method can be seamlessly applied to existing methods, and improve\ntheir performances with a clear margin. The code is available at\nhttps://github.com/lgzlIlIlI/Boosting-WTAL.\n", "date": "2023-05-01", "categories": []}
{"id": "2304.14499", "title": "Human activity recognition using deep learning approaches and single\n  frame cnn and convolutional lstm", "abstract": "  Human activity recognition is one of the most important tasks in computer\nvision and has proved useful in different fields such as healthcare, sports\ntraining and security. There are a number of approaches that have been explored\nto solve this task, some of them involving sensor data, and some involving\nvideo data. In this paper, we aim to explore two deep learning-based\napproaches, namely single frame Convolutional Neural Networks (CNNs) and\nconvolutional Long Short-Term Memory to recognise human actions from videos.\nUsing a convolutional neural networks-based method is advantageous as CNNs can\nextract features automatically and Long Short-Term Memory networks are great\nwhen it comes to working on sequence data such as video. The two models were\ntrained and evaluated on a benchmark action recognition dataset, UCF50, and\nanother dataset that was created for the experimentation. Though both models\nexhibit good accuracies, the single frame CNN model outperforms the\nConvolutional LSTM model by having an accuracy of 99.8% with the UCF50 dataset.\n", "date": "2023-04-18", "categories": []}
{"id": "2304.10335", "title": "A baseline on continual learning methods for video action recognition", "abstract": "  Continual learning has recently attracted attention from the research\ncommunity, as it aims to solve long-standing limitations of classic\nsupervisedly-trained models. However, most research on this subject has tackled\ncontinual learning in simple image classification scenarios. In this paper, we\npresent a benchmark of state-of-the-art continual learning methods on video\naction recognition. Besides the increased complexity due to the temporal\ndimension, the video setting imposes stronger requirements on computing\nresources for top-performing rehearsal methods. To counteract the increased\nmemory requirements, we present two method-agnostic variants for rehearsal\nmethods, exploiting measures of either model confidence or data information to\nselect memorable samples. Our experiments show that, as expected from the\nliterature, rehearsal methods outperform other approaches; moreover, the\nproposed memory-efficient variants are shown to be effective at retaining a\ncertain level of performance with a smaller buffer size.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.12616", "title": "Weakly-Supervised Temporal Action Localization with Bidirectional\n  Semantic Consistency Constraint", "abstract": "  Weakly Supervised Temporal Action Localization (WTAL) aims to classify and\nlocalize temporal boundaries of actions for the video, given only video-level\ncategory labels in the training datasets. Due to the lack of boundary\ninformation during training, existing approaches formulate WTAL as a\nclassificationproblem, i.e., generating the temporal class activation map\n(T-CAM) for localization. However, with only classification loss, the model\nwould be sub-optimized, i.e., the action-related scenes are enough to\ndistinguish different class labels. Regarding other actions in the\naction-related scene ( i.e., the scene same as positive actions) as co-scene\nactions, this sub-optimized model would misclassify the co-scene actions as\npositive actions. To address this misclassification, we propose a simple yet\nefficient method, named bidirectional semantic consistency constraint (Bi-SCC),\nto discriminate the positive actions from co-scene actions. The proposed Bi-SCC\nfirstly adopts a temporal context augmentation to generate an augmented video\nthat breaks the correlation between positive actions and their co-scene actions\nin the inter-video; Then, a semantic consistency constraint (SCC) is used to\nenforce the predictions of the original video and augmented video to be\nconsistent, hence suppressing the co-scene actions. However, we find that this\naugmented video would destroy the original temporal context. Simply applying\nthe consistency constraint would affect the completeness of localized positive\nactions. Hence, we boost the SCC in a bidirectional way to suppress co-scene\nactions while ensuring the integrity of positive actions, by cross-supervising\nthe original and augmented videos. Finally, our proposed Bi-SCC can be applied\nto current WTAL approaches, and improve their performance. Experimental results\nshow that our approach outperforms the state-of-the-art methods on THUMOS14 and\nActivityNet.\n", "date": "2023-04-25", "categories": []}
{"id": "2212.11694", "title": "Timestamp-Supervised Action Segmentation from the Perspective of\n  Clustering", "abstract": "  Video action segmentation under timestamp supervision has recently received\nmuch attention due to lower annotation costs. Most existing methods generate\npseudo-labels for all frames in each video to train the segmentation model.\nHowever, these methods suffer from incorrect pseudo-labels, especially for the\nsemantically unclear frames in the transition region between two consecutive\nactions, which we call ambiguous intervals. To address this issue, we propose a\nnovel framework from the perspective of clustering, which includes the\nfollowing two parts. First, pseudo-label ensembling generates incomplete but\nhigh-quality pseudo-label sequences, where the frames in ambiguous intervals\nhave no pseudo-labels. Second, iterative clustering iteratively propagates the\npseudo-labels to the ambiguous intervals by clustering, and thus updates the\npseudo-label sequences to train the model. We further introduce a clustering\nloss, which encourages the features of frames within the same action segment\nmore compact. Extensive experiments show the effectiveness of our method.\n", "date": "2022-12-22", "categories": []}
{"id": "2303.00138", "title": "Texture-Based Input Feature Selection for Action Recognition", "abstract": "  The performance of video action recognition has been significantly boosted by\nusing motion representations within a two-stream Convolutional Neural Network\n(CNN) architecture. However, there are a few challenging problems in action\nrecognition in real scenarios, e.g., the variations in viewpoints and poses,\nand the changes in backgrounds. The domain discrepancy between the training\ndata and the test data causes the performance drop. To improve the model\nrobustness, we propose a novel method to determine the task-irrelevant content\nin inputs which increases the domain discrepancy. The method is based on a\nhuman parsing model (HP model) which jointly conducts dense correspondence\nlabelling and semantic part segmentation. The predictions from the HP model\nalso function as re-rendering the human regions in each video using the same\nset of textures to make humans appearances in all classes be the same. A\nrevised dataset is generated for training and testing and makes the action\nrecognition model exhibit invariance to the irrelevant content in the inputs.\nMoreover, the predictions from the HP model are used to enrich the inputs to\nthe AR model during both training and testing. Experimental results show that\nour proposed model is superior to existing models for action recognition on the\nHMDB-51 dataset and the Penn Action dataset.\n", "date": "2023-02-28", "categories": []}
{"id": "2304.00325", "title": "SVT: Supertoken Video Transformer for Efficient Video Understanding", "abstract": "  Whether by processing videos with fixed resolution from start to end or\nincorporating pooling and down-scaling strategies, existing video transformers\nprocess the whole video content throughout the network without specially\nhandling the large portions of redundant information. In this paper, we present\na Supertoken Video Transformer (SVT) that incorporates a Semantic Pooling\nModule (SPM) to aggregate latent representations along the depth of visual\ntransformer based on their semantics, and thus, reduces redundancy inherent in\nvideo inputs.~Qualitative results show that our method can effectively reduce\nredundancy by merging latent representations with similar semantics and thus\nincrease the proportion of salient information for downstream\ntasks.~Quantitatively, our method improves the performance of both ViT and MViT\nwhile requiring significantly less computations on the Kinectics and\nSomething-Something-V2 benchmarks.~More specifically, with our SPM, we improve\nthe accuracy of MAE-pretrained ViT-B and ViT-L by 1.5% with 33% less GFLOPs and\nby 0.2% with 55% less FLOPs, respectively, on the Kinectics-400 benchmark, and\nimprove the accuracy of MViTv2-B by 0.2% and 0.3% with 22% less GFLOPs on\nKinectics-400 and Something-Something-V2, respectively.\n", "date": "2023-04-01", "categories": []}
{"id": "2304.11975", "title": "MRSN: Multi-Relation Support Network for Video Action Detection", "abstract": "  Action detection is a challenging video understanding task, requiring\nmodeling spatio-temporal and interaction relations. Current methods usually\nmodel actor-actor and actor-context relations separately, ignoring their\ncomplementarity and mutual support. To solve this problem, we propose a novel\nnetwork called Multi-Relation Support Network (MRSN). In MRSN, Actor-Context\nRelation Encoder (ACRE) and Actor-Actor Relation Encoder (AARE) model the\nactor-context and actor-actor relation separately. Then Relation Support\nEncoder (RSE) computes the supports between the two relations and performs\nrelation-level interactions. Finally, Relation Consensus Module (RCM) enhances\ntwo relations with the long-term relations from the Long-term Relation Bank\n(LRB) and yields a consensus. Our experiments demonstrate that modeling\nrelations separately and performing relation-level interactions can achieve and\noutperformer state-of-the-art results on two challenging video datasets: AVA\nand UCF101-24.\n", "date": "2023-04-24", "categories": []}
{"id": "2304.12160", "title": "End-to-End Spatio-Temporal Action Localisation with Video Transformers", "abstract": "  The most performant spatio-temporal action localisation models use external\nperson proposals and complex external memory banks. We propose a fully\nend-to-end, purely-transformer based model that directly ingests an input\nvideo, and outputs tubelets -- a sequence of bounding boxes and the action\nclasses at each frame. Our flexible model can be trained with either sparse\nbounding-box supervision on individual frames, or full tubelet annotations. And\nin both cases, it predicts coherent tubelets as the output. Moreover, our\nend-to-end model requires no additional pre-processing in the form of\nproposals, or post-processing in terms of non-maximal suppression. We perform\nextensive ablation experiments, and significantly advance the state-of-the-art\nresults on four different spatio-temporal action localisation benchmarks with\nboth sparse keyframes and full tubelet annotations.\n", "date": "2023-04-24", "categories": []}
{"id": "2304.10073", "title": "Video-based Contrastive Learning on Decision Trees: from Action\n  Recognition to Autism Diagnosis", "abstract": "  How can we teach a computer to recognize 10,000 different actions? Deep\nlearning has evolved from supervised and unsupervised to self-supervised\napproaches. In this paper, we present a new contrastive learning-based\nframework for decision tree-based classification of actions, including\nhuman-human interactions (HHI) and human-object interactions (HOI). The key\nidea is to translate the original multi-class action recognition into a series\nof binary classification tasks on a pre-constructed decision tree. Under the\nnew framework of contrastive learning, we present the design of an interaction\nadjacent matrix (IAM) with skeleton graphs as the backbone for modeling various\naction-related attributes such as periodicity and symmetry. Through the\nconstruction of various pretext tasks, we obtain a series of binary\nclassification nodes on the decision tree that can be combined to support\nhigher-level recognition tasks. Experimental justification for the potential of\nour approach in real-world applications ranges from interaction recognition to\nsymmetry detection. In particular, we have demonstrated the promising\nperformance of video-based autism spectrum disorder (ASD) diagnosis on the\nCalTech interview video database.\n", "date": "2023-04-20", "categories": []}
{"id": "2304.10316", "title": "Search-Map-Search: A Frame Selection Paradigm for Action Recognition", "abstract": "  Despite the success of deep learning in video understanding tasks, processing\nevery frame in a video is computationally expensive and often unnecessary in\nreal-time applications. Frame selection aims to extract the most informative\nand representative frames to help a model better understand video content.\nExisting frame selection methods either individually sample frames based on\nper-frame importance prediction, without considering interaction among frames,\nor adopt reinforcement learning agents to find representative frames in\nsuccession, which are costly to train and may lead to potential stability\nissues. To overcome the limitations of existing methods, we propose a\nSearch-Map-Search learning paradigm which combines the advantages of heuristic\nsearch and supervised learning to select the best combination of frames from a\nvideo as one entity. By combining search with learning, the proposed method can\nbetter capture frame interactions while incurring a low inference overhead.\nSpecifically, we first propose a hierarchical search method conducted on each\ntraining video to search for the optimal combination of frames with the lowest\nerror on the downstream task. A feature mapping function is then learned to map\nthe frames of a video to the representation of its target optimal frame\ncombination. During inference, another search is performed on an unseen video\nto select a combination of frames whose feature representation is close to the\nprojected feature representation. Extensive experiments based on several action\nrecognition benchmarks demonstrate that our frame selection method effectively\nimproves performance of action recognition models, and significantly\noutperforms a number of competitive baselines.\n", "date": "2023-04-20", "categories": []}
{"id": "2210.13605", "title": "GliTr: Glimpse Transformers with Spatiotemporal Consistency for Online\n  Action Prediction", "abstract": "  Many online action prediction models observe complete frames to locate and\nattend to informative subregions in the frames called glimpses and recognize an\nongoing action based on global and local information. However, in applications\nwith constrained resources, an agent may not be able to observe the complete\nframe, yet must still locate useful glimpses to predict an incomplete action\nbased on local information only. In this paper, we develop Glimpse Transformers\n(GliTr), which observe only narrow glimpses at all times, thus predicting an\nongoing action and the following most informative glimpse location based on the\npartial spatiotemporal information collected so far. In the absence of a ground\ntruth for the optimal glimpse locations for action recognition, we train GliTr\nusing a novel spatiotemporal consistency objective: We require GliTr to attend\nto the glimpses with features similar to the corresponding complete frames\n(i.e. spatial consistency) and the resultant class logits at time $t$\nequivalent to the ones predicted using whole frames up to $t$ (i.e. temporal\nconsistency). Inclusion of our proposed consistency objective yields ~10%\nhigher accuracy on the Something-Something-v2 (SSv2) dataset than the baseline\ncross-entropy objective. Overall, despite observing only ~33% of the total area\nper frame, GliTr achieves 53.02% and 93.91% accuracy on the SSv2 and Jester\ndatasets, respectively.\n", "date": "2022-10-24", "categories": []}
{"id": "2212.04638", "title": "FLAG3D: A 3D Fitness Activity Dataset with Language Instruction", "abstract": "  With the continuously thriving popularity around the world, fitness activity\nanalytic has become an emerging research topic in computer vision. While a\nvariety of new tasks and algorithms have been proposed recently, there are\ngrowing hunger for data resources involved in high-quality data, fine-grained\nlabels, and diverse environments. In this paper, we present FLAG3D, a\nlarge-scale 3D fitness activity dataset with language instruction containing\n180K sequences of 60 categories. FLAG3D features the following three aspects:\n1) accurate and dense 3D human pose captured from advanced MoCap system to\nhandle the complex activity and large movement, 2) detailed and professional\nlanguage instruction to describe how to perform a specific activity, 3)\nversatile video resources from a high-tech MoCap system, rendering software,\nand cost-effective smartphones in natural environments. Extensive experiments\nand in-depth analysis show that FLAG3D contributes great research value for\nvarious challenges, such as cross-domain human action recognition, dynamic\nhuman mesh recovery, and language-guided human action generation. Our dataset\nand source code are publicly available at https://andytang15.github.io/FLAG3D.\n", "date": "2022-12-09", "categories": []}
{"id": "2304.09530", "title": "SelfAct: Personalized Activity Recognition based on Self-Supervised and\n  Active Learning", "abstract": "  Supervised Deep Learning (DL) models are currently the leading approach for\nsensor-based Human Activity Recognition (HAR) on wearable and mobile devices.\nHowever, training them requires large amounts of labeled data whose collection\nis often time-consuming, expensive, and error-prone. At the same time, due to\nthe intra- and inter-variability of activity execution, activity models should\nbe personalized for each user. In this work, we propose SelfAct: a novel\nframework for HAR combining self-supervised and active learning to mitigate\nthese problems. SelfAct leverages a large pool of unlabeled data collected from\nmany users to pre-train through self-supervision a DL model, with the goal of\nlearning a meaningful and efficient latent representation of sensor data. The\nresulting pre-trained model can be locally used by new users, which will\nfine-tune it thanks to a novel unsupervised active learning strategy. Our\nexperiments on two publicly available HAR datasets demonstrate that SelfAct\nachieves results that are close to or even better than the ones of fully\nsupervised approaches with a small number of active learning queries.\n", "date": "2023-04-19", "categories": []}
{"id": "2304.08803", "title": "MLP-AIR: An Efficient MLP-Based Method for Actor Interaction Relation\n  Learning in Group Activity Recognition", "abstract": "  The task of Group Activity Recognition (GAR) aims to predict the activity\ncategory of the group by learning the actor spatial-temporal interaction\nrelation in the group. Therefore, an effective actor relation learning method\nis crucial for the GAR task. The previous works mainly learn the interaction\nrelation by the well-designed GCNs or Transformers. For example, to infer the\nactor interaction relation, GCNs need a learnable adjacency, and Transformers\nneed to calculate the self-attention. Although the above methods can model the\ninteraction relation effectively, they also increase the complexity of the\nmodel (the number of parameters and computations). In this paper, we design a\nnovel MLP-based method for Actor Interaction Relation learning (MLP-AIR) in\nGAR. Compared with GCNs and Transformers, our method has a competitive but\nconceptually and technically simple alternative, significantly reducing the\ncomplexity. Specifically, MLP-AIR includes three sub-modules: MLP-based Spatial\nrelation modeling module (MLP-S), MLP-based Temporal relation modeling module\n(MLP-T), and MLP-based Relation refining module (MLP-R). MLP-S is used to model\nthe spatial relation between different actors in each frame. MLP-T is used to\nmodel the temporal relation between different frames for each actor. MLP-R is\nused further to refine the relation between different dimensions of relation\nfeatures to improve the feature's expression ability. To evaluate the MLP-AIR,\nwe conduct extensive experiments on two widely used benchmarks, including the\nVolleyball and Collective Activity datasets. Experimental results demonstrate\nthat MLP-AIR can get competitive results but with low complexity.\n", "date": "2023-04-18", "categories": []}
{"id": "2304.07978", "title": "Improving Weakly Supervised Temporal Action Localization by Bridging\n  Train-Test Gap in Pseudo Labels", "abstract": "  The task of weakly supervised temporal action localization targets at\ngenerating temporal boundaries for actions of interest, meanwhile the action\ncategory should also be classified. Pseudo-label-based methods, which serve as\nan effective solution, have been widely studied recently. However, existing\nmethods generate pseudo labels during training and make predictions during\ntesting under different pipelines or settings, resulting in a gap between\ntraining and testing. In this paper, we propose to generate high-quality pseudo\nlabels from the predicted action boundaries. Nevertheless, we note that\nexisting post-processing, like NMS, would lead to information loss, which is\ninsufficient to generate high-quality action boundaries. More importantly,\ntransforming action boundaries into pseudo labels is quite challenging, since\nthe predicted action instances are generally overlapped and have different\nconfidence scores. Besides, the generated pseudo-labels can be fluctuating and\ninaccurate at the early stage of training. It might repeatedly strengthen the\nfalse predictions if there is no mechanism to conduct self-correction. To\ntackle these issues, we come up with an effective pipeline for learning better\npseudo labels. Firstly, we propose a Gaussian weighted fusion module to\npreserve information of action instances and obtain high-quality action\nboundaries. Second, we formulate the pseudo-label generation as an optimization\nproblem under the constraints in terms of the confidence scores of action\ninstances. Finally, we introduce the idea of $\\Delta$ pseudo labels, which\nenables the model with the ability of self-correction. Our method achieves\nsuperior performance to existing methods on two benchmarks, THUMOS14 and\nActivityNet1.3, achieving gains of 1.9\\% on THUMOS14 and 3.7\\% on\nActivityNet1.3 in terms of average mAP.\n", "date": "2023-04-17", "categories": []}
{"id": "2304.08210", "title": "ATTACH Dataset: Annotated Two-Handed Assembly Actions for Human Action\n  Understanding", "abstract": "  With the emergence of collaborative robots (cobots), human-robot\ncollaboration in industrial manufacturing is coming into focus. For a cobot to\nact autonomously and as an assistant, it must understand human actions during\nassembly. To effectively train models for this task, a dataset containing\nsuitable assembly actions in a realistic setting is crucial. For this purpose,\nwe present the ATTACH dataset, which contains 51.6 hours of assembly with 95.2k\nannotated fine-grained actions monitored by three cameras, which represent\npotential viewpoints of a cobot. Since in an assembly context workers tend to\nperform different actions simultaneously with their two hands, we annotated the\nperformed actions for each hand separately. Therefore, in the ATTACH dataset,\nmore than 68% of annotations overlap with other annotations, which is many\ntimes more than in related datasets, typically featuring more simplistic\nassembly tasks. For better generalization with respect to the background of the\nworking area, we did not only record color and depth images, but also used the\nAzure Kinect body tracking SDK for estimating 3D skeletons of the worker. To\ncreate a first baseline, we report the performance of state-of-the-art methods\nfor action recognition as well as action detection on video and\nskeleton-sequence inputs. The dataset is available at\nhttps://www.tu-ilmenau.de/neurob/data-sets-code/attach-dataset .\n", "date": "2023-04-17", "categories": []}
{"id": "2301.07944", "title": "Revisiting the Spatial and Temporal Modeling for Few-shot Action\n  Recognition", "abstract": "  Spatial and temporal modeling is one of the most core aspects of few-shot\naction recognition. Most previous works mainly focus on long-term temporal\nrelation modeling based on high-level spatial representations, without\nconsidering the crucial low-level spatial features and short-term temporal\nrelations. Actually, the former feature could bring rich local semantic\ninformation, and the latter feature could represent motion characteristics of\nadjacent frames, respectively. In this paper, we propose SloshNet, a new\nframework that revisits the spatial and temporal modeling for few-shot action\nrecognition in a finer manner. First, to exploit the low-level spatial\nfeatures, we design a feature fusion architecture search module to\nautomatically search for the best combination of the low-level and high-level\nspatial features. Next, inspired by the recent transformer, we introduce a\nlong-term temporal modeling module to model the global temporal relations based\non the extracted spatial appearance features. Meanwhile, we design another\nshort-term temporal modeling module to encode the motion characteristics\nbetween adjacent frame representations. After that, the final predictions can\nbe obtained by feeding the embedded rich spatial-temporal features to a common\nframe-level class prototype matcher. We extensively validate the proposed\nSloshNet on four few-shot action recognition datasets, including\nSomething-Something V2, Kinetics, UCF101, and HMDB51. It achieves favorable\nresults against state-of-the-art methods in all datasets.\n", "date": "2023-01-19", "categories": []}
{"id": "2207.01398", "title": "Large-scale Robustness Analysis of Video Action Recognition Models", "abstract": "  We have seen a great progress in video action recognition in recent years.\nThere are several models based on convolutional neural network (CNN) and some\nrecent transformer based approaches which provide top performance on existing\nbenchmarks. In this work, we perform a large-scale robustness analysis of these\nexisting models for video action recognition. We focus on robustness against\nreal-world distribution shift perturbations instead of adversarial\nperturbations. We propose four different benchmark datasets, HMDB51-P,\nUCF101-P, Kinetics400-P, and SSv2-P to perform this analysis. We study\nrobustness of six state-of-the-art action recognition models against 90\ndifferent perturbations. The study reveals some interesting findings, 1)\ntransformer based models are consistently more robust compared to CNN based\nmodels, 2) Pretraining improves robustness for Transformer based models more\nthan CNN based models, and 3) All of the studied models are robust to temporal\nperturbations for all datasets but SSv2; suggesting the importance of temporal\ninformation for action recognition varies based on the dataset and activities.\nNext, we study the role of augmentations in model robustness and present a\nreal-world dataset, UCF101-DS, which contains realistic distribution shifts, to\nfurther validate some of these findings. We believe this study will serve as a\nbenchmark for future research in robust video action recognition.\n", "date": "2022-07-04", "categories": []}
{"id": "2304.03307", "title": "Vita-CLIP: Video and text adaptive CLIP via Multimodal Prompting", "abstract": "  Adopting contrastive image-text pretrained models like CLIP towards video\nclassification has gained attention due to its cost-effectiveness and\ncompetitive performance. However, recent works in this area face a trade-off.\nFinetuning the pretrained model to achieve strong supervised performance\nresults in low zero-shot generalization. Similarly, freezing the backbone to\nretain zero-shot capability causes significant drop in supervised accuracy.\nBecause of this, recent works in literature typically train separate models for\nsupervised and zero-shot action recognition. In this work, we propose a\nmultimodal prompt learning scheme that works to balance the supervised and\nzero-shot performance under a single unified training. Our prompting approach\non the vision side caters for three aspects: 1) Global video-level prompts to\nmodel the data distribution; 2) Local frame-level prompts to provide per-frame\ndiscriminative conditioning; and 3) a summary prompt to extract a condensed\nvideo representation. Additionally, we define a prompting scheme on the text\nside to augment the textual context. Through this prompting scheme, we can\nachieve state-of-the-art zero-shot performance on Kinetics-600, HMDB51 and\nUCF101 while remaining competitive in the supervised setting. By keeping the\npretrained backbone frozen, we optimize a much lower number of parameters and\nretain the existing general representation which helps achieve the strong\nzero-shot performance. Our codes/models are released at\nhttps://github.com/TalalWasim/Vita-CLIP.\n", "date": "2023-04-06", "categories": []}
{"id": "2304.03631", "title": "Therbligs in Action: Video Understanding through Motion Primitives", "abstract": "  In this paper we introduce a rule-based, compositional, and hierarchical\nmodeling of action using Therbligs as our atoms. Introducing these atoms\nprovides us with a consistent, expressive, contact-centered representation of\naction. Over the atoms we introduce a differentiable method of rule-based\nreasoning to regularize for logical consistency. Our approach is complementary\nto other approaches in that the Therblig-based representations produced by our\narchitecture augment rather than replace existing architectures'\nrepresentations. We release the first Therblig-centered annotations over two\npopular video datasets - EPIC Kitchens 100 and 50-Salads. We also broadly\ndemonstrate benefits to adopting Therblig representations through evaluation on\nthe following tasks: action segmentation, action anticipation, and action\nrecognition - observing an average 10.5\\%/7.53\\%/6.5\\% relative improvement,\nrespectively, over EPIC Kitchens and an average 8.9\\%/6.63\\%/4.8\\% relative\nimprovement, respectively, over 50 Salads. Code and data will be made publicly\navailable.\n", "date": "2023-04-06", "categories": []}
{"id": "2211.13222", "title": "SVFormer: Semi-supervised Video Transformer for Action Recognition", "abstract": "  Semi-supervised action recognition is a challenging but critical task due to\nthe high cost of video annotations. Existing approaches mainly use\nconvolutional neural networks, yet current revolutionary vision transformer\nmodels have been less explored. In this paper, we investigate the use of\ntransformer models under the SSL setting for action recognition. To this end,\nwe introduce SVFormer, which adopts a steady pseudo-labeling framework (ie,\nEMA-Teacher) to cope with unlabeled video samples. While a wide range of data\naugmentations have been shown effective for semi-supervised image\nclassification, they generally produce limited results for video recognition.\nWe therefore introduce a novel augmentation strategy, Tube TokenMix, tailored\nfor video data where video clips are mixed via a mask with consistent masked\ntokens over the temporal axis. In addition, we propose a temporal warping\naugmentation to cover the complex temporal variation in videos, which stretches\nselected frames to various temporal durations in the clip. Extensive\nexperiments on three datasets Kinetics-400, UCF-101, and HMDB-51 verify the\nadvantage of SVFormer. In particular, SVFormer outperforms the state-of-the-art\nby 31.5% with fewer training epochs under the 1% labeling rate of Kinetics-400.\nOur method can hopefully serve as a strong benchmark and encourage future\nsearch on semi-supervised action recognition with Transformer networks.\n", "date": "2022-11-23", "categories": []}
{"id": "2304.02110", "title": "DIR-AS: Decoupling Individual Identification and Temporal Reasoning for\n  Action Segmentation", "abstract": "  Fully supervised action segmentation works on frame-wise action recognition\nwith dense annotations and often suffers from the over-segmentation issue.\nExisting works have proposed a variety of solutions such as boundary-aware\nnetworks, multi-stage refinement, and temporal smoothness losses. However, most\nof them take advantage of frame-wise supervision, which cannot effectively\ntackle the evaluation metrics with different granularities. In this paper, for\nthe desirable large receptive field, we first develop a novel local-global\nattention mechanism with temporal pyramid dilation and temporal pyramid pooling\nfor efficient multi-scale attention. Then we decouple two inherent goals in\naction segmentation, ie, (1) individual identification solved by frame-wise\nsupervision, and (2) temporal reasoning tackled by action set prediction.\nAfterward, an action alignment module fuses these different granularity\npredictions, leading to more accurate and smoother action segmentation. We\nachieve state-of-the-art accuracy, eg, 82.8% (+2.6%) on GTEA and 74.7% (+1.2%)\non Breakfast, which demonstrates the effectiveness of our proposed method,\naccompanied by extensive ablation studies. The code will be made available\nlater.\n", "date": "2023-04-04", "categories": []}
{"id": "2301.00436", "title": "Hierarchical Explanations for Video Action Recognition", "abstract": "  To interpret deep neural networks, one main approach is to dissect the visual\ninput and find the prototypical parts responsible for the classification.\nHowever, existing methods often ignore the hierarchical relationship between\nthese prototypes, and thus can not explain semantic concepts at both higher\nlevel (e.g., water sports) and lower level (e.g., swimming). In this paper\ninspired by human cognition system, we leverage hierarchal information to deal\nwith uncertainty: When we observe water and human activity, but no definitive\naction it can be recognized as the water sports parent class. Only after\nobserving a person swimming can we definitively refine it to the swimming\naction. To this end, we propose HIerarchical Prototype Explainer (HIPE) to\nbuild hierarchical relations between prototypes and classes. HIPE enables a\nreasoning process for video action classification by dissecting the input video\nframes on multiple levels of the class hierarchy, our method is also applicable\nto other video tasks. The faithfulness of our method is verified by reducing\naccuracy-explainability trade off on ActivityNet and UCF-101 while providing\nmulti-level explanations.\n", "date": "2023-01-01", "categories": []}
{"id": "2304.00254", "title": "DOAD: Decoupled One Stage Action Detection Network", "abstract": "  Localizing people and recognizing their actions from videos is a challenging\ntask towards high-level video understanding. Existing methods are mostly\ntwo-stage based, with one stage for person bounding box generation and the\nother stage for action recognition. However, such two-stage methods are\ngenerally with low efficiency. We observe that directly unifying detection and\naction recognition normally suffers from (i) inferior learning due to different\ndesired properties of context representation for detection and action\nrecognition; (ii) optimization difficulty with insufficient training data. In\nthis work, we present a decoupled one-stage network dubbed DOAD, to mitigate\nabove issues and improve the efficiency for spatio-temporal action detection.\nTo achieve it, we decouple detection and action recognition into two branches.\nSpecifically, one branch focuses on detection representation for actor\ndetection, and the other one for action recognition. For the action branch, we\ndesign a transformer-based module (TransPC) to model pairwise relationships\nbetween people and context. Different from commonly used vector-based dot\nproduct in self-attention, it is built upon a novel matrix-based key and value\nfor Hadamard attention to model person-context information. It not only\nexploits relationships between person pairs but also takes into account context\nand relative position information. The results on AVA and UCF101-24 datasets\nshow that our method is competitive with two-stage state-of-the-art methods\nwith significant efficiency improvement.\n", "date": "2023-04-01", "categories": []}
{"id": "2304.01110", "title": "AutoLabel: CLIP-based framework for Open-set Video Domain Adaptation", "abstract": "  Open-set Unsupervised Video Domain Adaptation (OUVDA) deals with the task of\nadapting an action recognition model from a labelled source domain to an\nunlabelled target domain that contains \"target-private\" categories, which are\npresent in the target but absent in the source. In this work we deviate from\nthe prior work of training a specialized open-set classifier or weighted\nadversarial learning by proposing to use pre-trained Language and Vision Models\n(CLIP). The CLIP is well suited for OUVDA due to its rich representation and\nthe zero-shot recognition capabilities. However, rejecting target-private\ninstances with the CLIP's zero-shot protocol requires oracle knowledge about\nthe target-private label names. To circumvent the impossibility of the\nknowledge of label names, we propose AutoLabel that automatically discovers and\ngenerates object-centric compositional candidate target-private class names.\nDespite its simplicity, we show that CLIP when equipped with AutoLabel can\nsatisfactorily reject the target-private instances, thereby facilitating better\nalignment between the shared classes of the two domains. The code is available.\n", "date": "2023-04-03", "categories": []}
{"id": "2304.01567", "title": "A real-time algorithm for human action recognition in RGB and thermal\n  video", "abstract": "  Monitoring the movement and actions of humans in video in real-time is an\nimportant task. We present a deep learning based algorithm for human action\nrecognition for both RGB and thermal cameras. It is able to detect and track\nhumans and recognize four basic actions (standing, walking, running, lying) in\nreal-time on a notebook with a NVIDIA GPU. For this, it combines state of the\nart components for object detection (Scaled YoloV4), optical flow (RAFT) and\npose estimation (EvoSkeleton). Qualitative experiments on a set of tunnel\nvideos show that the proposed algorithm works robustly for both RGB and thermal\nvideo.\n", "date": "2023-04-04", "categories": []}
{"id": "2304.00387", "title": "HaLP: Hallucinating Latent Positives for Skeleton-based Self-Supervised\n  Learning of Actions", "abstract": "  Supervised learning of skeleton sequence encoders for action recognition has\nreceived significant attention in recent times. However, learning such encoders\nwithout labels continues to be a challenging problem. While prior works have\nshown promising results by applying contrastive learning to pose sequences, the\nquality of the learned representations is often observed to be closely tied to\ndata augmentations that are used to craft the positives. However, augmenting\npose sequences is a difficult task as the geometric constraints among the\nskeleton joints need to be enforced to make the augmentations realistic for\nthat action. In this work, we propose a new contrastive learning approach to\ntrain models for skeleton-based action recognition without labels. Our key\ncontribution is a simple module, HaLP - to Hallucinate Latent Positives for\ncontrastive learning. Specifically, HaLP explores the latent space of poses in\nsuitable directions to generate new positives. To this end, we present a novel\noptimization formulation to solve for the synthetic positives with an explicit\ncontrol on their hardness. We propose approximations to the objective, making\nthem solvable in closed form with minimal overhead. We show via experiments\nthat using these generated positives within a standard contrastive learning\nframework leads to consistent improvements across benchmarks such as NTU-60,\nNTU-120, and PKU-II on tasks like linear evaluation, transfer learning, and kNN\nevaluation. Our code will be made available at\nhttps://github.com/anshulbshah/HaLP.\n", "date": "2023-04-01", "categories": []}
{"id": "2304.00858", "title": "Focalized Contrastive View-invariant Learning for Self-supervised\n  Skeleton-based Action Recognition", "abstract": "  Learning view-invariant representation is a key to improving feature\ndiscrimination power for skeleton-based action recognition. Existing approaches\ncannot effectively remove the impact of viewpoint due to the implicit\nview-dependent representations. In this work, we propose a self-supervised\nframework called Focalized Contrastive View-invariant Learning (FoCoViL), which\nsignificantly suppresses the view-specific information on the representation\nspace where the viewpoints are coarsely aligned. By maximizing mutual\ninformation with an effective contrastive loss between multi-view sample pairs,\nFoCoViL associates actions with common view-invariant properties and\nsimultaneously separates the dissimilar ones. We further propose an adaptive\nfocalization method based on pairwise similarity to enhance contrastive\nlearning for a clearer cluster boundary in the learned space. Different from\nmany existing self-supervised representation learning work that rely heavily on\nsupervised classifiers, FoCoViL performs well on both unsupervised and\nsupervised classifiers with superior recognition performance. Extensive\nexperiments also show that the proposed contrastive-based focalization\ngenerates a more discriminative latent representation.\n", "date": "2023-04-03", "categories": []}
{"id": "2304.00946", "title": "MoLo: Motion-augmented Long-short Contrastive Learning for Few-shot\n  Action Recognition", "abstract": "  Current state-of-the-art approaches for few-shot action recognition achieve\npromising performance by conducting frame-level matching on learned visual\nfeatures. However, they generally suffer from two limitations: i) the matching\nprocedure between local frames tends to be inaccurate due to the lack of\nguidance to force long-range temporal perception; ii) explicit motion learning\nis usually ignored, leading to partial information loss. To address these\nissues, we develop a Motion-augmented Long-short Contrastive Learning (MoLo)\nmethod that contains two crucial components, including a long-short contrastive\nobjective and a motion autodecoder. Specifically, the long-short contrastive\nobjective is to endow local frame features with long-form temporal awareness by\nmaximizing their agreement with the global token of videos belonging to the\nsame class. The motion autodecoder is a lightweight architecture to reconstruct\npixel motions from the differential features, which explicitly embeds the\nnetwork with motion dynamics. By this means, MoLo can simultaneously learn\nlong-range temporal context and motion cues for comprehensive few-shot\nmatching. To demonstrate the effectiveness, we evaluate MoLo on five standard\nbenchmarks, and the results show that MoLo favorably outperforms recent\nadvanced methods. The source code is available at\nhttps://github.com/alibaba-mmai-research/MoLo.\n", "date": "2023-04-03", "categories": []}
{"id": "2211.13916", "title": "Towards Good Practices for Missing Modality Robust Action Recognition", "abstract": "  Standard multi-modal models assume the use of the same modalities in training\nand inference stages. However, in practice, the environment in which\nmulti-modal models operate may not satisfy such assumption. As such, their\nperformances degrade drastically if any modality is missing in the inference\nstage. We ask: how can we train a model that is robust to missing modalities?\nThis paper seeks a set of good practices for multi-modal action recognition,\nwith a particular interest in circumstances where some modalities are not\navailable at an inference time. First, we study how to effectively regularize\nthe model during training (e.g., data augmentation). Second, we investigate on\nfusion methods for robustness to missing modalities: we find that\ntransformer-based fusion shows better robustness for missing modality than\nsummation or concatenation. Third, we propose a simple modular network,\nActionMAE, which learns missing modality predictive coding by randomly dropping\nmodality features and tries to reconstruct them with the remaining modality\nfeatures. Coupling these good practices, we build a model that is not only\neffective in multi-modal action recognition but also robust to modality\nmissing. Our model achieves the state-of-the-arts on multiple benchmarks and\nmaintains competitive performances even in missing modality scenarios. Codes\nare available at https://github.com/sangminwoo/ActionMAE.\n", "date": "2022-11-25", "categories": []}
{"id": "2303.17285", "title": "Decomposed Cross-modal Distillation for RGB-based Temporal Action\n  Detection", "abstract": "  Temporal action detection aims to predict the time intervals and the classes\nof action instances in the video. Despite the promising performance, existing\ntwo-stream models exhibit slow inference speed due to their reliance on\ncomputationally expensive optical flow. In this paper, we introduce a\ndecomposed cross-modal distillation framework to build a strong RGB-based\ndetector by transferring knowledge of the motion modality. Specifically,\ninstead of direct distillation, we propose to separately learn RGB and motion\nrepresentations, which are in turn combined to perform action localization. The\ndual-branch design and the asymmetric training objectives enable effective\nmotion knowledge transfer while preserving RGB information intact. In addition,\nwe introduce a local attentive fusion to better exploit the multimodal\ncomplementarity. It is designed to preserve the local discriminability of the\nfeatures that is important for action localization. Extensive experiments on\nthe benchmarks verify the effectiveness of the proposed method in enhancing\nRGB-based action detectors. Notably, our framework is agnostic to backbones and\ndetection heads, bringing consistent gains across different model combinations.\n", "date": "2023-03-30", "categories": []}
{"id": "2303.17294", "title": "JCDNet: Joint of Common and Definite phases Network for Weakly\n  Supervised Temporal Action Localization", "abstract": "  Weakly-supervised temporal action localization aims to localize action\ninstances in untrimmed videos with only video-level supervision. We witness\nthat different actions record common phases, e.g., the run-up in the HighJump\nand LongJump. These different actions are defined as conjoint actions, whose\nrest parts are definite phases, e.g., leaping over the bar in a HighJump.\nCompared with the common phases, the definite phases are more easily localized\nin existing researches. Most of them formulate this task as a Multiple Instance\nLearning paradigm, in which the common phases are tended to be confused with\nthe background, and affect the localization completeness of the conjoint\nactions. To tackle this challenge, we propose a Joint of Common and Definite\nphases Network (JCDNet) by improving feature discriminability of the conjoint\nactions. Specifically, we design a Class-Aware Discriminative module to enhance\nthe contribution of the common phases in classification by the guidance of the\ncoarse definite-phase features. Besides, we introduce a temporal attention\nmodule to learn robust action-ness scores via modeling temporal dependencies,\ndistinguishing the common phases from the background. Extensive experiments on\nthree datasets (THUMOS14, ActivityNetv1.2, and a conjoint-action subset)\ndemonstrate that JCDNet achieves competitive performance against the\nstate-of-the-art methods. Keywords: weakly-supervised learning, temporal action\nlocalization, conjoint action\n", "date": "2023-03-30", "categories": []}
{"id": "2303.16268", "title": "TimeBalance: Temporally-Invariant and Temporally-Distinctive Video\n  Representations for Semi-Supervised Action Recognition", "abstract": "  Semi-Supervised Learning can be more beneficial for the video domain compared\nto images because of its higher annotation cost and dimensionality. Besides,\nany video understanding task requires reasoning over both spatial and temporal\ndimensions. In order to learn both the static and motion related features for\nthe semi-supervised action recognition task, existing methods rely on hard\ninput inductive biases like using two-modalities (RGB and Optical-flow) or\ntwo-stream of different playback rates. Instead of utilizing unlabeled videos\nthrough diverse input streams, we rely on self-supervised video\nrepresentations, particularly, we utilize temporally-invariant and\ntemporally-distinctive representations. We observe that these representations\ncomplement each other depending on the nature of the action. Based on this\nobservation, we propose a student-teacher semi-supervised learning framework,\nTimeBalance, where we distill the knowledge from a temporally-invariant and a\ntemporally-distinctive teacher. Depending on the nature of the unlabeled video,\nwe dynamically combine the knowledge of these two teachers based on a novel\ntemporal similarity-based reweighting scheme. Our method achieves\nstate-of-the-art performance on three action recognition benchmarks: UCF101,\nHMDB51, and Kinetics400. Code: https://github.com/DAVEISHAN/TimeBalance\n", "date": "2023-03-28", "categories": []}
{"id": "2303.16867", "title": "A Video-based End-to-end Pipeline for Non-nutritive Sucking Action\n  Recognition and Segmentation in Young Infants", "abstract": "  We present an end-to-end computer vision pipeline to detect non-nutritive\nsucking (NNS) -- an infant sucking pattern with no nutrition delivered -- as a\npotential biomarker for developmental delays, using off-the-shelf baby monitor\nvideo footage. One barrier to clinical (or algorithmic) assessment of NNS stems\nfrom its sparsity, requiring experts to wade through hours of footage to find\nminutes of relevant activity. Our NNS activity segmentation algorithm solves\nthis problem by identifying periods of NNS with high certainty -- up to 94.0\\%\naverage precision and 84.9\\% average recall across 30 heterogeneous 60 s clips,\ndrawn from our manually annotated NNS clinical in-crib dataset of 183 hours of\novernight baby monitor footage from 19 infants. Our method is based on an\nunderlying NNS action recognition algorithm, which uses spatiotemporal deep\nlearning networks and infant-specific pose estimation, achieving 94.9\\%\naccuracy in binary classification of 960 2.5 s balanced NNS vs. non-NNS clips.\nTested on our second, independent, and public NNS in-the-wild dataset, NNS\nrecognition classification reaches 92.3\\% accuracy, and NNS segmentation\nachieves 90.8\\% precision and 84.2\\% recall.\n", "date": "2023-03-29", "categories": []}
{"id": "2211.14053", "title": "Re^2TAL: Rewiring Pretrained Video Backbones for Reversible Temporal\n  Action Localization", "abstract": "  Temporal action localization (TAL) requires long-form reasoning to predict\nactions of various durations and complex content. Given limited GPU memory,\ntraining TAL end to end (i.e., from videos to predictions) on long videos is a\nsignificant challenge. Most methods can only train on pre-extracted features\nwithout optimizing them for the localization problem, consequently limiting\nlocalization performance. In this work, to extend the potential in TAL\nnetworks, we propose a novel end-to-end method Re2TAL, which rewires pretrained\nvideo backbones for reversible TAL. Re2TAL builds a backbone with reversible\nmodules, where the input can be recovered from the output such that the bulky\nintermediate activations can be cleared from memory during training. Instead of\ndesigning one single type of reversible module, we propose a network rewiring\nmechanism, to transform any module with a residual connection to a reversible\nmodule without changing any parameters. This provides two benefits: (1) a large\nvariety of reversible networks are easily obtained from existing and even\nfuture model designs, and (2) the reversible models require much less training\neffort as they reuse the pre-trained parameters of their original\nnon-reversible versions. Re2TAL, only using the RGB modality, reaches 37.01%\naverage mAP on ActivityNet-v1.3, a new state-of-the-art record, and mAP 64.9%\nat tIoU=0.5 on THUMOS-14, outperforming all other RGB-only methods.\n", "date": "2022-11-25", "categories": []}
