Unsupervised action segmentation is a challenging task. It involves finding action categories and boundaries in videos without labels. Existing Optimal Transport (OT) methods use global constraints. This causes them to overlook the use of local information. Furthermore, existing Optimal transport architectures are prone to confirmation bias because they overly trust the pseudo-labels they generate. This causes models to learn from noise in the early training stages. To address these issues, we propose FIS-OT. It is a novel Feature-Induced Structured Optimal Transport framework. First, we introduce a Feature Enhanced Generator (FEG) module. It serves as an internal regularizer. By using triplet loss, FEG captures local consistency. It provides robust supervision that is independent of noisy pseudo-labels. Second, we propose a Feature-Induced Residual Structural Prior. This combines a fixed temporal backbone with dynamic feature similarities. This design ensures temporal continuity. It also allows the solver to adapt to complex action structures. Finally, we establish a cyclic optimization loop. This aligns local feature learning with global structural alignment. Extensive experiments on the three datasets show the effectiveness of our method.
Tingyu Lin, Christian Stippel, Armin Dadras +4cs.CV
Shot boundary detection (SBD) is widely treated as the localisation of local visual discontinuities, yet many false positives such as hand-held shake, illumination flicker, motion blur, occlusion, and damaged archival material produce equally sharp local change without introducing a new shot. We reformulate SBD as boundary semantic discrimination: a frame is favoured as a boundary only when its local change evidence is accompanied by a persistent update of the video's latent temporal state, rather than a transient excursion that returns to the surrounding trend. This persistence test is operationalised with a continuous latent state from a FiLM-conditioned sinusoidal representation network and a structured discriminator that combines three semantic cues, local change, transient impulse, and return-to-trend, into a single interpretable per-frame signal over a dual-rate temporal backbone. The resulting framework, PERSIST, turns every decision into an inspectable one: the persistence criterion is trained into the classifier, its per-frame effect stays readable from the gate triple, and its learned latent state is measurably boundary-discriminative. On a 2,727-video per-subtype diagnostic it removes 33-80% of flash, text-overlay, and archival false positives relative to an identically trained cue detector, and at matched true-transition recall it roughly halves TransNetV2's pseudo-event false positives on that diagnostic and cuts its false positives on ClipShots footage by about a quarter, while preserving recall. It does so while reaching parity with the strongest public detector across online, broadcast, short-form, and historical-archive transfer evaluations, under markedly stricter training: it learns from ClipShots real transitions only, whereas the anchor draws on additional corpora whose transitions are 85% synthetic. Code is available at https://github.com/linty5/PERSIST.
Etienne Casanova, Sevan Brodjian, Pietro Peronacs.CV
Videos are expensive to analyze frame by frame, yet many video understanding tasks depend on knowing where relevant moments occur. A system may need to find when an action changes, locate the segment described by a sentence, or choose a few frames for a vision-language model. Existing methods often solve these problems separately, using task-specific training data or specialized architectures. We study whether a pretrained video-text model can provide enough temporal structure to support several of these tasks at once. We present STITCH, a training-free method that divides a video into semantically meaningful temporal chunks. STITCH embeds short video windows with a frozen video-text backbone and detects changes in the resulting embedding sequence. These chunks are computed once per video and reused across tasks. We evaluate STITCH on generic event boundary detection, language-based moment retrieval, and frame selection for long-video VLM reasoning. Across all three settings, STITCH remains competitive with more specialized methods while requiring no task-specific training, with especially clear gains when only a small number of frames or tokens can be processed. These results suggest that reusable temporal abstraction is a promising direction for general video understanding, allowing dense video streams to be converted once into semantic units that can be localized, retrieved, sampled, or reasoned over by downstream systems.
Articulated human pose provides detailed body-configuration information beyond coarse spatial relationships, but whether this detail yields greater discriminative information when the downstream pipeline is held fixed remains unclear. We examine this through early violence detection. Holding the tracker, temporal head, supervision, folds, and evaluation fixed, we compare five interaction representations spanning coarse bounding-box geometry, a matched handcrafted pose analogue, enriched pose descriptors, and a matched-capacity encoder learned from raw joints, under video-level evaluation with cluster-bootstrap intervals. No pose-based representation outperforms coarse geometry, though with fifteen anomalous videos this subset cannot rule out small effects. Extending the pipeline to frozen visual encoders, and repeating the comparison on XD-Violence (137 anomalous videos, nine times our UCF-Crime sample), person-crop appearance and whole-frame context both exceed geometry by a wide margin, yet context matches appearance on UCF-Crime and exceeds it on the larger split: cropping to the interacting people yields no advantage over encoding the whole frame. This prompts a direct test of what the benchmark measures. Scoring anomalous videos using only frames preceding the annotated onset, under a control removing sequence length as a cue, retains 39-91% of above-chance separation on both benchmarks, including for seven hand-designed geometric channels. Inspection of the tightest pre-onset windows identifies concrete provenance artifacts: editorial title cards and platform watermarks absent from the surveillance footage supplying the normal class. Video-level AUC here is thus a composite of event evidence and pre-event source cues, a shared source of discrimination that can obscure differences between representations. The diagnostic requires only annotations these benchmarks already ship.