{"id": "2303.03729", "title": "Learning Discriminative Representations for Skeleton Based Action\n  Recognition", "abstract": "  Human action recognition aims at classifying the category of human action\nfrom a segment of a video. Recently, people have dived into designing GCN-based\nmodels to extract features from skeletons for performing this task, because\nskeleton representations are much more efficient and robust than other\nmodalities such as RGB frames. However, when employing the skeleton data, some\nimportant clues like related items are also discarded. It results in some\nambiguous actions that are hard to be distinguished and tend to be\nmisclassified. To alleviate this problem, we propose an auxiliary feature\nrefinement head (FR Head), which consists of spatial-temporal decoupling and\ncontrastive feature refinement, to obtain discriminative representations of\nskeletons. Ambiguous samples are dynamically discovered and calibrated in the\nfeature space. Furthermore, FR Head could be imposed on different stages of\nGCNs to build a multi-level refinement for stronger supervision. Extensive\nexperiments are conducted on NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets.\nOur proposed models obtain competitive results from state-of-the-art methods\nand can help to discriminate those ambiguous samples. Codes are available at\nhttps://github.com/zhysora/FR-Head.\n", "date": "2023-03-07", "categories": []}
{"id": "2303.15467", "title": "Enlarging Instance-specific and Class-specific Information for Open-set\n  Action Recognition", "abstract": "  Open-set action recognition is to reject unknown human action cases which are\nout of the distribution of the training set. Existing methods mainly focus on\nlearning better uncertainty scores but dismiss the importance of feature\nrepresentations. We find that features with richer semantic diversity can\nsignificantly improve the open-set performance under the same uncertainty\nscores. In this paper, we begin with analyzing the feature representation\nbehavior in the open-set action recognition (OSAR) problem based on the\ninformation bottleneck (IB) theory, and propose to enlarge the\ninstance-specific (IS) and class-specific (CS) information contained in the\nfeature for better performance. To this end, a novel Prototypical Similarity\nLearning (PSL) framework is proposed to keep the instance variance within the\nsame class to retain more IS information. Besides, we notice that unknown\nsamples sharing similar appearances to known samples are easily misclassified\nas known classes. To alleviate this issue, video shuffling is further\nintroduced in our PSL to learn distinct temporal information between original\nand shuffled samples, which we find enlarges the CS information. Extensive\nexperiments demonstrate that the proposed PSL can significantly boost both the\nopen-set and closed-set performance and achieves state-of-the-art results on\nmultiple benchmarks. Code is available at https://github.com/Jun-CEN/PSL.\n", "date": "2023-03-25", "categories": []}
{"id": "2303.15879", "title": "STMixer: A One-Stage Sparse Action Detector", "abstract": "  Traditional video action detectors typically adopt the two-stage pipeline,\nwhere a person detector is first employed to generate actor boxes and then 3D\nRoIAlign is used to extract actor-specific features for classification. This\ndetection paradigm requires multi-stage training and inference, and cannot\ncapture context information outside the bounding box. Recently, a few\nquery-based action detectors are proposed to predict action instances in an\nend-to-end manner. However, they still lack adaptability in feature sampling\nand decoding, thus suffering from the issues of inferior performance or slower\nconvergence. In this paper, we propose a new one-stage sparse action detector,\ntermed STMixer. STMixer is based on two core designs. First, we present a\nquery-based adaptive feature sampling module, which endows our STMixer with the\nflexibility of mining a set of discriminative features from the entire\nspatiotemporal domain. Second, we devise a dual-branch feature mixing module,\nwhich allows our STMixer to dynamically attend to and mix video features along\nthe spatial and the temporal dimension respectively for better feature\ndecoding. Coupling these two designs with a video backbone yields an efficient\nend-to-end action detector. Without bells and whistles, our STMixer obtains the\nstate-of-the-art results on the datasets of AVA, UCF101-24, and JHMDB.\n", "date": "2023-03-28", "categories": []}
{"id": "2303.16084", "title": "Rethinking matching-based few-shot action recognition", "abstract": "  Few-shot action recognition, i.e. recognizing new action classes given only a\nfew examples, benefits from incorporating temporal information. Prior work\neither encodes such information in the representation itself and learns\nclassifiers at test time, or obtains frame-level features and performs pairwise\ntemporal matching. We first evaluate a number of matching-based approaches\nusing features from spatio-temporal backbones, a comparison missing from the\nliterature, and show that the gap in performance between simple baselines and\nmore complicated methods is significantly reduced. Inspired by this, we propose\nChamfer++, a non-temporal matching function that achieves state-of-the-art\nresults in few-shot action recognition. We show that, when starting from\ntemporal features, our parameter-free and interpretable approach can outperform\nall other matching-based and classifier methods for one-shot action recognition\non three common datasets without using temporal information in the matching\nstage. Project page: https://jbertrand89.github.io/matching-based-fsar\n", "date": "2023-03-28", "categories": []}
{"id": "2303.16118", "title": "CycleACR: Cycle Modeling of Actor-Context Relations for Video Action\n  Detection", "abstract": "  The relation modeling between actors and scene context advances video action\ndetection where the correlation of multiple actors makes their action\nrecognition challenging. Existing studies model each actor and scene relation\nto improve action recognition. However, the scene variations and background\ninterference limit the effectiveness of this relation modeling. In this paper,\nwe propose to select actor-related scene context, rather than directly leverage\nraw video scenario, to improve relation modeling. We develop a Cycle\nActor-Context Relation network (CycleACR) where there is a symmetric graph that\nmodels the actor and context relations in a bidirectional form. Our CycleACR\nconsists of the Actor-to-Context Reorganization (A2C-R) that collects actor\nfeatures for context feature reorganizations, and the Context-to-Actor\nEnhancement (C2A-E) that dynamically utilizes reorganized context features for\nactor feature enhancement. Compared to existing designs that focus on C2A-E,\nour CycleACR introduces A2C-R for a more effective relation modeling. This\nmodeling advances our CycleACR to achieve state-of-the-art performance on two\npopular action detection datasets (i.e., AVA and UCF101-24). We also provide\nablation studies and visualizations as well to show how our cycle actor-context\nrelation modeling improves video action detection. Code is available at\nhttps://github.com/MCG-NJU/CycleACR.\n", "date": "2023-03-28", "categories": []}
{"id": "2207.01297", "title": "Revisiting Classifier: Transferring Vision-Language Models for Video\n  Recognition", "abstract": "  Transferring knowledge from task-agnostic pre-trained deep models for\ndownstream tasks is an important topic in computer vision research. Along with\nthe growth of computational capacity, we now have open-source vision-language\npre-trained models in large scales of the model architecture and amount of\ndata. In this study, we focus on transferring knowledge for video\nclassification tasks. Conventional methods randomly initialize the linear\nclassifier head for vision classification, but they leave the usage of the text\nencoder for downstream visual recognition tasks undiscovered. In this paper, we\nrevise the role of the linear classifier and replace the classifier with the\ndifferent knowledge from pre-trained model. We utilize the well-pretrained\nlanguage model to generate good semantic target for efficient transferring\nlearning. The empirical study shows that our method improves both the\nperformance and the training speed of video classification, with a negligible\nchange in the model. Our simple yet effective tuning paradigm achieves\nstate-of-the-art performance and efficient training on various video\nrecognition scenarios, i.e., zero-shot, few-shot, general recognition. In\nparticular, our paradigm achieves the state-of-the-art accuracy of 87.8% on\nKinetics-400, and also surpasses previous methods by 20~50% absolute top-1\naccuracy under zero-shot, few-shot settings on five popular video datasets.\nCode and models can be found at https://github.com/whwu95/Text4Vis .\n", "date": "2022-07-04", "categories": []}
{"id": "2207.09644", "title": "Hierarchically Self-Supervised Transformer for Human Skeleton\n  Representation Learning", "abstract": "  Despite the success of fully-supervised human skeleton sequence modeling,\nutilizing self-supervised pre-training for skeleton sequence representation\nlearning has been an active field because acquiring task-specific skeleton\nannotations at large scales is difficult. Recent studies focus on learning\nvideo-level temporal and discriminative information using contrastive learning,\nbut overlook the hierarchical spatial-temporal nature of human skeletons.\nDifferent from such superficial supervision at the video level, we propose a\nself-supervised hierarchical pre-training scheme incorporated into a\nhierarchical Transformer-based skeleton sequence encoder (Hi-TRS), to\nexplicitly capture spatial, short-term, and long-term temporal dependencies at\nframe, clip, and video levels, respectively. To evaluate the proposed\nself-supervised pre-training scheme with Hi-TRS, we conduct extensive\nexperiments covering three skeleton-based downstream tasks including action\nrecognition, action detection, and motion prediction. Under both supervised and\nsemi-supervised evaluation protocols, our method achieves the state-of-the-art\nperformance. Additionally, we demonstrate that the prior knowledge learned by\nour model in the pre-training stage has strong transfer capability for\ndifferent downstream tasks.\n", "date": "2022-07-20", "categories": []}
{"id": "2211.14905", "title": "Multi-Modal Few-Shot Temporal Action Detection", "abstract": "  Few-shot (FS) and zero-shot (ZS) learning are two different approaches for\nscaling temporal action detection (TAD) to new classes. The former adapts a\npretrained vision model to a new task represented by as few as a single video\nper class, whilst the latter requires no training examples by exploiting a\nsemantic description of the new class. In this work, we introduce a new\nmulti-modality few-shot (MMFS) TAD problem, which can be considered as a\nmarriage of FS-TAD and ZS-TAD by leveraging few-shot support videos and new\nclass names jointly. To tackle this problem, we further introduce a novel\nMUlti-modality PromPt mETa-learning (MUPPET) method. This is enabled by\nefficiently bridging pretrained vision and language models whilst maximally\nreusing already learned capacity. Concretely, we construct multi-modal prompts\nby mapping support videos into the textual token space of a vision-language\nmodel using a meta-learned adapter-equipped visual semantics tokenizer. To\ntackle large intra-class variation, we further design a query feature\nregulation scheme. Extensive experiments on ActivityNetv1.3 and THUMOS14\ndemonstrate that our MUPPET outperforms state-of-the-art alternative methods,\noften by a large margin. We also show that our MUPPET can be easily extended to\ntackle the few-shot object detection problem and again achieves the\nstate-of-the-art performance on MS-COCO dataset. The code will be available in\nhttps://github.com/sauradip/MUPPET\n", "date": "2022-11-27", "categories": []}
{"id": "2301.00182", "title": "Bidirectional Cross-Modal Knowledge Exploration for Video Recognition\n  with Pre-trained Vision-Language Models", "abstract": "  Vision-language models (VLMs) pre-trained on large-scale image-text pairs\nhave demonstrated impressive transferability on various visual tasks.\nTransferring knowledge from such powerful VLMs is a promising direction for\nbuilding effective video recognition models. However, current exploration in\nthis field is still limited. We believe that the greatest value of pre-trained\nVLMs lies in building a bridge between visual and textual domains. In this\npaper, we propose a novel framework called BIKE, which utilizes the cross-modal\nbridge to explore bidirectional knowledge: i) We introduce the Video Attribute\nAssociation mechanism, which leverages the Video-to-Text knowledge to generate\ntextual auxiliary attributes for complementing video recognition. ii) We also\npresent a Temporal Concept Spotting mechanism that uses the Text-to-Video\nexpertise to capture temporal saliency in a parameter-free manner, leading to\nenhanced video representation. Extensive studies on six popular video datasets,\nincluding Kinetics-400 & 600, UCF-101, HMDB-51, ActivityNet and Charades, show\nthat our method achieves state-of-the-art performance in various recognition\nscenarios, such as general, zero-shot, and few-shot video recognition. Our best\nmodel achieves a state-of-the-art accuracy of 88.6% on the challenging\nKinetics-400 using the released CLIP model. The code is available at\nhttps://github.com/whwu95/BIKE .\n", "date": "2022-12-31", "categories": []}
{"id": "2303.14358", "title": "Multi-view knowledge distillation transformer for human action\n  recognition", "abstract": "  Recently, Transformer-based methods have been utilized to improve the\nperformance of human action recognition. However, most of these studies assume\nthat multi-view data is complete, which may not always be the case in\nreal-world scenarios. Therefore, this paper presents a novel Multi-view\nKnowledge Distillation Transformer (MKDT) framework that consists of a teacher\nnetwork and a student network. This framework aims to handle incomplete human\naction problems in real-world applications. Specifically, the multi-view\nknowledge distillation transformer uses a hierarchical vision transformer with\nshifted windows to capture more spatial-temporal information. Experimental\nresults demonstrate that our framework outperforms the CNN-based method on\nthree public datasets.\n", "date": "2023-03-25", "categories": []}
{"id": "2303.14474", "title": "3Mformer: Multi-order Multi-mode Transformer for Skeletal Action\n  Recognition", "abstract": "  Many skeletal action recognition models use GCNs to represent the human body\nby 3D body joints connected body parts. GCNs aggregate one- or few-hop graph\nneighbourhoods, and ignore the dependency between not linked body joints. We\npropose to form hypergraph to model hyper-edges between graph nodes (e.g.,\nthird- and fourth-order hyper-edges capture three and four nodes) which help\ncapture higher-order motion patterns of groups of body joints. We split action\nsequences into temporal blocks, Higher-order Transformer (HoT) produces\nembeddings of each temporal block based on (i) the body joints, (ii) pairwise\nlinks of body joints and (iii) higher-order hyper-edges of skeleton body\njoints. We combine such HoT embeddings of hyper-edges of orders 1, ..., r by a\nnovel Multi-order Multi-mode Transformer (3Mformer) with two modules whose\norder can be exchanged to achieve coupled-mode attention on coupled-mode tokens\nbased on 'channel-temporal block', 'order-channel-body joint',\n'channel-hyper-edge (any order)' and 'channel-only' pairs. The first module,\ncalled Multi-order Pooling (MP), additionally learns weighted aggregation along\nthe hyper-edge mode, whereas the second module, Temporal block Pooling (TP),\naggregates along the temporal block mode. Our end-to-end trainable network\nyields state-of-the-art results compared to GCN-, transformer- and\nhypergraph-based counterparts.\n", "date": "2023-03-25", "categories": []}
{"id": "2303.14817", "title": "Frame Flexible Network", "abstract": "  Existing video recognition algorithms always conduct different training\npipelines for inputs with different frame numbers, which requires repetitive\ntraining operations and multiplying storage costs. If we evaluate the model\nusing other frames which are not used in training, we observe the performance\nwill drop significantly (see Fig.1), which is summarized as Temporal Frequency\nDeviation phenomenon. To fix this issue, we propose a general framework, named\nFrame Flexible Network (FFN), which not only enables the model to be evaluated\nat different frames to adjust its computation, but also reduces the memory\ncosts of storing multiple models significantly. Concretely, FFN integrates\nseveral sets of training sequences, involves Multi-Frequency Alignment (MFAL)\nto learn temporal frequency invariant representations, and leverages\nMulti-Frequency Adaptation (MFAD) to further strengthen the representation\nabilities. Comprehensive empirical validations using various architectures and\npopular benchmarks solidly demonstrate the effectiveness and generalization of\nFFN (e.g., 7.08/5.15/2.17% performance gain at Frame 4/8/16 on\nSomething-Something V1 dataset over Uniformer). Code is available at\nhttps://github.com/BeSpontaneous/FFN.\n", "date": "2023-03-26", "categories": []}
{"id": "2303.15270", "title": "Unified Keypoint-based Action Recognition Framework via Structured\n  Keypoint Pooling", "abstract": "  This paper simultaneously addresses three limitations associated with\nconventional skeleton-based action recognition; skeleton detection and tracking\nerrors, poor variety of the targeted actions, as well as person-wise and\nframe-wise action recognition. A point cloud deep-learning paradigm is\nintroduced to the action recognition, and a unified framework along with a\nnovel deep neural network architecture called Structured Keypoint Pooling is\nproposed. The proposed method sparsely aggregates keypoint features in a\ncascaded manner based on prior knowledge of the data structure (which is\ninherent in skeletons), such as the instances and frames to which each keypoint\nbelongs, and achieves robustness against input errors. Its less constrained and\ntracking-free architecture enables time-series keypoints consisting of human\nskeletons and nonhuman object contours to be efficiently treated as an input 3D\npoint cloud and extends the variety of the targeted action. Furthermore, we\npropose a Pooling-Switching Trick inspired by Structured Keypoint Pooling. This\ntrick switches the pooling kernels between the training and inference phases to\ndetect person-wise and frame-wise actions in a weakly supervised manner using\nonly video-level action labels. This trick enables our training scheme to\nnaturally introduce novel data augmentation, which mixes multiple point clouds\nextracted from different videos. In the experiments, we comprehensively verify\nthe effectiveness of the proposed method against the limitations, and the\nmethod outperforms state-of-the-art skeleton-based action recognition and\nspatio-temporal action localization methods.\n", "date": "2023-03-27", "categories": []}
{"id": "2211.09590", "title": "Hypergraph Transformer for Skeleton-based Action Recognition", "abstract": "  Skeleton-based action recognition aims to recognize human actions given human\njoint coordinates with skeletal interconnections. By defining a graph with\njoints as vertices and their natural connections as edges, previous works\nsuccessfully adopted Graph Convolutional networks (GCNs) to model joint\nco-occurrences and achieved superior performance. More recently, a limitation\nof GCNs is identified, i.e., the topology is fixed after training. To relax\nsuch a restriction, Self-Attention (SA) mechanism has been adopted to make the\ntopology of GCNs adaptive to the input, resulting in the state-of-the-art\nhybrid models. Concurrently, attempts with plain Transformers have also been\nmade, but they still lag behind state-of-the-art GCN-based methods due to the\nlack of structural prior. Unlike hybrid models, we propose a more elegant\nsolution to incorporate the bone connectivity into Transformer via a graph\ndistance embedding. Our embedding retains the information of skeletal structure\nduring training, whereas GCNs merely use it for initialization. More\nimportantly, we reveal an underlying issue of graph models in general, i.e.,\npairwise aggregation essentially ignores the high-order kinematic dependencies\nbetween body joints. To fill this gap, we propose a new self-attention (SA)\nmechanism on hypergraph, termed Hypergraph Self-Attention (HyperSA), to\nincorporate intrinsic higher-order relations into the model. We name the\nresulting model Hyperformer, and it beats state-of-the-art graph models w.r.t.\naccuracy and efficiency on NTU RGB+D, NTU RGB+D 120, and Northwestern-UCLA\ndatasets.\n", "date": "2022-11-17", "categories": []}
{"id": "2303.12698", "title": "Open Set Action Recognition via Multi-Label Evidential Learning", "abstract": "  Existing methods for open-set action recognition focus on novelty detection\nthat assumes video clips show a single action, which is unrealistic in the real\nworld. We propose a new method for open set action recognition and novelty\ndetection via MUlti-Label Evidential learning (MULE), that goes beyond previous\nnovel action detection methods by addressing the more general problems of\nsingle or multiple actors in the same scene, with simultaneous action(s) by any\nactor. Our Beta Evidential Neural Network estimates multi-action uncertainty\nwith Beta densities based on actor-context-object relation representations. An\nevidence debiasing constraint is added to the objective function for\noptimization to reduce the static bias of video representations, which can\nincorrectly correlate predictions and static cues. We develop a learning\nalgorithm based on a primal-dual average scheme update to optimize the proposed\nproblem. Theoretical analysis of the optimization algorithm demonstrates the\nconvergence of the primal solution sequence and bounds for both the loss\nfunction and the debiasing constraint. Uncertainty and belief-based novelty\nestimation mechanisms are formulated to detect novel actions. Extensive\nexperiments on two real-world video datasets show that our proposed approach\nachieves promising performance in single/multi-actor, single/multi-action\nsettings.\n", "date": "2023-02-27", "categories": []}
{"id": "2207.13306", "title": "Object-ABN: Learning to Generate Sharp Attention Maps for Action\n  Recognition", "abstract": "  In this paper we propose an extension of the Attention Branch Network (ABN)\nby using instance segmentation for generating sharper attention maps for action\nrecognition. Methods for visual explanation such as Grad-CAM usually generate\nblurry maps which are not intuitive for humans to understand, particularly in\nrecognizing actions of people in videos. Our proposed method, Object-ABN,\ntackles this issue by introducing a new mask loss that makes the generated\nattention maps close to the instance segmentation result. Further the PC loss\nand multiple attention maps are introduced to enhance the sharpness of the maps\nand improve the performance of classification. Experimental results with UCF101\nand SSv2 shows that the generated maps by the proposed method are much clearer\nqualitatively and quantitatively than those of the original ABN.\n", "date": "2022-07-27", "categories": []}
{"id": "2210.11035", "title": "PointTAD: Multi-Label Temporal Action Detection with Learnable Query\n  Points", "abstract": "  Traditional temporal action detection (TAD) usually handles untrimmed videos\nwith small number of action instances from a single label (e.g., ActivityNet,\nTHUMOS). However, this setting might be unrealistic as different classes of\nactions often co-occur in practice. In this paper, we focus on the task of\nmulti-label temporal action detection that aims to localize all action\ninstances from a multi-label untrimmed video. Multi-label TAD is more\nchallenging as it requires for fine-grained class discrimination within a\nsingle video and precise localization of the co-occurring instances. To\nmitigate this issue, we extend the sparse query-based detection paradigm from\nthe traditional TAD and propose the multi-label TAD framework of PointTAD.\nSpecifically, our PointTAD introduces a small set of learnable query points to\nrepresent the important frames of each action instance. This point-based\nrepresentation provides a flexible mechanism to localize the discriminative\nframes at boundaries and as well the important frames inside the action.\nMoreover, we perform the action decoding process with the Multi-level\nInteractive Module to capture both point-level and instance-level action\nsemantics. Finally, our PointTAD employs an end-to-end trainable framework\nsimply based on RGB input for easy deployment. We evaluate our proposed method\non two popular benchmarks and introduce the new metric of detection-mAP for\nmulti-label TAD. Our model outperforms all previous methods by a large margin\nunder the detection-mAP metric, and also achieves promising results under the\nsegmentation-mAP metric. Code is available at\nhttps://github.com/MCG-NJU/PointTAD.\n", "date": "2022-10-20", "categories": []}
{"id": "2211.15393", "title": "Video Test-Time Adaptation for Action Recognition", "abstract": "  Although action recognition systems can achieve top performance when\nevaluated on in-distribution test points, they are vulnerable to unanticipated\ndistribution shifts in test data. However, test-time adaptation of video action\nrecognition models against common distribution shifts has so far not been\ndemonstrated. We propose to address this problem with an approach tailored to\nspatio-temporal models that is capable of adaptation on a single video sample\nat a step. It consists in a feature distribution alignment technique that\naligns online estimates of test set statistics towards the training statistics.\nWe further enforce prediction consistency over temporally augmented views of\nthe same test video sample. Evaluations on three benchmark action recognition\ndatasets show that our proposed technique is architecture-agnostic and able to\nsignificantly boost the performance on both, the state of the art convolutional\narchitecture TANet and the Video Swin Transformer. Our proposed method\ndemonstrates a substantial performance gain over existing test-time adaptation\napproaches in both evaluations of a single distribution shift and the\nchallenging case of random distribution shifts. Code will be available at\n\\url{https://github.com/wlin-at/ViTTA}.\n", "date": "2022-11-24", "categories": []}
{"id": "2303.10904", "title": "Actionlet-Dependent Contrastive Learning for Unsupervised Skeleton-Based\n  Action Recognition", "abstract": "  The self-supervised pretraining paradigm has achieved great success in\nskeleton-based action recognition. However, these methods treat the motion and\nstatic parts equally, and lack an adaptive design for different parts, which\nhas a negative impact on the accuracy of action recognition. To realize the\nadaptive action modeling of both parts, we propose an Actionlet-Dependent\nContrastive Learning method (ActCLR). The actionlet, defined as the\ndiscriminative subset of the human skeleton, effectively decomposes motion\nregions for better action modeling. In detail, by contrasting with the static\nanchor without motion, we extract the motion region of the skeleton data, which\nserves as the actionlet, in an unsupervised manner. Then, centering on\nactionlet, a motion-adaptive data transformation method is built. Different\ndata transformations are applied to actionlet and non-actionlet regions to\nintroduce more diversity while maintaining their own characteristics.\nMeanwhile, we propose a semantic-aware feature pooling method to build feature\nrepresentations among motion and static regions in a distinguished manner.\nExtensive experiments on NTU RGB+D and PKUMMD show that the proposed method\nachieves remarkable action recognition performance. More visualization and\nquantitative experiments demonstrate the effectiveness of our method. Our\nproject website is available at https://langlandslin.github.io/projects/ActCLR/\n", "date": "2023-03-20", "categories": []}
{"id": "2303.11732", "title": "Multi-modal Prompting for Low-Shot Temporal Action Localization", "abstract": "  In this paper, we consider the problem of temporal action localization under\nlow-shot (zero-shot & few-shot) scenario, with the goal of detecting and\nclassifying the action instances from arbitrary categories within some\nuntrimmed videos, even not seen at training time. We adopt a Transformer-based\ntwo-stage action localization architecture with class-agnostic action proposal,\nfollowed by open-vocabulary classification. We make the following\ncontributions. First, to compensate image-text foundation models with temporal\nmotions, we improve category-agnostic action proposal by explicitly aligning\nembeddings of optical flows, RGB and texts, which has largely been ignored in\nexisting low-shot methods. Second, to improve open-vocabulary action\nclassification, we construct classifiers with strong discriminative power,\ni.e., avoid lexical ambiguities. To be specific, we propose to prompt the\npre-trained CLIP text encoder either with detailed action descriptions\n(acquired from large-scale language models), or visually-conditioned\ninstance-specific prompt vectors. Third, we conduct thorough experiments and\nablation studies on THUMOS14 and ActivityNet1.3, demonstrating the superior\nperformance of our proposed model, outperforming existing state-of-the-art\napproaches by one significant margin.\n", "date": "2023-03-21", "categories": []}
{"id": "2303.12016", "title": "Automatic evaluation of herding behavior in towed fishing gear using\n  end-to-end training of CNN and attention-based networks", "abstract": "  This paper considers the automatic classification of herding behavior in the\ncluttered low-visibility environment that typically surrounds towed fishing\ngear. The paper compares three convolutional and attention-based deep action\nrecognition network architectures trained end-to-end on a small set of video\nsequences captured by a remotely controlled camera and classified by an expert\nin fishing technology. The sequences depict a scene in front of a fishing trawl\nwhere the conventional herding mechanism has been replaced by directed laser\nlight. The goal is to detect the presence of a fish in the sequence and\nclassify whether or not the fish reacts to the lasers. A two-stream CNN model,\na CNN-transformer hybrid, and a pure transformer model were trained end-to-end\nto achieve 63%, 54%, and 60% 10-fold classification accuracy on the three-class\ntask when compared to the human expert. Inspection of the activation maps\nlearned by the three networks raises questions about the attributes of the\nsequences the models may be learning, specifically whether changes in viewpoint\nintroduced by human camera operators that affect the position of laser lines in\nthe video frames may interfere with the classification. This underlines the\nimportance of careful experimental design when capturing scientific data for\nautomatic end-to-end evaluation and the usefulness of inspecting the trained\nmodels.\n", "date": "2023-03-21", "categories": []}
{"id": "2303.11100", "title": "A Multi-Task Deep Learning Approach for Sensor-based Human Activity\n  Recognition and Segmentation", "abstract": "  Sensor-based human activity segmentation and recognition are two important\nand challenging problems in many real-world applications and they have drawn\nincreasing attention from the deep learning community in recent years. Most of\nthe existing deep learning works were designed based on pre-segmented sensor\nstreams and they have treated activity segmentation and recognition as two\nseparate tasks. In practice, performing data stream segmentation is very\nchallenging. We believe that both activity segmentation and recognition may\nconvey unique information which can complement each other to improve the\nperformance of the two tasks. In this paper, we firstly proposes a new\nmultitask deep neural network to solve the two tasks simultaneously. The\nproposed neural network adopts selective convolution and features multiscale\nwindows to segment activities of long or short time durations. First, multiple\nwindows of different scales are generated to center on each unit of the feature\nsequence. Then, the model is trained to predict, for each window, the activity\nclass and the offset to the true activity boundaries. Finally, overlapping\nwindows are filtered out by non-maximum suppression, and adjacent windows of\nthe same activity are concatenated to complete the segmentation task. Extensive\nexperiments were conducted on eight popular benchmarking datasets, and the\nresults show that our proposed method outperforms the state-of-the-art methods\nboth for activity recognition and segmentation.\n", "date": "2023-03-20", "categories": []}
{"id": "2303.09558", "title": "Classification of Primitive Manufacturing Tasks from Filtered Event Data", "abstract": "  Collaborative robots are increasingly present in industry to support human\nactivities. However, to make the human-robot collaborative process more\neffective, there are several challenges to be addressed. Collaborative robotic\nsystems need to be aware of the human activities to (1) anticipate\ncollaborative/assistive actions, (2) learn by demonstration, and (3) activate\nsafety procedures in shared workspace. This study proposes an action\nclassification system to recognize primitive assembly tasks from human motion\nevents data captured by a Dynamic and Active-pixel Vision Sensor (DAVIS).\nSeveral filters are compared and combined to remove event data noise. Task\npatterns are classified from a continuous stream of event data using advanced\ndeep learning and recurrent networks to classify spatial and temporal features.\nExperiments were conducted on a novel dataset, the dataset of manufacturing\ntasks (DMT22), featuring 5 classes of representative manufacturing primitives\n(PickUp, Place, Screw, Hold, Idle) from 5 participants. Results show that the\nproposed filters remove about 65\\% of all events (noise) per recording,\nconducting to a classification accuracy up to 99,37\\% for subjects that trained\nthe system and 97.08\\% for new subjects. Data from a left-handed subject were\nsuccessfully classified using only right-handed training data. These results\nare object independent.\n", "date": "2023-03-15", "categories": []}
{"id": "2303.07347", "title": "TriDet: Temporal Action Detection with Relative Boundary Modeling", "abstract": "  In this paper, we present a one-stage framework TriDet for temporal action\ndetection. Existing methods often suffer from imprecise boundary predictions\ndue to the ambiguous action boundaries in videos. To alleviate this problem, we\npropose a novel Trident-head to model the action boundary via an estimated\nrelative probability distribution around the boundary. In the feature pyramid\nof TriDet, we propose an efficient Scalable-Granularity Perception (SGP) layer\nto mitigate the rank loss problem of self-attention that takes place in the\nvideo features and aggregate information across different temporal\ngranularities. Benefiting from the Trident-head and the SGP-based feature\npyramid, TriDet achieves state-of-the-art performance on three challenging\nbenchmarks: THUMOS14, HACS and EPIC-KITCHEN 100, with lower computational\ncosts, compared to previous methods. For example, TriDet hits an average mAP of\n$69.3\\%$ on THUMOS14, outperforming the previous best by $2.5\\%$, but with only\n$74.6\\%$ of its latency. The code is released to\nhttps://github.com/sssste/TriDet.\n", "date": "2023-03-13", "categories": []}
{"id": "2303.08450", "title": "PoseRAC: Pose Saliency Transformer for Repetitive Action Counting", "abstract": "  This paper presents a significant contribution to the field of repetitive\naction counting through the introduction of a new approach called Pose Saliency\nRepresentation. The proposed method efficiently represents each action using\nonly two salient poses instead of redundant frames, which significantly reduces\nthe computational cost while improving the performance. Moreover, we introduce\na pose-level method, PoseRAC, which is based on this representation and\nachieves state-of-the-art performance on two new version datasets by using Pose\nSaliency Annotation to annotate salient poses for training. Our lightweight\nmodel is highly efficient, requiring only 20 minutes for training on a GPU, and\ninfers nearly 10x faster compared to previous methods. In addition, our\napproach achieves a substantial improvement over the previous state-of-the-art\nTransRAC, achieving an OBO metric of 0.56 compared to 0.29 of TransRAC. The\ncode and new dataset are available at https://github.com/MiracleDance/PoseRAC\nfor further research and experimentation, making our proposed approach highly\naccessible to the research community.\n", "date": "2023-03-15", "categories": []}
{"id": "2303.08920", "title": "EgoViT: Pyramid Video Transformer for Egocentric Action Recognition", "abstract": "  Capturing interaction of hands with objects is important to autonomously\ndetect human actions from egocentric videos. In this work, we present a pyramid\nvideo transformer with a dynamic class token generator for egocentric action\nrecognition. Different from previous video transformers, which use the same\nstatic embedding as the class token for diverse inputs, we propose a dynamic\nclass token generator that produces a class token for each input video by\nanalyzing the hand-object interaction and the related motion information. The\ndynamic class token can diffuse such information to the entire model by\ncommunicating with other informative tokens in the subsequent transformer\nlayers. With the dynamic class token, dissimilarity between videos can be more\nprominent, which helps the model distinguish various inputs. In addition,\ntraditional video transformers explore temporal features globally, which\nrequires large amounts of computation. However, egocentric videos often have a\nlarge amount of background scene transition, which causes discontinuities\nacross distant frames. In this case, blindly reducing the temporal sampling\nrate will risk losing crucial information. Hence, we also propose a pyramid\narchitecture to hierarchically process the video from short-term high rate to\nlong-term low rate. With the proposed architecture, we significantly reduce the\ncomputational cost as well as the memory requirement without sacrificing from\nthe model performance. We perform comparisons with different baseline video\ntransformers on the EPIC-KITCHENS-100 and EGTEA Gaze+ datasets. Both\nquantitative and qualitative results show that the proposed model can\nefficiently improve the performance for egocentric action recognition.\n", "date": "2023-03-15", "categories": []}
{"id": "2303.09055", "title": "TemporalMaxer: Maximize Temporal Context with only Max Pooling for\n  Temporal Action Localization", "abstract": "  Temporal Action Localization (TAL) is a challenging task in video\nunderstanding that aims to identify and localize actions within a video\nsequence. Recent studies have emphasized the importance of applying long-term\ntemporal context modeling (TCM) blocks to the extracted video clip features\nsuch as employing complex self-attention mechanisms. In this paper, we present\nthe simplest method ever to address this task and argue that the extracted\nvideo clip features are already informative to achieve outstanding performance\nwithout sophisticated architectures. To this end, we introduce TemporalMaxer,\nwhich minimizes long-term temporal context modeling while maximizing\ninformation from the extracted video clip features with a basic,\nparameter-free, and local region operating max-pooling block. Picking out only\nthe most critical information for adjacent and local clip embeddings, this\nblock results in a more efficient TAL model. We demonstrate that TemporalMaxer\noutperforms other state-of-the-art methods that utilize long-term TCM such as\nself-attention on various TAL datasets while requiring significantly fewer\nparameters and computational resources. The code for our approach is publicly\navailable at https://github.com/TuanTNG/TemporalMaxer\n", "date": "2023-03-16", "categories": []}
{"id": "2303.07789", "title": "Activity Recognition From Newborn Resuscitation Videos", "abstract": "  Objective: Birth asphyxia is one of the leading causes of neonatal deaths. A\nkey for survival is performing immediate and continuous quality newborn\nresuscitation. A dataset of recorded signals during newborn resuscitation,\nincluding videos, has been collected in Haydom, Tanzania, and the aim is to\nanalyze the treatment and its effect on the newborn outcome. An important step\nis to generate timelines of relevant resuscitation activities, including\nventilation, stimulation, suction, etc., during the resuscitation episodes.\nMethods: We propose a two-step deep neural network system, ORAA-net, utilizing\nlow-quality video recordings of resuscitation episodes to do activity\nrecognition during newborn resuscitation. The first step is to detect and track\nrelevant objects using Convolutional Neural Networks (CNN) and post-processing,\nand the second step is to analyze the proposed activity regions from step 1 to\ndo activity recognition using 3D CNNs. Results: The system recognized the\nactivities newborn uncovered, stimulation, ventilation and suction with a mean\nprecision of 77.67 %, a mean recall of 77,64 %, and a mean accuracy of 92.40 %.\nMoreover, the accuracy of the estimated number of Health Care Providers (HCPs)\npresent during the resuscitation episodes was 68.32 %. Conclusion: The results\nindicate that the proposed CNN-based two-step ORAAnet could be used for object\ndetection and activity recognition in noisy low-quality newborn resuscitation\nvideos. Significance: A thorough analysis of the effect the different\nresuscitation activities have on the newborn outcome could potentially allow us\nto optimize treatment guidelines, training, debriefing, and local quality\nimprovement in newborn resuscitation.\n", "date": "2023-03-14", "categories": []}
{"id": "2211.17042", "title": "Spatio-Temporal Crop Aggregation for Video Representation Learning", "abstract": "  We propose Spatio-temporal Crop Aggregation for video representation LEarning\n(SCALE), a novel method that enjoys high scalability at both training and\ninference time. Our model builds long-range video features by learning from\nsets of video clip-level features extracted with a pre-trained backbone. To\ntrain the model, we propose a self-supervised objective consisting of masked\nclip feature prediction. We apply sparsity to both the input, by extracting a\nrandom set of video clips, and to the loss function, by only reconstructing the\nsparse inputs. Moreover, we use dimensionality reduction by working in the\nlatent space of a pre-trained backbone applied to single video clips. These\ntechniques make our method not only extremely efficient to train but also\nhighly effective in transfer learning. We demonstrate that our video\nrepresentation yields state-of-the-art performance with linear, non-linear, and\nKNN probing on common action classification and video understanding datasets.\n", "date": "2022-11-30", "categories": []}
{"id": "2301.03046", "title": "STPrivacy: Spatio-Temporal Privacy-Preserving Action Recognition", "abstract": "  Existing methods of privacy-preserving action recognition (PPAR) mainly focus\non frame-level (spatial) privacy removal through 2D CNNs. Unfortunately, they\nhave two major drawbacks. First, they may compromise temporal dynamics in input\nvideos, which are critical for accurate action recognition. Second, they are\nvulnerable to practical attacking scenarios where attackers probe for privacy\nfrom an entire video rather than individual frames. To address these issues, we\npropose a novel framework STPrivacy to perform video-level PPAR. For the first\ntime, we introduce vision Transformers into PPAR by treating a video as a\ntubelet sequence, and accordingly design two complementary mechanisms, i.e.,\nsparsification and anonymization, to remove privacy from a spatio-temporal\nperspective. In specific, our privacy sparsification mechanism applies adaptive\ntoken selection to abandon action-irrelevant tubelets. Then, our anonymization\nmechanism implicitly manipulates the remaining action-tubelets to erase privacy\nin the embedding space through adversarial learning. These mechanisms provide\nsignificant advantages in terms of privacy preservation for human eyes and\naction-privacy trade-off adjustment during deployment. We additionally\ncontribute the first two large-scale PPAR benchmarks, VP-HMDB51 and VP-UCF101,\nto the community. Extensive evaluations on them, as well as two other tasks,\nvalidate the effectiveness and generalization capability of our framework.\n", "date": "2023-01-08", "categories": []}
{"id": "2303.06242", "title": "HYperbolic Self-Paced Learning for Self-Supervised Skeleton-based Action\n  Representations", "abstract": "  Self-paced learning has been beneficial for tasks where some initial\nknowledge is available, such as weakly supervised learning and domain\nadaptation, to select and order the training sample sequence, from easy to\ncomplex. However its applicability remains unexplored in unsupervised learning,\nwhereby the knowledge of the task matures during training. We propose a novel\nHYperbolic Self-Paced model (HYSP) for learning skeleton-based action\nrepresentations. HYSP adopts self-supervision: it uses data augmentations to\ngenerate two views of the same sample, and it learns by matching one (named\nonline) to the other (the target). We propose to use hyperbolic uncertainty to\ndetermine the algorithmic learning pace, under the assumption that less\nuncertain samples should be more strongly driving the training, with a larger\nweight and pace. Hyperbolic uncertainty is a by-product of the adopted\nhyperbolic neural networks, it matures during training and it comes with no\nextra cost, compared to the established Euclidean SSL framework counterparts.\nWhen tested on three established skeleton-based action recognition datasets,\nHYSP outperforms the state-of-the-art on PKU-MMD I, as well as on 2 out of 3\ndownstream tasks on NTU-60 and NTU-120. Additionally, HYSP only uses positive\npairs and bypasses therefore the complex and computationally-demanding mining\nprocedures required for the negatives in contrastive techniques. Code is\navailable at https://github.com/paolomandica/HYSP.\n", "date": "2023-03-10", "categories": []}
{"id": "2211.03004", "title": "Bringing Online Egocentric Action Recognition into the wild", "abstract": "  To enable a safe and effective human-robot cooperation, it is crucial to\ndevelop models for the identification of human activities. Egocentric vision\nseems to be a viable solution to solve this problem, and therefore many works\nprovide deep learning solutions to infer human actions from first person\nvideos. However, although very promising, most of these do not consider the\nmajor challenges that comes with a realistic deployment, such as the\nportability of the model, the need for real-time inference, and the robustness\nwith respect to the novel domains (i.e., new spaces, users, tasks). With this\npaper, we set the boundaries that egocentric vision models should consider for\nrealistic applications, defining a novel setting of egocentric action\nrecognition in the wild, which encourages researchers to develop novel,\napplications-aware solutions. We also present a new model-agnostic technique\nthat enables the rapid repurposing of existing architectures in this new\ncontext, demonstrating the feasibility to deploy a model on a tiny device\n(Jetson Nano) and to perform the task directly on the edge with very low energy\nconsumption (2.4W on average at 50 fps). The code is publicly available at:\nhttps://github.com/EgocentricVision/EgoWild.\n", "date": "2022-11-06", "categories": []}
{"id": "2303.05166", "title": "TAEC: Unsupervised Action Segmentation with Temporal-Aware Embedding and\n  Clustering", "abstract": "  Temporal action segmentation in untrimmed videos has gained increased\nattention recently. However, annotating action classes and frame-wise\nboundaries is extremely time consuming and cost intensive, especially on\nlarge-scale datasets. To address this issue, we propose an unsupervised\napproach for learning action classes from untrimmed video sequences. In\nparticular, we propose a temporal embedding network that combines relative time\nprediction, feature reconstruction, and sequence-to-sequence learning, to\npreserve the spatial layout and sequential nature of the video features. A\ntwo-step clustering pipeline on these embedded feature representations then\nallows us to enforce temporal consistency within, as well as across videos.\nBased on the identified clusters, we decode the video into coherent temporal\nsegments that correspond to semantically meaningful action classes. Our\nevaluation on three challenging datasets shows the impact of each component\nand, furthermore, demonstrates our state-of-the-art unsupervised action\nsegmentation results.\n", "date": "2023-03-09", "categories": []}
{"id": "2303.03716", "title": "Challenges of the Creation of a Dataset for Vision Based Human Hand\n  Action Recognition in Industrial Assembly", "abstract": "  This work presents the Industrial Hand Action Dataset V1, an industrial\nassembly dataset consisting of 12 classes with 459,180 images in the basic\nversion and 2,295,900 images after spatial augmentation. Compared to other\nfreely available datasets tested, it has an above-average duration and, in\naddition, meets the technical and legal requirements for industrial assembly\nlines. Furthermore, the dataset contains occlusions, hand-object interaction,\nand various fine-grained human hand actions for industrial assembly tasks that\nwere not found in combination in examined datasets. The recorded ground truth\nassembly classes were selected after extensive observation of real-world use\ncases. A Gated Transformer Network, a state-of-the-art model from the\ntransformer domain was adapted, and proved with a test accuracy of 86.25%\nbefore hyperparameter tuning by 18,269,959 trainable parameters, that it is\npossible to train sequential deep learning models with this dataset.\n", "date": "2023-03-07", "categories": []}
{"id": "2303.03166", "title": "Faster Learning of Temporal Action Proposal via Sparse Multilevel\n  Boundary Generator", "abstract": "  Temporal action localization in videos presents significant challenges in the\nfield of computer vision. While the boundary-sensitive method has been widely\nadopted, its limitations include incomplete use of intermediate and global\ninformation, as well as an inefficient proposal feature generator. To address\nthese challenges, we propose a novel framework, Sparse Multilevel Boundary\nGenerator (SMBG), which enhances the boundary-sensitive method with boundary\nclassification and action completeness regression. SMBG features a multi-level\nboundary module that enables faster processing by gathering boundary\ninformation at different lengths. Additionally, we introduce a sparse\nextraction confidence head that distinguishes information inside and outside\nthe action, further optimizing the proposal feature generator. To improve the\nsynergy between multiple branches and balance positive and negative samples, we\npropose a global guidance loss. Our method is evaluated on two popular\nbenchmarks, ActivityNet-1.3 and THUMOS14, and is shown to achieve\nstate-of-the-art performance, with a better inference speed (2.47xBSN++,\n2.12xDBG). These results demonstrate that SMBG provides a more efficient and\nsimple solution for generating temporal action proposals. Our proposed\nframework has the potential to advance the field of computer vision and enhance\nthe accuracy and speed of temporal action localization in video analysis.The\ncode and models are made available at\n\\url{https://github.com/zhouyang-001/SMBG-for-temporal-action-proposal}.\n", "date": "2023-03-06", "categories": []}
{"id": "2211.14924", "title": "Post-Processing Temporal Action Detection", "abstract": "  Existing Temporal Action Detection (TAD) methods typically take a\npre-processing step in converting an input varying-length video into a\nfixed-length snippet representation sequence, before temporal boundary\nestimation and action classification. This pre-processing step would temporally\ndownsample the video, reducing the inference resolution and hampering the\ndetection performance in the original temporal resolution. In essence, this is\ndue to a temporal quantization error introduced during the resolution\ndownsampling and recovery. This could negatively impact the TAD performance,\nbut is largely ignored by existing methods. To address this problem, in this\nwork we introduce a novel model-agnostic post-processing method without model\nredesign and retraining. Specifically, we model the start and end points of\naction instances with a Gaussian distribution for enabling temporal boundary\ninference at a sub-snippet level. We further introduce an efficient\nTaylor-expansion based approximation, dubbed as Gaussian Approximated\nPost-processing (GAP). Extensive experiments demonstrate that our GAP can\nconsistently improve a wide variety of pre-trained off-the-shelf TAD models on\nthe challenging ActivityNet (+0.2% -0.7% in average mAP) and THUMOS (+0.2%\n-0.5% in average mAP) benchmarks. Such performance gains are already\nsignificant and highly comparable to those achieved by novel model designs.\nAlso, GAP can be integrated with model training for further performance gain.\nImportantly, GAP enables lower temporal resolutions for more efficient\ninference, facilitating low-resource applications. The code will be available\nin https://github.com/sauradip/GAP\n", "date": "2022-11-27", "categories": []}
{"id": "2212.03125", "title": "Self-supervised and Weakly Supervised Contrastive Learning for\n  Frame-wise Action Representations", "abstract": "  Previous work on action representation learning focused on global\nrepresentations for short video clips. In contrast, many practical\napplications, such as video alignment, strongly demand learning the intensive\nrepresentation of long videos. In this paper, we introduce a new framework of\ncontrastive action representation learning (CARL) to learn frame-wise action\nrepresentation in a self-supervised or weakly-supervised manner, especially for\nlong videos. Specifically, we introduce a simple but effective video encoder\nthat considers both spatial and temporal context by combining convolution and\ntransformer. Inspired by the recent massive progress in self-supervised\nlearning, we propose a new sequence contrast loss (SCL) applied to two related\nviews obtained by expanding a series of spatio-temporal data in two versions.\nOne is the self-supervised version that optimizes embedding space by minimizing\nKL-divergence between sequence similarity of two augmented views and prior\nGaussian distribution of timestamp distance. The other is the weakly-supervised\nversion that builds more sample pairs among videos using video-level labels by\ndynamic time wrapping (DTW). Experiments on FineGym, PennAction, and Pouring\ndatasets show that our method outperforms previous state-of-the-art by a large\nmargin for downstream fine-grained action classification and even faster\ninference. Surprisingly, although without training on paired videos like in\nprevious works, our self-supervised version also shows outstanding performance\nin video alignment and fine-grained frame retrieval tasks.\n", "date": "2022-12-06", "categories": []}
{"id": "2303.00977", "title": "Ego-Vehicle Action Recognition based on Semi-Supervised Contrastive\n  Learning", "abstract": "  In recent years, many automobiles have been equipped with cameras, which have\naccumulated an enormous amount of video footage of driving scenes. Autonomous\ndriving demands the highest level of safety, for which even unimaginably rare\ndriving scenes have to be collected in training data to improve the recognition\naccuracy for specific scenes. However, it is prohibitively costly to find very\nfew specific scenes from an enormous amount of videos. In this article, we show\nthat proper video-to-video distances can be defined by focusing on ego-vehicle\nactions. It is well known that existing methods based on supervised learning\ncannot handle videos that do not fall into predefined classes, though they work\nwell in defining video-to-video distances in the embedding space between\nlabeled videos. To tackle this problem, we propose a method based on\nsemi-supervised contrastive learning. We consider two related but distinct\ncontrastive learning: standard graph contrastive learning and our proposed\nSOIA-based contrastive learning. We observe that the latter approach can\nprovide more sensible video-to-video distances between unlabeled videos. Next,\nthe effectiveness of our method is quantified by evaluating the classification\nperformance of the ego-vehicle action recognition using HDD dataset, which\nshows that our method including unlabeled data in training significantly\noutperforms the existing methods using only labeled data in training.\n", "date": "2023-03-02", "categories": []}
{"id": "2302.14350", "title": "Knowledge Augmented Relation Inference for Group Activity Recognition", "abstract": "  Most existing group activity recognition methods construct spatial-temporal\nrelations merely based on visual representation. Some methods introduce extra\nknowledge, such as action labels, to build semantic relations and use them to\nrefine the visual presentation. However, the knowledge they explored just stay\nat the semantic-level, which is insufficient for pursing notable accuracy. In\nthis paper, we propose to exploit knowledge concretization for the group\nactivity recognition, and develop a novel Knowledge Augmented Relation\nInference framework that can effectively use the concretized knowledge to\nimprove the individual representations. Specifically, the framework consists of\na Visual Representation Module to extract individual appearance features, a\nKnowledge Augmented Semantic Relation Module explore semantic representations\nof individual actions, and a Knowledge-Semantic-Visual Interaction Module aims\nto integrate visual and semantic information by the knowledge. Benefiting from\nthese modules, the proposed framework can utilize knowledge to enhance the\nrelation inference process and the individual representations, thus improving\nthe performance of group activity recognition. Experimental results on two\npublic datasets show that the proposed framework achieves competitive\nperformance compared with state-of-the-art methods.\n", "date": "2023-02-28", "categories": []}
{"id": "2302.12967", "title": "Temporal-Channel Topology Enhanced Network for Skeleton-Based Action\n  Recognition", "abstract": "  Skeleton-based action recognition has become popular in recent years due to\nits efficiency and robustness. Most current methods adopt graph convolutional\nnetwork (GCN) for topology modeling, but GCN-based methods are limited in\nlong-distance correlation modeling and generalizability. In contrast, the\npotential of convolutional neural network (CNN) for topology modeling has not\nbeen fully explored. In this paper, we propose a novel CNN architecture,\nTemporal-Channel Topology Enhanced Network (TCTE-Net), to learn spatial and\ntemporal topologies for skeleton-based action recognition. The TCTE-Net\nconsists of two modules: the Temporal-Channel Focus module, which learns a\ntemporal-channel focus matrix to identify the most critical feature\nrepresentations, and the Dynamic Channel Topology Attention module, which\ndynamically learns spatial topological features, and fuses them with an\nattention mechanism to model long-distance channel-wise topology. We conduct\nexperiments on NTU RGB+D, NTU RGB+D 120, and FineGym datasets. TCTE-Net shows\nstate-of-the-art performance compared to CNN-based methods and achieves\nsuperior performance compared to GCN-based methods. The code is available at\nhttps://github.com/aikuniverse/TCTE-Net.\n", "date": "2023-02-25", "categories": []}
{"id": "2302.13074", "title": "Temporal Segment Transformer for Action Segmentation", "abstract": "  Recognizing human actions from untrimmed videos is an important task in\nactivity understanding, and poses unique challenges in modeling long-range\ntemporal relations. Recent works adopt a predict-and-refine strategy which\nconverts an initial prediction to action segments for global context modeling.\nHowever, the generated segment representations are often noisy and exhibit\ninaccurate segment boundaries, over-segmentation and other problems. To deal\nwith these issues, we propose an attention based approach which we call\n\\textit{temporal segment transformer}, for joint segment relation modeling and\ndenoising. The main idea is to denoise segment representations using attention\nbetween segment and frame representations, and also use inter-segment attention\nto capture temporal correlations between segments. The refined segment\nrepresentations are used to predict action labels and adjust segment\nboundaries, and a final action segmentation is produced based on voting from\nsegment masks. We show that this novel architecture achieves state-of-the-art\naccuracy on the popular 50Salads, GTEA and Breakfast benchmarks. We also\nconduct extensive ablations to demonstrate the effectiveness of different\ncomponents of our design.\n", "date": "2023-02-25", "categories": []}
{"id": "2302.12007", "title": "DMMG: Dual Min-Max Games for Self-Supervised Skeleton-Based Action\n  Recognition", "abstract": "  In this work, we propose a new Dual Min-Max Games (DMMG) based\nself-supervised skeleton action recognition method by augmenting unlabeled data\nin a contrastive learning framework. Our DMMG consists of a viewpoint variation\nmin-max game and an edge perturbation min-max game. These two min-max games\nadopt an adversarial paradigm to perform data augmentation on the skeleton\nsequences and graph-structured body joints, respectively. Our viewpoint\nvariation min-max game focuses on constructing various hard contrastive pairs\nby generating skeleton sequences from various viewpoints. These hard\ncontrastive pairs help our model learn representative action features, thus\nfacilitating model transfer to downstream tasks. Moreover, our edge\nperturbation min-max game specializes in building diverse hard contrastive\nsamples through perturbing connectivity strength among graph-based body joints.\nThe connectivity-strength varying contrastive pairs enable the model to capture\nminimal sufficient information of different actions, such as representative\ngestures for an action while preventing the model from overfitting. By fully\nexploiting the proposed DMMG, we can generate sufficient challenging\ncontrastive pairs and thus achieve discriminative action feature\nrepresentations from unlabeled skeleton data in a self-supervised manner.\nExtensive experiments demonstrate that our method achieves superior results\nunder various evaluation protocols on widely-used NTU-RGB+D and NTU120-RGB+D\ndatasets.\n", "date": "2023-02-22", "categories": []}
{"id": "2302.11027", "title": "Analysis of Real-Time Hostile Activitiy Detection from Spatiotemporal\n  Features Using Time Distributed Deep CNNs, RNNs and Attention-Based\n  Mechanisms", "abstract": "  Real-time video surveillance, through CCTV camera systems has become\nessential for ensuring public safety which is a priority today. Although CCTV\ncameras help a lot in increasing security, these systems require constant human\ninteraction and monitoring. To eradicate this issue, intelligent surveillance\nsystems can be built using deep learning video classification techniques that\ncan help us automate surveillance systems to detect violence as it happens. In\nthis research, we explore deep learning video classification techniques to\ndetect violence as they are happening. Traditional image classification\ntechniques fall short when it comes to classifying videos as they attempt to\nclassify each frame separately for which the predictions start to flicker.\nTherefore, many researchers are coming up with video classification techniques\nthat consider spatiotemporal features while classifying. However, deploying\nthese deep learning models with methods such as skeleton points obtained\nthrough pose estimation and optical flow obtained through depth sensors, are\nnot always practical in an IoT environment. Although these techniques ensure a\nhigher accuracy score, they are computationally heavier. Keeping these\nconstraints in mind, we experimented with various video classification and\naction recognition techniques such as ConvLSTM, LRCN (with both custom CNN\nlayers and VGG-16 as feature extractor) CNNTransformer and C3D. We achieved a\ntest accuracy of 80% on ConvLSTM, 83.33% on CNN-BiLSTM, 70% on VGG16-BiLstm\n,76.76% on CNN-Transformer and 80% on C3D.\n", "date": "2023-02-21", "categories": []}
{"id": "2208.11314", "title": "Modality Mixer for Multi-modal Action Recognition", "abstract": "  In multi-modal action recognition, it is important to consider not only the\ncomplementary nature of different modalities but also global action content. In\nthis paper, we propose a novel network, named Modality Mixer (M-Mixer) network,\nto leverage complementary information across modalities and temporal context of\nan action for multi-modal action recognition. We also introduce a simple yet\neffective recurrent unit, called Multi-modal Contextualization Unit (MCU),\nwhich is a core component of M-Mixer. Our MCU temporally encodes a sequence of\none modality (e.g., RGB) with action content features of other modalities\n(e.g., depth, IR). This process encourages M-Mixer to exploit global action\ncontent and also to supplement complementary information of other modalities.\nAs a result, our proposed method outperforms state-of-the-art methods on NTU\nRGB+D 60, NTU RGB+D 120, and NW-UCLA datasets. Moreover, we demonstrate the\neffectiveness of M-Mixer by conducting comprehensive ablation studies.\n", "date": "2022-08-24", "categories": []}
{"id": "2302.09657", "title": "Table Tennis Stroke Detection and Recognition Using Ball Trajectory Data", "abstract": "  In this work, the novel task of detecting and classifying table tennis\nstrokes solely using the ball trajectory has been explored. A single camera\nsetup positioned in the umpire's view has been employed to procure a dataset\nconsisting of six stroke classes executed by four professional table tennis\nplayers. Ball tracking using YOLOv4, a traditional object detection model, and\nTrackNetv2, a temporal heatmap based model, have been implemented on our\ndataset and their performances have been benchmarked. A mathematical approach\ndeveloped to extract temporal boundaries of strokes using the ball trajectory\ndata yielded a total of 2023 valid strokes in our dataset, while also detecting\nservices and missed strokes successfully. The temporal convolutional network\ndeveloped performed stroke recognition on completely unseen data with an\naccuracy of 87.155%. Several machine learning and deep learning based model\narchitectures have been trained for stroke recognition using ball trajectory\ninput and benchmarked based on their performances. While stroke recognition in\nthe field of table tennis has been extensively explored based on human action\nrecognition using video data focused on the player's actions, the use of ball\ntrajectory data for the same is an unexplored characteristic of the sport.\nHence, the motivation behind the work is to demonstrate that meaningful\ninferences such as stroke detection and recognition can be drawn using minimal\ninput information.\n", "date": "2023-02-19", "categories": []}
{"id": "2302.08689", "title": "Dynamic Spatial-temporal Hypergraph Convolutional Network for\n  Skeleton-based Action Recognition", "abstract": "  Skeleton-based action recognition relies on the extraction of\nspatial-temporal topological information. Hypergraphs can establish prior\nunnatural dependencies for the skeleton. However, the existing methods only\nfocus on the construction of spatial topology and ignore the time-point\ndependence. This paper proposes a dynamic spatial-temporal hypergraph\nconvolutional network (DST-HCN) to capture spatial-temporal information for\nskeleton-based action recognition. DST-HCN introduces a time-point hypergraph\n(TPH) to learn relationships at time points. With multiple spatial static\nhypergraphs and dynamic TPH, our network can learn more complete\nspatial-temporal features. In addition, we use the high-order information\nfusion module (HIF) to fuse spatial-temporal information synchronously.\nExtensive experiments on NTU RGB+D, NTU RGB+D 120, and NW-UCLA datasets show\nthat our model achieves state-of-the-art, especially compared with hypergraph\nmethods.\n", "date": "2023-02-17", "categories": []}
{"id": "2302.08493", "title": "Deep Multi-stream Network for Video-based Calving Sign Detection", "abstract": "  We have designed a deep multi-stream network for automatically detecting\ncalving signs from video. Calving sign detection from a camera, which is a\nnon-contact sensor, is expected to enable more efficient livestock management.\nAs large-scale, well-developed data cannot generally be assumed when\nestablishing calving detection systems, the basis for making the prediction\nneeds to be presented to farmers during operation, so black-box modeling (also\nknown as end-to-end modeling) is not appropriate. For practical operation of\ncalving detection systems, the present study aims to incorporate expert\nknowledge into a deep neural network. To this end, we propose a multi-stream\ncalving sign detection network in which multiple calving-related features are\nextracted from the corresponding feature extraction networks designed for each\nattribute with different characteristics, such as a cow's posture, rotation,\nand movement, known as calving signs, and are then integrated appropriately\ndepending on the cow's situation. Experimental comparisons conducted using\nvideos of 15 cows demonstrated that our multi-stream system yielded a\nsignificant improvement over the end-to-end system, and the multi-stream\narchitecture significantly contributed to a reduction in detection errors. In\naddition, the distinctive mixture weights we observed helped provide\ninterpretability of the system's behavior.\n", "date": "2023-01-10", "categories": []}
{"id": "2302.02316", "title": "Spatiotemporal Decouple-and-Squeeze Contrastive Learning for\n  Semi-Supervised Skeleton-based Action Recognition", "abstract": "  Contrastive learning has been successfully leveraged to learn action\nrepresentations for addressing the problem of semi-supervised skeleton-based\naction recognition. However, most contrastive learning-based methods only\ncontrast global features mixing spatiotemporal information, which confuses the\nspatial- and temporal-specific information reflecting different semantic at the\nframe level and joint level. Thus, we propose a novel Spatiotemporal\nDecouple-and-Squeeze Contrastive Learning (SDS-CL) framework to comprehensively\nlearn more abundant representations of skeleton-based actions by jointly\ncontrasting spatial-squeezing features, temporal-squeezing features, and global\nfeatures. In SDS-CL, we design a new Spatiotemporal-decoupling Intra-Inter\nAttention (SIIA) mechanism to obtain the spatiotemporal-decoupling attentive\nfeatures for capturing spatiotemporal specific information by calculating\nspatial- and temporal-decoupling intra-attention maps among joint/motion\nfeatures, as well as spatial- and temporal-decoupling inter-attention maps\nbetween joint and motion features. Moreover, we present a new Spatial-squeezing\nTemporal-contrasting Loss (STL), a new Temporal-squeezing Spatial-contrasting\nLoss (TSL), and the Global-contrasting Loss (GL) to contrast the\nspatial-squeezing joint and motion features at the frame level,\ntemporal-squeezing joint and motion features at the joint level, as well as\nglobal joint and motion features at the skeleton level. Extensive experimental\nresults on four public datasets show that the proposed SDS-CL achieves\nperformance gains compared with other competitive methods.\n", "date": "2023-02-05", "categories": []}
{"id": "2302.02327", "title": "Pyramid Self-attention Polymerization Learning for Semi-supervised\n  Skeleton-based Action Recognition", "abstract": "  Most semi-supervised skeleton-based action recognition approaches aim to\nlearn the skeleton action representations only at the joint level, but neglect\nthe crucial motion characteristics at the coarser-grained body (e.g., limb,\ntrunk) level that provide rich additional semantic information, though the\nnumber of labeled data is limited. In this work, we propose a novel Pyramid\nSelf-attention Polymerization Learning (dubbed as PSP Learning) framework to\njointly learn body-level, part-level, and joint-level action representations of\njoint and motion data containing abundant and complementary semantic\ninformation via contrastive learning covering coarse-to-fine granularity.\nSpecifically, to complement semantic information from coarse to fine\ngranularity in skeleton actions, we design a new Pyramid Polymerizing Attention\n(PPA) mechanism that firstly calculates the body-level attention map,\npart-level attention map, and joint-level attention map, as well as polymerizes\nthese attention maps in a level-by-level way (i.e., from body level to part\nlevel, and further to joint level). Moreover, we present a new Coarse-to-fine\nContrastive Loss (CCL) including body-level contrast loss, part-level contrast\nloss, and joint-level contrast loss to jointly measure the similarity between\nthe body/part/joint-level contrasting features of joint and motion data.\nFinally, extensive experiments are conducted on the NTU RGB+D and North-Western\nUCLA datasets to demonstrate the competitive performance of the proposed PSP\nLearning in the semi-supervised skeleton-based action recognition task. The\nsource codes of PSP Learning are publicly available at\nhttps://github.com/1xbq1/PSP-Learning.\n", "date": "2023-02-05", "categories": []}
{"id": "2302.02752", "title": "Baseline Method for the Sport Task of MediaEval 2022 with 3D CNNs using\n  Attention Mechanisms", "abstract": "  This paper presents the baseline method proposed for the Sports Video task\npart of the MediaEval 2022 benchmark. This task proposes two subtasks: stroke\nclassification from trimmed videos, and stroke detection from untrimmed videos.\nThis baseline addresses both subtasks. We propose two types of 3D-CNN\narchitectures to solve the two subtasks. Both 3D-CNNs use Spatio-temporal\nconvolutions and attention mechanisms. The architectures and the training\nprocess are tailored to solve the addressed subtask. This baseline method is\nshared publicly online to help the participants in their investigation and\nalleviate eventually some aspects of the task such as video processing,\ntraining method, evaluation and submission routine. The baseline method reaches\n86.4% of accuracy with our v2 model for the classification subtask. For the\ndetection subtask, the baseline reaches a mAP of 0.131 and IoU of 0.515 with\nour v1 model.\n", "date": "2023-02-06", "categories": []}
{"id": "2302.02755", "title": "Fine-Grained Action Detection with RGB and Pose Information using Two\n  Stream Convolutional Networks", "abstract": "  As participants of the MediaEval 2022 Sport Task, we propose a two-stream\nnetwork approach for the classification and detection of table tennis strokes.\nEach stream is a succession of 3D Convolutional Neural Network (CNN) blocks\nusing attention mechanisms. Each stream processes different 4D inputs. Our\nmethod utilizes raw RGB data and pose information computed from MMPose toolbox.\nThe pose information is treated as an image by applying the pose either on a\nblack background or on the original RGB frame it has been computed from. Best\nperformance is obtained by feeding raw RGB data to one stream, Pose + RGB\n(PRGB) information to the other stream and applying late fusion on the\nfeatures. The approaches were evaluated on the provided TTStroke-21 data sets.\nWe can report an improvement in stroke classification, reaching 87.3% of\naccuracy, while the detection does not outperform the baseline but still\nreaches an IoU of 0.349 and mAP of 0.110.\n", "date": "2023-02-06", "categories": []}
{"id": "2302.03024", "title": "AIM: Adapting Image Models for Efficient Video Action Recognition", "abstract": "  Recent vision transformer based video models mostly follow the ``image\npre-training then finetuning\" paradigm and have achieved great success on\nmultiple video benchmarks. However, full finetuning such a video model could be\ncomputationally expensive and unnecessary, given the pre-trained image\ntransformer models have demonstrated exceptional transferability. In this work,\nwe propose a novel method to Adapt pre-trained Image Models (AIM) for efficient\nvideo understanding. By freezing the pre-trained image model and adding a few\nlightweight Adapters, we introduce spatial adaptation, temporal adaptation and\njoint adaptation to gradually equip an image model with spatiotemporal\nreasoning capability. We show that our proposed AIM can achieve competitive or\neven better performance than prior arts with substantially fewer tunable\nparameters on four video action recognition benchmarks. Thanks to its\nsimplicity, our method is also generally applicable to different image\npre-trained models, which has the potential to leverage more powerful image\nfoundation models in the future. The project webpage is\n\\url{https://adapt-image-models.github.io/}.\n", "date": "2023-02-06", "categories": []}
{"id": "2302.01921", "title": "Transformers in Action Recognition: A Review on Temporal Modeling", "abstract": "  In vision-based action recognition, spatio-temporal features from different\nmodalities are used for recognizing activities. Temporal modeling is a long\nchallenge of action recognition. However, there are limited methods such as\npre-computed motion features, three-dimensional (3D) filters, and recurrent\nneural networks (RNN) for modeling motion information in deep-based approaches.\nRecently, transformers success in modeling long-range dependencies in natural\nlanguage processing (NLP) tasks has gotten great attention from other domains;\nincluding speech, image, and video, to rely entirely on self-attention without\nusing sequence-aligned RNNs or convolutions. Although the application of\ntransformers to action recognition is relatively new, the amount of research\nproposed on this topic within the last few years is astounding. This paper\nespecially reviews recent progress in deep learning methods for modeling\ntemporal variations. It focuses on action recognition methods that use\ntransformers for temporal modeling, discussing their main features, used\nmodalities, and identifying opportunities and challenges for future research.\n", "date": "2022-12-29", "categories": []}
{"id": "2301.13360", "title": "Skeleton-based Human Action Recognition via Convolutional Neural\n  Networks (CNN)", "abstract": "  Recently, there has been a remarkable increase in the interest towards\nskeleton-based action recognition within the research community, owing to its\nvarious advantageous features, including computational efficiency,\nrepresentative features, and illumination invariance. Despite this, researchers\ncontinue to explore and investigate the most optimal way to represent human\nactions through skeleton representation and the extracted features. As a\nresult, the growth and availability of human action recognition datasets have\nrisen substantially. In addition, deep learning-based algorithms have gained\nwidespread popularity due to the remarkable advancements in various computer\nvision tasks. Most state-of-the-art contributions in skeleton-based action\nrecognition incorporate a Graph Neural Network (GCN) architecture for\nrepresenting the human body and extracting features. Our research demonstrates\nthat Convolutional Neural Networks (CNNs) can attain comparable results to GCN,\nprovided that the proper training techniques, augmentations, and optimizers are\napplied. Our approach has been rigorously validated, and we have achieved a\nscore of 95% on the NTU-60 dataset\n", "date": "2023-01-31", "categories": []}
{"id": "2301.13576", "title": "Sport Task: Fine Grained Action Detection and Classification of Table\n  Tennis Strokes from Videos for MediaEval 2022", "abstract": "  Sports video analysis is a widespread research topic. Its applications are\nvery diverse, like events detection during a match, video summary, or\nfine-grained movement analysis of athletes. As part of the MediaEval 2022\nbenchmarking initiative, this task aims at detecting and classifying subtle\nmovements from sport videos. We focus on recordings of table tennis matches.\nConducted since 2019, this task provides a classification challenge from\nuntrimmed videos recorded under natural conditions with known temporal\nboundaries for each stroke. Since 2021, the task also provides a stroke\ndetection challenge from unannotated, untrimmed videos. This year, the\ntraining, validation, and test sets are enhanced to ensure that all strokes are\nrepresented in each dataset. The dataset is now similar to the one used in [1,\n2]. This research is intended to build tools for coaches and athletes who want\nto further evaluate their sport performances.\n", "date": "2023-01-31", "categories": []}
{"id": "2301.12436", "title": "Team VI-I2R Technical Report on EPIC-KITCHENS-100 Unsupervised Domain\n  Adaptation Challenge for Action Recognition 2022", "abstract": "  In this report, we present the technical details of our submission to the\nEPIC-KITCHENS-100 Unsupervised Domain Adaptation (UDA) Challenge for Action\nRecognition 2022. This task aims to adapt an action recognition model trained\non a labeled source domain to an unlabeled target domain. To achieve this goal,\nwe propose an action-aware domain adaptation framework that leverages the prior\nknowledge induced from the action recognition task during the adaptation.\nSpecifically, we disentangle the source features into action-relevant features\nand action-irrelevant features using the learned action classifier and then\nalign the target features with the action-relevant features. To further improve\nthe action prediction performance, we exploit the verb-noun co-occurrence\nmatrix to constrain and refine the action predictions. Our final submission\nachieved the first place in terms of top-1 action recognition accuracy.\n", "date": "2023-01-29", "categories": []}
{"id": "2301.11495", "title": "Skeleton-based Action Recognition through Contrasting Two-Stream\n  Spatial-Temporal Networks", "abstract": "  For pursuing accurate skeleton-based action recognition, most prior methods\nuse the strategy of combining Graph Convolution Networks (GCNs) with\nattention-based methods in a serial way. However, they regard the human\nskeleton as a complete graph, resulting in less variations between different\nactions (e.g., the connection between the elbow and head in action ``clapping\nhands''). For this, we propose a novel Contrastive GCN-Transformer Network\n(ConGT) which fuses the spatial and temporal modules in a parallel way. The\nConGT involves two parallel streams: Spatial-Temporal Graph Convolution stream\n(STG) and Spatial-Temporal Transformer stream (STT). The STG is designed to\nobtain action representations maintaining the natural topology structure of the\nhuman skeleton. The STT is devised to acquire action representations containing\nthe global relationships among joints. Since the action representations\nproduced from these two streams contain different characteristics, and each of\nthem knows little information of the other, we introduce the contrastive\nlearning paradigm to guide their output representations of the same sample to\nbe as close as possible in a self-supervised manner. Through the contrastive\nlearning, they can learn information from each other to enrich the action\nfeatures by maximizing the mutual information between the two types of action\nrepresentations. To further improve action recognition accuracy, we introduce\nthe Cyclical Focal Loss (CFL) which can focus on confident training samples in\nearly training epochs, with an increasing focus on hard samples during the\nmiddle epochs. We conduct experiments on three benchmark datasets, which\ndemonstrate that our model achieves state-of-the-art performance in action\nrecognition.\n", "date": "2023-01-27", "categories": []}
{"id": "2301.10161", "title": "Dataset Bias in Human Activity Recognition", "abstract": "  When creating multi-channel time-series datasets for Human Activity\nRecognition (HAR), researchers are faced with the issue of subject selection\ncriteria. It is unknown what physical characteristics and/or soft-biometrics,\nsuch as age, height, and weight, need to be taken into account to train a\nclassifier to achieve robustness towards heterogeneous populations in the\ntraining and testing data. This contribution statistically curates the training\ndata to assess to what degree the physical characteristics of humans influence\nHAR performance. We evaluate the performance of a state-of-the-art\nconvolutional neural network on two HAR datasets that vary in the sensors,\nactivities, and recording for time-series HAR. The training data is\nintentionally biased with respect to human characteristics to determine the\nfeatures that impact motion behaviour. The evaluations brought forth the impact\nof the subjects' characteristics on HAR. Thus, providing insights regarding the\nrobustness of the classifier with respect to heterogeneous populations. The\nstudy is a step forward in the direction of fair and trustworthy artificial\nintelligence by attempting to quantify representation bias in multi-channel\ntime series HAR data.\n", "date": "2023-01-19", "categories": []}
{"id": "2301.00986", "title": "Look, Listen, and Attack: Backdoor Attacks Against Video Action\n  Recognition", "abstract": "  Deep neural networks (DNNs) are vulnerable to a class of attacks called\n\"backdoor attacks\", which create an association between a backdoor trigger and\na target label the attacker is interested in exploiting. A backdoored DNN\nperforms well on clean test images, yet persistently predicts an\nattacker-defined label for any sample in the presence of the backdoor trigger.\nAlthough backdoor attacks have been extensively studied in the image domain,\nthere are very few works that explore such attacks in the video domain, and\nthey tend to conclude that image backdoor attacks are less effective in the\nvideo domain. In this work, we revisit the traditional backdoor threat model\nand incorporate additional video-related aspects to that model. We show that\npoisoned-label image backdoor attacks could be extended temporally in two ways,\nstatically and dynamically, leading to highly effective attacks in the video\ndomain. In addition, we explore natural video backdoors to highlight the\nseriousness of this vulnerability in the video domain. And, for the first time,\nwe study multi-modal (audiovisual) backdoor attacks against video action\nrecognition models, where we show that attacking a single modality is enough\nfor achieving a high attack success rate.\n", "date": "2023-01-03", "categories": []}
{"id": "2209.04780", "title": "MAiVAR: Multimodal Audio-Image and Video Action Recognizer", "abstract": "  Currently, action recognition is predominately performed on video data as\nprocessed by CNNs. We investigate if the representation process of CNNs can\nalso be leveraged for multimodal action recognition by incorporating\nimage-based audio representations of actions in a task. To this end, we propose\nMultimodal Audio-Image and Video Action Recognizer (MAiVAR), a CNN-based\naudio-image to video fusion model that accounts for video and audio modalities\nto achieve superior action recognition performance. MAiVAR extracts meaningful\nimage representations of audio and fuses it with video representation to\nachieve better performance as compared to both modalities individually on a\nlarge-scale action recognition dataset.\n", "date": "2022-09-11", "categories": []}
{"id": "2301.07565", "title": "Gated-ViGAT: Efficient Bottom-Up Event Recognition and Explanation Using\n  a New Frame Selection Policy and Gating Mechanism", "abstract": "  In this paper, Gated-ViGAT, an efficient approach for video event\nrecognition, utilizing bottom-up (object) information, a new frame sampling\npolicy and a gating mechanism is proposed. Specifically, the frame sampling\npolicy uses weighted in-degrees (WiDs), derived from the adjacency matrices of\ngraph attention networks (GATs), and a dissimilarity measure to select the most\nsalient and at the same time diverse frames representing the event in the\nvideo. Additionally, the proposed gating mechanism fetches the selected frames\nsequentially, and commits early-exiting when an adequately confident decision\nis achieved. In this way, only a few frames are processed by the\ncomputationally expensive branch of our network that is responsible for the\nbottom-up information extraction. The experimental evaluation on two large,\npublicly available video datasets (MiniKinetics, ActivityNet) demonstrates that\nGated-ViGAT provides a large computational complexity reduction in comparison\nto our previous approach (ViGAT), while maintaining the excellent event\nrecognition and explainability performance. Gated-ViGAT source code is made\npublicly available at https://github.com/bmezaris/Gated-ViGAT\n", "date": "2023-01-18", "categories": []}
{"id": "2301.06018", "title": "CMAE-V: Contrastive Masked Autoencoders for Video Action Recognition", "abstract": "  Contrastive Masked Autoencoder (CMAE), as a new self-supervised framework,\nhas shown its potential of learning expressive feature representations in\nvisual image recognition. This work shows that CMAE also trivially generalizes\nwell on video action recognition without modifying the architecture and the\nloss criterion. By directly replacing the original pixel shift with the\ntemporal shift, our CMAE for visual action recognition, CMAE-V for short, can\ngenerate stronger feature representations than its counterpart based on pure\nmasked autoencoders. Notably, CMAE-V, with a hybrid architecture, can achieve\n82.2% and 71.6% top-1 accuracy on the Kinetics-400 and Something-something V2\ndatasets, respectively. We hope this report could provide some informative\ninspiration for future works.\n", "date": "2023-01-15", "categories": []}
{"id": "2301.06082", "title": "A Survey on Human Action Recognition", "abstract": "  Human Action Recognition (HAR), one of the most important tasks in computer\nvision, has developed rapidly in the past decade and has a wide range of\napplications in health monitoring, intelligent surveillance, virtual reality,\nhuman computer interaction and so on. Human actions can be represented by a\nwide variety of modalities, such as RGB-D cameras, audio, inertial sensors,etc.\nConsequently, in addition to the mainstream single modality based HAR\napproaches, more and more research is devoted to the multimodal domain due to\nthe complementary properties between multimodal data. In this paper, we present\na survey of HAR methods in recent years according to the different input\nmodalities. Meanwhile, considering that most of the recent surveys on HAR focus\non the third perspective, while this survey aims to provide a more\ncomprehensive introduction to HAR novices and researchers, we therefore also\ninvestigate the actions recognition methods from the first perspective in\nrecent years. Finally, we give a brief introduction about the benchmark HAR\ndatasets and show the performance comparison of different methods on these\ndatasets.\n", "date": "2022-12-20", "categories": []}
{"id": "2301.06187", "title": "CNN-Based Action Recognition and Pose Estimation for Classifying Animal\n  Behavior from Videos: A Survey", "abstract": "  Classifying the behavior of humans or animals from videos is important in\nbiomedical fields for understanding brain function and response to stimuli.\nAction recognition, classifying activities performed by one or more subjects in\na trimmed video, forms the basis of many of these techniques. Deep learning\nmodels for human action recognition have progressed significantly over the last\ndecade. Recently, there is an increased interest in research that incorporates\ndeep learning-based action recognition for animal behavior classification.\nHowever, human action recognition methods are more developed. This survey\npresents an overview of human action recognition and pose estimation methods\nthat are based on convolutional neural network (CNN) architectures and have\nbeen adapted for animal behavior classification in neuroscience. Pose\nestimation, estimating joint positions from an image frame, is included because\nit is often applied before classifying animal behavior. First, we provide\nfoundational information on algorithms that learn spatiotemporal features\nthrough 2D, two-stream, and 3D CNNs. We explore motivating factors that\ndetermine optimizers, loss functions and training procedures, and compare their\nperformance on benchmark datasets. Next, we review animal behavior frameworks\nthat use or build upon these methods, organized by the level of supervision\nthey require. Our discussion is uniquely focused on the technical evolution of\nthe underlying CNN models and their architectural adaptations (which we\nillustrate), rather than their usability in a neuroscience lab. We conclude by\ndiscussing open research problems, and possible research directions. Our survey\nis designed to be a resource for researchers developing fully unsupervised\nanimal behavior classification systems of which there are only a few examples\nin the literature.\n", "date": "2023-01-15", "categories": []}
{"id": "2212.10596", "title": "Open-Vocabulary Temporal Action Detection with Off-the-Shelf Image-Text\n  Features", "abstract": "  Detecting actions in untrimmed videos should not be limited to a small,\nclosed set of classes. We present a simple, yet effective strategy for\nopen-vocabulary temporal action detection utilizing pretrained image-text\nco-embeddings. Despite being trained on static images rather than videos, we\nshow that image-text co-embeddings enable openvocabulary performance\ncompetitive with fully-supervised models. We show that the performance can be\nfurther improved by ensembling the image-text features with features encoding\nlocal motion, like optical flow based features, or other modalities, like\naudio. In addition, we propose a more reasonable open-vocabulary evaluation\nsetting for the ActivityNet data set, where the category splits are based on\nsimilarity rather than random assignment.\n", "date": "2022-12-20", "categories": []}
{"id": "2301.03330", "title": "HyRSM++: Hybrid Relation Guided Temporal Set Matching for Few-shot\n  Action Recognition", "abstract": "  Recent attempts mainly focus on learning deep representations for each video\nindividually under the episodic meta-learning regime and then performing\ntemporal alignment to match query and support videos. However, they still\nsuffer from two drawbacks: (i) learning individual features without considering\nthe entire task may result in limited representation capability, and (ii)\nexisting alignment strategies are sensitive to noises and misaligned instances.\nTo handle the two limitations, we propose a novel Hybrid Relation guided\ntemporal Set Matching (HyRSM++) approach for few-shot action recognition. The\ncore idea of HyRSM++ is to integrate all videos within the task to learn\ndiscriminative representations and involve a robust matching technique. To be\nspecific, HyRSM++ consists of two key components, a hybrid relation module and\na temporal set matching metric. Given the basic representations from the\nfeature extractor, the hybrid relation module is introduced to fully exploit\nassociated relations within and cross videos in an episodic task and thus can\nlearn task-specific embeddings. Subsequently, in the temporal set matching\nmetric, we carry out the distance measure between query and support videos from\na set matching perspective and design a Bi-MHM to improve the resilience to\nmisaligned instances. In addition, we explicitly exploit the temporal coherence\nin videos to regularize the matching process. Furthermore, we extend the\nproposed HyRSM++ to deal with the more challenging semi-supervised few-shot\naction recognition and unsupervised few-shot action recognition tasks.\nExperimental results on multiple benchmarks demonstrate that our method\nachieves state-of-the-art performance under various few-shot settings. The\nsource code is available at\nhttps://github.com/alibaba-mmai-research/HyRSMPlusPlus.\n", "date": "2023-01-09", "categories": []}
{"id": "2301.02642", "title": "Triple-stream Deep Metric Learning of Great Ape Behavioural Actions", "abstract": "  We propose the first metric learning system for the recognition of great ape\nbehavioural actions. Our proposed triple stream embedding architecture works on\ncamera trap videos taken directly in the wild and demonstrates that the\nutilisation of an explicit DensePose-C chimpanzee body part segmentation stream\neffectively complements traditional RGB appearance and optical flow streams. We\nevaluate system variants with different feature fusion techniques and long-tail\nrecognition approaches. Results and ablations show performance improvements of\n~12% in top-1 accuracy over previous results achieved on the PanAf-500 dataset\ncontaining 180,000 manually annotated frames across nine behavioural actions.\nFurthermore, we provide a qualitative analysis of our findings and augment the\nmetric learning system with long-tail recognition techniques showing that\naverage per class accuracy -- critical in the domain -- can be improved by ~23%\ncompared to the literature on that dataset. Finally, since our embedding spaces\nare constructed as metric, we provide first data-driven visualisations of the\ngreat ape behavioural action spaces revealing emerging geometry and topology.\nWe hope that the work sparks further interest in this vital application area of\ncomputer vision for the benefit of endangered great apes.\n", "date": "2023-01-06", "categories": []}
{"id": "2301.01216", "title": "An end-to-end multi-scale network for action prediction in videos", "abstract": "  In this paper, we develop an efficient multi-scale network to predict action\nclasses in partial videos in an end-to-end manner. Unlike most existing methods\nwith offline feature generation, our method directly takes frames as input and\nfurther models motion evolution on two different temporal scales.Therefore, we\nsolve the complexity problems of the two stages of modeling and the problem of\ninsufficient temporal and spatial information of a single scale. Our proposed\nEnd-to-End MultiScale Network (E2EMSNet) is composed of two scales which are\nnamed segment scale and observed global scale. The segment scale leverages\ntemporal difference over consecutive frames for finer motion patterns by\nsupplying 2D convolutions. For observed global scale, a Long Short-Term Memory\n(LSTM) is incorporated to capture motion features of observed frames. Our model\nprovides a simple and efficient modeling framework with a small computational\ncost. Our E2EMSNet is evaluated on three challenging datasets: BIT, HMDB51, and\nUCF101. The extensive experiments demonstrate the effectiveness of our method\nfor action prediction in videos.\n", "date": "2022-12-31", "categories": []}
{"id": "2210.00132", "title": "Alignment-guided Temporal Attention for Video Action Recognition", "abstract": "  Temporal modeling is crucial for various video learning tasks. Most recent\napproaches employ either factorized (2D+1D) or joint (3D) spatial-temporal\noperations to extract temporal contexts from the input frames. While the former\nis more efficient in computation, the latter often obtains better performance.\nIn this paper, we attribute this to a dilemma between the sufficiency and the\nefficiency of interactions among various positions in different frames. These\ninteractions affect the extraction of task-relevant information shared among\nframes. To resolve this issue, we prove that frame-by-frame alignments have the\npotential to increase the mutual information between frame representations,\nthereby including more task-relevant information to boost effectiveness. Then\nwe propose Alignment-guided Temporal Attention (ATA) to extend 1-dimensional\ntemporal attention with parameter-free patch-level alignments between\nneighboring frames. It can act as a general plug-in for image backbones to\nconduct the action recognition task without any model-specific design.\nExtensive experiments on multiple benchmarks demonstrate the superiority and\ngenerality of our module.\n", "date": "2022-09-30", "categories": []}
{"id": "2207.09759", "title": "Task-adaptive Spatial-Temporal Video Sampler for Few-shot Action\n  Recognition", "abstract": "  A primary challenge faced in few-shot action recognition is inadequate video\ndata for training. To address this issue, current methods in this field mainly\nfocus on devising algorithms at the feature level while little attention is\npaid to processing input video data. Moreover, existing frame sampling\nstrategies may omit critical action information in temporal and spatial\ndimensions, which further impacts video utilization efficiency. In this paper,\nwe propose a novel video frame sampler for few-shot action recognition to\naddress this issue, where task-specific spatial-temporal frame sampling is\nachieved via a temporal selector (TS) and a spatial amplifier (SA).\nSpecifically, our sampler first scans the whole video at a small computational\ncost to obtain a global perception of video frames. The TS plays its role in\nselecting top-T frames that contribute most significantly and subsequently. The\nSA emphasizes the discriminative information of each frame by amplifying\ncritical regions with the guidance of saliency maps. We further adopt\ntask-adaptive learning to dynamically adjust the sampling strategy according to\nthe episode task at hand. Both the implementations of TS and SA are\ndifferentiable for end-to-end optimization, facilitating seamless integration\nof our proposed sampler with most few-shot action recognition methods.\nExtensive experiments show a significant boost in the performances on various\nbenchmarks including long-term videos.The code is available at\nhttps://github.com/R00Kie-Liu/Sampler\n", "date": "2022-07-20", "categories": []}
{"id": "2212.11030", "title": "Deep set conditioned latent representations for action recognition", "abstract": "  In recent years multi-label, multi-class video action recognition has gained\nsignificant popularity. While reasoning over temporally connected atomic\nactions is mundane for intelligent species, standard artificial neural networks\n(ANN) still struggle to classify them. In the real world, atomic actions often\ntemporally connect to form more complex composite actions. The challenge lies\nin recognising composite action of varying durations while other distinct\ncomposite or atomic actions occur in the background. Drawing upon the success\nof relational networks, we propose methods that learn to reason over the\nsemantic concept of objects and actions. We empirically show how ANNs benefit\nfrom pretraining, relational inductive biases and unordered set-based latent\nrepresentations. In this paper we propose deep set conditioned I3D (SCI3D), a\ntwo stream relational network that employs latent representation of state and\nvisual representation for reasoning over events and actions. They learn to\nreason about temporally connected actions in order to identify all of them in\nthe video. The proposed method achieves an improvement of around 1.49% mAP in\natomic action recognition and 17.57% mAP in composite action recognition, over\na I3D-NL baseline, on the CATER dataset.\n", "date": "2022-12-21", "categories": []}
{"id": "2212.11078", "title": "C2F-TCN: A Framework for Semi and Fully Supervised Temporal Action\n  Segmentation", "abstract": "  Temporal action segmentation tags action labels for every frame in an input\nuntrimmed video containing multiple actions in a sequence. For the task of\ntemporal action segmentation, we propose an encoder-decoder-style architecture\nnamed C2F-TCN featuring a \"coarse-to-fine\" ensemble of decoder outputs. The\nC2F-TCN framework is enhanced with a novel model agnostic temporal feature\naugmentation strategy formed by the computationally inexpensive strategy of the\nstochastic max-pooling of segments. It produces more accurate and\nwell-calibrated supervised results on three benchmark action segmentation\ndatasets. We show that the architecture is flexible for both supervised and\nrepresentation learning. In line with this, we present a novel unsupervised way\nto learn frame-wise representation from C2F-TCN. Our unsupervised learning\napproach hinges on the clustering capabilities of the input features and the\nformation of multi-resolution features from the decoder's implicit structure.\nFurther, we provide the first semi-supervised temporal action segmentation\nresults by merging representation learning with conventional supervised\nlearning. Our semi-supervised learning scheme, called\n``Iterative-Contrastive-Classify (ICC)'', progressively improves in performance\nwith more labeled data. The ICC semi-supervised learning in C2F-TCN, with 40%\nlabeled videos, performs similar to fully supervised counterparts.\n", "date": "2022-12-20", "categories": []}