Multi-camera systems are foundational to modern media production, and multi-camera editing is a critical task. This involves the proper selection of the appropriate camera view at each moment. In this paper, we propose a novel Dual-Transformer architecture with Cross-Attention that heavily outperformed the current SOTA models over the TVMCE dataset (TV Shows Multicamera Editing dataset). Our model decouples these tasks: (1) a dedicated temporal encoder first processes the sequence of past frames to build a rich memory of the recent history, and (2) the candidate camera views then act as queries to this memory via a cross-attention module, allowing each candidate to independently interrogate the historical context and find the most relevant information for its own evaluation. Our approach achieved 56.60% Precision@0.5, representing a substantial improvement over the prior best result of 37.16%. We further conducted an ablation study exploring the use of lightweight backbone architectures, where the SwinV2 backbone yielded the best performance, achieving 69.65% Precision@0.5. Using this best-performing configuration, we then investigated the feasibility of adapting the model to replicate the editing style of a specific human editor. To this end, we fine-tuned the model using varying proportions of the initial segment of a target video. Our results demonstrate that even with only 20% of the video used for fine-tuning, the model exhibited measurable improvements in Precision@0.5, indicating strong potential for data-efficient personalization of editing style adapted to each individual TV show or producer.
Perceived risk in driving evolves over time and may be supported by specific scene entities, yet supervision is typically limited to coarse video-level judgments. Learning \emph{when} supporting evidence emerges and \emph{which entities} support a risk predictor would ordinarily require costly temporal- and entity-level annotations. We introduce \textbf{CoRE}, a weakly supervised coarse-to-fine framework that learns fine-grained prediction support from coarse video supervision. CoRE first trains a video-level predictor and then freezes it. Structured interventions over candidate temporal regions or entity tracks measure how each candidate changes the coarse prediction, producing graded prediction-effect targets. These targets are distilled into a student that directly predicts temporal and entity support from the original video, without requiring interventions at inference. We evaluate this learning principle across three complementary settings: RISEE tests perceived-risk support from subjective clip-level judgments without temporal or entity-level risk annotations; DoTA provides independent temporal event annotations for evaluating weakly supervised traffic-anomaly localization; and UCF-Crime tests whether the same coarse-to-fine mechanism extends to a standard non-driving anomaly-detection benchmark. Across these settings, CoRE learns informative fine-grained support from coarse supervision, with strong temporal localization on DoTA and competitive performance on UCF-Crime. These results show that coarse video predictions can provide useful supervision for recovering the fine-grained evidence supporting them, without requiring corresponding fine-grained labels.
Recognizing sequential construction activities is important for collaborative human-robot work; for example, robots are able to understand workers' current and upcoming actions and provide timely tool delivery or physical support. However, despite extensive research on construction worker activity recognition, existing studies have been limited to classifying activity categories, such as climbing, lifting, and walking, instead of recognizing fine-grained activity transitions from long-horizon sequences. Addressing this problem is challenging because annotating action temporal boundaries in long construction videos is time-consuming. In this study, we propose ConsensusTAS, a label-free, self-supervised learning approach to segment continuous video streams into distinct activity phases by exploiting the internal consensus of candidate segmentations. We evaluated our algorithm on three public datasets, where it outperformed state-of-the-art methods, achieving an F1@10 of 73.08 on GTEA, an F1@10 of 64.33 on Breakfast, and an F1@50 of 33.50 on static-camera videos from Assembly101. We also tested it on real-world construction videos, where post-hoc evaluation showed that the model successfully recognized and segmented actions within the composite activity of bricklaying, such as spreading mortar on a brick, placing the brick, pressing, and aligning. Compared with other temporal action segmentation models that require computationally intensive large vision-language models, our method can run on a CPU, which provides practical value for video surveillance and human-robot collaboration on mobile robotic platforms.