{"id": "2212.08859", "title": "iCub! Do you recognize what I am doing?: multimodal human action\n  recognition on multisensory-enabled iCub robot", "abstract": "  This study uses multisensory data (i.e., color and depth) to recognize human\nactions in the context of multimodal human-robot interaction. Here we employed\nthe iCub robot to observe the predefined actions of the human partners by using\nfour different tools on 20 objects. We show that the proposed multimodal\nensemble learning leverages complementary characteristics of three color\ncameras and one depth sensor that improves, in most cases, recognition accuracy\ncompared to the models trained with a single modality. The results indicate\nthat the proposed models can be deployed on the iCub robot that requires\nmultimodal action recognition, including social tasks such as partner-specific\nadaptation, and contextual behavior understanding, to mention a few.\n", "date": "2022-12-17", "categories": []}
{"id": "2212.09013", "title": "Graph Neural Network based Child Activity Recognition", "abstract": "  This paper presents an implementation on child activity recognition (CAR)\nwith a graph convolution network (GCN) based deep learning model since prior\nimplementations in this domain have been dominated by CNN, LSTM and other\nmethods despite the superior performance of GCN. To the best of our knowledge,\nwe are the first to use a GCN model in child activity recognition domain. In\novercoming the challenges of having small size publicly available child action\ndatasets, several learning methods such as feature extraction, fine-tuning and\ncurriculum learning were implemented to improve the model performance. Inspired\nby the contradicting claims made on the use of transfer learning in CAR, we\nconducted a detailed implementation and analysis on transfer learning together\nwith a study on negative transfer learning effect on CAR as it hasn't been\naddressed previously. As the principal contribution, we were able to develop a\nST-GCN based CAR model which, despite the small size of the dataset, obtained\naround 50% accuracy on vanilla implementations. With feature extraction and\nfine-tuning methods, accuracy was improved by 20%-30% with the highest accuracy\nbeing 82.24%. Furthermore, the results provided on activity datasets\nempirically demonstrate that with careful selection of pre-train model datasets\nthrough methods such as curriculum learning could enhance the accuracy levels.\nFinally, we provide preliminary evidence on possible frame rate effect on the\naccuracy of CAR models, a direction future research can explore.\n", "date": "2022-12-18", "categories": []}
{"id": "2212.09027", "title": "2D Pose Estimation based Child Action Recognition", "abstract": "  We present a graph convolutional network with 2D pose estimation for the\nfirst time on child action recognition task achieving on par results with an\nRGB modality based model on a novel benchmark dataset containing unconstrained\nenvironment based videos.\n", "date": "2022-12-18", "categories": []}
{"id": "2212.09335", "title": "Distilling Vision-Language Pre-training to Collaborate with\n  Weakly-Supervised Temporal Action Localization", "abstract": "  Weakly-supervised temporal action localization (WTAL) learns to detect and\nclassify action instances with only category labels. Most methods widely adopt\nthe off-the-shelf Classification-Based Pre-training (CBP) to generate video\nfeatures for action localization. However, the different optimization\nobjectives between classification and localization, make temporally localized\nresults suffer from the serious incomplete issue. To tackle this issue without\nadditional annotations, this paper considers to distill free action knowledge\nfrom Vision-Language Pre-training (VLP), since we surprisingly observe that the\nlocalization results of vanilla VLP have an over-complete issue, which is just\ncomplementary to the CBP results. To fuse such complementarity, we propose a\nnovel distillation-collaboration framework with two branches acting as CBP and\nVLP respectively. The framework is optimized through a dual-branch alternate\ntraining strategy. Specifically, during the B step, we distill the confident\nbackground pseudo-labels from the CBP branch; while during the F step, the\nconfident foreground pseudo-labels are distilled from the VLP branch. And as a\nresult, the dual-branch complementarity is effectively fused to promote a\nstrong alliance. Extensive experiments and ablation studies on THUMOS14 and\nActivityNet1.2 reveal that our method significantly outperforms\nstate-of-the-art methods.\n", "date": "2022-12-19", "categories": []}
{"id": "2212.08363", "title": "Fast Learning of Dynamic Hand Gesture Recognition with Few-Shot Learning\n  Models", "abstract": "  We develop Few-Shot Learning models trained to recognize five or ten\ndifferent dynamic hand gestures, respectively, which are arbitrarily\ninterchangeable by providing the model with one, two, or five examples per hand\ngesture. All models were built in the Few-Shot Learning architecture of the\nRelation Network (RN), in which Long-Short-Term Memory cells form the backbone.\nThe models use hand reference points extracted from RGB-video sequences of the\nJester dataset which was modified to contain 190 different types of hand\ngestures. Result show accuracy of up to 88.8% for recognition of five and up to\n81.2% for ten dynamic hand gestures. The research also sheds light on the\npotential effort savings of using a Few-Shot Learning approach instead of a\ntraditional Deep Learning approach to detect dynamic hand gestures. Savings\nwere defined as the number of additional observations required when a Deep\nLearning model is trained on new hand gestures instead of a Few Shot Learning\nmodel. The difference with respect to the total number of observations required\nto achieve approximately the same accuracy indicates potential savings of up to\n630 observations for five and up to 1260 observations for ten hand gestures to\nbe recognized. Since labeling video recordings of hand gestures implies\nsignificant effort, these savings can be considered substantial.\n", "date": "2022-12-16", "categories": []}
{"id": "2207.00712", "title": "Turning to a Teacher for Timestamp Supervised Temporal Action\n  Segmentation", "abstract": "  Temporal action segmentation in videos has drawn much attention recently.\nTimestamp supervision is a cost-effective way for this task. To obtain more\ninformation to optimize the model, the existing method generated pseudo\nframe-wise labels iteratively based on the output of a segmentation model and\nthe timestamp annotations. However, this practice may introduce noise and\noscillation during the training, and lead to performance degeneration. To\naddress this problem, we propose a new framework for timestamp supervised\ntemporal action segmentation by introducing a teacher model parallel to the\nsegmentation model to help stabilize the process of model optimization. The\nteacher model can be seen as an ensemble of the segmentation model, which helps\nto suppress the noise and to improve the stability of pseudo labels. We further\nintroduce a segmentally smoothing loss, which is more focused and cohesive, to\nenforce the smooth transition of the predicted probabilities within action\ninstances. The experiments on three datasets show that our method outperforms\nthe state-of-the-art method and performs comparably against the\nfully-supervised methods at a much lower annotation cost.\n", "date": "2022-07-02", "categories": []}
{"id": "2212.06348", "title": "Dilation-Erosion for Single-Frame Supervised Temporal Action\n  Localization", "abstract": "  To balance the annotation labor and the granularity of supervision,\nsingle-frame annotation has been introduced in temporal action localization. It\nprovides a rough temporal location for an action but implicitly overstates the\nsupervision from the annotated-frame during training, leading to the confusion\nbetween actions and backgrounds, i.e., action incompleteness and background\nfalse positives. To tackle the two challenges, in this work, we present the\nSnippet Classification model and the Dilation-Erosion module. In the\nDilation-Erosion module, we expand the potential action segments with a loose\ncriterion to alleviate the problem of action incompleteness and then remove the\nbackground from the potential action segments to alleviate the problem of\naction incompleteness. Relying on the single-frame annotation and the output of\nthe snippet classification, the Dilation-Erosion module mines pseudo\nsnippet-level ground-truth, hard backgrounds and evident backgrounds, which in\nturn further trains the Snippet Classification model. It forms a cyclic\ndependency. Furthermore, we propose a new embedding loss to aggregate the\nfeatures of action instances with the same label and separate the features of\nactions from backgrounds. Experiments on THUMOS14 and ActivityNet 1.2 validate\nthe effectiveness of the proposed method. Code has been made publicly available\n(https://github.com/LingJun123/single-frame-TAL).\n", "date": "2022-12-13", "categories": []}
{"id": "2212.06023", "title": "Reconstructing Humpty Dumpty: Multi-feature Graph Autoencoder for Open\n  Set Action Recognition", "abstract": "  Most action recognition datasets and algorithms assume a closed world, where\nall test samples are instances of the known classes. In open set problems, test\nsamples may be drawn from either known or unknown classes. Existing open set\naction recognition methods are typically based on extending closed set methods\nby adding post hoc analysis of classification scores or feature distances and\ndo not capture the relations among all the video clip elements. Our approach\nuses the reconstruction error to determine the novelty of the video since\nunknown classes are harder to put back together and thus have a higher\nreconstruction error than videos from known classes. We refer to our solution\nto the open set action recognition problem as \"Humpty Dumpty\", due to its\nreconstruction abilities. Humpty Dumpty is a novel graph-based autoencoder that\naccounts for contextual and semantic relations among the clip pieces for\nimproved reconstruction. A larger reconstruction error leads to an increased\nlikelihood that the action can not be reconstructed, i.e., can not put Humpty\nDumpty back together again, indicating that the action has never been seen\nbefore and is novel/unknown. Extensive experiments are performed on two\npublicly available action recognition datasets including HMDB-51 and UCF-101,\nshowing the state-of-the-art performance for open set action recognition.\n", "date": "2022-12-12", "categories": []}
{"id": "2209.09092", "title": "TASKED: Transformer-based Adversarial learning for human activity\n  recognition using wearable sensors via Self-KnowledgE Distillation", "abstract": "  Wearable sensor-based human activity recognition (HAR) has emerged as a\nprincipal research area and is utilized in a variety of applications. Recently,\ndeep learning-based methods have achieved significant improvement in the HAR\nfield with the development of human-computer interaction applications. However,\nthey are limited to operating in a local neighborhood in the process of a\nstandard convolution neural network, and correlations between different sensors\non body positions are ignored. In addition, they still face significant\nchallenging problems with performance degradation due to large gaps in the\ndistribution of training and test data, and behavioral differences between\nsubjects. In this work, we propose a novel Transformer-based Adversarial\nlearning framework for human activity recognition using wearable sensors via\nSelf-KnowledgE Distillation (TASKED), that accounts for individual sensor\norientations and spatial and temporal features. The proposed method is capable\nof learning cross-domain embedding feature representations from multiple\nsubjects datasets using adversarial learning and the maximum mean discrepancy\n(MMD) regularization to align the data distribution over multiple domains. In\nthe proposed method, we adopt the teacher-free self-knowledge distillation to\nimprove the stability of the training procedure and the performance of human\nactivity recognition. Experimental results show that TASKED not only\noutperforms state-of-the-art methods on the four real-world public HAR datasets\n(alone or combined) but also improves the subject generalization effectively.\n", "date": "2022-09-14", "categories": []}
{"id": "2212.03384", "title": "DroneAttention: Sparse Weighted Temporal Attention for Drone-Camera\n  Based Activity Recognition", "abstract": "  Human activity recognition (HAR) using drone-mounted cameras has attracted\nconsiderable interest from the computer vision research community in recent\nyears. A robust and efficient HAR system has a pivotal role in fields like\nvideo surveillance, crowd behavior analysis, sports analysis, and\nhuman-computer interaction. What makes it challenging are the complex poses,\nunderstanding different viewpoints, and the environmental scenarios where the\naction is taking place. To address such complexities, in this paper, we propose\na novel Sparse Weighted Temporal Attention (SWTA) module to utilize sparsely\nsampled video frames for obtaining global weighted temporal attention. The\nproposed SWTA is comprised of two parts. First, temporal segment network that\nsparsely samples a given set of frames. Second, weighted temporal attention,\nwhich incorporates a fusion of attention maps derived from optical flow, with\nraw RGB images. This is followed by a basenet network, which comprises a\nconvolutional neural network (CNN) module along with fully connected layers\nthat provide us with activity recognition. The SWTA network can be used as a\nplug-in module to the existing deep CNN architectures, for optimizing them to\nlearn temporal information by eliminating the need for a separate temporal\nstream. It has been evaluated on three publicly available benchmark datasets,\nnamely Okutama, MOD20, and Drone-Action. The proposed model has received an\naccuracy of 72.76%, 92.56%, and 78.86% on the respective datasets thereby\nsurpassing the previous state-of-the-art performances by a margin of 25.26%,\n18.56%, and 2.94%, respectively.\n", "date": "2022-12-07", "categories": []}
{"id": "2212.01638", "title": "VLG: General Video Recognition with Web Textual Knowledge", "abstract": "  Video recognition in an open and dynamic world is quite challenging, as we\nneed to handle different settings such as close-set, long-tail, few-shot and\nopen-set. By leveraging semantic knowledge from noisy text descriptions crawled\nfrom the Internet, we focus on the general video recognition (GVR) problem of\nsolving different recognition tasks within a unified framework. The core\ncontribution of this paper is twofold. First, we build a comprehensive video\nrecognition benchmark of Kinetics-GVR, including four sub-task datasets to\ncover the mentioned settings. To facilitate the research of GVR, we propose to\nutilize external textual knowledge from the Internet and provide multi-source\ntext descriptions for all action classes. Second, inspired by the flexibility\nof language representation, we present a unified visual-linguistic framework\n(VLG) to solve the problem of GVR by an effective two-stage training paradigm.\nOur VLG is first pre-trained on video and language datasets to learn a shared\nfeature space, and then devises a flexible bi-modal attention head to\ncollaborate high-level semantic concepts under different settings. Extensive\nresults show that our VLG obtains the state-of-the-art performance under four\nsettings. The superior performance demonstrates the effectiveness and\ngeneralization ability of our proposed framework. We hope our work makes a step\ntowards the general video recognition and could serve as a baseline for future\nresearch. The code and models will be available at\nhttps://github.com/MCG-NJU/VLG.\n", "date": "2022-12-03", "categories": []}
{"id": "2212.02082", "title": "Hierarchical Contrast for Unsupervised Skeleton-based Action\n  Representation Learning", "abstract": "  This paper targets unsupervised skeleton-based action representation learning\nand proposes a new Hierarchical Contrast (HiCo) framework. Different from the\nexisting contrastive-based solutions that typically represent an input skeleton\nsequence into instance-level features and perform contrast holistically, our\nproposed HiCo represents the input into multiple-level features and performs\ncontrast in a hierarchical manner. Specifically, given a human skeleton\nsequence, we represent it into multiple feature vectors of different\ngranularities from both temporal and spatial domains via sequence-to-sequence\n(S2S) encoders and unified downsampling modules. Besides, the hierarchical\ncontrast is conducted in terms of four levels: instance level, domain level,\nclip level, and part level. Moreover, HiCo is orthogonal to the S2S encoder,\nwhich allows us to flexibly embrace state-of-the-art S2S encoders. Extensive\nexperiments on four datasets, i.e., NTU-60, NTU-120, PKU-MMD I and II, show\nthat HiCo achieves a new state-of-the-art for unsupervised skeleton-based\naction representation learning in two downstream tasks including action\nrecognition and retrieval, and its learned action representation is of good\ntransferability. Besides, we also show that our framework is effective for\nsemi-supervised skeleton-based action recognition. Our code is available at\nhttps://github.com/HuiGuanLab/HiCo.\n", "date": "2022-12-05", "categories": []}
{"id": "2212.02266", "title": "Applications of human activity recognition in industrial processes --\n  Synergy of human and technology", "abstract": "  Human-technology collaboration relies on verbal and non-verbal communication.\nMachines must be able to detect and understand the movements of humans to\nfacilitate non-verbal communication. In this article, we introduce ongoing\nresearch on human activity recognition in intralogistics, and show how it can\nbe applied in industrial settings. We show how semantic attributes can be used\nto describe human activities flexibly and how context informantion increases\nthe performance of classifiers to recognise them automatically. Beyond that, we\npresent a concept based on a cyber-physical twin that can reduce the effort and\ntime necessary to create a training dataset for human activity recognition. In\nthe future, it will be possible to train a classifier solely with realistic\nsimulation data, while maintaining or even increasing the classification\nperformance.\n", "date": "2022-12-05", "categories": []}
{"id": "2208.14856", "title": "Active Learning with Effective Scoring Functions for Semi-Supervised\n  Temporal Action Localization", "abstract": "  Temporal Action Localization (TAL) aims to predict both action category and\ntemporal boundary of action instances in untrimmed videos, i.e., start and end\ntime. Fully-supervised solutions are usually adopted in most existing works,\nand proven to be effective. One of the practical bottlenecks in these solutions\nis the large amount of labeled training data required. To reduce expensive\nhuman label cost, this paper focuses on a rarely investigated yet practical\ntask named semi-supervised TAL and proposes an effective active learning\nmethod, named AL-STAL. We leverage four steps for actively selecting video\nsamples with high informativeness and training the localization model, named\n\\emph{Train, Query, Annotate, Append}. Two scoring functions that consider the\nuncertainty of localization model are equipped in AL-STAL, thus facilitating\nthe video sample rank and selection. One takes entropy of predicted label\ndistribution as measure of uncertainty, named Temporal Proposal Entropy (TPE).\nAnd the other introduces a new metric based on mutual information between\nadjacent action proposals and evaluates the informativeness of video samples,\nnamed Temporal Context Inconsistency (TCI). To validate the effectiveness of\nproposed method, we conduct extensive experiments on two benchmark datasets\nTHUMOS'14 and ActivityNet 1.3. Experiment results show that AL-STAL outperforms\nthe existing competitors and achieves satisfying performance compared with\nfully-supervised learning.\n", "date": "2022-08-31", "categories": []}
{"id": "2211.17045", "title": "From Actions to Events: A Transfer Learning Approach Using Improved Deep\n  Belief Networks", "abstract": "  In the last decade, exponential data growth supplied machine learning-based\nalgorithms' capacity and enabled their usage in daily-life activities.\nAdditionally, such an improvement is partially explained due to the advent of\ndeep learning techniques, i.e., stacks of simple architectures that end up in\nmore complex models. Although both factors produce outstanding results, they\nalso pose drawbacks regarding the learning process as training complex models\nover large datasets are expensive and time-consuming. Such a problem is even\nmore evident when dealing with video analysis. Some works have considered\ntransfer learning or domain adaptation, i.e., approaches that map the knowledge\nfrom one domain to another, to ease the training burden, yet most of them\noperate over individual or small blocks of frames. This paper proposes a novel\napproach to map the knowledge from action recognition to event recognition\nusing an energy-based model, denoted as Spectral Deep Belief Network. Such a\nmodel can process all frames simultaneously, carrying spatial and temporal\ninformation through the learning process. The experimental results conducted\nover two public video dataset, the HMDB-51 and the UCF-101, depict the\neffectiveness of the proposed model and its reduced computational burden when\ncompared to traditional energy-based models, such as Restricted Boltzmann\nMachines and Deep Belief Networks.\n", "date": "2022-11-30", "categories": []}
{"id": "2206.06346", "title": "Bringing Image Scene Structure to Video via Frame-Clip Consistency of\n  Object Tokens", "abstract": "  Recent action recognition models have achieved impressive results by\nintegrating objects, their locations and interactions. However, obtaining dense\nstructured annotations for each frame is tedious and time-consuming, making\nthese methods expensive to train and less scalable. At the same time, if a\nsmall set of annotated images is available, either within or outside the domain\nof interest, how could we leverage these for a video downstream task? We\npropose a learning framework StructureViT (SViT for short), which demonstrates\nhow utilizing the structure of a small number of images only available during\ntraining can improve a video model. SViT relies on two key insights. First, as\nboth images and videos contain structured information, we enrich a transformer\nmodel with a set of \\emph{object tokens} that can be used across images and\nvideos. Second, the scene representations of individual frames in video should\n\"align\" with those of still images. This is achieved via a \\emph{Frame-Clip\nConsistency} loss, which ensures the flow of structured information between\nimages and videos. We explore a particular instantiation of scene structure,\nnamely a \\emph{Hand-Object Graph}, consisting of hands and objects with their\nlocations as nodes, and physical relations of contact/no-contact as edges. SViT\nshows strong performance improvements on multiple video understanding tasks and\ndatasets. Furthermore, it won in the Ego4D CVPR'22 Object State Localization\nchallenge. For code and pretrained models, visit the project page at\n\\url{https://eladb3.github.io/SViT/}\n", "date": "2022-06-13", "categories": []}
{"id": "2209.12362", "title": "Multi-dataset Training of Transformers for Robust Action Recognition", "abstract": "  We study the task of robust feature representations, aiming to generalize\nwell on multiple datasets for action recognition. We build our method on\nTransformers for its efficacy. Although we have witnessed great progress for\nvideo action recognition in the past decade, it remains challenging yet\nvaluable how to train a single model that can perform well across multiple\ndatasets. Here, we propose a novel multi-dataset training paradigm, MultiTrain,\nwith the design of two new loss terms, namely informative loss and projection\nloss, aiming to learn robust representations for action recognition. In\nparticular, the informative loss maximizes the expressiveness of the feature\nembedding while the projection loss for each dataset mines the intrinsic\nrelations between classes across datasets. We verify the effectiveness of our\nmethod on five challenging datasets, Kinetics-400, Kinetics-700,\nMoments-in-Time, Activitynet and Something-something-v2 datasets. Extensive\nexperimental results show that our method can consistently improve\nstate-of-the-art performance. Code and models are released.\n", "date": "2022-09-26", "categories": []}
{"id": "2211.12748", "title": "Dynamic Appearance: A Video Representation for Action Recognition with\n  Joint Training", "abstract": "  Static appearance of video may impede the ability of a deep neural network to\nlearn motion-relevant features in video action recognition. In this paper, we\nintroduce a new concept, Dynamic Appearance (DA), summarizing the appearance\ninformation relating to movement in a video while filtering out the static\ninformation considered unrelated to motion. We consider distilling the dynamic\nappearance from raw video data as a means of efficient video understanding. To\nthis end, we propose the Pixel-Wise Temporal Projection (PWTP), which projects\nthe static appearance of a video into a subspace within its original vector\nspace, while the dynamic appearance is encoded in the projection residual\ndescribing a special motion pattern. Moreover, we integrate the PWTP module\nwith a CNN or Transformer into an end-to-end training framework, which is\noptimized by utilizing multi-objective optimization algorithms. We provide\nextensive experimental results on four action recognition benchmarks:\nKinetics400, Something-Something V1, UCF101 and HMDB51.\n", "date": "2022-11-23", "categories": []}
{"id": "2211.13471", "title": "Minority-Oriented Vicinity Expansion with Attentive Aggregation for\n  Video Long-Tailed Recognition", "abstract": "  A dramatic increase in real-world video volume with extremely diverse and\nemerging topics naturally forms a long-tailed video distribution in terms of\ntheir categories, and it spotlights the need for Video Long-Tailed Recognition\n(VLTR). In this work, we summarize the challenges in VLTR and explore how to\novercome them. The challenges are: (1) it is impractical to re-train the whole\nmodel for high-quality features, (2) acquiring frame-wise labels requires\nextensive cost, and (3) long-tailed data triggers biased training. Yet, most\nexisting works for VLTR unavoidably utilize image-level features extracted from\npretrained models which are task-irrelevant, and learn by video-level labels.\nTherefore, to deal with such (1) task-irrelevant features and (2) video-level\nlabels, we introduce two complementary learnable feature aggregators. Learnable\nlayers in each aggregator are to produce task-relevant representations, and\neach aggregator is to assemble the snippet-wise knowledge into a video\nrepresentative. Then, we propose Minority-Oriented Vicinity Expansion (MOVE)\nthat explicitly leverages the class frequency into approximating the vicinity\ndistributions to alleviate (3) biased training. By combining these solutions,\nour approach achieves state-of-the-art results on large-scale VideoLT and\nsynthetically induced Imbalanced-MiniKinetics200. With VideoLT features from\nResNet-50, it attains 18% and 58% relative improvements on head and tail\nclasses over the previous state-of-the-art method, respectively.\n", "date": "2022-11-24", "categories": []}
{"id": "2211.13694", "title": "Hand Guided High Resolution Feature Enhancement for Fine-Grained Atomic\n  Action Segmentation within Complex Human Assemblies", "abstract": "  Due to the rapid temporal and fine-grained nature of complex human assembly\natomic actions, traditional action segmentation approaches requiring the\nspatial (and often temporal) down sampling of video frames often loose vital\nfine-grained spatial and temporal information required for accurate\nclassification within the manufacturing domain. In order to fully utilise\nhigher resolution video data (often collected within the manufacturing domain)\nand facilitate real time accurate action segmentation - required for human\nrobot collaboration - we present a novel hand location guided high resolution\nfeature enhanced model. We also propose a simple yet effective method of\ndeploying offline trained action recognition models for real time action\nsegmentation on temporally short fine-grained actions, through the use of\nsurround sampling while training and temporally aware label cleaning at\ninference. We evaluate our model on a novel action segmentation dataset\ncontaining 24 (+background) atomic actions from video data of a real world\nrobotics assembly production line. Showing both high resolution hand features\nas well as traditional frame wide features improve fine-grained atomic action\nclassification, and that though temporally aware label clearing our model is\ncapable of surpassing similar encoder/decoder methods, while allowing for real\ntime classification.\n", "date": "2022-11-24", "categories": []}
{"id": "2211.13171", "title": "Query Efficient Cross-Dataset Transferable Black-Box Attack on Action\n  Recognition", "abstract": "  Black-box adversarial attacks present a realistic threat to action\nrecognition systems. Existing black-box attacks follow either a query-based\napproach where an attack is optimized by querying the target model, or a\ntransfer-based approach where attacks are generated using a substitute model.\nWhile these methods can achieve decent fooling rates, the former tends to be\nhighly query-inefficient while the latter assumes extensive knowledge of the\nblack-box model's training data. In this paper, we propose a new attack on\naction recognition that addresses these shortcomings by generating\nperturbations to disrupt the features learned by a pre-trained substitute model\nto reduce the number of queries. By using a nearly disjoint dataset to train\nthe substitute model, our method removes the requirement that the substitute\nmodel be trained using the same dataset as the target model, and leverages\nqueries to the target model to retain the fooling rate benefits provided by\nquery-based methods. This ultimately results in attacks which are more\ntransferable than conventional black-box attacks. Through extensive\nexperiments, we demonstrate highly query-efficient black-box attacks with the\nproposed framework. Our method achieves 8% and 12% higher deception rates\ncompared to state-of-the-art query-based and transfer-based attacks,\nrespectively.\n", "date": "2022-11-23", "categories": []}
{"id": "2211.12030", "title": "Knowledge Prompting for Few-shot Action Recognition", "abstract": "  Few-shot action recognition in videos is challenging for its lack of\nsupervision and difficulty in generalizing to unseen actions. To address this\ntask, we propose a simple yet effective method, called knowledge prompting,\nwhich leverages commonsense knowledge of actions from external resources to\nprompt a powerful pre-trained vision-language model for few-shot\nclassification. We first collect large-scale language descriptions of actions,\ndefined as text proposals, to build an action knowledge base. The collection of\ntext proposals is done by filling in handcraft sentence templates with external\naction-related corpus or by extracting action-related phrases from captions of\nWeb instruction videos.Then we feed these text proposals into the pre-trained\nvision-language model along with video frames to generate matching scores of\nthe proposals to each frame, and the scores can be treated as action semantics\nwith strong generalization. Finally, we design a lightweight temporal modeling\nnetwork to capture the temporal evolution of action semantics for\nclassification.Extensive experiments on six benchmark datasets demonstrate that\nour method generally achieves the state-of-the-art performance while reducing\nthe training overhead to 0.001 of existing methods.\n", "date": "2022-11-22", "categories": []}
{"id": "2211.12334", "title": "A Graph-Based Method for Soccer Action Spotting Using Unsupervised\n  Player Classification", "abstract": "  Action spotting in soccer videos is the task of identifying the specific time\nwhen a certain key action of the game occurs. Lately, it has received a large\namount of attention and powerful methods have been introduced. Action spotting\ninvolves understanding the dynamics of the game, the complexity of events, and\nthe variation of video sequences. Most approaches have focused on the latter,\ngiven that their models exploit the global visual features of the sequences. In\nthis work, we focus on the former by (a) identifying and representing the\nplayers, referees, and goalkeepers as nodes in a graph, and by (b) modeling\ntheir temporal interactions as sequences of graphs. For the player\nidentification, or player classification task, we obtain an accuracy of 97.72%\nin our annotated benchmark. For the action spotting task, our method obtains an\noverall performance of 57.83% average-mAP by combining it with other\naudiovisual modalities. This performance surpasses similar graph-based methods\nand has competitive results with heavy computing methods. Code and data are\navailable at https://github.com/IPCV/soccer_action_spotting.\n", "date": "2022-11-22", "categories": []}
{"id": "2211.11324", "title": "Slow Motion Matters: A Slow Motion Enhanced Network for Weakly\n  Supervised Temporal Action Localization", "abstract": "  Weakly supervised temporal action localization (WTAL) aims to localize\nactions in untrimmed videos with only weak supervision information (e.g.\nvideo-level labels). Most existing models handle all input videos with a fixed\ntemporal scale. However, such models are not sensitive to actions whose pace of\nthe movements is different from the ``normal\" speed, especially slow-motion\naction instances, which complete the movements with a much slower speed than\ntheir counterparts with a normal speed. Here arises the slow-motion blurred\nissue: It is hard to explore salient slow-motion information from videos at\n``normal\" speed. In this paper, we propose a novel framework termed Slow Motion\nEnhanced Network (SMEN) to improve the ability of a WTAL network by\ncompensating its sensitivity on slow-motion action segments. The proposed SMEN\ncomprises a Mining module and a Localization module. The mining module\ngenerates mask to mine slow-motion-related features by utilizing the\nrelationships between the normal motion and slow motion; while the localization\nmodule leverages the mined slow-motion features as complementary information to\nimprove the temporal action localization results. Our proposed framework can be\neasily adapted by existing WTAL networks and enable them be more sensitive to\nslow-motion actions. Extensive experiments on three benchmarks are conducted,\nwhich demonstrate the high performance of our proposed framework.\n", "date": "2022-11-21", "categories": []}
{"id": "2210.12686", "title": "Holistic Interaction Transformer Network for Action Detection", "abstract": "  Actions are about how we interact with the environment, including other\npeople, objects, and ourselves. In this paper, we propose a novel multi-modal\nHolistic Interaction Transformer Network (HIT) that leverages the largely\nignored, but critical hand and pose information essential to most human\nactions. The proposed \"HIT\" network is a comprehensive bi-modal framework that\ncomprises an RGB stream and a pose stream. Each of them separately models\nperson, object, and hand interactions. Within each sub-network, an\nIntra-Modality Aggregation module (IMA) is introduced that selectively merges\nindividual interaction units. The resulting features from each modality are\nthen glued using an Attentive Fusion Mechanism (AFM). Finally, we extract cues\nfrom the temporal context to better classify the occurring actions using cached\nmemory. Our method significantly outperforms previous approaches on the J-HMDB,\nUCF101-24, and MultiSports datasets. We also achieve competitive results on\nAVA. The code will be available at https://github.com/joslefaure/HIT.\n", "date": "2022-10-23", "categories": []}
{"id": "2211.09950", "title": "TempNet: Temporal Attention Towards the Detection of Animal Behaviour in\n  Videos", "abstract": "  Recent advancements in cabled ocean observatories have increased the quality\nand prevalence of underwater videos; this data enables the extraction of\nhigh-level biologically relevant information such as species' behaviours.\nDespite this increase in capability, most modern methods for the automatic\ninterpretation of underwater videos focus only on the detection and counting\norganisms. We propose an efficient computer vision- and deep learning-based\nmethod for the detection of biological behaviours in videos. TempNet uses an\nencoder bridge and residual blocks to maintain model performance with a\ntwo-staged, spatial, then temporal, encoder. TempNet also presents temporal\nattention during spatial encoding as well as Wavelet Down-Sampling\npre-processing to improve model accuracy. Although our system is designed for\napplications to diverse fish behaviours (i.e, is generic), we demonstrate its\napplication to the detection of sablefish (Anoplopoma fimbria) startle events.\nWe compare the proposed approach with a state-of-the-art end-to-end video\ndetection method (ReMotENet) and a hybrid method previously offered exclusively\nfor the detection of sablefish's startle events in videos from an existing\ndataset. Results show that our novel method comfortably outperforms the\ncomparison baselines in multiple metrics, reaching a per-clip accuracy and\nprecision of 80% and 0.81, respectively. This represents a relative improvement\nof 31% in accuracy and 27% in precision over the compared methods using this\ndataset. Our computational pipeline is also highly efficient, as it can process\neach 4-second video clip in only 38ms. Furthermore, since it does not employ\nfeatures specific to sablefish startle events, our system can be easily\nextended to other behaviours in future works.\n", "date": "2022-11-17", "categories": []}
{"id": "2211.09992", "title": "Look More but Care Less in Video Recognition", "abstract": "  Existing action recognition methods typically sample a few frames to\nrepresent each video to avoid the enormous computation, which often limits the\nrecognition performance. To tackle this problem, we propose Ample and Focal\nNetwork (AFNet), which is composed of two branches to utilize more frames but\nwith less computation. Specifically, the Ample Branch takes all input frames to\nobtain abundant information with condensed computation and provides the\nguidance for Focal Branch by the proposed Navigation Module; the Focal Branch\nsqueezes the temporal size to only focus on the salient frames at each\nconvolution block; in the end, the results of two branches are adaptively fused\nto prevent the loss of information. With this design, we can introduce more\nframes to the network but cost less computation. Besides, we demonstrate AFNet\ncan utilize fewer frames while achieving higher accuracy as the dynamic\nselection in intermediate features enforces implicit temporal modeling.\nFurther, we show that our method can be extended to reduce spatial redundancy\nwith even less cost. Extensive experiments on five datasets demonstrate the\neffectiveness and efficiency of our method.\n", "date": "2022-11-18", "categories": []}
{"id": "2211.09648", "title": "HARDVS: Revisiting Human Activity Recognition with Dynamic Vision\n  Sensors", "abstract": "  The main streams of human activity recognition (HAR) algorithms are developed\nbased on RGB cameras which are suffered from illumination, fast motion,\nprivacy-preserving, and large energy consumption. Meanwhile, the biologically\ninspired event cameras attracted great interest due to their unique features,\nsuch as high dynamic range, dense temporal but sparse spatial resolution, low\nlatency, low power, etc. As it is a newly arising sensor, even there is no\nrealistic large-scale dataset for HAR. Considering its great practical value,\nin this paper, we propose a large-scale benchmark dataset to bridge this gap,\ntermed HARDVS, which contains 300 categories and more than 100K event\nsequences. We evaluate and report the performance of multiple popular HAR\nalgorithms, which provide extensive baselines for future works to compare. More\nimportantly, we propose a novel spatial-temporal feature learning and fusion\nframework, termed ESTF, for event stream based human activity recognition. It\nfirst projects the event streams into spatial and temporal embeddings using\nStemNet, then, encodes and fuses the dual-view representations using\nTransformer networks. Finally, the dual features are concatenated and fed into\na classification head for activity prediction. Extensive experiments on\nmultiple datasets fully validated the effectiveness of our model. Both the\ndataset and source code will be released on\n\\url{https://github.com/Event-AHU/HARDVS}.\n", "date": "2022-11-17", "categories": []}
{"id": "2211.09074", "title": "Where a Strong Backbone Meets Strong Features -- ActionFormer for Ego4D\n  Moment Queries Challenge", "abstract": "  This report describes our submission to the Ego4D Moment Queries Challenge\n2022. Our submission builds on ActionFormer, the state-of-the-art backbone for\ntemporal action localization, and a trio of strong video features from\nSlowFast, Omnivore and EgoVLP. Our solution is ranked 2nd on the public\nleaderboard with 21.76% average mAP on the test set, which is nearly three\ntimes higher than the official baseline. Further, we obtain 42.54% Recall@1x at\ntIoU=0.5 on the test set, outperforming the top-ranked solution by a\nsignificant margin of 1.41 absolute percentage points. Our code is available at\nhttps://github.com/happyharrycn/actionformer_release.\n", "date": "2022-11-16", "categories": []}
{"id": "2211.08252", "title": "Dynamic Temporal Filtering in Video Models", "abstract": "  Video temporal dynamics is conventionally modeled with 3D spatial-temporal\nkernel or its factorized version comprised of 2D spatial kernel and 1D temporal\nkernel. The modeling power, nevertheless, is limited by the fixed window size\nand static weights of a kernel along the temporal dimension. The pre-determined\nkernel size severely limits the temporal receptive fields and the fixed weights\ntreat each spatial location across frames equally, resulting in sub-optimal\nsolution for long-range temporal modeling in natural scenes. In this paper, we\npresent a new recipe of temporal feature learning, namely Dynamic Temporal\nFilter (DTF), that novelly performs spatial-aware temporal modeling in\nfrequency domain with large temporal receptive field. Specifically, DTF\ndynamically learns a specialized frequency filter for every spatial location to\nmodel its long-range temporal dynamics. Meanwhile, the temporal feature of each\nspatial location is also transformed into frequency feature spectrum via 1D\nFast Fourier Transform (FFT). The spectrum is modulated by the learnt frequency\nfilter, and then transformed back to temporal domain with inverse FFT. In\naddition, to facilitate the learning of frequency filter in DTF, we perform\nframe-wise aggregation to enhance the primary temporal feature with its\ntemporal neighbors by inter-frame correlation. It is feasible to plug DTF block\ninto ConvNets and Transformer, yielding DTF-Net and DTF-Transformer. Extensive\nexperiments conducted on three datasets demonstrate the superiority of our\nproposals. More remarkably, DTF-Transformer achieves an accuracy of 83.5% on\nKinetics-400 dataset. Source code is available at\n\\url{https://github.com/FuchenUSTC/DTF}.\n", "date": "2022-11-15", "categories": []}
{"id": "2211.06023", "title": "Soft-Landing Strategy for Alleviating the Task Discrepancy Problem in\n  Temporal Action Localization Tasks", "abstract": "  Temporal Action Localization (TAL) methods typically operate on top of\nfeature sequences from a frozen snippet encoder that is pretrained with the\nTrimmed Action Classification (TAC) tasks, resulting in a task discrepancy\nproblem. While existing TAL methods mitigate this issue either by retraining\nthe encoder with a pretext task or by end-to-end fine-tuning, they commonly\nrequire an overload of high memory and computation. In this work, we introduce\nSoft-Landing (SoLa) strategy, an efficient yet effective framework to bridge\nthe transferability gap between the pretrained encoder and the downstream tasks\nby incorporating a light-weight neural network, i.e., a SoLa module, on top of\nthe frozen encoder. We also propose an unsupervised training scheme for the\nSoLa module; it learns with inter-frame Similarity Matching that uses the frame\ninterval as its supervisory signal, eliminating the need for temporal\nannotations. Experimental evaluation on various benchmarks for downstream TAL\ntasks shows that our method effectively alleviates the task discrepancy problem\nwith remarkable computational efficiency.\n", "date": "2022-11-11", "categories": []}
{"id": "2211.06173", "title": "Investigating Enhancements to Contrastive Predictive Coding for Human\n  Activity Recognition", "abstract": "  The dichotomy between the challenging nature of obtaining annotations for\nactivities, and the more straightforward nature of data collection from\nwearables, has resulted in significant interest in the development of\ntechniques that utilize large quantities of unlabeled data for learning\nrepresentations. Contrastive Predictive Coding (CPC) is one such method,\nlearning effective representations by leveraging properties of time-series data\nto setup a contrastive future timestep prediction task. In this work, we\npropose enhancements to CPC, by systematically investigating the encoder\narchitecture, the aggregator network, and the future timestep prediction,\nresulting in a fully convolutional architecture, thereby improving\nparallelizability. Across sensor positions and activities, our method shows\nsubstantial improvements on four of six target datasets, demonstrating its\nability to empower a wide range of application scenarios. Further, in the\npresence of very limited labeled data, our technique significantly outperforms\nboth supervised and self-supervised baselines, positively impacting situations\nwhere collecting only a few seconds of labeled data may be possible. This is\npromising, as CPC does not require specialized data transformations or\nreconstructions for learning effective representations.\n", "date": "2022-11-11", "categories": []}
{"id": "2209.03561", "title": "Video Vision Transformers for Violence Detection", "abstract": "  Law enforcement and city safety are significantly impacted by detecting\nviolent incidents in surveillance systems. Although modern (smart) cameras are\nwidely available and affordable, such technological solutions are impotent in\nmost instances. Furthermore, personnel monitoring CCTV recordings frequently\nshow a belated reaction, resulting in the potential cause of catastrophe to\npeople and property. Thus automated detection of violence for swift actions is\nvery crucial. The proposed solution uses a novel end-to-end deep learning-based\nvideo vision transformer (ViViT) that can proficiently discern fights, hostile\nmovements, and violent events in video sequences. The study presents utilizing\na data augmentation strategy to overcome the downside of weaker inductive\nbiasness while training vision transformers on a smaller training datasets. The\nevaluated results can be subsequently sent to local concerned authority, and\nthe captured video can be analyzed. In comparison to state-of-theart (SOTA)\napproaches the proposed method achieved auspicious performance on some of the\nchallenging benchmark datasets.\n", "date": "2022-09-08", "categories": []}
{"id": "2211.05299", "title": "Prior-enhanced Temporal Action Localization using Subject-aware Spatial\n  Attention", "abstract": "  Temporal action localization (TAL) aims to detect the boundary and identify\nthe class of each action instance in a long untrimmed video. Current approaches\ntreat video frames homogeneously, and tend to give background and key objects\nexcessive attention. This limits their sensitivity to localize action\nboundaries. To this end, we propose a prior-enhanced temporal action\nlocalization method (PETAL), which only takes in RGB input and incorporates\naction subjects as priors. This proposal leverages action subjects' information\nwith a plug-and-play subject-aware spatial attention module (SA-SAM) to\ngenerate an aggregated and subject-prioritized representation. Experimental\nresults on THUMOS-14 and ActivityNet-1.3 datasets demonstrate that the proposed\nPETAL achieves competitive performance using only RGB features, e.g., boosting\nmAP by 2.41% or 0.25% over the state-of-the-art approach that uses RGB features\nor with additional optical flow features on the THUMOS-14 dataset.\n", "date": "2022-11-10", "categories": []}
{"id": "2211.05531", "title": "SWTF: Sparse Weighted Temporal Fusion for Drone-Based Activity\n  Recognition", "abstract": "  Drone-camera based human activity recognition (HAR) has received significant\nattention from the computer vision research community in the past few years. A\nrobust and efficient HAR system has a pivotal role in fields like video\nsurveillance, crowd behavior analysis, sports analysis, and human-computer\ninteraction. What makes it challenging are the complex poses, understanding\ndifferent viewpoints, and the environmental scenarios where the action is\ntaking place. To address such complexities, in this paper, we propose a novel\nSparse Weighted Temporal Fusion (SWTF) module to utilize sparsely sampled video\nframes for obtaining global weighted temporal fusion outcome. The proposed SWTF\nis divided into two components. First, a temporal segment network that sparsely\nsamples a given set of frames. Second, weighted temporal fusion, that\nincorporates a fusion of feature maps derived from optical flow, with raw RGB\nimages. This is followed by base-network, which comprises a convolutional\nneural network module along with fully connected layers that provide us with\nactivity recognition. The SWTF network can be used as a plug-in module to the\nexisting deep CNN architectures, for optimizing them to learn temporal\ninformation by eliminating the need for a separate temporal stream. It has been\nevaluated on three publicly available benchmark datasets, namely Okutama,\nMOD20, and Drone-Action. The proposed model has received an accuracy of 72.76%,\n92.56%, and 78.86% on the respective datasets thereby surpassing the previous\nstate-of-the-art performances by a significant margin.\n", "date": "2022-11-10", "categories": []}
{"id": "2211.04888", "title": "Extending Temporal Data Augmentation for Video Action Recognition", "abstract": "  Pixel space augmentation has grown in popularity in many Deep Learning areas,\ndue to its effectiveness, simplicity, and low computational cost. Data\naugmentation for videos, however, still remains an under-explored research\ntopic, as most works have been treating inputs as stacks of static images\nrather than temporally linked series of data. Recently, it has been shown that\ninvolving the time dimension when designing augmentations can be superior to\nits spatial-only variants for video action recognition. In this paper, we\npropose several novel enhancements to these techniques to strengthen the\nrelationship between the spatial and temporal domains and achieve a deeper\nlevel of perturbations. The video action recognition results of our techniques\noutperform their respective variants in Top-1 and Top-5 settings on the UCF-101\nand the HMDB-51 datasets.\n", "date": "2022-11-09", "categories": []}
{"id": "2211.04905", "title": "SimOn: A Simple Framework for Online Temporal Action Localization", "abstract": "  Online Temporal Action Localization (On-TAL) aims to immediately provide\naction instances from untrimmed streaming videos. The model is not allowed to\nutilize future frames and any processing techniques to modify past predictions,\nmaking On-TAL much more challenging. In this paper, we propose a simple yet\neffective framework, termed SimOn, that learns to predict action instances\nusing the popular Transformer architecture in an end-to-end manner.\nSpecifically, the model takes the current frame feature as a query and a set of\npast context information as keys and values of the Transformer. Different from\nthe prior work that uses a set of outputs of the model as past contexts, we\nleverage the past visual context and the learnable context embedding for the\ncurrent query. Experimental results on the THUMOS14 and ActivityNet1.3 datasets\nshow that our model remarkably outperforms the previous methods, achieving a\nnew state-of-the-art On-TAL performance. In addition, the evaluation for Online\nDetection of Action Start (ODAS) demonstrates the effectiveness and robustness\nof our method in the online setting. The code is available at\nhttps://github.com/TuanTNG/SimOn\n", "date": "2022-11-08", "categories": []}
{"id": "2211.04255", "title": "Two-stream Multi-dimensional Convolutional Network for Real-time\n  Violence Detection", "abstract": "  The increasing number of surveillance cameras and security concerns have made\nautomatic violent activity detection from surveillance footage an active area\nfor research. Modern deep learning methods have achieved good accuracy in\nviolence detection and proved to be successful because of their applicability\nin intelligent surveillance systems. However, the models are computationally\nexpensive and large in size because of their inefficient methods for feature\nextraction. This work presents a novel architecture for violence detection\ncalled Two-stream Multi-dimensional Convolutional Network (2s-MDCN), which uses\nRGB frames and optical flow to detect violence. Our proposed method extracts\ntemporal and spatial information independently by 1D, 2D, and 3D convolutions.\nDespite combining multi-dimensional convolutional networks, our models are\nlightweight and efficient due to reduced channel capacity, yet they learn to\nextract meaningful spatial and temporal information. Additionally, combining\nRGB frames and optical flow yields 2.2% more accuracy than a single RGB stream.\nRegardless of having less complexity, our models obtained state-of-the-art\naccuracy of 89.7% on the largest violence detection benchmark dataset.\n", "date": "2022-11-08", "categories": []}
{"id": "2211.04331", "title": "Multi-Stage Based Feature Fusion of Multi-Modal Data for Human Activity\n  Recognition", "abstract": "  To properly assist humans in their needs, human activity recognition (HAR)\nsystems need the ability to fuse information from multiple modalities. Our\nhypothesis is that multimodal sensors, visual and non-visual tend to provide\ncomplementary information, addressing the limitations of other modalities. In\nthis work, we propose a multi-modal framework that learns to effectively\ncombine features from RGB Video and IMU sensors, and show its robustness for\nMMAct and UTD-MHAD datasets. Our model is trained in two-stage, where in the\nfirst stage, each input encoder learns to effectively extract features, and in\nthe second stage, learns to combine these individual features. We show\nsignificant improvements of 22% and 11% compared to video only and IMU only\nsetup on UTD-MHAD dataset, and 20% and 12% on MMAct datasets. Through extensive\nexperimentation, we show the robustness of our model on zero shot setting, and\nlimited annotated data setting. We further compare with state-of-the-art\nmethods that use more input modalities and show that our method outperforms\nsignificantly on the more difficult MMact dataset, and performs comparably in\nUTD-MHAD dataset.\n", "date": "2022-11-08", "categories": []}
{"id": "2211.02965", "title": "Can Ensemble of Classifiers Provide Better Recognition Results in\n  Packaging Activity?", "abstract": "  Skeleton-based Motion Capture (MoCap) systems have been widely used in the\ngame and film industry for mimicking complex human actions for a long time.\nMoCap data has also proved its effectiveness in human activity recognition\ntasks. However, it is a quite challenging task for smaller datasets. The lack\nof such data for industrial activities further adds to the difficulties. In\nthis work, we have proposed an ensemble-based machine learning methodology that\nis targeted to work better on MoCap datasets. The experiments have been\nperformed on the MoCap data given in the Bento Packaging Activity Recognition\nChallenge 2021. Bento is a Japanese word that resembles lunch-box. Upon\nprocessing the raw MoCap data at first, we have achieved an astonishing\naccuracy of 98% on 10-fold Cross-Validation and 82% on\nLeave-One-Out-Cross-Validation by using the proposed ensemble model.\n", "date": "2022-11-05", "categories": []}
{"id": "2211.01311", "title": "Distill and Collect for Semi-Supervised Temporal Action Segmentation", "abstract": "  Recent temporal action segmentation approaches need frame annotations during\ntraining to be effective. These annotations are very expensive and\ntime-consuming to obtain. This limits their performances when only limited\nannotated data is available. In contrast, we can easily collect a large corpus\nof in-domain unannotated videos by scavenging through the internet. Thus, this\npaper proposes an approach for the temporal action segmentation task that can\nsimultaneously leverage knowledge from annotated and unannotated video\nsequences. Our approach uses multi-stream distillation that repeatedly refines\nand finally combines their frame predictions. Our model also predicts the\naction order, which is later used as a temporal constraint while estimating\nframes labels to counter the lack of supervision for unannotated videos. In the\nend, our evaluation of the proposed approach on two different datasets\ndemonstrates its capability to achieve comparable performance to the full\nsupervision despite limited annotation.\n", "date": "2022-11-02", "categories": []}
{"id": "2211.01342", "title": "Fine-grained Human Activity Recognition Using Virtual On-body\n  Acceleration Data", "abstract": "  Previous work has demonstrated that virtual accelerometry data, extracted\nfrom videos using cross-modality transfer approaches like IMUTube, is\nbeneficial for training complex and effective human activity recognition (HAR)\nmodels. Systems like IMUTube were originally designed to cover activities that\nare based on substantial body (part) movements. Yet, life is complex, and a\nrange of activities of daily living is based on only rather subtle movements,\nwhich bears the question to what extent systems like IMUTube are of value also\nfor fine-grained HAR, i.e., When does IMUTube break? In this work we first\nintroduce a measure to quantitatively assess the subtlety of human movements\nthat are underlying activities of interest--the motion subtlety index\n(MSI)--which captures local pixel movements and pose changes in the vicinity of\ntarget virtual sensor locations, and correlate it to the eventual activity\nrecognition accuracy. We then perform a \"stress-test\" on IMUTube and explore\nfor which activities with underlying subtle movements a cross-modality transfer\napproach works, and for which not. As such, the work presented in this paper\nallows us to map out the landscape for IMUTube applications in practical\nscenarios.\n", "date": "2022-11-02", "categories": []}
{"id": "2210.16820", "title": "Temporal-Viewpoint Transportation Plan for Skeletal Few-shot Action\n  Recognition", "abstract": "  We propose a Few-shot Learning pipeline for 3D skeleton-based action\nrecognition by Joint tEmporal and cAmera viewpoiNt alIgnmEnt (JEANIE). To\nfactor out misalignment between query and support sequences of 3D body joints,\nwe propose an advanced variant of Dynamic Time Warping which jointly models\neach smooth path between the query and support frames to achieve simultaneously\nthe best alignment in the temporal and simulated camera viewpoint spaces for\nend-to-end learning under the limited few-shot training data. Sequences are\nencoded with a temporal block encoder based on Simple Spectral Graph\nConvolution, a lightweight linear Graph Neural Network backbone. We also\ninclude a setting with a transformer. Finally, we propose a similarity-based\nloss which encourages the alignment of sequences of the same class while\npreventing the alignment of unrelated sequences. We show state-of-the-art\nresults on NTU-60, NTU-120, Kinetics-skeleton and UWA3D Multiview Activity II.\n", "date": "2022-10-30", "categories": []}
{"id": "2210.17412", "title": "Adversarial Domain Adaptation for Action Recognition Around the Clock", "abstract": "  Due to the numerous potential applications in visual surveillance and\nnighttime driving, recognizing human action in low-light conditions remains a\ndifficult problem in computer vision. Existing methods separate action\nrecognition and dark enhancement into two distinct steps to accomplish this\ntask. However, isolating the recognition and enhancement impedes end-to-end\nlearning of the space-time representation for video action classification. This\npaper presents a domain adaptation-based action recognition approach that uses\nadversarial learning in cross-domain settings to learn cross-domain action\nrecognition. Supervised learning can train it on a large amount of labeled data\nfrom the source domain (daytime action sequences). However, it uses deep domain\ninvariant features to perform unsupervised learning on many unlabelled data\nfrom the target domain (night-time action sequences). The resulting augmented\nmodel, named 3D-DiNet can be trained using standard backpropagation with an\nadditional layer. It achieves SOTA performance on InFAR and XD145 actions\ndatasets.\n", "date": "2022-10-25", "categories": []}
{"id": "2206.03064", "title": "A Simple and Efficient Pipeline to Build an End-to-End Spatial-Temporal\n  Action Detector", "abstract": "  Spatial-temporal action detection is a vital part of video understanding.\nCurrent spatial-temporal action detection methods mostly use an object detector\nto obtain person candidates and classify these person candidates into different\naction categories. So-called two-stage methods are heavy and hard to apply in\nreal-world applications. Some existing methods build one-stage pipelines, But a\nlarge performance drop exists with the vanilla one-stage pipeline and extra\nclassification modules are needed to achieve comparable performance. In this\npaper, we explore a simple and effective pipeline to build a strong one-stage\nspatial-temporal action detector. The pipeline is composed by two parts: one is\na simple end-to-end spatial-temporal action detector. The proposed end-to-end\ndetector has minor architecture changes to current proposal-based detectors and\ndoes not add extra action classification modules. The other part is a novel\nlabeling strategy to utilize unlabeled frames in sparse annotated data. We\nnamed our model as SE-STAD. The proposed SE-STAD achieves around 2% mAP boost\nand around 80% FLOPs reduction. Our code will be released at\nhttps://github.com/4paradigm-CV/SE-STAD.\n", "date": "2022-06-07", "categories": []}
{"id": "2210.15187", "title": "Learning Joint Representation of Human Motion and Language", "abstract": "  In this work, we present MoLang (a Motion-Language connecting model) for\nlearning joint representation of human motion and language, leveraging both\nunpaired and paired datasets of motion and language modalities. To this end, we\npropose a motion-language model with contrastive learning, empowering our model\nto learn better generalizable representations of the human motion domain.\nEmpirical results show that our model learns strong representations of human\nmotion data through navigating language modality. Our proposed method is able\nto perform both action recognition and motion retrieval tasks with a single\nmodel where it outperforms state-of-the-art approaches on a number of action\nrecognition benchmarks.\n", "date": "2022-10-27", "categories": []}
{"id": "2209.07474", "title": "On the Surprising Effectiveness of Transformers in Low-Labeled Video\n  Recognition", "abstract": "  Recently vision transformers have been shown to be competitive with\nconvolution-based methods (CNNs) broadly across multiple vision tasks. The less\nrestrictive inductive bias of transformers endows greater representational\ncapacity in comparison with CNNs. However, in the image classification setting\nthis flexibility comes with a trade-off with respect to sample efficiency,\nwhere transformers require ImageNet-scale training. This notion has carried\nover to video where transformers have not yet been explored for video\nclassification in the low-labeled or semi-supervised settings. Our work\nempirically explores the low data regime for video classification and discovers\nthat, surprisingly, transformers perform extremely well in the low-labeled\nvideo setting compared to CNNs. We specifically evaluate video vision\ntransformers across two contrasting video datasets (Kinetics-400 and\nSomethingSomething-V2) and perform thorough analysis and ablation studies to\nexplain this observation using the predominant features of video transformer\narchitectures. We even show that using just the labeled data, transformers\nsignificantly outperform complex semi-supervised CNN methods that leverage\nlarge-scale unlabeled data as well. Our experiments inform our recommendation\nthat semi-supervised learning video work should consider the use of video\ntransformers in the future.\n", "date": "2022-09-15", "categories": []}
{"id": "2210.14284", "title": "Refining Action Boundaries for One-stage Detection", "abstract": "  Current one-stage action detection methods, which simultaneously predict\naction boundaries and the corresponding class, do not estimate or use a measure\nof confidence in their boundary predictions, which can lead to inaccurate\nboundaries. We incorporate the estimation of boundary confidence into one-stage\nanchor-free detection, through an additional prediction head that predicts the\nrefined boundaries with higher confidence. We obtain state-of-the-art\nperformance on the challenging EPIC-KITCHENS-100 action detection as well as\nthe standard THUMOS14 action detection benchmarks, and achieve improvement on\nthe ActivityNet-1.3 benchmark.\n", "date": "2022-10-25", "categories": []}
{"id": "2209.02250", "title": "Spatio-Temporal Action Detection Under Large Motion", "abstract": "  Current methods for spatiotemporal action tube detection often extend a\nbounding box proposal at a given keyframe into a 3D temporal cuboid and pool\nfeatures from nearby frames. However, such pooling fails to accumulate\nmeaningful spatiotemporal features if the position or shape of the actor shows\nlarge 2D motion and variability through the frames, due to large camera motion,\nlarge actor shape deformation, fast actor action and so on. In this work, we\naim to study the performance of cuboid-aware feature aggregation in action\ndetection under large action. Further, we propose to enhance actor feature\nrepresentation under large motion by tracking actors and performing temporal\nfeature aggregation along the respective tracks. We define the actor motion\nwith intersection-over-union (IoU) between the boxes of action tubes/tracks at\nvarious fixed time scales. The action having a large motion would result in\nlower IoU over time, and slower actions would maintain higher IoU. We find that\ntrack-aware feature aggregation consistently achieves a large improvement in\naction detection performance, especially for actions under large motion\ncompared to the cuboid-aware baseline. As a result, we also report\nstate-of-the-art on the large-scale MultiSports dataset. The Code is available\nat https://github.com/gurkirt/ActionTrackDetectron.\n", "date": "2022-09-06", "categories": []}
{"id": "2207.05182", "title": "Fine-grained Activities of People Worldwide", "abstract": "  Every day, humans perform many closely related activities that involve subtle\ndiscriminative motions, such as putting on a shirt vs. putting on a jacket, or\nshaking hands vs. giving a high five. Activity recognition by ethical visual AI\ncould provide insights into our patterns of daily life, however existing\nactivity recognition datasets do not capture the massive diversity of these\nhuman activities around the world. To address this limitation, we introduce\nCollector, a free mobile app to record video while simultaneously annotating\nobjects and activities of consented subjects. This new data collection platform\nwas used to curate the Consented Activities of People (CAP) dataset, the first\nlarge-scale, fine-grained activity dataset of people worldwide. The CAP dataset\ncontains 1.45M video clips of 512 fine grained activity labels of daily life,\ncollected by 780 subjects in 33 countries. We provide activity classification\nand activity detection benchmarks for this dataset, and analyze baseline\nresults to gain insight into how people around with world perform common\nactivities. The dataset, benchmarks, evaluation tools, public leaderboards and\nmobile apps are available for use at visym.github.io/cap.\n", "date": "2022-07-11", "categories": []}
{"id": "2210.11387", "title": "Transformer-based Action recognition in hand-object interacting\n  scenarios", "abstract": "  This report describes the 2nd place solution to the ECCV 2022 Human Body,\nHands, and Activities (HBHA) from Egocentric and Multi-view Cameras Challenge:\nAction Recognition. This challenge aims to recognize hand-object interaction in\nan egocentric view. We propose a framework that estimates keypoints of two\nhands and an object with a Transformer-based keypoint estimator and recognizes\nactions based on the estimated keypoints. We achieved a top-1 accuracy of\n87.19% on the testset.\n", "date": "2022-10-20", "categories": []}
{"id": "2210.10247", "title": "Performance of different machine learning methods on activity\n  recognition and pose estimation datasets", "abstract": "  With advancements in computer vision taking place day by day, recently a lot\nof light is being shed on activity recognition. With the range for real-world\napplications utilizing this field of study increasing across a multitude of\nindustries such as security and healthcare, it becomes crucial for businesses\nto distinguish which machine learning methods perform better than others in the\narea. This paper strives to aid in this predicament i.e. building upon previous\nrelated work, it employs both classical and ensemble approaches on rich pose\nestimation (OpenPose) and HAR datasets. Making use of appropriate metrics to\nevaluate the performance for each model, the results show that overall, random\nforest yields the highest accuracy in classifying ADLs. Relatively all the\nmodels have excellent performance across both datasets, except for logistic\nregression and AdaBoost perform poorly in the HAR one. With the limitations of\nthis paper also discussed in the end, the scope for further research is vast,\nwhich can use this paper as a base in aims of producing better results.\n", "date": "2022-10-19", "categories": []}
{"id": "2210.07400", "title": "Real-time Action Recognition for Fine-Grained Actions and The Hand Wash\n  Dataset", "abstract": "  In this paper we present a three-stream algorithm for real-time action\nrecognition and a new dataset of handwash videos, with the intent of aligning\naction recognition with real-world constraints to yield effective conclusions.\nA three-stream fusion algorithm is proposed, which runs both accurately and\nefficiently, in real-time even on low-powered systems such as a Raspberry Pi.\nThe cornerstone of the proposed algorithm is the incorporation of both spatial\nand temporal information, as well as the information of the objects in a video\nwhile using an efficient architecture, and Optical Flow computation to achieve\ncommendable results in real-time. The results achieved by this algorithm are\nbenchmarked on the UCF-101 as well as the HMDB-51 datasets, achieving an\naccuracy of 92.7% and 64.9% respectively. An important point to note is that\nthe algorithm is novel in the aspect that it is also able to learn the\nintricate differences between extremely similar actions, which would be\ndifficult even for the human eye. Additionally, noticing a dearth in the number\nof datasets for the recognition of very similar or fine-grained actions, this\npaper also introduces a new dataset that is made publicly available, the Hand\nWash Dataset with the intent of introducing a new benchmark for fine-grained\naction recognition tasks in the future.\n", "date": "2022-10-13", "categories": []}
{"id": "2210.07503", "title": "STAR-Transformer: A Spatio-temporal Cross Attention Transformer for\n  Human Action Recognition", "abstract": "  In action recognition, although the combination of spatio-temporal videos and\nskeleton features can improve the recognition performance, a separate model and\nbalancing feature representation for cross-modal data are required. To solve\nthese problems, we propose Spatio-TemporAl cRoss (STAR)-transformer, which can\neffectively represent two cross-modal features as a recognizable vector. First,\nfrom the input video and skeleton sequence, video frames are output as global\ngrid tokens and skeletons are output as joint map tokens, respectively. These\ntokens are then aggregated into multi-class tokens and input into\nSTAR-transformer. The STAR-transformer encoder layer consists of a full\nself-attention (FAttn) module and a proposed zigzag spatio-temporal attention\n(ZAttn) module. Similarly, the continuous decoder consists of a FAttn module\nand a proposed binary spatio-temporal attention (BAttn) module.\nSTAR-transformer learns an efficient multi-feature representation of the\nspatio-temporal features by properly arranging pairings of the FAttn, ZAttn,\nand BAttn modules. Experimental results on the Penn-Action, NTU RGB+D 60, and\n120 datasets show that the proposed method achieves a promising improvement in\nperformance in comparison to previous state-of-the-art methods.\n", "date": "2022-10-14", "categories": []}
{"id": "2210.06501", "title": "Robust Action Segmentation from Timestamp Supervision", "abstract": "  Action segmentation is the task of predicting an action label for each frame\nof an untrimmed video. As obtaining annotations to train an approach for action\nsegmentation in a fully supervised way is expensive, various approaches have\nbeen proposed to train action segmentation models using different forms of weak\nsupervision, e.g., action transcripts, action sets, or more recently\ntimestamps. Timestamp supervision is a promising type of weak supervision as\nobtaining one timestamp per action is less expensive than annotating all\nframes, but it provides more information than other forms of weak supervision.\nHowever, previous works assume that every action instance is annotated with a\ntimestamp, which is a restrictive assumption since it assumes that annotators\ndo not miss any action. In this work, we relax this restrictive assumption and\ntake missing annotations for some action instances into account. We show that\nour approach is more robust to missing annotations compared to other approaches\nand various baselines.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.06682", "title": "Application-Driven AI Paradigm for Hand-Held Action Detection", "abstract": "  In practical applications especially with safety requirement, some hand-held\nactions need to be monitored closely, including smoking cigarettes, dialing,\neating, etc. Taking smoking cigarettes as example, existing smoke detection\nalgorithms usually detect the cigarette or cigarette with hand as the target\nobject only, which leads to low accuracy. In this paper, we propose an\napplication-driven AI paradigm for hand-held action detection based on\nhierarchical object detection. It is a coarse-to-fine hierarchical detection\nframework composed of two modules. The first one is a coarse detection module\nwith the human pose consisting of the whole hand, cigarette and head as target\nobject. The followed second one is a fine detection module with the fingers\nholding cigarette, mouth area and the whole cigarette as target. Some\nexperiments are done with the dataset collected from real-world scenarios, and\nthe results show that the proposed framework achieve higher detection rate with\ngood adaptation and robustness in complex environments.\n", "date": "2022-10-13", "categories": []}
{"id": "2208.04950", "title": "BabyNet: A Lightweight Network for Infant Reaching Action Recognition in\n  Unconstrained Environments to Support Future Pediatric Rehabilitation\n  Applications", "abstract": "  Action recognition is an important component to improve autonomy of physical\nrehabilitation devices, such as wearable robotic exoskeletons. Existing human\naction recognition algorithms focus on adult applications rather than pediatric\nones. In this paper, we introduce BabyNet, a light-weight (in terms of\ntrainable parameters) network structure to recognize infant reaching action\nfrom off-body stationary cameras. We develop an annotated dataset that includes\ndiverse reaches performed while in a sitting posture by different infants in\nunconstrained environments (e.g., in home settings, etc.). Our approach uses\nthe spatial and temporal connection of annotated bounding boxes to interpret\nonset and offset of reaching, and to detect a complete reaching action. We\nevaluate the efficiency of our proposed approach and compare its performance\nagainst other learning-based network structures in terms of capability of\ncapturing temporal inter-dependencies and accuracy of detection of reaching\nonset and offset. Results indicate our BabyNet can attain solid performance in\nterms of (average) testing accuracy that exceeds that of other larger networks,\nand can hence serve as a light-weight data-driven framework for video-based\ninfant reaching action recognition.\n", "date": "2022-08-09", "categories": []}
{"id": "2210.05895", "title": "DG-STGCN: Dynamic Spatial-Temporal Modeling for Skeleton-based Action\n  Recognition", "abstract": "  Graph convolution networks (GCN) have been widely used in skeleton-based\naction recognition. We note that existing GCN-based approaches primarily rely\non prescribed graphical structures (ie., a manually defined topology of\nskeleton joints), which limits their flexibility to capture complicated\ncorrelations between joints. To move beyond this limitation, we propose a new\nframework for skeleton-based action recognition, namely Dynamic Group\nSpatio-Temporal GCN (DG-STGCN). It consists of two modules, DG-GCN and DG-TCN,\nrespectively, for spatial and temporal modeling. In particular, DG-GCN uses\nlearned affinity matrices to capture dynamic graphical structures instead of\nrelying on a prescribed one, while DG-TCN performs group-wise temporal\nconvolutions with varying receptive fields and incorporates a dynamic\njoint-skeleton fusion module for adaptive multi-level temporal modeling. On a\nwide range of benchmarks, including NTURGB+D, Kinetics-Skeleton, BABEL, and\nToyota SmartHome, DG-STGCN consistently outperforms state-of-the-art methods,\noften by a notable margin.\n", "date": "2022-10-12", "categories": []}
{"id": "2210.06192", "title": "Pose-Guided Graph Convolutional Networks for Skeleton-Based Action\n  Recognition", "abstract": "  Graph convolutional networks (GCNs), which can model the human body skeletons\nas spatial and temporal graphs, have shown remarkable potential in\nskeleton-based action recognition. However, in the existing GCN-based methods,\ngraph-structured representation of the human skeleton makes it difficult to be\nfused with other modalities, especially in the early stages. This may limit\ntheir scalability and performance in action recognition tasks. In addition, the\npose information, which naturally contains informative and discriminative clues\nfor action recognition, is rarely explored together with skeleton data in\nexisting methods. In this work, we propose pose-guided GCN (PG-GCN), a\nmulti-modal framework for high-performance human action recognition. In\nparticular, a multi-stream network is constructed to simultaneously explore the\nrobust features from both the pose and skeleton data, while a dynamic attention\nmodule is designed for early-stage feature fusion. The core idea of this module\nis to utilize a trainable graph to aggregate features from the skeleton stream\nwith that of the pose stream, which leads to a network with more robust feature\nrepresentation ability. Extensive experiments show that the proposed PG-GCN can\nachieve state-of-the-art performance on the NTU RGB+D 60 and NTU RGB+D 120\ndatasets.\n", "date": "2022-10-10", "categories": []}
{"id": "2209.00638", "title": "Unified Fully and Timestamp Supervised Temporal Action Segmentation via\n  Sequence to Sequence Translation", "abstract": "  This paper introduces a unified framework for video action segmentation via\nsequence to sequence (seq2seq) translation in a fully and timestamp supervised\nsetup. In contrast to current state-of-the-art frame-level prediction methods,\nwe view action segmentation as a seq2seq translation task, i.e., mapping a\nsequence of video frames to a sequence of action segments. Our proposed method\ninvolves a series of modifications and auxiliary loss functions on the standard\nTransformer seq2seq translation model to cope with long input sequences opposed\nto short output sequences and relatively few videos. We incorporate an\nauxiliary supervision signal for the encoder via a frame-wise loss and propose\na separate alignment decoder for an implicit duration prediction. Finally, we\nextend our framework to the timestamp supervised setting via our proposed\nconstrained k-medoids algorithm to generate pseudo-segmentations. Our proposed\nframework performs consistently on both fully and timestamp supervised\nsettings, outperforming or competing state-of-the-art on several datasets. Our\ncode is publicly available at https://github.com/boschresearch/UVAST.\n", "date": "2022-09-01", "categories": []}
{"id": "2210.04933", "title": "An Action Is Worth Multiple Words: Handling Ambiguity in Action\n  Recognition", "abstract": "  Precisely naming the action depicted in a video can be a challenging and\noftentimes ambiguous task. In contrast to object instances represented as nouns\n(e.g. dog, cat, chair, etc.), in the case of actions, human annotators\ntypically lack a consensus as to what constitutes a specific action (e.g.\njogging versus running). In practice, a given video can contain multiple valid\npositive annotations for the same action. As a result, video datasets often\ncontain significant levels of label noise and overlap between the atomic action\nclasses. In this work, we address the challenge of training multi-label action\nrecognition models from only single positive training labels. We propose two\napproaches that are based on generating pseudo training examples sampled from\nsimilar instances within the train set. Unlike other approaches that use\nmodel-derived pseudo-labels, our pseudo-labels come from human annotations and\nare selected based on feature similarity. To validate our approaches, we create\na new evaluation benchmark by manually annotating a subset of\nEPIC-Kitchens-100's validation set with multiple verb labels. We present\nresults on this new test set along with additional results on a new version of\nHMDB-51, called Confusing-HMDB-102, where we outperform existing methods in\nboth cases. Data and code are available at\nhttps://github.com/kiyoon/verb_ambiguity\n", "date": "2022-10-10", "categories": []}
{"id": "2207.10075", "title": "Is an Object-Centric Video Representation Beneficial for Transfer?", "abstract": "  The objective of this work is to learn an object-centric video\nrepresentation, with the aim of improving transferability to novel tasks, i.e.,\ntasks different from the pre-training task of action classification. To this\nend, we introduce a new object-centric video recognition model based on a\ntransformer architecture. The model learns a set of object-centric summary\nvectors for the video, and uses these vectors to fuse the visual and\nspatio-temporal trajectory 'modalities' of the video clip. We also introduce a\nnovel trajectory contrast loss to further enhance objectness in these summary\nvectors. With experiments on four datasets -- SomethingSomething-V2,\nSomethingElse, Action Genome and EpicKitchens -- we show that the\nobject-centric model outperforms prior video representations (both\nobject-agnostic and object-aware), when: (1) classifying actions on unseen\nobjects and unseen environments; (2) low-shot learning of novel classes; (3)\nlinear probe to other downstream tasks; as well as (4) for standard action\nclassification.\n", "date": "2022-07-20", "categories": []}
{"id": "2210.04331", "title": "Students taught by multimodal teachers are superior action recognizers", "abstract": "  The focal point of egocentric video understanding is modelling hand-object\ninteractions. Standard models -- CNNs, Vision Transformers, etc. -- which\nreceive RGB frames as input perform well, however, their performance improves\nfurther by employing additional modalities such as object detections, optical\nflow, audio, etc. as input. The added complexity of the required\nmodality-specific modules, on the other hand, makes these models impractical\nfor deployment. The goal of this work is to retain the performance of such\nmultimodal approaches, while using only the RGB images as input at inference\ntime. Our approach is based on multimodal knowledge distillation, featuring a\nmultimodal teacher (in the current experiments trained only using object\ndetections, optical flow and RGB frames) and a unimodal student (using only RGB\nframes as input). We present preliminary results which demonstrate that the\nresulting model -- distilled from a multimodal teacher -- significantly\noutperforms the baseline RGB model (trained without knowledge distillation), as\nwell as an omnivorous version of itself (trained on all modalities jointly), in\nboth standard and compositional action recognition.\n", "date": "2022-10-09", "categories": []}
{"id": "2210.03382", "title": "Temporal Feature Alignment in Contrastive Self-Supervised Learning for\n  Human Activity Recognition", "abstract": "  Automated Human Activity Recognition has long been a problem of great\ninterest in human-centered and ubiquitous computing. In the last years, a\nplethora of supervised learning algorithms based on deep neural networks has\nbeen suggested to address this problem using various modalities. While every\nmodality has its own limitations, there is one common challenge. Namely,\nsupervised learning requires vast amounts of annotated data which is\npractically hard to collect. In this paper, we benefit from the self-supervised\nlearning paradigm (SSL) that is typically used to learn deep feature\nrepresentations from unlabeled data. Moreover, we upgrade a contrastive SSL\nframework, namely SimCLR, widely used in various applications by introducing a\ntemporal feature alignment procedure for Human Activity Recognition.\nSpecifically, we propose integrating a dynamic time warping (DTW) algorithm in\na latent space to force features to be aligned in a temporal dimension.\nExtensive experiments have been conducted for the unimodal scenario with\ninertial modality as well as in multimodal settings using inertial and skeleton\ndata. According to the obtained results, the proposed approach has a great\npotential in learning robust feature representations compared to the recent SSL\nbaselines, and clearly outperforms supervised models in semi-supervised\nlearning. The code for the unimodal case is available via the following link:\nhttps://github.com/bulatkh/csshar_tfa.\n", "date": "2022-10-07", "categories": []}
{"id": "2210.02693", "title": "Focal and Global Spatial-Temporal Transformer for Skeleton-based Action\n  Recognition", "abstract": "  Despite great progress achieved by transformer in various vision tasks, it is\nstill underexplored for skeleton-based action recognition with only a few\nattempts. Besides, these methods directly calculate the pair-wise global\nself-attention equally for all the joints in both the spatial and temporal\ndimensions, undervaluing the effect of discriminative local joints and the\nshort-range temporal dynamics. In this work, we propose a novel Focal and\nGlobal Spatial-Temporal Transformer network (FG-STFormer), that is equipped\nwith two key components: (1) FG-SFormer: focal joints and global parts coupling\nspatial transformer. It forces the network to focus on modelling correlations\nfor both the learned discriminative spatial joints and human body parts\nrespectively. The selective focal joints eliminate the negative effect of\nnon-informative ones during accumulating the correlations. Meanwhile, the\ninteractions between the focal joints and body parts are incorporated to\nenhance the spatial dependencies via mutual cross-attention. (2) FG-TFormer:\nfocal and global temporal transformer. Dilated temporal convolution is\nintegrated into the global self-attention mechanism to explicitly capture the\nlocal temporal motion patterns of joints or body parts, which is found to be\nvital important to make temporal transformer work. Extensive experimental\nresults on three benchmarks, namely NTU-60, NTU-120 and NW-UCLA, show our\nFG-STFormer surpasses all existing transformer-based methods, and compares\nfavourably with state-of-the art GCN-based methods.\n", "date": "2022-10-06", "categories": []}
{"id": "2210.00507", "title": "Fast and Robust Video-Based Exercise Classification via Body Pose\n  Tracking and Scalable Multivariate Time Series Classifiers", "abstract": "  Technological advancements have spurred the usage of machine learning based\napplications in sports science. Physiotherapists, sports coaches and athletes\nactively look to incorporate the latest technologies in order to further\nimprove performance and avoid injuries. While wearable sensors are very\npopular, their use is hindered by constraints on battery power and sensor\ncalibration, especially for use cases which require multiple sensors to be\nplaced on the body. Hence, there is renewed interest in video-based data\ncapture and analysis for sports science. In this paper, we present the\napplication of classifying S\\&C exercises using video. We focus on the popular\nMilitary Press exercise, where the execution is captured with a video-camera\nusing a mobile device, such as a mobile phone, and the goal is to classify the\nexecution into different types. Since video recordings need a lot of storage\nand computation, this use case requires data reduction, while preserving the\nclassification accuracy and enabling fast prediction. To this end, we propose\nan approach named BodyMTS to turn video into time series by employing body pose\ntracking, followed by training and prediction using multivariate time series\nclassifiers. We analyze the accuracy and robustness of BodyMTS and show that it\nis robust to different types of noise caused by either video quality or pose\nestimation factors. We compare BodyMTS to state-of-the-art deep learning\nmethods which classify human activity directly from videos and show that\nBodyMTS achieves similar accuracy, but with reduced running time and model\nengineering effort. Finally, we discuss some of the practical aspects of\nemploying BodyMTS in this application in terms of accuracy and robustness under\nreduced data quality and size. We show that BodyMTS achieves an average\naccuracy of 87\\%, which is significantly higher than the accuracy of human\ndomain experts.\n", "date": "2022-10-02", "categories": []}
{"id": "2209.15271", "title": "Application-Driven AI Paradigm for Human Action Recognition", "abstract": "  Human action recognition in computer vision has been widely studied in recent\nyears. However, most algorithms consider only certain action specially with\neven high computational cost. That is not suitable for practical applications\nwith multiple actions to be identified with low computational cost. To meet\nvarious application scenarios, this paper presents a unified human action\nrecognition framework composed of two modules, i.e., multi-form human detection\nand corresponding action classification. Among them, an open-source dataset is\nconstructed to train a multi-form human detection model that distinguishes a\nhuman being's whole body, upper body or part body, and the followed action\nclassification model is adopted to recognize such action as falling, sleeping\nor on-duty, etc. Some experimental results show that the unified framework is\neffective for various application scenarios. It is expected to be a new\napplication-driven AI paradigm for human action recognition.\n", "date": "2022-09-30", "categories": []}
{"id": "2209.14757", "title": "Speeding Up Action Recognition Using Dynamic Accumulation of Residuals\n  in Compressed Domain", "abstract": "  With the widespread use of installed cameras, video-based monitoring\napproaches have seized considerable attention for different purposes like\nassisted living. Temporal redundancy and the sheer size of raw videos are the\ntwo most common problematic issues related to video processing algorithms. Most\nof the existing methods mainly focused on increasing accuracy by exploring\nconsecutive frames, which is laborious and cannot be considered for real-time\napplications. Since videos are mostly stored and transmitted in compressed\nformat, these kinds of videos are available on many devices. Compressed videos\ncontain a multitude of beneficial information, such as motion vectors and\nquantized coefficients. Proper use of this available information can greatly\nimprove the video understanding methods' performance. This paper presents an\napproach for using residual data, available in compressed videos directly,\nwhich can be obtained by a light partially decoding procedure. In addition, a\nmethod for accumulating similar residuals is proposed, which dramatically\nreduces the number of processed frames for action recognition. Applying neural\nnetworks exclusively for accumulated residuals in the compressed domain\naccelerates performance, while the classification results are highly\ncompetitive with raw video approaches.\n", "date": "2022-09-29", "categories": []}
{"id": "2209.15000", "title": "REST: REtrieve & Self-Train for generative action recognition", "abstract": "  This work is on training a generative action/video recognition model whose\noutput is a free-form action-specific caption describing the video (rather than\nan action class label). A generative approach has practical advantages like\nproducing more fine-grained and human-readable output, and being naturally\nopen-world. To this end, we propose to adapt a pre-trained generative Vision &\nLanguage (V&L) Foundation Model for video/action recognition. While recently\nthere have been a few attempts to adapt V&L models trained with contrastive\nlearning (e.g. CLIP) for video/action, to the best of our knowledge, we propose\nthe very first method that sets outs to accomplish this goal for a generative\nmodel. We firstly show that direct fine-tuning of a generative model to produce\naction classes suffers from severe overfitting. To alleviate this, we introduce\nREST, a training framework consisting of two key components: an unsupervised\nmethod for adapting the generative model to action/video by means of\npseudo-caption generation and Self-training, i.e. without using any\naction-specific labels; (b) a Retrieval approach based on CLIP for discovering\na diverse set of pseudo-captions for each video to train the model.\nImportantly, we show that both components are necessary to obtain high\naccuracy. We evaluate REST on the problem of zero-shot action recognition where\nwe show that our approach is very competitive when compared to contrastive\nlearning-based methods. Code will be made available.\n", "date": "2022-09-29", "categories": []}
{"id": "2209.13465", "title": "AdaFocusV3: On Unified Spatial-temporal Dynamic Video Recognition", "abstract": "  Recent research has revealed that reducing the temporal and spatial\nredundancy are both effective approaches towards efficient video recognition,\ne.g., allocating the majority of computation to a task-relevant subset of\nframes or the most valuable image regions of each frame. However, in most\nexisting works, either type of redundancy is typically modeled with another\nabsent. This paper explores the unified formulation of spatial-temporal dynamic\ncomputation on top of the recently proposed AdaFocusV2 algorithm, contributing\nto an improved AdaFocusV3 framework. Our method reduces the computational cost\nby activating the expensive high-capacity network only on some small but\ninformative 3D video cubes. These cubes are cropped from the space formed by\nframe height, width, and video duration, while their locations are adaptively\ndetermined with a light-weighted policy network on a per-sample basis. At test\ntime, the number of the cubes corresponding to each video is dynamically\nconfigured, i.e., video cubes are processed sequentially until a sufficiently\nreliable prediction is produced. Notably, AdaFocusV3 can be effectively trained\nby approximating the non-differentiable cropping operation with the\ninterpolation of deep features. Extensive empirical results on six benchmark\ndatasets (i.e., ActivityNet, FCVID, Mini-Kinetics, Something-Something V1&V2\nand Diving48) demonstrate that our model is considerably more efficient than\ncompetitive baselines.\n", "date": "2022-09-27", "categories": []}
{"id": "2209.11870", "title": "Leveraging Self-Supervised Training for Unintentional Action Recognition", "abstract": "  Unintentional actions are rare occurrences that are difficult to define\nprecisely and that are highly dependent on the temporal context of the action.\nIn this work, we explore such actions and seek to identify the points in videos\nwhere the actions transition from intentional to unintentional. We propose a\nmulti-stage framework that exploits inherent biases such as motion speed,\nmotion direction, and order to recognize unintentional actions. To enhance\nrepresentations via self-supervised training for the task of unintentional\naction recognition we propose temporal transformations, called Temporal\nTransformations of Inherent Biases of Unintentional Actions (T2IBUA). The\nmulti-stage approach models the temporal information on both the level of\nindividual frames and full clips. These enhanced representations show strong\nperformance for unintentional action recognition tasks. We provide an extensive\nablation study of our framework and report results that significantly improve\nover the state-of-the-art.\n", "date": "2022-09-23", "categories": []}
{"id": "2209.12061", "title": "Global Semantic Descriptors for Zero-Shot Action Recognition", "abstract": "  The success of Zero-shot Action Recognition (ZSAR) methods is intrinsically\nrelated to the nature of semantic side information used to transfer knowledge,\nalthough this aspect has not been primarily investigated in the literature.\nThis work introduces a new ZSAR method based on the relationships of\nactions-objects and actions-descriptive sentences. We demonstrate that\nrepresenting all object classes using descriptive sentences generates an\naccurate object-action affinity estimation when a paraphrase estimation method\nis used as an embedder. We also show how to estimate probabilities over the set\nof action classes based only on a set of sentences without hard human labeling.\nIn our method, the probabilities from these two global classifiers (i.e., which\nuse features computed over the entire video) are combined, producing an\nefficient transfer knowledge model for action classification. Our results are\nstate-of-the-art in the Kinetics-400 dataset and are competitive on UCF-101\nunder the ZSAR evaluation. Our code is available at\nhttps://github.com/valterlej/objsentzsar\n", "date": "2022-09-24", "categories": []}
{"id": "2209.12074", "title": "Self-supervised Learning for Unintentional Action Prediction", "abstract": "  Distinguishing if an action is performed as intended or if an intended action\nfails is an important skill that not only humans have, but that is also\nimportant for intelligent systems that operate in human environments.\nRecognizing if an action is unintentional or anticipating if an action will\nfail, however, is not straightforward due to lack of annotated data. While\nvideos of unintentional or failed actions can be found in the Internet in\nabundance, high annotation costs are a major bottleneck for learning networks\nfor these tasks. In this work, we thus study the problem of self-supervised\nrepresentation learning for unintentional action prediction. While previous\nworks learn the representation based on a local temporal neighborhood, we show\nthat the global context of a video is needed to learn a good representation for\nthe three downstream tasks: unintentional action classification, localization\nand anticipation. In the supplementary material, we show that the learned\nrepresentation can be used for detecting anomalies in videos as well.\n", "date": "2022-09-24", "categories": []}
{"id": "2209.11634", "title": "View-Invariant Skeleton-based Action Recognition via Global-Local\n  Contrastive Learning", "abstract": "  Skeleton-based human action recognition has been drawing more interest\nrecently due to its low sensitivity to appearance changes and the accessibility\nof more skeleton data. However, even the 3D skeletons captured in practice are\nstill sensitive to the viewpoint and direction gave the occlusion of different\nhuman-body joints and the errors in human joint localization. Such view\nvariance of skeleton data may significantly affect the performance of action\nrecognition. To address this issue, we propose in this paper a new\nview-invariant representation learning approach, without any manual action\nlabeling, for skeleton-based human action recognition. Specifically, we\nleverage the multi-view skeleton data simultaneously taken for the same person\nin the network training, by maximizing the mutual information between the\nrepresentations extracted from different views, and then propose a global-local\ncontrastive loss to model the multi-scale co-occurrence relationships in both\nspatial and temporal domains. Extensive experimental results show that the\nproposed method is robust to the view difference of the input skeleton data and\nsignificantly boosts the performance of unsupervised skeleton-based human\naction methods, resulting in new state-of-the-art accuracies on two challenging\nmulti-view benchmarks of PKUMMD and NTU RGB+D.\n", "date": "2022-09-23", "categories": []}
{"id": "2209.11680", "title": "An Overview of Violence Detection Techniques: Current Challenges and\n  Future Directions", "abstract": "  The Big Video Data generated in today's smart cities has raised concerns from\nits purposeful usage perspective, where surveillance cameras, among many others\nare the most prominent resources to contribute to the huge volumes of data,\nmaking its automated analysis a difficult task in terms of computation and\npreciseness. Violence Detection (VD), broadly plunging under Action and\nActivity recognition domain, is used to analyze Big Video data for anomalous\nactions incurred due to humans. The VD literature is traditionally based on\nmanually engineered features, though advancements to deep learning based\nstandalone models are developed for real-time VD analysis. This paper focuses\non overview of deep sequence learning approaches along with localization\nstrategies of the detected violence. This overview also dives into the initial\nimage processing and machine learning-based VD literature and their possible\nadvantages such as efficiency against the current complex models.\nFurthermore,the datasets are discussed, to provide an analysis of the current\nmodels, explaining their pros and cons with future directions in VD domain\nderived from an in-depth analysis of the previous methods.\n", "date": "2022-09-21", "categories": []}
{"id": "2209.10073", "title": "Adaptive Local-Component-aware Graph Convolutional Network for One-shot\n  Skeleton-based Action Recognition", "abstract": "  Skeleton-based action recognition receives increasing attention because the\nskeleton representations reduce the amount of training data by eliminating\nvisual information irrelevant to actions. To further improve the sample\nefficiency, meta-learning-based one-shot learning solutions were developed for\nskeleton-based action recognition. These methods find the nearest neighbor\naccording to the similarity between instance-level global average embedding.\nHowever, such measurement holds unstable representativity due to inadequate\ngeneralized learning on local invariant and noisy features, while intuitively,\nmore fine-grained recognition usually relies on determining key local body\nmovements. To address this limitation, we present the Adaptive\nLocal-Component-aware Graph Convolutional Network, which replaces the\ncomparison metric with a focused sum of similarity measurements on aligned\nlocal embedding of action-critical spatial/temporal segments. Comprehensive\none-shot experiments on the public benchmark of NTU-RGB+D 120 indicate that our\nmethod provides a stronger representation than the global embedding and helps\nour model reach state-of-the-art.\n", "date": "2022-09-21", "categories": []}
{"id": "2209.10126", "title": "Exploring Modulated Detection Transformer as a Tool for Action\n  Recognition in Videos", "abstract": "  During recent years transformers architectures have been growing in\npopularity. Modulated Detection Transformer (MDETR) is an end-to-end\nmulti-modal understanding model that performs tasks such as phase grounding,\nreferring expression comprehension, referring expression segmentation, and\nvisual question answering. One remarkable aspect of the model is the capacity\nto infer over classes that it was not previously trained for. In this work we\nexplore the use of MDETR in a new task, action detection, without any previous\ntraining. We obtain quantitative results using the Atomic Visual Actions\ndataset. Although the model does not report the best performance in the task,\nwe believe that it is an interesting finding. We show that it is possible to\nuse a multi-modal model to tackle a task that it was not designed for. Finally,\nwe believe that this line of research may lead into the generalization of MDETR\nin additional downstream tasks.\n", "date": "2022-09-21", "categories": []}