Surgical action triplet recognition constitutes a critical task in context-aware robot-assisted surgery, facilitating automatic surgical action perception by identifying instrument, verb, target, and their association. However, existing works struggle to analyze such complex surgical scenes due to three main issues: (1) component-level optimization conflicts caused by entangled feature spaces, (2) category-level optimization conflicts arising from severe data imbalance, and (3) lack of domain knowledge guidance that limits model interpretability and robustness. To address these challenges, we propose a Mixture-of-Experts-guided Co-Optimization (\textit{MoeCo}) framework powered by knowledge-driven learning. Within the co-optimization pipeline, to first mitigate component-level conflicts, we introduce a component-tailored adapter that disentangles task-specific features across spatial-temporal regimes, facilitating effective component specialization. Next, we develop a coordinated gradient learning strategy to handle category-level conflicts, which adaptively rebalances positive-negative gradients to enhance the perception of rare categories. Notably, inspired by surgical domain expertise, we introduce a knowledge-driven mixture-of-experts mechanism that dynamically integrates multimodal large language model-guided knowledge via activated experts, thereby enriching the co-optimization pipeline with more expressive and robust representations. Extensive experiments on the public CholecT45 and CholecT50 datasets confirm the effectiveness of the proposed co-optimization pipeline and the superiority of dynamic priors integration via the knowledge-driven mixture-of-experts mechanism.
Machines that understand humans should perceive the present and anticipate the future. Existing human-centric vision model are pretrained on human images, set the state of the art in static dense perception, so motion and anticipation are out of reach. Here we present Human-JEPA, a human-centric vision model trained on video by anchored forecasting: dense targets are pinned to a frozen copy of the initialization, preventing a silent collapse of dense perception, and block masks are replaced by a pure past-to-future split, avoiding a five-point action tax and a seventeen-point re-identification collapse. Under frozen probes, Human-JEPA leads the pixel-anchored specialists on pose and person re-identification at 2.7 times fewer parameters, conceding high-resolution dense parsing, and its released predictor head is the first that does not degrade anticipation. A single safely adapted model thus serves both halves of understanding humans.
Visual understanding in sports has emerged as a hot topic in computer vision in recent years. Most existing basketball video datasets adopt single action or activity as sample, which can neither preserve the temporal continuity of game events nor support complex tasks such as action anticipation. To address this issue, this paper constructs the first possession-level basketball video dataset (PL-NBA), in which each sample is composed of a complete NBA offensive possession. Collected from 60 NBA games, PL-NBA contains 11,000 valid offensive possession clips and 31,567 annotated events with player names, captions, event types and timestamps. Each video clip includes multiple events and preserves the continuity of events, which is helpful for analysis of tactic. Experiment is conducted on multiple visual understanding tasks, including event recognition, video captioning, temporal action localization and action anticipation. Experimental results show that existing methods achieve limited performance on above four tasks, demonstrating that PL-NBA is a challenging benchmark for sports video understanding.
Vision encoders are a critical component of vision-language models, and scaling their capacity effectively improves performance. However, dense scaling increases compute cost and inference latency. Mixture-of-Experts (MoE) architectures offer a compelling alternative, having enabled efficient scaling in LLMs, yet the MoE design space for CLIP-style vision encoders remains underexplored at State-of-the-Art (SOTA) levels. In this work, we systematically study MoE designs for vision encoder scaling and find that fine-grained MoE topologies yield substantial gains over both dense and standard MoE counterparts. We further propose an auxiliary-loss-free balancing variant for better expert utilization, and design a specialized MoE kernel to mitigate inference latency overhead. To enhance video capabilities while preserving image knowledge, we introduce frame-level distillation paired with a novel freezing mechanism. We pretrain a series of Mixture-of-Experts Vision Encoders (MoE-ViE) across a range of sizes, all consistently outperforming their dense counterparts. Our largest model matches the zero-shot performance of a SOTA encoder 1.7x its size at 76% of its latency. When aligned with an LLM, MoE-ViE surpasses all compared encoders on image and video benchmarks, including those with up to 5x more activated parameters. Code is available at https://github.com/facebookresearch/moe_vie.
Training query-propagation end-to-end multi-object tracking (MOT) models requires dense bounding-box and identity annotations across video sequences, making dataset construction expensive. Clip-level active learning reduces this cost by selecting video clips for annotation, but prior acquisition criteria based on output-level temporal uncertainty may miss clips whose informativeness comes from association instability in propagated track states. We propose QPID (Query-Propagation Instability and Diversity), a clip acquisition method for query-propagation MOT that targets association instability in propagated track states. QPID estimates this instability by applying two-sided perturbations to internal track states and measuring prediction differences from a clean reference branch. The key idea is that, in stable clips, each propagated track should continue to follow the same target under small perturbations, whereas in ambiguous clips, small changes in the track state can alter which target the track follows, leading to changes in localization or confidence. QPID measures these perturbation-induced prediction differences with two metrics: Localization Drift and Entropy-Weighted Confidence Discrepancy. These metrics are aggregated into a clip-level association-instability score. To avoid redundant uncertainty-only selection, QPID selects a representative annotation batch from high-instability clips using Uncertainty-Weighted Visual Coverage with track-level visual prototypes. Experiments on DanceTrack and SportsMOT with MeMOTR and SambaMOTR show that QPID achieves strong performance compared with active learning baselines under the same annotation budget.