{"id": "2209.09393", "title": "Mitigating Representation Bias in Action Recognition: Algorithms and\n  Benchmarks", "abstract": "  Deep learning models have achieved excellent recognition results on\nlarge-scale video benchmarks. However, they perform poorly when applied to\nvideos with rare scenes or objects, primarily due to the bias of existing video\ndatasets. We tackle this problem from two different angles: algorithm and\ndataset. From the perspective of algorithms, we propose Spatial-aware\nMulti-Aspect Debiasing (SMAD), which incorporates both explicit debiasing with\nmulti-aspect adversarial training and implicit debiasing with the spatial\nactionness reweighting module, to learn a more generic representation invariant\nto non-action aspects. To neutralize the intrinsic dataset bias, we propose\nOmniDebias to leverage web data for joint training selectively, which can\nachieve higher performance with far fewer web data. To verify the\neffectiveness, we establish evaluation protocols and perform extensive\nexperiments on both re-distributed splits of existing datasets and a new\nevaluation dataset focusing on the action with rare scenes. We also show that\nthe debiased representation can generalize better when transferred to other\ndatasets and tasks.\n", "date": "2022-09-20", "categories": []}
{"id": "2209.09236", "title": "Real-time Online Video Detection with Temporal Smoothing Transformers", "abstract": "  Streaming video recognition reasons about objects and their actions in every\nframe of a video. A good streaming recognition model captures both long-term\ndynamics and short-term changes of video. Unfortunately, in most existing\nmethods, the computational complexity grows linearly or quadratically with the\nlength of the considered dynamics. This issue is particularly pronounced in\ntransformer-based architectures. To address this issue, we reformulate the\ncross-attention in a video transformer through the lens of kernel and apply two\nkinds of temporal smoothing kernel: A box kernel or a Laplace kernel. The\nresulting streaming attention reuses much of the computation from frame to\nframe, and only requires a constant time update each frame. Based on this idea,\nwe build TeSTra, a Temporal Smoothing Transformer, that takes in arbitrarily\nlong inputs with constant caching and computing overhead. Specifically, it runs\n$6\\times$ faster than equivalent sliding-window based transformers with 2,048\nframes in a streaming setting. Furthermore, thanks to the increased temporal\nspan, TeSTra achieves state-of-the-art results on THUMOS'14 and\nEPIC-Kitchen-100, two standard online action detection and action anticipation\ndatasets. A real-time version of TeSTra outperforms all but one prior\napproaches on the THUMOS'14 dataset.\n", "date": "2022-09-19", "categories": []}
{"id": "2208.11908", "title": "Adaptive Perception Transformer for Temporal Action Localization", "abstract": "  Temporal action localization aims to predict the boundary and category of\neach action instance in untrimmed long videos. Most of previous methods based\non anchors or proposals neglect the global-local context interaction in entire\nvideo sequences. Besides, their multi-stage designs cannot generate action\nboundaries and categories straightforwardly. To address the above issues, this\npaper proposes a end-to-end model, called Adaptive Perception transformer\n(AdaPerFormer for short). Specifically, AdaPerFormer explores a dual-branch\nattention mechanism. One branch takes care of the global perception attention,\nwhich can model entire video sequences and aggregate global relevant contexts.\nWhile the other branch concentrates on the local convolutional shift to\naggregate intra-frame and inter-frame information through our bidirectional\nshift operation. The end-to-end nature produces the boundaries and categories\nof video actions without extra steps. Extensive experiments together with\nablation studies are provided to reveal the effectiveness of our design. Our\nmethod obtains competitive performance on the THUMOS14 and ActivityNet-1.3\ndataset.\n", "date": "2022-08-25", "categories": []}
{"id": "2208.08599", "title": "Spatial Temporal Graph Attention Network for Skeleton-Based Action\n  Recognition", "abstract": "  It's common for current methods in skeleton-based action recognition to\nmainly consider capturing long-term temporal dependencies as skeleton sequences\nare typically long (>128 frames), which forms a challenging problem for\nprevious approaches. In such conditions, short-term dependencies are few\nformally considered, which are critical for classifying similar actions. Most\ncurrent approaches are consisted of interleaving spatial-only modules and\ntemporal-only modules, where direct information flow among joints in adjacent\nframes are hindered, thus inferior to capture short-term motion and distinguish\nsimilar action pairs. To handle this limitation, we propose a general\nframework, coined as STGAT, to model cross-spacetime information flow. It\nequips the spatial-only modules with spatial-temporal modeling for regional\nperception. While STGAT is theoretically effective for spatial-temporal\nmodeling, we propose three simple modules to reduce local spatial-temporal\nfeature redundancy and further release the potential of STGAT, which (1) narrow\nthe scope of self-attention mechanism, (2) dynamically weight joints along\ntemporal dimension, and (3) separate subtle motion from static features,\nrespectively. As a robust feature extractor, STGAT generalizes better upon\nclassifying similar actions than previous methods, witnessed by both\nqualitative and quantitative results. STGAT achieves state-of-the-art\nperformance on three large-scale datasets: NTU RGB+D 60, NTU RGB+D 120, and\nKinetics Skeleton 400. Code is released.\n", "date": "2022-08-18", "categories": []}
{"id": "2209.05700", "title": "Vision Transformers for Action Recognition: A Survey", "abstract": "  Vision transformers are emerging as a powerful tool to solve computer vision\nproblems. Recent techniques have also proven the efficacy of transformers\nbeyond the image domain to solve numerous video-related tasks. Among those,\nhuman action recognition is receiving special attention from the research\ncommunity due to its widespread applications. This article provides the first\ncomprehensive survey of vision transformer techniques for action recognition.\nWe analyze and summarize the existing and emerging literature in this direction\nwhile highlighting the popular trends in adapting transformers for action\nrecognition. Due to their specialized application, we collectively refer to\nthese methods as ``action transformers''. Our literature review provides\nsuitable taxonomies for action transformers based on their architecture,\nmodality, and intended objective. Within the context of action transformers, we\nexplore the techniques to encode spatio-temporal data, dimensionality\nreduction, frame patch and spatio-temporal cube construction, and various\nrepresentation methods. We also investigate the optimization of spatio-temporal\nattention in transformer layers to handle longer sequences, typically by\nreducing the number of tokens in a single attention operation. Moreover, we\nalso investigate different network learning strategies, such as self-supervised\nand zero-shot learning, along with their associated losses for\ntransformer-based action recognition. This survey also summarizes the progress\ntowards gaining grounds on evaluation metric scores on important benchmarks\nwith action transformers. Finally, it provides a discussion on the challenges,\noutlook, and future avenues for this research direction.\n", "date": "2022-09-13", "categories": []}
{"id": "2209.04525", "title": "PoliTO-IIT-CINI Submission to the EPIC-KITCHENS-100 Unsupervised Domain\n  Adaptation Challenge for Action Recognition", "abstract": "  In this report, we describe the technical details of our submission to the\nEPIC-Kitchens-100 Unsupervised Domain Adaptation (UDA) Challenge in Action\nRecognition. To tackle the domain-shift which exists under the UDA setting, we\nfirst exploited a recent Domain Generalization (DG) technique, called Relative\nNorm Alignment (RNA). Secondly, we extended this approach to work on unlabelled\ntarget data, enabling a simpler adaptation of the model to the target\ndistribution in an unsupervised fashion. To this purpose, we included in our\nframework UDA algorithms, such as multi-level adversarial alignment and\nattentive entropy. By analyzing the challenge setting, we notice the presence\nof a secondary concurrence shift in the data, which is usually called\nenvironmental bias. It is caused by the existence of different environments,\ni.e., kitchens. To deal with these two shifts (environmental and temporal), we\nextended our system to perform Multi-Source Multi-Target Domain Adaptation.\nFinally, we employed distinct models in our final proposal to leverage the\npotential of popular video architectures, and we introduced two more losses for\nthe ensemble adaptation. Our submission (entry 'plnet') is visible on the\nleaderboard and ranked in 2nd position for 'verb', and in 3rd position for both\n'noun' and 'action'.\n", "date": "2022-09-09", "categories": []}
{"id": "2209.05077", "title": "BON: An extended public domain dataset for human activity recognition", "abstract": "  Body-worn first-person vision (FPV) camera enables to extract a rich source\nof information on the environment from the subject's viewpoint. However, the\nresearch progress in wearable camera-based egocentric office activity\nunderstanding is slow compared to other activity environments (e.g., kitchen\nand outdoor ambulatory), mainly due to the lack of adequate datasets to train\nmore sophisticated (e.g., deep learning) models for human activity recognition\nin office environments. This paper provides details of a large and publicly\navailable office activity dataset (BON) collected in different office settings\nacross three geographical locations: Barcelona (Spain), Oxford (UK) and Nairobi\n(Kenya), using a chest-mounted GoPro Hero camera. The BON dataset contains\neighteen common office activities that can be categorised into person-to-person\ninteractions (e.g., Chat with colleagues), person-to-object (e.g., Writing on a\nwhiteboard), and proprioceptive (e.g., Walking). Annotation is provided for\neach segment of video with 5-seconds duration. Generally, BON contains 25\nsubjects and 2639 total segments. In order to facilitate further research in\nthe sub-domain, we have also provided results that could be used as baselines\nfor future studies.\n", "date": "2022-09-12", "categories": []}
{"id": "2209.05194", "title": "Graphing the Future: Activity and Next Active Object Prediction using\n  Graph-based Activity Representations", "abstract": "  We present a novel approach for the visual prediction of human-object\ninteractions in videos. Rather than forecasting the human and object motion or\nthe future hand-object contact points, we aim at predicting (a)the class of the\non-going human-object interaction and (b) the class(es) of the next active\nobject(s) (NAOs), i.e., the object(s) that will be involved in the interaction\nin the near future as well as the time the interaction will occur. Graph\nmatching relies on the efficient Graph Edit distance (GED) method. The\nexperimental evaluation of the proposed approach was conducted using two\nwell-established video datasets that contain human-object interactions, namely\nthe MSR Daily Activities and the CAD120. High prediction accuracy was obtained\nfor both action prediction and NAO forecasting.\n", "date": "2022-09-12", "categories": []}
{"id": "2208.12940", "title": "Actor-identified Spatiotemporal Action Detection -- Detecting Who Is\n  Doing What in Videos", "abstract": "  The success of deep learning on video Action Recognition (AR) has motivated\nresearchers to progressively promote related tasks from the coarse level to the\nfine-grained level. Compared with conventional AR which only predicts an action\nlabel for the entire video, Temporal Action Detection (TAD) has been\ninvestigated for estimating the start and end time for each action in videos.\nTaking TAD a step further, Spatiotemporal Action Detection (SAD) has been\nstudied for localizing the action both spatially and temporally in videos.\nHowever, who performs the action, is generally ignored in SAD, while\nidentifying the actor could also be important. To this end, we propose a novel\ntask, Actor-identified Spatiotemporal Action Detection (ASAD), to bridge the\ngap between SAD and actor identification.\n  In ASAD, we not only detect the spatiotemporal boundary for instance-level\naction but also assign the unique ID to each actor. To approach ASAD, Multiple\nObject Tracking (MOT) and Action Classification (AC) are two fundamental\nelements. By using MOT, the spatiotemporal boundary of each actor is obtained\nand assigned to a unique actor identity. By using AC, the action class is\nestimated within the corresponding spatiotemporal boundary. Since ASAD is a new\ntask, it poses many new challenges that cannot be addressed by existing\nmethods: i) no dataset is specifically created for ASAD, ii) no evaluation\nmetrics are designed for ASAD, iii) current MOT performance is the bottleneck\nto obtain satisfactory ASAD results. To address those problems, we contribute\nto i) annotate a new ASAD dataset, ii) propose ASAD evaluation metrics by\nconsidering multi-label actions and actor identification, iii) improve the data\nassociation strategies in MOT to boost the MOT performance, which leads to\nbetter ASAD results. The code is available at https://github.com/fandulu/ASAD.\n", "date": "2022-08-27", "categories": []}
{"id": "2209.01311", "title": "A Novel Self-Knowledge Distillation Approach with Siamese Representation\n  Learning for Action Recognition", "abstract": "  Knowledge distillation is an effective transfer of knowledge from a heavy\nnetwork (teacher) to a small network (student) to boost students' performance.\nSelf-knowledge distillation, the special case of knowledge distillation, has\nbeen proposed to remove the large teacher network training process while\npreserving the student's performance. This paper introduces a novel\nSelf-knowledge distillation approach via Siamese representation learning, which\nminimizes the difference between two representation vectors of the two\ndifferent views from a given sample. Our proposed method, SKD-SRL, utilizes\nboth soft label distillation and the similarity of representation vectors.\nTherefore, SKD-SRL can generate more consistent predictions and representations\nin various views of the same data point. Our benchmark has been evaluated on\nvarious standard datasets. The experimental results have shown that SKD-SRL\nsignificantly improves the accuracy compared to existing supervised learning\nand knowledge distillation methods regardless of the networks.\n", "date": "2022-09-03", "categories": []}
{"id": "2209.01425", "title": "Dynamic Spatio-Temporal Specialization Learning for Fine-Grained Action\n  Recognition", "abstract": "  The goal of fine-grained action recognition is to successfully discriminate\nbetween action categories with subtle differences. To tackle this, we derive\ninspiration from the human visual system which contains specialized regions in\nthe brain that are dedicated towards handling specific tasks. We design a novel\nDynamic Spatio-Temporal Specialization (DSTS) module, which consists of\nspecialized neurons that are only activated for a subset of samples that are\nhighly similar. During training, the loss forces the specialized neurons to\nlearn discriminative fine-grained differences to distinguish between these\nsimilar samples, improving fine-grained recognition. Moreover, a\nspatio-temporal specialization method further optimizes the architectures of\nthe specialized neurons to capture either more spatial or temporal fine-grained\ninformation, to better tackle the large range of spatio-temporal variations in\nthe videos. Lastly, we design an Upstream-Downstream Learning algorithm to\noptimize our model's dynamic decisions during training, improving the\nperformance of our DSTS module. We obtain state-of-the-art performance on two\nwidely-used fine-grained action recognition datasets.\n", "date": "2022-09-03", "categories": []}
{"id": "2209.00760", "title": "Temporal Contrastive Learning with Curriculum", "abstract": "  We present ConCur, a contrastive video representation learning method that\nuses curriculum learning to impose a dynamic sampling strategy in contrastive\ntraining. More specifically, ConCur starts the contrastive training with easy\npositive samples (temporally close and semantically similar clips), and as the\ntraining progresses, it increases the temporal span effectively sampling hard\npositives (temporally away and semantically dissimilar). To learn better\ncontext-aware representations, we also propose an auxiliary task of predicting\nthe temporal distance between a positive pair of clips. We conduct extensive\nexperiments on two popular action recognition datasets, UCF101 and HMDB51, on\nwhich our proposed method achieves state-of-the-art performance on two\nbenchmark tasks of video action recognition and video retrieval. We explore the\nimpact of encoder backbones and pre-training strategies by using R(2+1)D and\nC3D encoders and pre-training on Kinetics-400 and Kinetics-200 datasets.\nMoreover, a detailed ablation study shows the effectiveness of each of the\ncomponents of our proposed method.\n", "date": "2022-09-02", "categories": []}
{"id": "2209.00065", "title": "ViA: View-invariant Skeleton Action Representation Learning via Motion\n  Retargeting", "abstract": "  Current self-supervised approaches for skeleton action representation\nlearning often focus on constrained scenarios, where videos and skeleton data\nare recorded in laboratory settings. When dealing with estimated skeleton data\nin real-world videos, such methods perform poorly due to the large variations\nacross subjects and camera viewpoints. To address this issue, we introduce ViA,\na novel View-Invariant Autoencoder for self-supervised skeleton action\nrepresentation learning. ViA leverages motion retargeting between different\nhuman performers as a pretext task, in order to disentangle the latent\naction-specific `Motion' features on top of the visual representation of a 2D\nor 3D skeleton sequence. Such `Motion' features are invariant to skeleton\ngeometry and camera view and allow ViA to facilitate both, cross-subject and\ncross-view action classification tasks. We conduct a study focusing on\ntransfer-learning for skeleton-based action recognition with self-supervised\npre-training on real-world data (e.g., Posetics). Our results showcase that\nskeleton representations learned from ViA are generic enough to improve upon\nstate-of-the-art action classification accuracy, not only on 3D laboratory\ndatasets such as NTU-RGB+D 60 and NTU-RGB+D 120, but also on real-world\ndatasets where only 2D data are accurately estimated, e.g., Toyota Smarthome,\nUAV-Human and Penn Action.\n", "date": "2022-08-31", "categories": []}
{"id": "2209.00407", "title": "MAPLE: Masked Pseudo-Labeling autoEncoder for Semi-supervised Point\n  Cloud Action Recognition", "abstract": "  Recognizing human actions from point cloud videos has attracted tremendous\nattention from both academia and industry due to its wide applications like\nautomatic driving, robotics, and so on. However, current methods for point\ncloud action recognition usually require a huge amount of data with manual\nannotations and a complex backbone network with high computation costs, which\nmakes it impractical for real-world applications. Therefore, this paper\nconsiders the task of semi-supervised point cloud action recognition. We\npropose a Masked Pseudo-Labeling autoEncoder (\\textbf{MAPLE}) framework to\nlearn effective representations with much fewer annotations for point cloud\naction recognition. In particular, we design a novel and efficient\n\\textbf{De}coupled \\textbf{s}patial-\\textbf{t}emporal Trans\\textbf{Former}\n(\\textbf{DestFormer}) as the backbone of MAPLE. In DestFormer, the spatial and\ntemporal dimensions of the 4D point cloud videos are decoupled to achieve\nefficient self-attention for learning both long-term and short-term features.\nMoreover, to learn discriminative features from fewer annotations, we design a\nmasked pseudo-labeling autoencoder structure to guide the DestFormer to\nreconstruct features of masked frames from the available frames. More\nimportantly, for unlabeled data, we exploit the pseudo-labels from the\nclassification head as the supervision signal for the reconstruction of\nfeatures from the masked frames. Finally, comprehensive experiments demonstrate\nthat MAPLE achieves superior results on three public benchmarks and outperforms\nthe state-of-the-art method by 8.08\\% accuracy on the MSR-Action3D dataset.\n", "date": "2022-09-01", "categories": []}
{"id": "2209.00451", "title": "Group Activity Recognition in Basketball Tracking Data -- Neural\n  Embeddings in Team Sports (NETS)", "abstract": "  Like many team sports, basketball involves two groups of players who engage\nin collaborative and adversarial activities to win a game. Players and teams\nare executing various complex strategies to gain an advantage over their\nopponents. Defining, identifying, and analyzing different types of activities\nis an important task in sports analytics, as it can lead to better strategies\nand decisions by the players and coaching staff. The objective of this paper is\nto automatically recognize basketball group activities from tracking data\nrepresenting locations of players and the ball during a game. We propose a\nnovel deep learning approach for group activity recognition (GAR) in team\nsports called NETS. To efficiently model the player relations in team sports,\nwe combined a Transformer-based architecture with LSTM embedding, and a\nteam-wise pooling layer to recognize the group activity. Training such a neural\nnetwork generally requires a large amount of annotated data, which incurs high\nlabeling cost. To address scarcity of manual labels, we generate weak-labels\nand pretrain the neural network on a self-supervised trajectory prediction\ntask. We used a large tracking data set from 632 NBA games to evaluate our\napproach. The results show that NETS is capable of learning group activities\nwith high accuracy, and that self- and weak-supervised training in NETS have a\npositive impact on GAR accuracy.\n", "date": "2022-08-31", "categories": []}
{"id": "2208.14847", "title": "Attentive pooling for Group Activity Recognition", "abstract": "  In group activity recognition, hierarchical framework is widely adopted to\nrepresent the relationships between individuals and their corresponding group,\nand has achieved promising performance. However, the existing methods simply\nemployed max/average pooling in this framework, which ignored the distinct\ncontributions of different individuals to the group activity recognition. In\nthis paper, we propose a new contextual pooling scheme, named attentive\npooling, which enables the weighted information transition from individual\nactions to group activity. By utilizing the attention mechanism, the attentive\npooling is intrinsically interpretable and able to embed member context into\nthe existing hierarchical model. In order to verify the effectiveness of the\nproposed scheme, two specific attentive pooling methods, i.e., global attentive\npooling (GAP) and hierarchical attentive pooling (HAP) are designed. GAP\nrewards the individuals that are significant to group activity, while HAP\nfurther considers the hierarchical division by introducing subgroup structure.\nThe experimental results on the benchmark dataset demonstrate that our proposal\nis significantly superior beyond the baseline and is comparable to the\nstate-of-the-art methods.\n", "date": "2022-08-31", "categories": []}
{"id": "2208.14209", "title": "A Circular Window-based Cascade Transformer for Online Action Detection", "abstract": "  Online action detection aims at the accurate action prediction of the current\nframe based on long historical observations. Meanwhile, it demands real-time\ninference on online streaming videos. In this paper, we advocate a novel and\nefficient principle for online action detection. It merely updates the latest\nand oldest historical representations in one window but reuses the intermediate\nones, which have been already computed. Based on this principle, we introduce a\nwindow-based cascade Transformer with a circular historical queue, where it\nconducts multi-stage attentions and cascade refinement on each window. We also\nexplore the association between online action detection and its counterpart\noffline action segmentation as an auxiliary task. We find that such an extra\nsupervision helps discriminative history clustering and acts as feature\naugmentation for better training the classifier and cascade refinement. Our\nproposed method achieves the state-of-the-art performances on three challenging\ndatasets THUMOS'14, TVSeries, and HDD. Codes will be available after\nacceptance.\n", "date": "2022-08-30", "categories": []}
{"id": "2208.12673", "title": "Enabling Weakly-Supervised Temporal Action Localization from On-Device\n  Learning of the Video Stream", "abstract": "  Detecting actions in videos have been widely applied in on-device\napplications. Practical on-device videos are always untrimmed with both action\nand background. It is desirable for a model to both recognize the class of\naction and localize the temporal position where the action happens. Such a task\nis called temporal action location (TAL), which is always trained on the cloud\nwhere multiple untrimmed videos are collected and labeled. It is desirable for\na TAL model to continuously and locally learn from new data, which can directly\nimprove the action detection precision while protecting customers' privacy.\nHowever, it is non-trivial to train a TAL model, since tremendous video samples\nwith temporal annotations are required. However, annotating videos frame by\nframe is exorbitantly time-consuming and expensive. Although weakly-supervised\nTAL (W-TAL) has been proposed to learn from untrimmed videos with only\nvideo-level labels, such an approach is also not suitable for on-device\nlearning scenarios. In practical on-device learning applications, data are\ncollected in streaming. Dividing such a long video stream into multiple video\nsegments requires lots of human effort, which hinders the exploration of\napplying the TAL tasks to realistic on-device learning applications. To enable\nW-TAL models to learn from a long, untrimmed streaming video, we propose an\nefficient video learning approach that can directly adapt to new environments.\nWe first propose a self-adaptive video dividing approach with a contrast\nscore-based segment merging approach to convert the video stream into multiple\nsegments. Then, we explore different sampling strategies on the TAL tasks to\nrequest as few labels as possible. To the best of our knowledge, we are the\nfirst attempt to directly learn from the on-device, long video stream.\n", "date": "2022-08-25", "categories": []}
{"id": "2208.07929", "title": "ViT-ReT: Vision and Recurrent Transformer Neural Networks for Human\n  Activity Recognition in Videos", "abstract": "  Human activity recognition is an emerging and important area in computer\nvision which seeks to determine the activity an individual or group of\nindividuals are performing. The applications of this field ranges from\ngenerating highlight videos in sports, to intelligent surveillance and gesture\nrecognition. Most activity recognition systems rely on a combination of\nconvolutional neural networks (CNNs) to perform feature extraction from the\ndata and recurrent neural networks (RNNs) to determine the time dependent\nnature of the data. This paper proposes and designs two transformer neural\nnetworks for human activity recognition: a recurrent transformer (ReT), a\nspecialized neural network used to make predictions on sequences of data, as\nwell as a vision transformer (ViT), a transformer optimized for extracting\nsalient features from images, to improve speed and scalability of activity\nrecognition. We have provided an extensive comparison of the proposed\ntransformer neural networks with the contemporary CNN and RNN-based human\nactivity recognition models in terms of speed and accuracy.\n", "date": "2022-08-16", "categories": []}
{"id": "2208.12027", "title": "Two-stage Fall Events Classification with Human Skeleton Data", "abstract": "  Fall detection and classification become an imper- ative problem for\nhealthcare applications particularity with the increasingly ageing population.\nCurrently, most of the fall clas- sification algorithms provide binary fall or\nno-fall classification. For better healthcare, it is thus not enough to do\nbinary fall classification but to extend it to multiple fall events\nclassification. In this work, we utilize the privacy mitigating human skeleton\ndata for multiple fall events classification. The skeleton features are\nextracted from the original RGB images to not only mitigate the personal\nprivacy, but also to reduce the impact of the dynamic illuminations. The\nproposed fall events classification method is divided into two stages. In the\nfirst stage, the model is trained to achieve the binary classification to\nfilter out the no-fall events. Then, in the second stage, the deep neural\nnetwork (DNN) model is trained to further classify the five types of fall\nevents. In order to confirm the efficiency of the proposed method, the\nexperiments on the UP-Fall dataset outperform the state-of-the-art.\n", "date": "2022-08-25", "categories": []}
{"id": "2208.12257", "title": "Video Mobile-Former: Video Recognition with Efficient Global\n  Spatial-temporal Modeling", "abstract": "  Transformer-based models have achieved top performance on major video\nrecognition benchmarks. Benefiting from the self-attention mechanism, these\nmodels show stronger ability of modeling long-range dependencies compared to\nCNN-based models. However, significant computation overheads, resulted from the\nquadratic complexity of self-attention on top of a tremendous number of tokens,\nlimit the use of existing video transformers in applications with limited\nresources like mobile devices. In this paper, we extend Mobile-Former to Video\nMobile-Former, which decouples the video architecture into a lightweight\n3D-CNNs for local context modeling and a Transformer modules for global\ninteraction modeling in a parallel fashion. To avoid significant computational\ncost incurred by computing self-attention between the large number of local\npatches in videos, we propose to use very few global tokens (e.g., 6) for a\nwhole video in Transformers to exchange information with 3D-CNNs with a\ncross-attention mechanism. Through efficient global spatial-temporal modeling,\nVideo Mobile-Former significantly improves the video recognition performance of\nalternative lightweight baselines, and outperforms other efficient CNN-based\nmodels at the low FLOP regime from 500M to 6G total FLOPs on various video\nrecognition tasks. It is worth noting that Video Mobile-Former is the first\nTransformer-based video model which constrains the computational budget within\n1G FLOPs.\n", "date": "2022-08-25", "categories": []}
{"id": "2208.11108", "title": "Efficient Attention-free Video Shift Transformers", "abstract": "  This paper tackles the problem of efficient video recognition. In this area,\nvideo transformers have recently dominated the efficiency (top-1 accuracy vs\nFLOPs) spectrum. At the same time, there have been some attempts in the image\ndomain which challenge the necessity of the self-attention operation within the\ntransformer architecture, advocating the use of simpler approaches for token\nmixing. However, there are no results yet for the case of video recognition,\nwhere the self-attention operator has a significantly higher impact (compared\nto the case of images) on efficiency. To address this gap, in this paper, we\nmake the following contributions: (a) we construct a highly efficient \\&\naccurate attention-free block based on the shift operator, coined Affine-Shift\nblock, specifically designed to approximate as closely as possible the\noperations in the MHSA block of a Transformer layer. Based on our Affine-Shift\nblock, we construct our Affine-Shift Transformer and show that it already\noutperforms all existing shift/MLP--based architectures for ImageNet\nclassification. (b) We extend our formulation in the video domain to construct\nVideo Affine-Shift Transformer (VAST), the very first purely attention-free\nshift-based video transformer. (c) We show that VAST significantly outperforms\nrecent state-of-the-art transformers on the most popular action recognition\nbenchmarks for the case of models with low computational and memory footprint.\nCode will be made available.\n", "date": "2022-08-23", "categories": []}
{"id": "2208.09588", "title": "Review on Action Recognition for Accident Detection in Smart City\n  Transportation Systems", "abstract": "  Action detection and public traffic safety are crucial aspects of a safe\ncommunity and a better society. Monitoring traffic flows in a smart city using\ndifferent surveillance cameras can play a significant role in recognizing\naccidents and alerting first responders. The utilization of action recognition\n(AR) in computer vision tasks has contributed towards high-precision\napplications in video surveillance, medical imaging, and digital signal\nprocessing. This paper presents an intensive review focusing on action\nrecognition in accident detection and autonomous transportation systems for a\nsmart city. In this paper, we focused on AR systems that used diverse sources\nof traffic video capturing, such as static surveillance cameras on traffic\nintersections, highway monitoring cameras, drone cameras, and dash-cams.\nThrough this review, we identified the primary techniques, taxonomies, and\nalgorithms used in AR for autonomous transportation and accident detection. We\nalso examined data sets utilized in the AR tasks, identifying the main sources\nof datasets and features of the datasets. This paper provides potential\nresearch direction to develop and integrate accident detection systems for\nautonomous cars and public traffic safety systems by alerting emergency\npersonnel and law enforcement in the event of road accidents to minimize human\nerror in accident reporting and provide a spontaneous response to victims\n", "date": "2022-08-20", "categories": []}
{"id": "2207.06580", "title": "Proposal-Free Temporal Action Detection via Global Segmentation Mask\n  Learning", "abstract": "  Existing temporal action detection (TAD) methods rely on generating an\noverwhelmingly large number of proposals per video. This leads to complex model\ndesigns due to proposal generation and/or per-proposal action instance\nevaluation and the resultant high computational cost. In this work, for the\nfirst time, we propose a proposal-free Temporal Action detection model with\nGlobal Segmentation mask (TAGS). Our core idea is to learn a global\nsegmentation mask of each action instance jointly at the full video length. The\nTAGS model differs significantly from the conventional proposal-based methods\nby focusing on global temporal representation learning to directly detect local\nstart and end points of action instances without proposals. Further, by\nmodeling TAD holistically rather than locally at the individual proposal level,\nTAGS needs a much simpler model architecture with lower computational cost.\nExtensive experiments show that despite its simpler design, TAGS outperforms\nexisting TAD methods, achieving new state-of-the-art performance on two\nbenchmarks. Importantly, it is ~ 20x faster to train and ~1.6x more efficient\nfor inference. Our PyTorch implementation of TAGS is available at\nhttps://github.com/sauradip/TAGS .\n", "date": "2022-07-14", "categories": []}
{"id": "2208.09150", "title": "Part-aware Prototypical Graph Network for One-shot Skeleton-based Action\n  Recognition", "abstract": "  In this paper, we study the problem of one-shot skeleton-based action\nrecognition, which poses unique challenges in learning transferable\nrepresentation from base classes to novel classes, particularly for\nfine-grained actions. Existing meta-learning frameworks typically rely on the\nbody-level representations in spatial dimension, which limits the\ngeneralisation to capture subtle visual differences in the fine-grained label\nspace. To overcome the above limitation, we propose a part-aware prototypical\nrepresentation for one-shot skeleton-based action recognition. Our method\ncaptures skeleton motion patterns at two distinctive spatial levels, one for\nglobal contexts among all body joints, referred to as body level, and the other\nattends to local spatial regions of body parts, referred to as the part level.\nWe also devise a class-agnostic attention mechanism to highlight important\nparts for each action class. Specifically, we develop a part-aware prototypical\ngraph network consisting of three modules: a cascaded embedding module for our\ndual-level modelling, an attention-based part fusion module to fuse parts and\ngenerate part-aware prototypes, and a matching module to perform classification\nwith the part-aware representations. We demonstrate the effectiveness of our\nmethod on two public skeleton-based action recognition datasets: NTU RGB+D 120\nand NW-UCLA.\n", "date": "2022-08-19", "categories": []}
{"id": "2208.08090", "title": "Progressive Cross-modal Knowledge Distillation for Human Action\n  Recognition", "abstract": "  Wearable sensor-based Human Action Recognition (HAR) has achieved remarkable\nsuccess recently. However, the accuracy performance of wearable sensor-based\nHAR is still far behind the ones from the visual modalities-based system (i.e.,\nRGB video, skeleton, and depth). Diverse input modalities can provide\ncomplementary cues and thus improve the accuracy performance of HAR, but how to\ntake advantage of multi-modal data on wearable sensor-based HAR has rarely been\nexplored. Currently, wearable devices, i.e., smartwatches, can only capture\nlimited kinds of non-visual modality data. This hinders the multi-modal HAR\nassociation as it is unable to simultaneously use both visual and non-visual\nmodality data. Another major challenge lies in how to efficiently utilize\nmultimodal data on wearable devices with their limited computation resources.\nIn this work, we propose a novel Progressive Skeleton-to-sensor Knowledge\nDistillation (PSKD) model which utilizes only time-series data, i.e.,\naccelerometer data, from a smartwatch for solving the wearable sensor-based HAR\nproblem. Specifically, we construct multiple teacher models using data from\nboth teacher (human skeleton sequence) and student (time-series accelerometer\ndata) modalities. In addition, we propose an effective progressive learning\nscheme to eliminate the performance gap between teacher and student models. We\nalso designed a novel loss function called Adaptive-Confidence Semantic (ACS),\nto allow the student model to adaptively select either one of the teacher\nmodels or the ground-truth label it needs to mimic. To demonstrate the\neffectiveness of our proposed PSKD method, we conduct extensive experiments on\nBerkeley-MHAD, UTD-MHAD, and MMAct datasets. The results confirm that the\nproposed PSKD method has competitive performance compared to the previous mono\nsensor-based HAR methods.\n", "date": "2022-08-17", "categories": []}
{"id": "2208.07493", "title": "Temporal Action Localization with Multi-temporal Scales", "abstract": "  Temporal action localization plays an important role in video analysis, which\naims to localize and classify actions in untrimmed videos. The previous methods\noften predict actions on a feature space of a single-temporal scale. However,\nthe temporal features of a low-level scale lack enough semantics for action\nclassification while a high-level scale cannot provide rich details of the\naction boundaries. To address this issue, we propose to predict actions on a\nfeature space of multi-temporal scales. Specifically, we use refined feature\npyramids of different scales to pass semantics from high-level scales to\nlow-level scales. Besides, to establish the long temporal scale of the entire\nvideo, we use a spatial-temporal transformer encoder to capture the long-range\ndependencies of video frames. Then the refined features with long-range\ndependencies are fed into a classifier for the coarse action prediction.\nFinally, to further improve the prediction accuracy, we propose to use a\nframe-level self attention module to refine the classification and boundaries\nof each action instance. Extensive experiments show that the proposed method\ncan outperform state-of-the-art approaches on the THUMOS14 dataset and achieves\ncomparable performance on the ActivityNet1.3 dataset. Compared with A2Net\n(TIP20, Avg\\{0.3:0.7\\}), Sub-Action (CSVT2022, Avg\\{0.1:0.5\\}), and AFSD\n(CVPR21, Avg\\{0.3:0.7\\}) on the THUMOS14 dataset, the proposed method can\nachieve improvements of 12.6\\%, 17.4\\% and 2.2\\%, respectively\n", "date": "2022-08-16", "categories": []}
{"id": "2208.07344", "title": "Action Recognition based on Cross-Situational Action-object Statistics", "abstract": "  Machine learning models of visual action recognition are typically trained\nand tested on data from specific situations where actions are associated with\ncertain objects. It is an open question how action-object associations in the\ntraining set influence a model's ability to generalize beyond trained\nsituations. We set out to identify properties of training data that lead to\naction recognition models with greater generalization ability. To do this, we\ntake inspiration from a cognitive mechanism called cross-situational learning,\nwhich states that human learners extract the meaning of concepts by observing\ninstances of the same concept across different situations. We perform\ncontrolled experiments with various types of action-object associations, and\nidentify key properties of action-object co-occurrence in training data that\nlead to better classifiers. Given that these properties are missing in the\ndatasets that are typically used to train action classifiers in the computer\nvision literature, our work provides useful insights on how we should best\nconstruct datasets for efficiently training for better generalization.\n", "date": "2022-08-15", "categories": []}
{"id": "2208.05775", "title": "PSUMNet: Unified Modality Part Streams are All You Need for Efficient\n  Pose-based Action Recognition", "abstract": "  Pose-based action recognition is predominantly tackled by approaches which\ntreat the input skeleton in a monolithic fashion, i.e. joints in the pose tree\nare processed as a whole. However, such approaches ignore the fact that action\ncategories are often characterized by localized action dynamics involving only\nsmall subsets of part joint groups involving hands (e.g. `Thumbs up') or legs\n(e.g. `Kicking'). Although part-grouping based approaches exist, each part\ngroup is not considered within the global pose frame, causing such methods to\nfall short. Further, conventional approaches employ independent modality\nstreams (e.g. joint, bone, joint velocity, bone velocity) and train their\nnetwork multiple times on these streams, which massively increases the number\nof training parameters. To address these issues, we introduce PSUMNet, a novel\napproach for scalable and efficient pose-based action recognition. At the\nrepresentation level, we propose a global frame based part stream approach as\nopposed to conventional modality based streams. Within each part stream, the\nassociated data from multiple modalities is unified and consumed by the\nprocessing pipeline. Experimentally, PSUMNet achieves state of the art\nperformance on the widely used NTURGB+D 60/120 dataset and dense joint skeleton\ndataset NTU 60-X/120-X. PSUMNet is highly efficient and outperforms competing\nmethods which use 100%-400% more parameters. PSUMNet also generalizes to the\nSHREC hand gesture dataset with competitive performance. Overall, PSUMNet's\nscalability, performance and efficiency makes it an attractive choice for\naction recognition and for deployment on compute-restricted embedded and edge\ndevices. Code and pretrained models can be accessed at\nhttps://github.com/skelemoa/psumnet\n", "date": "2022-08-11", "categories": []}
{"id": "2208.05034", "title": "Human Activity Recognition Using Cascaded Dual Attention CNN and\n  Bi-Directional GRU Framework", "abstract": "  Vision-based human activity recognition has emerged as one of the essential\nresearch areas in video analytics domain. Over the last decade, numerous\nadvanced deep learning algorithms have been introduced to recognize complex\nhuman actions from video streams. These deep learning algorithms have shown\nimpressive performance for the human activity recognition task. However, these\nnewly introduced methods either exclusively focus on model performance or the\neffectiveness of these models in terms of computational efficiency and\nrobustness, resulting in a biased tradeoff in their proposals to deal with\nchallenging human activity recognition problem. To overcome the limitations of\ncontemporary deep learning models for human activity recognition, this paper\npresents a computationally efficient yet generic spatial-temporal cascaded\nframework that exploits the deep discriminative spatial and temporal features\nfor human activity recognition. For efficient representation of human actions,\nwe have proposed an efficient dual attentional convolutional neural network\n(CNN) architecture that leverages a unified channel-spatial attention mechanism\nto extract human-centric salient features in video frames. The dual\nchannel-spatial attention layers together with the convolutional layers learn\nto be more attentive in the spatial receptive fields having objects over the\nnumber of feature maps. The extracted discriminative salient features are then\nforwarded to stacked bi-directional gated recurrent unit (Bi-GRU) for long-term\ntemporal modeling and recognition of human actions using both forward and\nbackward pass gradient learning. Extensive experiments are conducted, where the\nobtained results show that the proposed framework attains an improvement in\nexecution time up to 167 times in terms of frames per second as compared to\nmost of the contemporary action recognition methods.\n", "date": "2022-08-09", "categories": []}
{"id": "2208.03444", "title": "AFE-CNN: 3D Skeleton-based Action Recognition with Action Feature\n  Enhancement", "abstract": "  Existing 3D skeleton-based action recognition approaches reach impressive\nperformance by encoding handcrafted action features to image format and\ndecoding by CNNs. However, such methods are limited in two ways: a) the\nhandcrafted action features are difficult to handle challenging actions, and b)\nthey generally require complex CNN models to improve action recognition\naccuracy, which usually occur heavy computational burden. To overcome these\nlimitations, we introduce a novel AFE-CNN, which devotes to enhance the\nfeatures of 3D skeleton-based actions to adapt to challenging actions. We\npropose feature enhance modules from key joint, bone vector, key frame and\ntemporal perspectives, thus the AFE-CNN is more robust to camera views and body\nsizes variation, and significantly improve the recognition accuracy on\nchallenging actions. Moreover, our AFE-CNN adopts a light-weight CNN model to\ndecode images with action feature enhanced, which ensures a much lower\ncomputational burden than the state-of-the-art methods. We evaluate the AFE-CNN\non three benchmark skeleton-based action datasets: NTU RGB+D, NTU RGB+D 120,\nand UTKinect-Action3D, with extensive experimental results demonstrate our\noutstanding performance of AFE-CNN.\n", "date": "2022-08-06", "categories": []}
{"id": "2208.03497", "title": "Contrastive Positive Mining for Unsupervised 3D Action Representation\n  Learning", "abstract": "  Recent contrastive based 3D action representation learning has made great\nprogress. However, the strict positive/negative constraint is yet to be relaxed\nand the use of non-self positive is yet to be explored. In this paper, a\nContrastive Positive Mining (CPM) framework is proposed for unsupervised\nskeleton 3D action representation learning. The CPM identifies non-self\npositives in a contextual queue to boost learning. Specifically, the siamese\nencoders are adopted and trained to match the similarity distributions of the\naugmented instances in reference to all instances in the contextual queue. By\nidentifying the non-self positive instances in the queue, a positive-enhanced\nlearning strategy is proposed to leverage the knowledge of mined positives to\nboost the robustness of the learned latent space against intra-class and\ninter-class diversity. Experimental results have shown that the proposed CPM is\neffective and outperforms the existing state-of-the-art unsupervised methods on\nthe challenging NTU and PKU-MMD datasets.\n", "date": "2022-08-06", "categories": []}
{"id": "2208.03550", "title": "Frozen CLIP Models are Efficient Video Learners", "abstract": "  Video recognition has been dominated by the end-to-end learning paradigm --\nfirst initializing a video recognition model with weights of a pretrained image\nmodel and then conducting end-to-end training on videos. This enables the video\nnetwork to benefit from the pretrained image model. However, this requires\nsubstantial computation and memory resources for finetuning on videos and the\nalternative of directly using pretrained image features without finetuning the\nimage backbone leads to subpar results. Fortunately, recent advances in\nContrastive Vision-Language Pre-training (CLIP) pave the way for a new route\nfor visual recognition tasks. Pretrained on large open-vocabulary image-text\npair data, these models learn powerful visual representations with rich\nsemantics. In this paper, we present Efficient Video Learning (EVL) -- an\nefficient framework for directly training high-quality video recognition models\nwith frozen CLIP features. Specifically, we employ a lightweight Transformer\ndecoder and learn a query token to dynamically collect frame-level spatial\nfeatures from the CLIP image encoder. Furthermore, we adopt a local temporal\nmodule in each decoder layer to discover temporal clues from adjacent frames\nand their attention maps. We show that despite being efficient to train with a\nfrozen backbone, our models learn high quality video representations on a\nvariety of video recognition datasets. Code is available at\nhttps://github.com/OpenGVLab/efficient-video-recognition.\n", "date": "2022-08-06", "categories": []}
{"id": "2208.03648", "title": "Weakly Supervised Online Action Detection for Infant General Movements", "abstract": "  To make the earlier medical intervention of infants' cerebral palsy (CP),\nearly diagnosis of brain damage is critical. Although general movements\nassessment(GMA) has shown promising results in early CP detection, it is\nlaborious. Most existing works take videos as input to make fidgety\nmovements(FMs) classification for the GMA automation. Those methods require a\ncomplete observation of videos and can not localize video frames containing\nnormal FMs. Therefore we propose a novel approach named WO-GMA to perform FMs\nlocalization in the weakly supervised online setting. Infant body keypoints are\nfirst extracted as the inputs to WO-GMA. Then WO-GMA performs local\nspatio-temporal extraction followed by two network branches to generate pseudo\nclip labels and model online actions. With the clip-level pseudo labels, the\naction modeling branch learns to detect FMs in an online fashion. Experimental\nresults on a dataset with 757 videos of different infants show that WO-GMA can\nget state-of-the-art video-level classification and cliplevel detection\nresults. Moreover, only the first 20% duration of the video is needed to get\nclassification results as good as fully observed, implying a significantly\nshortened FMs diagnosis time. Code is available at:\nhttps://github.com/scofiedluo/WO-GMA.\n", "date": "2022-08-07", "categories": []}
{"id": "2208.03775", "title": "Video-based Human Action Recognition using Deep Learning: A Review", "abstract": "  Human action recognition is an important application domain in computer\nvision. Its primary aim is to accurately describe human actions and their\ninteractions from a previously unseen data sequence acquired by sensors. The\nability to recognize, understand, and predict complex human actions enables the\nconstruction of many important applications such as intelligent surveillance\nsystems, human-computer interfaces, health care, security, and military\napplications. In recent years, deep learning has been given particular\nattention by the computer vision community. This paper presents an overview of\nthe current state-of-the-art in action recognition using video analysis with\ndeep learning techniques. We present the most important deep learning models\nfor recognizing human actions, and analyze them to provide the current progress\nof deep learning algorithms applied to solve human action recognition problems\nin realistic videos highlighting their advantages and disadvantages. Based on\nthe quantitative analysis using recognition accuracies reported in the\nliterature, our study identifies state-of-the-art deep architectures in action\nrecognition and then provides current trends and open problems for future works\nin this field.\n", "date": "2022-08-07", "categories": []}
{"id": "2208.03040", "title": "Blockwise Temporal-Spatial Pathway Network", "abstract": "  Algorithms for video action recognition should consider not only spatial\ninformation but also temporal relations, which remains challenging. We propose\na 3D-CNN-based action recognition model, called the blockwise temporal-spatial\npath-way network (BTSNet), which can adjust the temporal and spatial receptive\nfields by multiple pathways. We designed a novel model inspired by an adaptive\nkernel selection-based model, which is an architecture for effective feature\nencoding that adaptively chooses spatial receptive fields for image\nrecognition. Expanding this approach to the temporal domain, our model extracts\ntemporal and channel-wise attention and fuses information on various candidate\noperations. For evaluation, we tested our proposed model on UCF-101, HMDB-51,\nSVW, and Epic-Kitchen datasets and showed that it generalized well without\npretraining. BTSNet also provides interpretable visualization based on\nspatiotemporal channel-wise attention. We confirm that the blockwise\ntemporal-spatial pathway supports a better representation for 3D convolutional\nblocks based on this visualization.\n", "date": "2022-08-05", "categories": []}
{"id": "2208.02459", "title": "Privacy-Preserving Action Recognition via Motion Difference Quantization", "abstract": "  The widespread use of smart computer vision systems in our personal spaces\nhas led to an increased consciousness about the privacy and security risks that\nthese systems pose. On the one hand, we want these systems to assist in our\ndaily lives by understanding their surroundings, but on the other hand, we want\nthem to do so without capturing any sensitive information. Towards this\ndirection, this paper proposes a simple, yet robust privacy-preserving encoder\ncalled BDQ for the task of privacy-preserving human action recognition that is\ncomposed of three modules: Blur, Difference, and Quantization. First, the input\nscene is passed to the Blur module to smoothen the edges. This is followed by\nthe Difference module to apply a pixel-wise intensity subtraction between\nconsecutive frames to highlight motion features and suppress obvious high-level\nprivacy attributes. Finally, the Quantization module is applied to the motion\ndifference frames to remove the low-level privacy attributes. The BDQ\nparameters are optimized in an end-to-end fashion via adversarial training such\nthat it learns to allow action recognition attributes while inhibiting privacy\nattributes. Our experiments on three benchmark datasets show that the proposed\nencoder design can achieve state-of-the-art trade-off when compared with\nprevious works. Furthermore, we show that the trade-off achieved is at par with\nthe DVS sensor-based event cameras. Code available at:\nhttps://github.com/suakaw/BDQ_PrivacyAR.\n", "date": "2022-08-04", "categories": []}
{"id": "2206.07846", "title": "Action Spotting using Dense Detection Anchors Revisited: Submission to\n  the SoccerNet Challenge 2022", "abstract": "  This brief technical report describes our submission to the Action Spotting\nSoccerNet Challenge 2022. The challenge was part of the CVPR 2022 ActivityNet\nWorkshop. Our submission was based on a recently proposed method which focuses\non increasing temporal precision via a densely sampled set of detection\nanchors. Due to its emphasis on temporal precision, this approach had shown\nsignificant improvements in the tight average-mAP metric. Tight average-mAP was\nused as the evaluation criterion for the challenge, and is defined using small\ntemporal evaluation tolerances, thus being more sensitive to small temporal\nerrors. In order to further improve results, here we introduce small changes in\nthe pre- and post-processing steps, and also combine different input feature\ntypes via late fusion. These changes brought improvements that helped us\nachieve the first place in the challenge and also led to a new state-of-the-art\non SoccerNet's test set when using the dataset's standard experimental\nprotocol. This report briefly reviews the action spotting method based on dense\ndetection anchors, then focuses on the modifications introduced for the\nchallenge. We also describe the experimental protocols and training procedures\nwe used, and finally present our results.\n", "date": "2022-06-15", "categories": []}
{"id": "2208.01897", "title": "Combined CNN Transformer Encoder for Enhanced Fine-grained Human Action\n  Recognition", "abstract": "  Fine-grained action recognition is a challenging task in computer vision. As\nfine-grained datasets have small inter-class variations in spatial and temporal\nspace, fine-grained action recognition model requires good temporal reasoning\nand discrimination of attribute action semantics. Leveraging on CNN's ability\nin capturing high level spatial-temporal feature representations and\nTransformer's modeling efficiency in capturing latent semantics and global\ndependencies, we investigate two frameworks that combine CNN vision backbone\nand Transformer Encoder to enhance fine-grained action recognition: 1) a\nvision-based encoder to learn latent temporal semantics, and 2) a multi-modal\nvideo-text cross encoder to exploit additional text input and learn cross\nassociation between visual and text semantics. Our experimental results show\nthat both our Transformer encoder frameworks effectively learn latent temporal\nsemantics and cross-modality association, with improved recognition performance\nover CNN vision model. We achieve new state-of-the-art performance on the\nFineGym benchmark dataset for both proposed architectures.\n", "date": "2022-08-03", "categories": []}
{"id": "2206.15031", "title": "Timestamp-Supervised Action Segmentation with Graph Convolutional\n  Networks", "abstract": "  We introduce a novel approach for temporal activity segmentation with\ntimestamp supervision. Our main contribution is a graph convolutional network,\nwhich is learned in an end-to-end manner to exploit both frame features and\nconnections between neighboring frames to generate dense framewise labels from\nsparse timestamp labels. The generated dense framewise labels can then be used\nto train the segmentation model. In addition, we propose a framework for\nalternating learning of both the segmentation model and the graph convolutional\nmodel, which first initializes and then iteratively refines the learned models.\nDetailed experiments on four public datasets, including 50 Salads, GTEA,\nBreakfast, and Desktop Assembly, show that our method is superior to the\nmulti-layer perceptron baseline, while performing on par with or better than\nthe state of the art in temporal activity segmentation with timestamp\nsupervision.\n", "date": "2022-06-30", "categories": []}
{"id": "2207.13259", "title": "Spatiotemporal Self-attention Modeling with Temporal Patch Shift for\n  Action Recognition", "abstract": "  Transformer-based methods have recently achieved great advancement on 2D\nimage-based vision tasks. For 3D video-based tasks such as action recognition,\nhowever, directly applying spatiotemporal transformers on video data will bring\nheavy computation and memory burdens due to the largely increased number of\npatches and the quadratic complexity of self-attention computation. How to\nefficiently and effectively model the 3D self-attention of video data has been\na great challenge for transformers. In this paper, we propose a Temporal Patch\nShift (TPS) method for efficient 3D self-attention modeling in transformers for\nvideo-based action recognition. TPS shifts part of patches with a specific\nmosaic pattern in the temporal dimension, thus converting a vanilla spatial\nself-attention operation to a spatiotemporal one with little additional cost.\nAs a result, we can compute 3D self-attention using nearly the same computation\nand memory cost as 2D self-attention. TPS is a plug-and-play module and can be\ninserted into existing 2D transformer models to enhance spatiotemporal feature\nlearning. The proposed method achieves competitive performance with\nstate-of-the-arts on Something-something V1 & V2, Diving-48, and Kinetics400\nwhile being much more efficient on computation and memory cost. The source code\nof TPS can be found at https://github.com/MartinXM/TPS.\n", "date": "2022-07-27", "categories": []}
{"id": "2207.12648", "title": "Efficient and Accurate Skeleton-Based Two-Person Interaction Recognition\n  Using Inter- and Intra-body Graphs", "abstract": "  Skeleton-based two-person interaction recognition has been gaining increasing\nattention as advancements are made in pose estimation and graph convolutional\nnetworks. Although the accuracy has been gradually improving, the increasing\ncomputational complexity makes it more impractical for a real-world\nenvironment. There is still room for accuracy improvement as the conventional\nmethods do not fully represent the relationship between inter-body joints. In\nthis paper, we propose a lightweight model for accurately recognizing\ntwo-person interactions. In addition to the architecture, which incorporates\nmiddle fusion, we introduce a factorized convolution technique to reduce the\nweight parameters of the model. We also introduce a network stream that\naccounts for relative distance changes between inter-body joints to improve\naccuracy. Experiments using two large-scale datasets, NTU RGB+D 60 and 120,\nshow that our method simultaneously achieved the highest accuracy and\nrelatively low computational complexity compared with the conventional methods.\n", "date": "2022-07-26", "categories": []}
{"id": "2207.12842", "title": "Unsupervised Domain Adaptation for Video Transformers in Action\n  Recognition", "abstract": "  Over the last few years, Unsupervised Domain Adaptation (UDA) techniques have\nacquired remarkable importance and popularity in computer vision. However, when\ncompared to the extensive literature available for images, the field of videos\nis still relatively unexplored. On the other hand, the performance of a model\nin action recognition is heavily affected by domain shift. In this paper, we\npropose a simple and novel UDA approach for video action recognition. Our\napproach leverages recent advances on spatio-temporal transformers to build a\nrobust source model that better generalises to the target domain. Furthermore,\nour architecture learns domain invariant features thanks to the introduction of\na novel alignment loss term derived from the Information Bottleneck principle.\nWe report results on two video action recognition benchmarks for UDA, showing\nstate-of-the-art performance on HMDB$\\leftrightarrow$UCF, as well as on\nKinetics$\\rightarrow$NEC-Drone, which is more challenging. This demonstrates\nthe effectiveness of our method in handling different levels of domain shift.\nThe source code is available at https://github.com/vturrisi/UDAVT.\n", "date": "2022-07-26", "categories": []}
{"id": "2206.04790", "title": "Learn2Augment: Learning to Composite Videos for Data Augmentation in\n  Action Recognition", "abstract": "  We address the problem of data augmentation for video action recognition.\nStandard augmentation strategies in video are hand-designed and sample the\nspace of possible augmented data points either at random, without knowing which\naugmented points will be better, or through heuristics. We propose to learn\nwhat makes a good video for action recognition and select only high-quality\nsamples for augmentation. In particular, we choose video compositing of a\nforeground and a background video as the data augmentation process, which\nresults in diverse and realistic new samples. We learn which pairs of videos to\naugment without having to actually composite them. This reduces the space of\npossible augmentations, which has two advantages: it saves computational cost\nand increases the accuracy of the final trained classifier, as the augmented\npairs are of higher quality than average. We present experimental results on\nthe entire spectrum of training settings: few-shot, semi-supervised and fully\nsupervised. We observe consistent improvements across all of them over prior\nwork and baselines on Kinetics, UCF101, HMDB51, and achieve a new\nstate-of-the-art on settings with limited data. We see improvements of up to\n8.6% in the semi-supervised setting.\n", "date": "2022-06-09", "categories": []}
{"id": "2207.11504", "title": "Intelligent 3D Network Protocol for Multimedia Data Classification using\n  Deep Learning", "abstract": "  In videos, the human's actions are of three-dimensional (3D) signals. These\nvideos investigate the spatiotemporal knowledge of human behavior. The\npromising ability is investigated using 3D convolution neural networks (CNNs).\nThe 3D CNNs have not yet achieved high output for their well-established\ntwo-dimensional (2D) equivalents in still photographs. Board 3D Convolutional\nMemory and Spatiotemporal fusion face training difficulty preventing 3D CNN\nfrom accomplishing remarkable evaluation. In this paper, we implement Hybrid\nDeep Learning Architecture that combines STIP and 3D CNN features to enhance\nthe performance of 3D videos effectively. After implementation, the more\ndetailed and deeper charting for training in each circle of space-time fusion.\nThe training model further enhances the results after handling complicated\nevaluations of models. The video classification model is used in this\nimplemented model. Intelligent 3D Network Protocol for Multimedia Data\nClassification using Deep Learning is introduced to further understand\nspacetime association in human endeavors. In the implementation of the result,\nthe well-known dataset, i.e., UCF101 to, evaluates the performance of the\nproposed hybrid technique. The results beat the proposed hybrid technique that\nsubstantially beats the initial 3D CNNs. The results are compared with\nstate-of-the-art frameworks from literature for action recognition on UCF101\nwith an accuracy of 95%.\n", "date": "2022-07-23", "categories": []}
{"id": "2207.11660", "title": "MAR: Masked Autoencoders for Efficient Action Recognition", "abstract": "  Standard approaches for video recognition usually operate on the full input\nvideos, which is inefficient due to the widely present spatio-temporal\nredundancy in videos. Recent progress in masked video modelling, i.e.,\nVideoMAE, has shown the ability of vanilla Vision Transformers (ViT) to\ncomplement spatio-temporal contexts given only limited visual contents.\nInspired by this, we propose propose Masked Action Recognition (MAR), which\nreduces the redundant computation by discarding a proportion of patches and\noperating only on a part of the videos. MAR contains the following two\nindispensable components: cell running masking and bridging classifier.\nSpecifically, to enable the ViT to perceive the details beyond the visible\npatches easily, cell running masking is presented to preserve the\nspatio-temporal correlations in videos, which ensures the patches at the same\nspatial location can be observed in turn for easy reconstructions.\nAdditionally, we notice that, although the partially observed features can\nreconstruct semantically explicit invisible patches, they fail to achieve\naccurate classification. To address this, a bridging classifier is proposed to\nbridge the semantic gap between the ViT encoded features for reconstruction and\nthe features specialized for classification. Our proposed MAR reduces the\ncomputational cost of ViT by 53% and extensive experiments show that MAR\nconsistently outperforms existing ViT models with a notable margin. Especially,\nwe found a ViT-Large trained by MAR outperforms the ViT-Huge trained by a\nstandard training scheme by convincing margins on both Kinetics-400 and\nSomething-Something v2 datasets, while our computation overhead of ViT-Large is\nonly 14.5% of ViT-Huge.\n", "date": "2022-07-24", "categories": []}
{"id": "2207.11805", "title": "Weakly-Supervised Temporal Action Detection for Fine-Grained Videos with\n  Hierarchical Atomic Actions", "abstract": "  Action understanding has evolved into the era of fine granularity, as most\nhuman behaviors in real life have only minor differences. To detect these\nfine-grained actions accurately in a label-efficient way, we tackle the problem\nof weakly-supervised fine-grained temporal action detection in videos for the\nfirst time. Without the careful design to capture subtle differences between\nfine-grained actions, previous weakly-supervised models for general action\ndetection cannot perform well in the fine-grained setting. We propose to model\nactions as the combinations of reusable atomic actions which are automatically\ndiscovered from data through self-supervised clustering, in order to capture\nthe commonality and individuality of fine-grained actions. The learnt atomic\nactions, represented by visual concepts, are further mapped to fine and coarse\naction labels leveraging the semantic label hierarchy. Our approach constructs\na visual representation hierarchy of four levels: clip level, atomic action\nlevel, fine action class level and coarse action class level, with supervision\nat each level. Extensive experiments on two large-scale fine-grained video\ndatasets, FineAction and FineGym, show the benefit of our proposed\nweakly-supervised model for fine-grained action detection, and it achieves\nstate-of-the-art results.\n", "date": "2022-07-24", "categories": []}
{"id": "2207.12100", "title": "IGFormer: Interaction Graph Transformer for Skeleton-based Human\n  Interaction Recognition", "abstract": "  Human interaction recognition is very important in many applications. One\ncrucial cue in recognizing an interaction is the interactive body parts. In\nthis work, we propose a novel Interaction Graph Transformer (IGFormer) network\nfor skeleton-based interaction recognition via modeling the interactive body\nparts as graphs. More specifically, the proposed IGFormer constructs\ninteraction graphs according to the semantic and distance correlations between\nthe interactive body parts, and enhances the representation of each person by\naggregating the information of the interactive body parts based on the learned\ngraphs. Furthermore, we propose a Semantic Partition Module to transform each\nhuman skeleton sequence into a Body-Part-Time sequence to better capture the\nspatial and temporal information of the skeleton sequence for learning the\ngraphs. Extensive experiments on three benchmark datasets demonstrate that our\nmodel outperforms the state-of-the-art with a significant margin.\n", "date": "2022-07-25", "categories": []}
{"id": "2207.09675", "title": "ERA: Expert Retrieval and Assembly for Early Action Prediction", "abstract": "  Early action prediction aims to successfully predict the class label of an\naction before it is completely performed. This is a challenging task because\nthe beginning stages of different actions can be very similar, with only minor\nsubtle differences for discrimination. In this paper, we propose a novel Expert\nRetrieval and Assembly (ERA) module that retrieves and assembles a set of\nexperts most specialized at using discriminative subtle differences, to\ndistinguish an input sample from other highly similar samples. To encourage our\nmodel to effectively use subtle differences for early action prediction, we\npush experts to discriminate exclusively between samples that are highly\nsimilar, forcing these experts to learn to use subtle differences that exist\nbetween those samples. Additionally, we design an effective Expert Learning\nRate Optimization method that balances the experts' optimization and leads to\nbetter performance. We evaluate our ERA module on four public action datasets\nand achieve state-of-the-art performance.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.11221", "title": "Domain Generalization for Activity Recognition via Adaptive Feature\n  Fusion", "abstract": "  Human activity recognition requires the efforts to build a generalizable\nmodel using the training datasets with the hope to achieve good performance in\ntest datasets. However, in real applications, the training and testing datasets\nmay have totally different distributions due to various reasons such as\ndifferent body shapes, acting styles, and habits, damaging the model's\ngeneralization performance. While such a distribution gap can be reduced by\nexisting domain adaptation approaches, they typically assume that the test data\ncan be accessed in the training stage, which is not realistic. In this paper,\nwe consider a more practical and challenging scenario: domain-generalized\nactivity recognition (DGAR) where the test dataset \\emph{cannot} be accessed\nduring training. To this end, we propose \\emph{Adaptive Feature Fusion for\nActivity Recognition~(AFFAR)}, a domain generalization approach that learns to\nfuse the domain-invariant and domain-specific representations to improve the\nmodel's generalization performance. AFFAR takes the best of both worlds where\ndomain-invariant representations enhance the transferability across domains and\ndomain-specific representations leverage the model discrimination power from\neach domain. Extensive experiments on three public HAR datasets show its\neffectiveness. Furthermore, we apply AFFAR to a real application, i.e., the\ndiagnosis of Children's Attention Deficit Hyperactivity Disorder~(ADHD), which\nalso demonstrates the superiority of our approach.\n", "date": "2022-07-21", "categories": []}
{"id": "2207.08653", "title": "Leveraging Action Affinity and Continuity for Semi-supervised Temporal\n  Action Segmentation", "abstract": "  We present a semi-supervised learning approach to the temporal action\nsegmentation task. The goal of the task is to temporally detect and segment\nactions in long, untrimmed procedural videos, where only a small set of videos\nare densely labelled, and a large collection of videos are unlabelled. To this\nend, we propose two novel loss functions for the unlabelled data: an action\naffinity loss and an action continuity loss. The action affinity loss guides\nthe unlabelled samples learning by imposing the action priors induced from the\nlabelled set. Action continuity loss enforces the temporal continuity of\nactions, which also provides frame-wise classification supervision. In\naddition, we propose an Adaptive Boundary Smoothing (ABS) approach to build\ncoarser action boundaries for more robust and reliable learning. The proposed\nloss functions and ABS were evaluated on three benchmarks. Results show that\nthey significantly improved action segmentation performance with a low amount\n(5% and 10%) of labelled data and achieved comparable results to full\nsupervision with 50% labelled data. Furthermore, ABS succeeded in boosting\nperformance when integrated into fully-supervised learning.\n", "date": "2022-07-18", "categories": []}
{"id": "2207.09662", "title": "HTNet: Anchor-free Temporal Action Localization with Hierarchical\n  Transformers", "abstract": "  Temporal action localization (TAL) is a task of identifying a set of actions\nin a video, which involves localizing the start and end frames and classifying\neach action instance. Existing methods have addressed this task by using\npredefined anchor windows or heuristic bottom-up boundary-matching strategies,\nwhich are major bottlenecks in inference time. Additionally, the main challenge\nis the inability to capture long-range actions due to a lack of global\ncontextual information. In this paper, we present a novel anchor-free\nframework, referred to as HTNet, which predicts a set of <start time, end time,\nclass> triplets from a video based on a Transformer architecture. After the\nprediction of coarse boundaries, we refine it through a background feature\nsampling (BFS) module and hierarchical Transformers, which enables our model to\naggregate global contextual information and effectively exploit the inherent\nsemantic relationships in a video. We demonstrate how our method localizes\naccurate action instances and achieves state-of-the-art performance on two TAL\nbenchmark datasets: THUMOS14 and ActivityNet 1.3.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.10137", "title": "A Generalized & Robust Framework For Timestamp Supervision in Temporal\n  Action Segmentation", "abstract": "  In temporal action segmentation, Timestamp supervision requires only a\nhandful of labelled frames per video sequence. For unlabelled frames, previous\nworks rely on assigning hard labels, and performance rapidly collapses under\nsubtle violations of the annotation assumptions. We propose a novel\nExpectation-Maximization (EM) based approach that leverages the label\nuncertainty of unlabelled frames and is robust enough to accommodate possible\nannotation errors. With accurate timestamp annotations, our proposed method\nproduces SOTA results and even exceeds the fully-supervised setup in several\nmetrics and datasets. When applied to timestamp annotations with missing action\nsegments, our method presents stable performance. To further test our\nformulation's robustness, we introduce the new challenging annotation setup of\nSkip-tag supervision. This setup relaxes constraints and requires annotations\nof any fixed number of random frames in a video, making it more flexible than\nTimestamp supervision while remaining competitive.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.10379", "title": "Temporal Saliency Query Network for Efficient Video Recognition", "abstract": "  Efficient video recognition is a hot-spot research topic with the explosive\ngrowth of multimedia data on the Internet and mobile devices. Most existing\nmethods select the salient frames without awareness of the class-specific\nsaliency scores, which neglect the implicit association between the saliency of\nframes and its belonging category. To alleviate this issue, we devise a novel\nTemporal Saliency Query (TSQ) mechanism, which introduces class-specific\ninformation to provide fine-grained cues for saliency measurement.\nSpecifically, we model the class-specific saliency measuring process as a\nquery-response task. For each category, the common pattern of it is employed as\na query and the most salient frames are responded to it. Then, the calculated\nsimilarities are adopted as the frame saliency scores. To achieve it, we\npropose a Temporal Saliency Query Network (TSQNet) that includes two\ninstantiations of the TSQ mechanism based on visual appearance similarities and\ntextual event-object relations. Afterward, cross-modality interactions are\nimposed to promote the information exchange between them. Finally, we use the\nclass-specific saliencies of the most confident categories generated by two\nmodalities to perform the selection of salient frames. Extensive experiments\ndemonstrate the effectiveness of our method by achieving state-of-the-art\nresults on ActivityNet, FCVID and Mini-Kinetics datasets. Our project page is\nat https://lawrencexia2008.github.io/projects/tsqnet .\n", "date": "2022-07-21", "categories": []}
{"id": "2207.10448", "title": "An Efficient Spatio-Temporal Pyramid Transformer for Action Detection", "abstract": "  The task of action detection aims at deducing both the action category and\nlocalization of the start and end moment for each action instance in a long,\nuntrimmed video. While vision Transformers have driven the recent advances in\nvideo understanding, it is non-trivial to design an efficient architecture for\naction detection due to the prohibitively expensive self-attentions over a long\nsequence of video clips. To this end, we present an efficient hierarchical\nSpatio-Temporal Pyramid Transformer (STPT) for action detection, building upon\nthe fact that the early self-attention layers in Transformers still focus on\nlocal patterns. Specifically, we propose to use local window attention to\nencode rich local spatio-temporal representations in the early stages while\napplying global attention modules to capture long-term space-time dependencies\nin the later stages. In this way, our STPT can encode both locality and\ndependency with largely reduced redundancy, delivering a promising trade-off\nbetween accuracy and efficiency. For example, with only RGB input, the proposed\nSTPT achieves 53.6% mAP on THUMOS14, surpassing I3D+AFSD RGB model by over 10%\nand performing favorably against state-of-the-art AFSD that uses additional\nflow features with 31% fewer GFLOPs, which serves as an effective and efficient\nend-to-end Transformer-based framework for action detection.\n", "date": "2022-07-21", "categories": []}
{"id": "2207.10519", "title": "Real-Time Elderly Monitoring for Senior Safety by Lightweight Human\n  Action Recognition", "abstract": "  With an increasing number of elders living alone, care-giving from a distance\nbecomes a compelling need, particularly for safety. Real-time monitoring and\naction recognition are essential to raise an alert timely when abnormal\nbehaviors or unusual activities occur. While wearable sensors are widely\nrecognized as a promising solution, highly depending on user's ability and\nwillingness makes them inefficient. In contrast, video streams collected\nthrough non-contact optical cameras provide richer information and release the\nburden on elders. In this paper, leveraging the Independently-Recurrent neural\nNetwork (IndRNN) we propose a novel Real-time Elderly Monitoring for senior\nSafety (REMS) based on lightweight human action recognition (HAR) technology.\nUsing captured skeleton images, the REMS scheme is able to recognize abnormal\nbehaviors or actions and preserve the user's privacy. To achieve high accuracy,\nthe HAR module is trained and fine-tuned using multiple databases. An extensive\nexperimental study verified that REMS system performs action recognition\naccurately and timely. REMS meets the design goals as a privacy-preserving\nelderly safety monitoring system and possesses the potential to be adopted in\nvarious smart monitoring systems.\n", "date": "2022-07-21", "categories": []}
{"id": "2207.09767", "title": "Collaborating Domain-shared and Target-specific Feature Clustering for\n  Cross-domain 3D Action Recognition", "abstract": "  In this work, we consider the problem of cross-domain 3D action recognition\nin the open-set setting, which has been rarely explored before. Specifically,\nthere is a source domain and a target domain that contain the skeleton\nsequences with different styles and categories, and our purpose is to cluster\nthe target data by utilizing the labeled source data and unlabeled target data.\nFor such a challenging task, this paper presents a novel approach dubbed CoDT\nto collaboratively cluster the domain-shared features and target-specific\nfeatures. CoDT consists of two parallel branches. One branch aims to learn\ndomain-shared features with supervised learning in the source domain, while the\nother is to learn target-specific features using contrastive learning in the\ntarget domain. To cluster the features, we propose an online clustering\nalgorithm that enables simultaneous promotion of robust pseudo label generation\nand feature clustering. Furthermore, to leverage the complementarity of\ndomain-shared features and target-specific features, we propose a novel\ncollaborative clustering strategy to enforce pair-wise relationship consistency\nbetween the two branches. We conduct extensive experiments on multiple\ncross-domain 3D action recognition datasets, and the results demonstrate the\neffectiveness of our method.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.09925", "title": "An Efficient Framework for Few-shot Skeleton-based Temporal Action\n  Segmentation", "abstract": "  Temporal action segmentation (TAS) aims to classify and locate actions in the\nlong untrimmed action sequence. With the success of deep learning, many deep\nmodels for action segmentation have emerged. However, few-shot TAS is still a\nchallenging problem. This study proposes an efficient framework for the\nfew-shot skeleton-based TAS, including a data augmentation method and an\nimproved model. The data augmentation approach based on motion interpolation is\npresented here to solve the problem of insufficient data, and can increase the\nnumber of samples significantly by synthesizing action sequences. Besides, we\nconcatenate a Connectionist Temporal Classification (CTC) layer with a network\ndesigned for skeleton-based TAS to obtain an optimized model. Leveraging CTC\ncan enhance the temporal alignment between prediction and ground truth and\nfurther improve the segment-wise metrics of segmentation results. Extensive\nexperiments on both public and self-constructed datasets, including two\nsmall-scale datasets and one large-scale dataset, show the effectiveness of two\nproposed methods in improving the performance of the few-shot skeleton-based\nTAS task.\n", "date": "2022-07-20", "categories": []}
{"id": "2207.09067", "title": "Time Is MattEr: Temporal Self-supervision for Video Transformers", "abstract": "  Understanding temporal dynamics of video is an essential aspect of learning\nbetter video representations. Recently, transformer-based architectural designs\nhave been extensively explored for video tasks due to their capability to\ncapture long-term dependency of input sequences. However, we found that these\nVideo Transformers are still biased to learn spatial dynamics rather than\ntemporal ones, and debiasing the spurious correlation is critical for their\nperformance. Based on the observations, we design simple yet effective\nself-supervised tasks for video models to learn temporal dynamics better.\nSpecifically, for debiasing the spatial bias, our method learns the temporal\norder of video frames as extra self-supervision and enforces the randomly\nshuffled frames to have low-confidence outputs. Also, our method learns the\ntemporal flow direction of video tokens among consecutive frames for enhancing\nthe correlation toward temporal dynamics. Under various video action\nrecognition tasks, we demonstrate the effectiveness of our method and its\ncompatibility with state-of-the-art Video Transformers.\n", "date": "2022-07-19", "categories": []}
{"id": "2207.09425", "title": "Geometric Features Informed Multi-person Human-object Interaction\n  Recognition in Videos", "abstract": "  Human-Object Interaction (HOI) recognition in videos is important for\nanalyzing human activity. Most existing work focusing on visual features\nusually suffer from occlusion in the real-world scenarios. Such a problem will\nbe further complicated when multiple people and objects are involved in HOIs.\nConsider that geometric features such as human pose and object position provide\nmeaningful information to understand HOIs, we argue to combine the benefits of\nboth visual and geometric features in HOI recognition, and propose a novel\nTwo-level Geometric feature-informed Graph Convolutional Network (2G-GCN). The\ngeometric-level graph models the interdependency between geometric features of\nhumans and objects, while the fusion-level graph further fuses them with visual\nfeatures of humans and objects. To demonstrate the novelty and effectiveness of\nour method in challenging scenarios, we propose a new multi-person HOI dataset\n(MPHOI-72). Extensive experiments on MPHOI-72 (multi-person HOI), CAD-120\n(single-human HOI) and Bimanual Actions (two-hand HOI) datasets demonstrate our\nsuperior performance compared to state-of-the-arts.\n", "date": "2022-07-19", "categories": []}
{"id": "2207.07806", "title": "CHARM: A Hierarchical Deep Learning Model for Classification of Complex\n  Human Activities Using Motion Sensors", "abstract": "  In this paper, we report a hierarchical deep learning model for\nclassification of complex human activities using motion sensors. In contrast to\ntraditional Human Activity Recognition (HAR) models used for event-based\nactivity recognition, such as step counting, fall detection, and gesture\nidentification, this new deep learning model, which we refer to as CHARM\n(Complex Human Activity Recognition Model), is aimed for recognition of\nhigh-level human activities that are composed of multiple different low-level\nactivities in a non-deterministic sequence, such as meal preparation, house\nchores, and daily routines. CHARM not only quantitatively outperforms\nstate-of-the-art supervised learning approaches for high-level activity\nrecognition in terms of average accuracy and F1 scores, but also automatically\nlearns to recognize low-level activities, such as manipulation gestures and\nlocomotion modes, without any explicit labels for such activities. This opens\nnew avenues for Human-Machine Interaction (HMI) modalities using wearable\nsensors, where the user can choose to associate an automated task with a\nhigh-level activity, such as controlling home automation (e.g., robotic vacuum\ncleaners, lights, and thermostats) or presenting contextually relevant\ninformation at the right time (e.g., reminders, status updates, and\nweather/news reports). In addition, the ability to learn low-level user\nactivities when trained using only high-level activity labels may pave the way\nto semi-supervised learning of HAR tasks that are inherently difficult to\nlabel.\n", "date": "2022-07-16", "categories": []}
{"id": "2207.08095", "title": "Learning from Temporal Spatial Cubism for Cross-Dataset Skeleton-based\n  Action Recognition", "abstract": "  Rapid progress and superior performance have been achieved for skeleton-based\naction recognition recently. In this article, we investigate this problem under\na cross-dataset setting, which is a new, pragmatic, and challenging task in\nreal-world scenarios. Following the unsupervised domain adaptation (UDA)\nparadigm, the action labels are only available on a source dataset, but\nunavailable on a target dataset in the training stage. Different from the\nconventional adversarial learning-based approaches for UDA, we utilize a\nself-supervision scheme to reduce the domain shift between two skeleton-based\naction datasets. Our inspiration is drawn from Cubism, an art genre from the\nearly 20th century, which breaks and reassembles the objects to convey a\ngreater context. By segmenting and permuting temporal segments or human body\nparts, we design two self-supervised learning classification tasks to explore\nthe temporal and spatial dependency of a skeleton-based action and improve the\ngeneralization ability of the model. We conduct experiments on six datasets for\nskeleton-based action recognition, including three large-scale datasets (NTU\nRGB+D, PKU-MMD, and Kinetics) where new cross-dataset settings and benchmarks\nare established. Extensive results demonstrate that our method outperforms\nstate-of-the-art approaches. The source codes of our model and all the compared\nmethods are available at https://github.com/shanice-l/st-cubism.\n", "date": "2022-07-17", "categories": []}
{"id": "2207.08184", "title": "Zero-Shot Temporal Action Detection via Vision-Language Prompting", "abstract": "  Existing temporal action detection (TAD) methods rely on large training data\nincluding segment-level annotations, limited to recognizing previously seen\nclasses alone during inference. Collecting and annotating a large training set\nfor each class of interest is costly and hence unscalable. Zero-shot TAD\n(ZS-TAD) resolves this obstacle by enabling a pre-trained model to recognize\nany unseen action classes. Meanwhile, ZS-TAD is also much more challenging with\nsignificantly less investigation. Inspired by the success of zero-shot image\nclassification aided by vision-language (ViL) models such as CLIP, we aim to\ntackle the more complex TAD task. An intuitive method is to integrate an\noff-the-shelf proposal detector with CLIP style classification. However, due to\nthe sequential localization (e.g, proposal generation) and classification\ndesign, it is prone to localization error propagation. To overcome this\nproblem, in this paper we propose a novel zero-Shot Temporal Action detection\nmodel via Vision-LanguagE prompting (STALE). Such a novel design effectively\neliminates the dependence between localization and classification by breaking\nthe route for error propagation in-between. We further introduce an interaction\nmechanism between classification and localization for improved optimization.\nExtensive experiments on standard ZS-TAD video benchmarks show that our STALE\nsignificantly outperforms state-of-the-art alternatives. Besides, our model\nalso yields superior results on supervised TAD over recent strong competitors.\nThe PyTorch implementation of STALE is available at\nhttps://github.com/sauradip/STALE.\n", "date": "2022-07-17", "categories": []}
{"id": "2207.05249", "title": "Efficient Human Vision Inspired Action Recognition using Adaptive\n  Spatiotemporal Sampling", "abstract": "  Adaptive sampling that exploits the spatiotemporal redundancy in videos is\ncritical for always-on action recognition on wearable devices with limited\ncomputing and battery resources. The commonly used fixed sampling strategy is\nnot context-aware and may under-sample the visual content, and thus adversely\nimpacts both computation efficiency and accuracy. Inspired by the concepts of\nfoveal vision and pre-attentive processing from the human visual perception\nmechanism, we introduce a novel adaptive spatiotemporal sampling scheme for\nefficient action recognition. Our system pre-scans the global scene context at\nlow-resolution and decides to skip or request high-resolution features at\nsalient regions for further processing. We validate the system on EPIC-KITCHENS\nand UCF-101 datasets for action recognition, and show that our proposed\napproach can greatly speed up inference with a tolerable loss of accuracy\ncompared with those from state-of-the-art baselines. Source code is available\nin https://github.com/knmac/adaptive_spatiotemporal.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.07646", "title": "Multimodal Open-Vocabulary Video Classification via Pre-Trained Vision\n  and Language Models", "abstract": "  Utilizing vision and language models (VLMs) pre-trained on large-scale\nimage-text pairs is becoming a promising paradigm for open-vocabulary visual\nrecognition. In this work, we extend this paradigm by leveraging motion and\naudio that naturally exist in video. We present \\textbf{MOV}, a simple yet\neffective method for \\textbf{M}ultimodal \\textbf{O}pen-\\textbf{V}ocabulary\nvideo classification. In MOV, we directly use the vision encoder from\npre-trained VLMs with minimal modifications to encode video, optical flow and\naudio spectrogram. We design a cross-modal fusion mechanism to aggregate\ncomplimentary multimodal information. Experiments on Kinetics-700 and VGGSound\nshow that introducing flow or audio modality brings large performance gains\nover the pre-trained VLM and existing methods. Specifically, MOV greatly\nimproves the accuracy on base classes, while generalizes better on novel\nclasses. MOV achieves state-of-the-art results on UCF and HMDB zero-shot video\nclassification benchmarks, significantly outperforming both traditional\nzero-shot methods and recent methods based on VLMs. Code and models will be\nreleased.\n", "date": "2022-07-15", "categories": []}
{"id": "2207.06659", "title": "Forcing the Whole Video as Background: An Adversarial Learning Strategy\n  for Weakly Temporal Action Localization", "abstract": "  With video-level labels, weakly supervised temporal action localization\n(WTAL) applies a localization-by-classification paradigm to detect and classify\nthe action in untrimmed videos. Due to the characteristic of classification,\nclass-specific background snippets are inevitably mis-activated to improve the\ndiscriminability of the classifier in WTAL. To alleviate the disturbance of\nbackground, existing methods try to enlarge the discrepancy between action and\nbackground through modeling background snippets with pseudo-snippet-level\nannotations, which largely rely on artificial hypotheticals. Distinct from the\nprevious works, we present an adversarial learning strategy to break the\nlimitation of mining pseudo background snippets. Concretely, the background\nclassification loss forces the whole video to be regarded as the background by\na background gradient reinforcement strategy, confusing the recognition model.\nReversely, the foreground(action) loss guides the model to focus on action\nsnippets under such conditions. As a result, competition between the two\nclassification losses drives the model to boost its ability for action\nmodeling. Simultaneously, a novel temporal enhancement network is designed to\nfacilitate the model to construct temporal relation of affinity snippets based\non the proposed strategy, for further improving the performance of action\nlocalization. Finally, extensive experiments conducted on THUMOS14 and\nActivityNet1.2 demonstrate the effectiveness of the proposed method.\n", "date": "2022-07-14", "categories": []}
{"id": "2207.07059", "title": "Semi-Supervised Temporal Action Detection with Proposal-Free Masking", "abstract": "  Existing temporal action detection (TAD) methods rely on a large number of\ntraining data with segment-level annotations. Collecting and annotating such a\ntraining set is thus highly expensive and unscalable. Semi-supervised TAD\n(SS-TAD) alleviates this problem by leveraging unlabeled videos freely\navailable at scale. However, SS-TAD is also a much more challenging problem\nthan supervised TAD, and consequently much under-studied. Prior SS-TAD methods\ndirectly combine an existing proposal-based TAD method and a SSL method. Due to\ntheir sequential localization (e.g, proposal generation) and classification\ndesign, they are prone to proposal error propagation. To overcome this\nlimitation, in this work we propose a novel Semi-supervised Temporal action\ndetection model based on PropOsal-free Temporal mask (SPOT) with a parallel\nlocalization (mask generation) and classification architecture. Such a novel\ndesign effectively eliminates the dependence between localization and\nclassification by cutting off the route for error propagation in-between. We\nfurther introduce an interaction mechanism between classification and\nlocalization for prediction refinement, and a new pretext task for\nself-supervised model pre-training. Extensive experiments on two standard\nbenchmarks show that our SPOT outperforms state-of-the-art alternatives, often\nby a large margin. The PyTorch implementation of SPOT is available at\nhttps://github.com/sauradip/SPOT\n", "date": "2022-07-14", "categories": []}
{"id": "2207.07097", "title": "ReAct: Temporal Action Detection with Relational Queries", "abstract": "  This work aims at advancing temporal action detection (TAD) using an\nencoder-decoder framework with action queries, similar to DETR, which has shown\ngreat success in object detection. However, the framework suffers from several\nproblems if directly applied to TAD: the insufficient exploration of\ninter-query relation in the decoder, the inadequate classification training due\nto a limited number of training samples, and the unreliable classification\nscores at inference. To this end, we first propose a relational attention\nmechanism in the decoder, which guides the attention among queries based on\ntheir relations. Moreover, we propose two losses to facilitate and stabilize\nthe training of action classification. Lastly, we propose to predict the\nlocalization quality of each action query at inference in order to distinguish\nhigh-quality queries. The proposed method, named ReAct, achieves the\nstate-of-the-art performance on THUMOS14, with much lower computational costs\nthan previous methods. Besides, extensive ablation studies are conducted to\nverify the effectiveness of each proposed component. The code is available at\nhttps://github.com/sssste/React.\n", "date": "2022-07-14", "categories": []}
{"id": "2207.06101", "title": "Global-local Motion Transformer for Unsupervised Skeleton-based Action\n  Learning", "abstract": "  We propose a new transformer model for the task of unsupervised learning of\nskeleton motion sequences. The existing transformer model utilized for\nunsupervised skeleton-based action learning is learned the instantaneous\nvelocity of each joint from adjacent frames without global motion information.\nThus, the model has difficulties in learning the attention globally over\nwhole-body motions and temporally distant joints. In addition, person-to-person\ninteractions have not been considered in the model. To tackle the learning of\nwhole-body motion, long-range temporal dynamics, and person-to-person\ninteractions, we design a global and local attention mechanism, where, global\nbody motions and local joint motions pay attention to each other. In addition,\nwe propose a novel pretraining strategy, multi-interval pose displacement\nprediction, to learn both global and local attention in diverse time ranges.\nThe proposed model successfully learns local dynamics of the joints and\ncaptures global context from the motion sequences. Our model outperforms\nstate-of-the-art models by notable margins in the representative benchmarks.\nCodes are available at https://github.com/Boeun-Kim/GL-Transformer.\n", "date": "2022-07-13", "categories": []}
{"id": "2207.06261", "title": "Is Appearance Free Action Recognition Possible?", "abstract": "  Intuition might suggest that motion and dynamic information are key to\nvideo-based action recognition. In contrast, there is evidence that\nstate-of-the-art deep-learning video understanding architectures are biased\ntoward static information available in single frames. Presently, a methodology\nand corresponding dataset to isolate the effects of dynamic information in\nvideo are missing. Their absence makes it difficult to understand how well\ncontemporary architectures capitalize on dynamic vs. static information. We\nrespond with a novel Appearance Free Dataset (AFD) for action recognition. AFD\nis devoid of static information relevant to action recognition in a single\nframe. Modeling of the dynamics is necessary for solving the task, as the\naction is only apparent through consideration of the temporal dimension. We\nevaluated 11 contemporary action recognition architectures on AFD as well as\nits related RGB video. Our results show a notable decrease in performance for\nall architectures on AFD compared to RGB. We also conducted a complimentary\nstudy with humans that shows their recognition accuracy on AFD and RGB is very\nsimilar and much better than the evaluated architectures on AFD. Our results\nmotivate a novel architecture that revives explicit recovery of optical flow,\nwithin a contemporary design for best performance on AFD and RGB.\n", "date": "2022-07-13", "categories": []}
{"id": "2207.05254", "title": "Hunting Group Clues with Transformers for Social Group Activity\n  Recognition", "abstract": "  This paper presents a novel framework for social group activity recognition.\nAs an expanded task of group activity recognition, social group activity\nrecognition requires recognizing multiple sub-group activities and identifying\ngroup members. Most existing methods tackle both tasks by refining region\nfeatures and then summarizing them into activity features. Such heuristic\nfeature design renders the effectiveness of features susceptible to incomplete\nperson localization and disregards the importance of scene contexts.\nFurthermore, region features are sub-optimal to identify group members because\nthe features may be dominated by those of people in the regions and have\ndifferent semantics. To overcome these drawbacks, we propose to leverage\nattention modules in transformers to generate effective social group features.\nOur method is designed in such a way that the attention modules identify and\nthen aggregate features relevant to social group activities, generating an\neffective feature for each social group. Group member information is embedded\ninto the features and thus accessed by feed-forward networks. The outputs of\nfeed-forward networks represent groups so concisely that group members can be\nidentified with simple Hungarian matching between groups and individuals.\nExperimental results show that our method outperforms state-of-the-art methods\non the Volleyball and Collective Activity datasets.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.05493", "title": "Skeletal Human Action Recognition using Hybrid Attention based Graph\n  Convolutional Network", "abstract": "  In skeleton-based action recognition, Graph Convolutional Networks model\nhuman skeletal joints as vertices and connect them through an adjacency matrix,\nwhich can be seen as a local attention mask. However, in most existing Graph\nConvolutional Networks, the local attention mask is defined based on natural\nconnections of human skeleton joints and ignores the dynamic relations for\nexample between head, hands and feet joints. In addition, the attention\nmechanism has been proven effective in Natural Language Processing and image\ndescription, which is rarely investigated in existing methods. In this work, we\nproposed a new adaptive spatial attention layer that extends local attention\nmap to global based on relative distance and relative angle information.\nMoreover, we design a new initial graph adjacency matrix that connects head,\nhands and feet, which shows visible improvement in terms of action recognition\naccuracy. The proposed model is evaluated on two large-scale and challenging\ndatasets in the field of human activities in daily life: NTU-RGB+D and Kinetics\nskeleton. The results demonstrate that our model has strong performance on both\ndataset.\n", "date": "2022-07-12", "categories": []}
{"id": "2207.05730", "title": "1st Place Solution to the EPIC-Kitchens Action Anticipation Challenge\n  2022", "abstract": "  In this report, we describe the technical details of our submission to the\nEPIC-Kitchens Action Anticipation Challenge 2022. In this competition, we\ndevelop the following two approaches. 1) Anticipation Time Knowledge\nDistillation using the soft labels learned by the teacher model as knowledge to\nguide the student network to learn the information of anticipation time; 2)\nVerb-Noun Relation Module for building the relationship between verbs and\nnouns. Our method achieves state-of-the-art results on the testing set of\nEPIC-Kitchens Action Anticipation Challenge 2022.\n", "date": "2022-07-10", "categories": []}
{"id": "2207.03782", "title": "VidConv: A modernized 2D ConvNet for Efficient Video Recognition", "abstract": "  Since being introduced in 2020, Vision Transformers (ViT) has been steadily\nbreaking the record for many vision tasks and are often described as\n``all-you-need\" to replace ConvNet. Despite that, ViTs are generally\ncomputational, memory-consuming, and unfriendly for embedded devices. In\naddition, recent research shows that standard ConvNet if redesigned and trained\nappropriately can compete favorably with ViT in terms of accuracy and\nscalability. In this paper, we adopt the modernized structure of ConvNet to\ndesign a new backbone for action recognition. Particularly, our main target is\nto serve for industrial product deployment, such as FPGA boards in which only\nstandard operations are supported. Therefore, our network simply consists of 2D\nconvolutions, without using any 3D convolution, long-range attention plugin, or\nTransformer blocks. While being trained with much fewer epochs (5x-10x), our\nbackbone surpasses the methods using (2+1)D and 3D convolution, and achieve\ncomparable results with ViT on two benchmark datasets.\n", "date": "2022-07-08", "categories": []}
{"id": "2207.03807", "title": "Beyond Transfer Learning: Co-finetuning for Action Localisation", "abstract": "  Transfer learning is the predominant paradigm for training deep networks on\nsmall target datasets. Models are typically pretrained on large ``upstream''\ndatasets for classification, as such labels are easy to collect, and then\nfinetuned on ``downstream'' tasks such as action localisation, which are\nsmaller due to their finer-grained annotations. In this paper, we question this\napproach, and propose co-finetuning -- simultaneously training a single model\non multiple ``upstream'' and ``downstream'' tasks. We demonstrate that\nco-finetuning outperforms traditional transfer learning when using the same\ntotal amount of data, and also show how we can easily extend our approach to\nmultiple ``upstream'' datasets to further improve performance. In particular,\nco-finetuning significantly improves the performance on rare classes in our\ndownstream task, as it has a regularising effect, and enables the network to\nlearn feature representations that transfer between different datasets.\nFinally, we observe how co-finetuning with public, video classification\ndatasets, we are able to achieve state-of-the-art results for spatio-temporal\naction localisation on the challenging AVA and AVA-Kinetics datasets,\noutperforming recent works which develop intricate models.\n", "date": "2022-07-08", "categories": []}
{"id": "2207.03065", "title": "Contrastive Learning from Spatio-Temporal Mixed Skeleton Sequences for\n  Self-Supervised Skeleton-Based Action Recognition", "abstract": "  Self-supervised skeleton-based action recognition with contrastive learning\nhas attracted much attention. Recent literature shows that data augmentation\nand large sets of contrastive pairs are crucial in learning such\nrepresentations. In this paper, we found that directly extending contrastive\npairs based on normal augmentations brings limited returns in terms of\nperformance, because the contribution of contrastive pairs from the normal data\naugmentation to the loss get smaller as training progresses. Therefore, we\ndelve into hard contrastive pairs for contrastive learning. Motivated by the\nsuccess of mixing augmentation strategy which improves the performance of many\ntasks by synthesizing novel samples, we propose SkeleMixCLR: a contrastive\nlearning framework with a spatio-temporal skeleton mixing augmentation\n(SkeleMix) to complement current contrastive learning approaches by providing\nhard contrastive samples. First, SkeleMix utilizes the topological information\nof skeleton data to mix two skeleton sequences by randomly combing the cropped\nskeleton fragments (the trimmed view) with the remaining skeleton sequences\n(the truncated view). Second, a spatio-temporal mask pooling is applied to\nseparate these two views at the feature level. Third, we extend contrastive\npairs with these two views. SkeleMixCLR leverages the trimmed and truncated\nviews to provide abundant hard contrastive pairs since they involve some\ncontext information from each other due to the graph convolution operations,\nwhich allows the model to learn better motion representations for action\nrecognition. Extensive experiments on NTU-RGB+D, NTU120-RGB+D, and PKU-MMD\ndatasets show that SkeleMixCLR achieves state-of-the-art performance. Codes are\navailable at https://github.com/czhaneva/SkeleMixCLR.\n", "date": "2022-07-07", "categories": []}
{"id": "2207.01708", "title": "Disentangled Action Recognition with Knowledge Bases", "abstract": "  Action in video usually involves the interaction of human with objects.\nAction labels are typically composed of various combinations of verbs and\nnouns, but we may not have training data for all possible combinations. In this\npaper, we aim to improve the generalization ability of the compositional action\nrecognition model to novel verbs or novel nouns that are unseen during training\ntime, by leveraging the power of knowledge graphs. Previous work utilizes\nverb-noun compositional action nodes in the knowledge graph, making it\ninefficient to scale since the number of compositional action nodes grows\nquadratically with respect to the number of verbs and nouns. To address this\nissue, we propose our approach: Disentangled Action Recognition with\nKnowledge-bases (DARK), which leverages the inherent compositionality of\nactions. DARK trains a factorized model by first extracting disentangled\nfeature representations for verbs and nouns, and then predicting classification\nweights using relations in external knowledge graphs. The type constraint\nbetween verb and noun is extracted from external knowledge bases and finally\napplied when composing actions. DARK has better scalability in the number of\nobjects and verbs, and achieves state-of-the-art performance on the Charades\ndataset. We further propose a new benchmark split based on the Epic-kitchen\ndataset which is an order of magnitude bigger in the numbers of classes and\nsamples, and benchmark various models on this benchmark.\n", "date": "2022-07-04", "categories": []}
{"id": "2207.02042", "title": "MVP: Robust Multi-View Practice for Driving Action Localization", "abstract": "  Distracted driving causes thousands of deaths per year, and how to apply\ndeep-learning methods to prevent these tragedies has become a crucial problem.\nIn Track3 of the 6th AI City Challenge, researchers provide a high-quality\nvideo dataset with densely action annotations. Due to the small data scale and\nunclear action boundary, the dataset presents a unique challenge to precisely\nlocalize all the different actions and classify their categories. In this\npaper, we make good use of the multi-view synchronization among videos, and\nconduct robust Multi-View Practice (MVP) for driving action localization. To\navoid overfitting, we fine-tune SlowFast with Kinetics-700 pre-training as the\nfeature extractor. Then the features of different views are passed to\nActionFormer to generate candidate action proposals. For precisely localizing\nall the actions, we design elaborate post-processing, including model voting,\nthreshold filtering and duplication removal. The results show that our MVP is\nrobust for driving action localization, which achieves 28.49% F1-score in the\nTrack3 test set.\n", "date": "2022-07-05", "categories": []}
{"id": "2206.15002", "title": "Spatial Transformer Network with Transfer Learning for Small-scale\n  Fine-grained Skeleton-based Tai Chi Action Recognition", "abstract": "  Human action recognition is a quite hugely investigated area where most\nremarkable action recognition networks usually use large-scale coarse-grained\naction datasets of daily human actions as inputs to state the superiority of\ntheir networks. We intend to recognize our small-scale fine-grained Tai Chi\naction dataset using neural networks and propose a transfer-learning method\nusing NTU RGB+D dataset to pre-train our network. More specifically, the\nproposed method first uses a large-scale NTU RGB+D dataset to pre-train the\nTransformer-based network for action recognition to extract common features\namong human motion. Then we freeze the network weights except for the fully\nconnected (FC) layer and take our Tai Chi actions as inputs only to train the\ninitialized FC weights. Experimental results show that our general model\npipeline can reach a high accuracy of small-scale fine-grained Tai Chi action\nrecognition with even few inputs and demonstrate that our method achieves the\nstate-of-the-art performance compared with previous Tai Chi action recognition\nmethods.\n", "date": "2022-06-30", "categories": []}
{"id": "2206.15085", "title": "Skeleton-based Action Recognition via Adaptive Cross-Form Learning", "abstract": "  Skeleton-based action recognition aims to project skeleton sequences to\naction categories, where skeleton sequences are derived from multiple forms of\npre-detected points. Compared with earlier methods that focus on exploring\nsingle-form skeletons via Graph Convolutional Networks (GCNs), existing methods\ntend to improve GCNs by leveraging multi-form skeletons due to their\ncomplementary cues. However, these methods (either adapting structure of GCNs\nor model ensemble) require the co-existence of all forms of skeletons during\nboth training and inference stages, while a typical situation in real life is\nthe existence of only partial forms for inference. To tackle this issue, we\npresent Adaptive Cross-Form Learning (ACFL), which empowers well-designed GCNs\nto generate complementary representation from single-form skeletons without\nchanging model capacity. Specifically, each GCN model in ACFL not only learns\naction representation from the single-form skeletons, but also adaptively\nmimics useful representations derived from other forms of skeletons. In this\nway, each GCN can learn how to strengthen what has been learned, thus\nexploiting model potential and facilitating action recognition as well.\nExtensive experiments conducted on three challenging benchmarks, i.e.,\nNTU-RGB+D 120, NTU-RGB+D 60 and UAV-Human, demonstrate the effectiveness and\ngeneralizability of the proposed method. Specifically, the ACFL significantly\nimproves various GCN models (i.e., CTR-GCN, MS-G3D, and Shift-GCN), achieving a\nnew record for skeleton-based action recognition.\n", "date": "2022-06-30", "categories": []}
{"id": "2206.14344", "title": "A New Adjacency Matrix Configuration in GCN-based Models for\n  Skeleton-based Action Recognition", "abstract": "  Human skeleton data has received increasing attention in action recognition\ndue to its background robustness and high efficiency. In skeleton-based action\nrecognition, graph convolutional network (GCN) has become the mainstream\nmethod. This paper analyzes the fundamental factor for GCN-based models -- the\nadjacency matrix. We notice that most GCN-based methods conduct their adjacency\nmatrix based on the human natural skeleton structure. Based on our former work\nand analysis, we propose that the human natural skeleton structure adjacency\nmatrix is not proper for skeleton-based action recognition. We propose a new\nadjacency matrix that abandons all rigid neighbor connections but lets the\nmodel adaptively learn the relationships of joints. We conduct extensive\nexperiments and analysis with a validation model on two skeleton-based action\nrecognition datasets (NTURGBD60 and FineGYM). Comprehensive experimental\nresults and analysis reveals that 1) the most widely used human natural\nskeleton structure adjacency matrix is unsuitable in skeleton-based action\nrecognition; 2) The proposed adjacency matrix is superior in model performance,\nnoise robustness and transferability.\n", "date": "2022-06-29", "categories": []}
{"id": "2206.13028", "title": "Multi-Scale Spatial Temporal Graph Convolutional Network for\n  Skeleton-Based Action Recognition", "abstract": "  Graph convolutional networks have been widely used for skeleton-based action\nrecognition due to their excellent modeling ability of non-Euclidean data. As\nthe graph convolution is a local operation, it can only utilize the short-range\njoint dependencies and short-term trajectory but fails to directly model the\ndistant joints relations and long-range temporal information that are vital to\ndistinguishing various actions. To solve this problem, we present a multi-scale\nspatial graph convolution (MS-GC) module and a multi-scale temporal graph\nconvolution (MT-GC) module to enrich the receptive field of the model in\nspatial and temporal dimensions. Concretely, the MS-GC and MT-GC modules\ndecompose the corresponding local graph convolution into a set of sub-graph\nconvolution, forming a hierarchical residual architecture. Without introducing\nadditional parameters, the features will be processed with a series of\nsub-graph convolutions, and each node could complete multiple spatial and\ntemporal aggregations with its neighborhoods. The final equivalent receptive\nfield is accordingly enlarged, which is capable of capturing both short- and\nlong-range dependencies in spatial and temporal domains. By coupling these two\nmodules as a basic block, we further propose a multi-scale spatial temporal\ngraph convolutional network (MST-GCN), which stacks multiple blocks to learn\neffective motion representations for action recognition. The proposed MST-GCN\nachieves remarkable performance on three challenging benchmark datasets, NTU\nRGB+D, NTU-120 RGB+D and Kinetics-Skeleton, for skeleton-based action\nrecognition.\n", "date": "2022-06-27", "categories": []}
{"id": "2206.11493", "title": "Learning to Refactor Action and Co-occurrence Features for Temporal\n  Action Localization", "abstract": "  The main challenge of Temporal Action Localization is to retrieve subtle\nhuman actions from various co-occurring ingredients, e.g., context and\nbackground, in an untrimmed video. While prior approaches have achieved\nsubstantial progress through devising advanced action detectors, they still\nsuffer from these co-occurring ingredients which often dominate the actual\naction content in videos. In this paper, we explore two orthogonal but\ncomplementary aspects of a video snippet, i.e., the action features and the\nco-occurrence features. Especially, we develop a novel auxiliary task by\ndecoupling these two types of features within a video snippet and recombining\nthem to generate a new feature representation with more salient action\ninformation for accurate action localization. We term our method RefactorNet,\nwhich first explicitly factorizes the action content and regularizes its\nco-occurrence features, and then synthesizes a new action-dominated video\nrepresentation. Extensive experimental results and ablation studies on THUMOS14\nand ActivityNet v1.3 demonstrate that our new representation, combined with a\nsimple action detector, can significantly improve the action localization\nperformance.\n", "date": "2022-06-23", "categories": []}
{"id": "2206.08632", "title": "Learning Using Privileged Information for Zero-Shot Action Recognition", "abstract": "  Zero-Shot Action Recognition (ZSAR) aims to recognize video actions that have\nnever been seen during training. Most existing methods assume a shared semantic\nspace between seen and unseen actions and intend to directly learn a mapping\nfrom a visual space to the semantic space. This approach has been challenged by\nthe semantic gap between the visual space and semantic space. This paper\npresents a novel method that uses object semantics as privileged information to\nnarrow the semantic gap and, hence, effectively, assist the learning. In\nparticular, a simple hallucination network is proposed to implicitly extract\nobject semantics during testing without explicitly extracting objects and a\ncross-attention module is developed to augment visual feature with the object\nsemantics. Experiments on the Olympic Sports, HMDB51 and UCF101 datasets have\nshown that the proposed method outperforms the state-of-the-art methods by a\nlarge margin.\n", "date": "2022-06-17", "categories": []}
{"id": "2206.10869", "title": "NVIDIA-UNIBZ Submission for EPIC-KITCHENS-100 Action Anticipation\n  Challenge 2022", "abstract": "  In this report, we describe the technical details of our submission for the\nEPIC-Kitchen-100 action anticipation challenge. Our modelings, the higher-order\nrecurrent space-time transformer and the message-passing neural network with\nedge learning, are both recurrent-based architectures which observe only 2.5\nseconds inference context to form the action anticipation prediction. By\naveraging the prediction scores from a set of models compiled with our proposed\ntraining pipeline, we achieved strong performance on the test set, which is\n19.61% overall mean top-5 recall, recorded as second place on the public\nleaderboard.\n", "date": "2022-06-22", "categories": []}
{"id": "2206.09082", "title": "Context-aware Proposal Network for Temporal Action Detection", "abstract": "  This technical report presents our first place winning solution for temporal\naction detection task in CVPR-2022 AcitivityNet Challenge. The task aims to\nlocalize temporal boundaries of action instances with specific classes in long\nuntrimmed videos. Recent mainstream attempts are based on dense boundary\nmatchings and enumerate all possible combinations to produce proposals. We\nargue that the generated proposals contain rich contextual information, which\nmay benefits detection confidence prediction. To this end, our method mainly\nconsists of the following three steps: 1) action classification and feature\nextraction by Slowfast, CSN, TimeSformer, TSP, I3D-flow, VGGish-audio, TPN and\nViViT; 2) proposal generation. Our proposed Context-aware Proposal Network\n(CPN) builds on top of BMN, GTAD and PRN to aggregate contextual information by\nrandomly masking some proposal features. 3) action detection. The final\ndetection prediction is calculated by assigning the proposals with\ncorresponding video-level classifcation results. Finally, we ensemble the\nresults under different feature combination settings and achieve 45.8%\nperformance on the test set, which improves the champion result in CVPR-2021\nActivityNet Challenge by 1.1% in terms of average mAP.\n", "date": "2022-06-18", "categories": []}
{"id": "2206.09852", "title": "M&M Mix: A Multimodal Multiview Transformer Ensemble", "abstract": "  This report describes the approach behind our winning solution to the 2022\nEpic-Kitchens Action Recognition Challenge. Our approach builds upon our recent\nwork, Multiview Transformer for Video Recognition (MTV), and adapts it to\nmultimodal inputs. Our final submission consists of an ensemble of Multimodal\nMTV (M&M) models varying backbone sizes and input modalities. Our approach\nachieved 52.8% Top-1 accuracy on the test set in action classes, which is 4.1%\nhigher than last year's winning entry.\n", "date": "2022-06-20", "categories": []}
{"id": "2206.09867", "title": "WiFi-based Spatiotemporal Human Action Perception", "abstract": "  WiFi-based sensing for human activity recognition (HAR) has recently become a\nhot topic as it brings great benefits when compared with video-based HAR, such\nas eliminating the demands of line-of-sight (LOS) and preserving privacy.\nMaking the WiFi signals to 'see' the action, however, is quite coarse and thus\nstill in its infancy. An end-to-end spatiotemporal WiFi signal neural network\n(STWNN) is proposed to enable WiFi-only sensing in both line-of-sight and\nnon-line-of-sight scenarios. Especially, the 3D convolution module is able to\nexplore the spatiotemporal continuity of WiFi signals, and the feature\nself-attention module can explicitly maintain dominant features. In addition, a\nnovel 3D representation for WiFi signals is designed to preserve multi-scale\nspatiotemporal information. Furthermore, a small wireless-vision dataset (WVAR)\nis synchronously collected to extend the potential of STWNN to 'see' through\nocclusions. Quantitative and qualitative results on WVAR and the other three\npublic benchmark datasets demonstrate the effectiveness of our approach on both\naccuracy and shift consistency.\n", "date": "2022-06-20", "categories": []}
{"id": "2206.10080", "title": "One-stage Action Detection Transformer", "abstract": "  In this work, we introduce our solution to the EPIC-KITCHENS-100 2022 Action\nDetection challenge. One-stage Action Detection Transformer (OADT) is proposed\nto model the temporal connection of video segments. With the help of OADT, both\nthe category and time boundary can be recognized simultaneously. After\nensembling multiple OADT models trained from different features, our model can\nreach 21.28\\% action mAP and ranks the 1st on the test-set of the Action\ndetection challenge.\n", "date": "2022-06-21", "categories": []}
{"id": "2206.10095", "title": "Pyramid Region-based Slot Attention Network for Temporal Action Proposal\n  Generation", "abstract": "  It has been found that temporal action proposal generation, which aims to\ndiscover the temporal action instances within the range of the start and end\nframes in the untrimmed videos, can largely benefit from proper temporal and\nsemantic context exploitation. The latest efforts were dedicated to considering\nthe temporal context and similarity-based semantic contexts through\nself-attention modules. However, they still suffer from cluttered background\ninformation and limited contextual feature learning. In this paper, we propose\na novel Pyramid Region-based Slot Attention (PRSlot) module to address these\nissues. Instead of using the similarity computation, our PRSlot module directly\nlearns the local relations in an encoder-decoder manner and generates the\nrepresentation of a local region enhanced based on the attention over input\nfeatures called \\textit{slot}. Specifically, upon the input snippet-level\nfeatures, PRSlot module takes the target snippet as \\textit{query}, its\nsurrounding region as \\textit{key} and then generates slot representations for\neach \\textit{query-key} slot by aggregating the local snippet context with a\nparallel pyramid strategy. Based on PRSlot modules, we present a novel Pyramid\nRegion-based Slot Attention Network termed PRSA-Net to learn a unified visual\nrepresentation with rich temporal and semantic context for better proposal\ngeneration. Extensive experiments are conducted on two widely adopted THUMOS14\nand ActivityNet-1.3 benchmarks. Our PRSA-Net outperforms other state-of-the-art\nmethods. In particular, we improve the AR@100 from the previous best 50.67% to\n56.12% for proposal generation and raise the mAP under 0.5 tIoU from 51.9\\% to\n58.7\\% for action detection on THUMOS14. \\textit{Code is available at}\n\\url{https://github.com/handhand123/PRSA-Net}\n", "date": "2022-06-21", "categories": []}
{"id": "2206.06629", "title": "Semantic-Discriminative Mixup for Generalizable Sensor-based\n  Cross-domain Activity Recognition", "abstract": "  It is expensive and time-consuming to collect sufficient labeled data to\nbuild human activity recognition (HAR) models. Training on existing data often\nmakes the model biased towards the distribution of the training data, thus the\nmodel might perform terribly on test data with different distributions.\nAlthough existing efforts on transfer learning and domain adaptation try to\nsolve the above problem, they still need access to unlabeled data on the target\ndomain, which may not be possible in real scenarios. Few works pay attention to\ntraining a model that can generalize well to unseen target domains for HAR. In\nthis paper, we propose a novel method called Semantic-Discriminative Mixup\n(SDMix) for generalizable cross-domain HAR. Firstly, we introduce\nsemantic-aware Mixup that considers the activity semantic ranges to overcome\nthe semantic inconsistency brought by domain differences. Secondly, we\nintroduce the large margin loss to enhance the discrimination of Mixup to\nprevent misclassification brought by noisy virtual labels. Comprehensive\ngeneralization experiments on five public datasets demonstrate that our SDMix\nsubstantially outperforms the state-of-the-art approaches with 6% average\naccuracy improvement on cross-person, cross-dataset, and cross-position HAR.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.06931", "title": "Stand-Alone Inter-Frame Attention in Video Models", "abstract": "  Motion, as the uniqueness of a video, has been critical to the development of\nvideo understanding models. Modern deep learning models leverage motion by\neither executing spatio-temporal 3D convolutions, factorizing 3D convolutions\ninto spatial and temporal convolutions separately, or computing self-attention\nalong temporal dimension. The implicit assumption behind such successes is that\nthe feature maps across consecutive frames can be nicely aggregated.\nNevertheless, the assumption may not always hold especially for the regions\nwith large deformation. In this paper, we present a new recipe of inter-frame\nattention block, namely Stand-alone Inter-Frame Attention (SIFA), that novelly\ndelves into the deformation across frames to estimate local self-attention on\neach spatial location. Technically, SIFA remoulds the deformable design via\nre-scaling the offset predictions by the difference between two frames. Taking\neach spatial location in the current frame as the query, the locally deformable\nneighbors in the next frame are regarded as the keys/values. Then, SIFA\nmeasures the similarity between query and keys as stand-alone attention to\nweighted average the values for temporal aggregation. We further plug SIFA\nblock into ConvNets and Vision Transformer, respectively, to devise SIFA-Net\nand SIFA-Transformer. Extensive experiments conducted on four video datasets\ndemonstrate the superiority of SIFA-Net and SIFA-Transformer as stronger\nbackbones. More remarkably, SIFA-Transformer achieves an accuracy of 83.1% on\nKinetics-400 dataset. Source code is available at\n\\url{https://github.com/FuchenUSTC/SIFA}.\n", "date": "2022-06-14", "categories": []}
{"id": "2206.06292", "title": "MLP-3D: A MLP-like 3D Architecture with Grouped Time Mixing", "abstract": "  Convolutional Neural Networks (CNNs) have been regarded as the go-to models\nfor visual recognition. More recently, convolution-free networks, based on\nmulti-head self-attention (MSA) or multi-layer perceptrons (MLPs), become more\nand more popular. Nevertheless, it is not trivial when utilizing these\nnewly-minted networks for video recognition due to the large variations and\ncomplexities in video data. In this paper, we present MLP-3D networks, a novel\nMLP-like 3D architecture for video recognition. Specifically, the architecture\nconsists of MLP-3D blocks, where each block contains one MLP applied across\ntokens (i.e., token-mixing MLP) and one MLP applied independently to each token\n(i.e., channel MLP). By deriving the novel grouped time mixing (GTM)\noperations, we equip the basic token-mixing MLP with the ability of temporal\nmodeling. GTM divides the input tokens into several temporal groups and\nlinearly maps the tokens in each group with the shared projection matrix.\nFurthermore, we devise several variants of GTM with different grouping\nstrategies, and compose each variant in different blocks of MLP-3D network by\ngreedy architecture search. Without the dependence on convolutions or attention\nmechanisms, our MLP-3D networks achieves 68.5\\%/81.4\\% top-1 accuracy on\nSomething-Something V2 and Kinetics-400 datasets, respectively. Despite with\nfewer computations, the results are comparable to state-of-the-art widely-used\n3D CNNs and video transformers. Source code is available at\nhttps://github.com/ZhaofanQiu/MLP-3D.\n", "date": "2022-06-13", "categories": []}
{"id": "2206.04668", "title": "GateHUB: Gated History Unit with Background Suppression for Online\n  Action Detection", "abstract": "  Online action detection is the task of predicting the action as soon as it\nhappens in a streaming video. A major challenge is that the model does not have\naccess to the future and has to solely rely on the history, i.e., the frames\nobserved so far, to make predictions. It is therefore important to accentuate\nparts of the history that are more informative to the prediction of the current\nframe. We present GateHUB, Gated History Unit with Background Suppression, that\ncomprises a novel position-guided gated cross-attention mechanism to enhance or\nsuppress parts of the history as per how informative they are for current frame\nprediction. GateHUB further proposes Future-augmented History (FaH) to make\nhistory features more informative by using subsequently observed frames when\navailable. In a single unified framework, GateHUB integrates the transformer's\nability of long-range temporal modeling and the recurrent model's capacity to\nselectively encode relevant information. GateHUB also introduces a background\nsuppression objective to further mitigate false positive background frames that\nclosely resemble the action frames. Extensive validation on three benchmark\ndatasets, THUMOS, TVSeries, and HDD, demonstrates that GateHUB significantly\noutperforms all existing methods and is also more efficient than the existing\nbest work. Furthermore, a flow-free version of GateHUB is able to achieve\nhigher or close accuracy at 2.8x higher frame rate compared to all existing\nmethods that require both RGB and optical flow information for prediction.\n", "date": "2022-06-09", "categories": []}
{"id": "2206.01910", "title": "The Spike Gating Flow: A Hierarchical Structure Based Spiking Neural\n  Network for Online Gesture Recognition", "abstract": "  Action recognition is an exciting research avenue for artificial intelligence\nsince it may be a game changer in the emerging industrial fields such as\nrobotic visions and automobiles. However, current deep learning faces major\nchallenges for such applications because of the huge computational cost and the\ninefficient learning. Hence, we develop a novel brain-inspired Spiking Neural\nNetwork (SNN) based system titled Spiking Gating Flow (SGF) for online action\nlearning. The developed system consists of multiple SGF units which assembled\nin a hierarchical manner. A single SGF unit involves three layers: a feature\nextraction layer, an event-driven layer and a histogram-based training layer.\nTo demonstrate the developed system capabilities, we employ a standard Dynamic\nVision Sensor (DVS) gesture classification as a benchmark. The results indicate\nthat we can achieve 87.5% accuracy which is comparable with Deep Learning (DL),\nbut at smaller training/inference data number ratio 1.5:1. And only a single\ntraining epoch is required during the learning process. Meanwhile, to the best\nof our knowledge, this is the highest accuracy among the non-backpropagation\nalgorithm based SNNs. At last, we conclude the few-shot learning paradigm of\nthe developed network: 1) a hierarchical structure-based network design\ninvolves human prior knowledge; 2) SNNs for content based global dynamic\nfeature detection.\n", "date": "2022-06-04", "categories": []}
{"id": "2206.02203", "title": "3D Convolutional with Attention for Action Recognition", "abstract": "  Human action recognition is one of the challenging tasks in computer vision.\nThe current action recognition methods use computationally expensive models for\nlearning spatio-temporal dependencies of the action. Models utilizing RGB\nchannels and optical flow separately, models using a two-stream fusion\ntechnique, and models consisting of both convolutional neural network (CNN) and\nlong-short term memory (LSTM) network are few examples of such complex models.\nMoreover, fine-tuning such complex models is computationally expensive as well.\nThis paper proposes a deep neural network architecture for learning such\ndependencies consisting of a 3D convolutional layer, fully connected (FC)\nlayers, and attention layer, which is simpler to implement and gives a\ncompetitive performance on the UCF-101 dataset. The proposed method first\nlearns spatial and temporal features of actions through 3D-CNN, and then the\nattention mechanism helps the model to locate attention to essential features\nfor recognition.\n", "date": "2022-06-05", "categories": []}
{"id": "2206.01038", "title": "A Survey on Video Action Recognition in Sports: Datasets, Methods and\n  Applications", "abstract": "  To understand human behaviors, action recognition based on videos is a common\napproach. Compared with image-based action recognition, videos provide much\nmore information. Reducing the ambiguity of actions and in the last decade,\nmany works focused on datasets, novel models and learning approaches have\nimproved video action recognition to a higher level. However, there are\nchallenges and unsolved problems, in particular in sports analytics where data\ncollection and labeling are more sophisticated, requiring sport professionals\nto annotate data. In addition, the actions could be extremely fast and it\nbecomes difficult to recognize them. Moreover, in team sports like football and\nbasketball, one action could involve multiple players, and to correctly\nrecognize them, we need to analyse all players, which is relatively\ncomplicated. In this paper, we present a survey on video action recognition for\nsports analytics. We introduce more than ten types of sports, including team\nsports, such as football, basketball, volleyball, hockey and individual sports,\nsuch as figure skating, gymnastics, table tennis, tennis, diving and badminton.\nThen we compare numerous existing frameworks for sports analysis to present\nstatus quo of video action recognition in both team sports and individual\nsports. Finally, we discuss the challenges and unsolved problems in this area\nand to facilitate sports analytics, we develop a toolbox using PaddlePaddle,\nwhich supports football, basketball, table tennis and figure skating action\nrecognition.\n", "date": "2022-06-02", "categories": []}
{"id": "2606.21358", "title": "LEViL: Label-Efficient Video Learning via Zero-Shot Distillation over VLM-Generated Pseudo-Label Spaces", "abstract": "Supervised video pretraining is a common transfer learning practice for improving downstream action recognition performance. However, it requires large-scale labeled source datasets, and the effectiveness of the learned initialization is influenced by the similarity between the source and target domains. Constructing such labeled pretraining datasets for different target domains is costly and difficult to scale. To address these limitations, this study proposes a label-efficient video learning framework that combines annotation-free video pretraining with target-label-set-aware fine-tuning. During pretraining, a vision-language model (VLM) generates textual descriptions of unlabeled videos, which are processed to construct an interpretable semantic pseudo-label space. A frozen video-language model then produces zero-shot soft target distributions over this space, allowing a student video encoder to learn semantically rich representations without manual source annotations. During downstream adaptation, target-label-set-aware fine-tuning combines supervised learning from labeled target videos with zero-shot distillation over the actual target label set, helping preserve VLM-derived semantic guidance while adapting the pretrained encoder to the target task. Experiments on UCF101 and HMDB51 show that the proposed framework outperforms the compared semi-supervised video action recognition methods across all evaluated limited-label regimes. Moreover, the annotation-free pretraining stage learns transferable representations that provide an effective initialization for full-data fine-tuning, despite relying on a comparatively modest unlabeled pretraining pool.", "date": "2026-06-19", "categories": []}
{"id": "2606.21607", "title": "T-MOR: Learning Motion-Aware Skeleton Representations for Human Action Recognition", "abstract": "Vision-language models such as CLIP have recently achieved strong performance on a wide range of visual understanding tasks. However, most existing models rely primarily on appearance-level supervision from images or videos, and do not explicitly model human motion, which is essential for fine-grained and human-centric action recognition task as actions are defined by temporally structured and physically grounded body movements. To address this problem, we propose Transferable skeleton MOtion Representation (T-MOR), a motion-aware framework that learns transferable action representations from skeleton sequences with the aid of video and language supervision during training. T-MOR adopts a multi-modal contrastive learning scheme that aligns skeleton motion with visual and textual representations, while performing inference using only lightweight skeleton inputs. To support large-scale pre-training, we construct PoseCap-1M, a new dataset that contains over one million synchronized video, skeleton, and text triplets covering diverse human activities. We evaluate T-MOR on a range of human-centric action recognition benchmarks, including action classification and frame-wise temporal detection. Experimental results show that T-MOR consistently improves performance across multiple datasets, such as Toyota Smarthome, Penn Action, UAV-Human, TSU, and Charades. In addition, T-MOR demonstrates strong generalization ability in few-shot and zero-shot settings, highlighting the effectiveness of motion-centric and embodied representations for transferable action understanding.", "date": "2026-06-19", "categories": []}
{"id": "2606.28377", "title": "Memory-Augmented LSTM Autoencoder for Unsupervised Activity Recognition with IMU Sensor Fusion", "abstract": "HAR using Inertial Measurement Unit (IMU) sensors is vital for healthcare monitoring and rehabilitation. Despite deep learning advancements, major challenges remain: reliance on labeled data, multi-sensor fusion complexity, and the limited ability of unsupervised methods to capture spatiotemporal dependencies. These issues are pronounced in real-world scenarios with noisy data, overlapping activities, and missing labels. We propose a fully unsupervised spatiotemporal feature fusion framework using a memory-augmented autoencoder. It enhances activity representations via short temporal windows of multi-sensor IMU data, enabling real-time applications. Our framework extracts hierarchical static features via a Stacked Autoencoder, fusing them within and across sensors. A sequence-to-sequence LSTM Autoencoder then temporally refines these features, incorporating historical motion patterns without labels. We analyze key hyperparameters to identify configurations that maximize feature separability under short-window constraints. Evaluated on DaLiAc and PAMAP2 using realistic inter-class window segmentation, our method achieves 96.6% and 98.4% accuracy, respectively, surpassing supervised baselines and unsupervised approaches. Our method improves feature separability by up to 9% despite shorter temporal windows. While our realistic inter-class segmentation reduces accuracy by ~7%, it was intentionally adopted to better reflect real-world activity transitions and practical relevance.", "date": "2026-06-19", "categories": []}
{"id": "2606.22416", "title": "Gen2Balance: Generative Balancing for Long-Tailed Video Action Recognition", "abstract": "We address the problem of training on long-tailed data for video action recognition. We propose to augment the training set using a text-to-video generative model, conditioned on diverse text prompts grounded in action profiles and training exemplars. Our approach, called Gen2Balance, converts an imbalanced training set into a balanced combination of real and generated video clips. To effectively learn from such data, we employ a two-stage training strategy that mitigates domain shift and yields significant improvements. We evaluate on long-tailed versions of standard benchmarks: UCF-101 (UCF-LT) and a 100-class subset of Kinetics (K100-LT) selected to prioritise temporally challenging actions. Gen2Balance improves accuracy over the strongest baselines for long-tailed learning by 5.1% and 7.0% on the respective datasets. On rare actions from the RareAct dataset (e.g., cut keyboard), Gen2Balance improves accuracy by 31.9%, demonstrating effectiveness for scarce actions. By varying the amount of synthetic data added, we show that partial balancing already achieves 79% of the performance gains at 27% of the compute cost on K100-LT, highlighting the practical scalability of Gen2Balance.", "date": "2026-06-21", "categories": []}
{"id": "2606.23118", "title": "LUMINA-26: Low-Light Understanding for Modeling and Interpreting Night-time Actions", "abstract": "Low-light human action recognition remains a challenging problem due to poor illumination, amplified noise, motion ambiguity, and diverse real-world scenes. Existing low-light datasets often lack sufficient action diversity, capture realism, or balanced class distribution, limiting the development of robust models. To address this, we introduce LUMINA-26: Low-Light Understanding for Modeling and Interpreting Night-time Actions, comprising 6,784 clips across 26 action classes, recorded from 22 subjects across 20 indoor and outdoor locations under naturally occurring low-light conditions. We also propose Illumi-Net: An Illumination-Adaptive Mixture-of-Experts Network, which leverages video-level illumination cues to guide adaptive enhancement and transformer-based spatio-temporal feature extraction, with expert-conditioned decision fusion. Our method surpasses previous state-of-the-art performance on ELLAR (Top-1: 55.13%, Top-5: 78.87%) and establishes a strong baseline on LUMINA-26 (Top-1: 75.95%, Top-5: 93.58%), offering a practical benchmark for future low-light action recognition research.", "date": "2026-06-22", "categories": []}
{"id": "2606.28389", "title": "SoccerNet 2026 Player-Centric Ball Action Spotting: Per-Player Attention with Agreement-Based Ensembling", "abstract": "We present our submission to the SoccerNet 2026 Player-Centric Ball Action Spotting challenge, which uses a two-stage pipeline: a Track-Aware Action Detector (TAAD) produces per-player action logits from broadcast video, and a Denoising Sequence Transduction (DST) transformer converts game-state features and TAAD logits into structured event sequences. We improve the TAAD with a temporal transformer that adds cross-frame context, alongside several training fixes. For the DST stage, we introduce a two-stage per-player attention mechanism operating on game-state features, and show that a spatial-first attention ordering (cross-player attention before temporal attention) improves validation Macro-F1 by 1.87%. To exploit architectural diversity, we train four model variants and combine them with a Weighted Event Fusion ensemble that applies agreement filtering to suppress single-model false positives while preserving recall, plus a dedicated exception for the rare tackle class. Our final system improves the challenge Macro-F1 from a baseline of 48.6 to 58.94.", "date": "2026-06-23", "categories": []}
{"id": "2606.25298", "title": "KidRisk: Benchmark Dataset for Children Dangerous Action Recognition", "abstract": "Children are naturally energetic, and during their spontaneous activities, they often encounter potentially dangerous situations, especially when lacking parental supervision. Identifying actions that pose risks plays a crucial role in ensuring their safety. This paper build a novel challenging dataset, namely KidRisk, including 2,500 short videos of children's actions and 10,000 images for dangerous action of children. We also introduce a benchmark on our newly constructs dataset and find that traditional deep learning models demonstrated limited effectiveness on these tasks. Therefore, we develop vision-language based baselines with exceptional context understanding of visual information. Our proposed methods achieved an accuracy of 83.53% in classifying children's actions and 96.14% in recognizing children's dangerous actions, significantly outperforming traditional approaches. These results confirm that vision-language models are not only feasible but also highly effective in detecting hazardous actions, contributing positively to safeguarding children's safety.", "date": "2026-06-24", "categories": []}
{"id": "2606.25317", "title": "ESTANet: Efficient Online Error Detection in Procedural Videos via Prediction Inconsistency", "abstract": "An efficient and accurate system for detecting errors in procedural tasks is crucial for supporting human needs in daily life, as it can provide instant notifications and guide people to correct mistakes. In this work, we study real-time online error detection in procedural videos from a simple but overlooked perspective: the prediction behavior of action detectors themselves. Instead of designing complex architectures or specialized supervision, we observe that action detectors naturally exhibit different prediction characteristics depending on their sensitivity to input dynamics and temporal context. We therefore propose ESTANet (Error-Sensitive and Temporally-vArying Network), a lightweight framework that detects errors by exploiting inconsistencies among action predictions produced by a small set of action detectors. We construct standard and error-sensitive action detectors that behave similarly on correct executions but respond differently when errors occur. Meanwhile, detectors operating with different temporal contexts further amplify prediction inconsistencies when the procedure deviates from the intended sequence. During inference, we detect errors by aggregating mismatches between standard and error-sensitive predictions through majority voting to flag frames that contain errors. Extensive experiments on EgoPER, Assembly-101-O, and EPIC-Tent-O demonstrate that ESTANet achieves state-of-the-art performance in online error detection while maintaining real-time efficiency with a lightweight architecture. Our results highlight that leveraging the intrinsic properties of action detectors can yield a powerful and practical solution for online error detection without increasing architectural design complexity.", "date": "2026-06-24", "categories": []}
{"id": "2606.25478", "title": "TACO: Towards Task-Consistent Open-Vocabulary Adaptation in Video Recognition", "abstract": "Adapting CLIP for open-vocabulary video recognition necessitates a delicate balance between newly acquired video knowledge and the pretrained generalization. While existing studies pursue this generalization-specialization trade-off with additional regularizations or constraints, we argue that they overlook the deviation of representations beyond the fine-tuning data distribution, resulting in suboptimal adaptation effects. We believe such deviation is inherited from the inconsistency between the fine-tuning and evaluation objectives, where model optimization is restricted to the known training distribution but evaluated on unseen ones. In this paper, we introduce \\emph{TACO}, a simple yet effective framework to mitigate the potential negative effects induced by this inconsistency. Our key insight is that adaptation should preserve OOD-relevant alignment beyond the training distribution. To this end, we propose \\emph{Relative Structure Distillation}, which regularizes the relative geometry of the representation space and suppresses harmful alignment shift during training. We further decouple the representation space from the optimization space with a lightweight specialization projection, allowing task-specific adaptation without directly overspecializing the representations used at test time. \\emph{TACO} establishes state-of-the-art performance on diverse benchmarks under cross-dataset and base-to-novel settings. Code will be released at https://github.com/ZMHH-H/TACO.", "date": "2026-06-24", "categories": []}
{"id": "2606.27886", "title": "A Comparison of Fusion Techniques for Multi-Modal Human Activity Recognition on the HARMES Dataset", "abstract": "Recent advances in Human Activity Recognition (HAR) from wearable sensors have shown that multi-modal deep learning models consistently outperform their uni-modal counterparts. Modalities can include IMUs, RGB cameras, audio signals, and others. One important aspect of multi-modal deep learning is the sensor fusion approach we apply. Over recent years, multiple fusion paradigms have been proposed for multi-modal HAR. However, to the best of our knowledge, no head-to-head comparison of these paradigms exists on a common multi-modal HAR benchmark dataset. To address this research gap, we systematically compare seven state-of-the-art sensor fusion methods on the recently released HARMES dataset, which comprises 61 hours of fully labeled IMU, audio, and ambient humidity data. The chosen dataset focuses on 15 household and personal hygiene activities of daily living (ADLs). By applying the seven different fusion techniques to a state-of-the-art multi-modal model architecture, we show that Gated Multi-modal Fusion achieves the highest macro F1-score (0.82), surpassing the concatenation-based late fusion HARMES paper baseline of 0.76 by +6pp under leave-one-participant-out evaluation. All code used in our experiments is made publicly available on GitHub.", "date": "2026-06-26", "categories": []}
{"id": "2606.31127", "title": "SkillSpotter: Pose-Aware Multi-View Skilled Action Detection and Grading in Ego-Exo Videos", "abstract": "To enable personalized, real-time coaching using Augmented Reality glasses or fixed camera setups in domains such as sports, cooking, or music, a system must understand not just what a person does, but how well they execute an activity. In an ego-exo video setting, this requires simultaneously detecting individual skilled actions and classifying each as correct or needing improvement, which Ego-Exo4D's proficiency demonstration benchmark formalized. We first adapt seven state-of-the-art temporal action detection architectures to this task, extend the evaluation protocol to disentangle detection from grading, and show that existing methods grade near-randomly. We then introduce SkillSpotter, a pose-aware multi-view architecture that jointly detects and grades skilled actions through three task-specific modules: (1) adaptive temporal suppression to handle the varying density of skilled actions across diverse activities, (2) gated 3D body pose fusion to leverage body kinematics as a complementary signal to visual features, and (3) bidirectional cross-view attention to combine ego and exo views effectively. SkillSpotter improves class-specific mAP from 12.40 to 21.82 (+76%) and balanced accuracy from 55.99% to 60.40% over the best baseline. SkillSpotter's modules transfer to other temporal action detection models with consistent gains, and our method generalizes beyond Ego-Exo4D to HoloAssist. Code: https://github.com/eth-siplab/SkillSpotter", "date": "2026-06-30", "categories": []}