Dynamic scene graphs (DSGs) capture spatio-temporal interactions across videos as $\langle$subject, predicate, object$\rangle$ triplets, and underpin downstream tasks such as video captioning, video question answering, and action analysis. However, end-to-end dynamic scene graph generation (DSGG) methods are closed-set: they recognize only objects and predicates from a fixed training vocabulary and struggle with the long-tailed distribution of rare concepts, severely limiting their real-world applicability. Existing open-vocabulary models typically inherit pretrained large language models, resulting in multi-stage training and inference with substantial cost. We introduce OvDSGG, the first end-to-end framework for open-vocabulary DSGG. OvDSGG builds on top of an open-vocabulary Spatial Backbone and a Temporal Backbone; we further propose a Triplet Feature Extraction Module that bridges them, and a Visual-Language Alignment Module that preserves open-vocabulary recognition by learning an adaptive decision boundary in the joint visual-language feature space, without expensive knowledge distillation in existing methods. We further introduce a rigorous open-vocabulary DSGG benchmark adapted from Action Genome, with disjoint Base/Novel splits for both objects and predicates. OvDSGG significantly outperforms open-vocabulary baselines across all metrics, with zero-shot Recall@$K$ scores 10.0--20.4 percentage point higher than the next-best baseline, while on closed-set DSGG remaining competitive with state-of-the-art models. Code and benchmark are publicly available at https://github.com/jhelsby/OvDSGG/.
Convolutional Neural Networks (CNNs) capture local features efficiently but struggle with global context due to their limited receptive field. On the other hand, transformers effectively capture global dependencies through self-attention but suffer from high redundancy and computational costs. Thus, to leverage the advantages of both CNNs and transformers, we propose a unified model (UniCon-Former) that aims to provide robust and efficient performance on dynamic hand gesture recognition. The unified approach helps the model to learn both local and global features. At the beginning of each transformer stage, the convolution projections help in decreasing the dimension of the input vectors of the transformer block. This creates a pyramidal structure at each transformer stage. These features enable the UniCon-Former to reduce resource usage than vanilla transformers, making it flexible for learning multi-scale and high-resolution features, which is required in hand gesture recognition. We have performed experiments with NVGesture and Briareo datasets and achieved state-of-the-art results with fewer parameters and MACs.
Face forgery detection is crucial for preserving the security and integrity of facial data given the rapid developments in face manipulation techniques and deep generative models. Existing methods for video face forgery detection typically assume that all frames in a forged video are manipulated, while detecting partially forged videos that contain only a subset of altered frames remains challenging. To address this issue, we propose a novel framework, UVIF, that utilizes additional annotated images to provide fine-grained supervision for detecting partial forgeries in videos. UVIF employs a unified encoder and a multi-task learning paradigm to jointly model facial videos and images for boosted video face forgery detection. A 2D backbone with temporal fusion modules is employed as the unified encoder. A pseudo labeling process is designed for video frames to bridge their representations with those of static images. A video-oriented feature alignment strategy is further introduced to reduce the distribution gap between videos and images. Extensive experiments on benchmark datasets demonstrate the effectiveness of our framework, which outperforms state-of-theart methods in detecting partially forged videos while introducing no additional computational overhead. Our code is available at https://github.com/haotianll/UVIF.
Understanding camera motion is fundamental to video perception, with applications in spatial intelligence and controllable video generation. Multimodal large language models (MLLMs) provide a natural interface for this task, but existing work typically assigns one or more labels to an entire clip. Such clip-level recognition overlooks two defining properties of real camera motion: it can change within a shot, and multiple movements can occur simultaneously. We therefore formulate camera-motion understanding as temporally grounded, compositional recognition, which requires a model to localize motion-consistent intervals and identify every movement active within each interval. We introduce CamChoreo, a benchmark of 4,229 real single-shot clips with expert-annotated temporal segments. Its annotations use a compact vocabulary of 20 direction-aware labels, and nearly half of the segments contain compound camera motion, with multiple movement primitives active simultaneously. Recognizing such fine-grained, compositional motion is hard for current MLLMs, whose visual encoders emphasize semantic content rather than the geometric evidence on which camera motion depends. Directly injecting features from a frozen 3D foundation model addresses this gap, but requires running the expensive geometry model on every input; we refer to this baseline as CamInject. We instead propose CamDistill, which distills the same geometric knowledge into lightweight camera tokens during training and removes the 3D model at inference. CamDistill matches the accuracy of direct feature injection without running the 3D teacher at inference. Together, CamChoreo and CamDistill advance camera-motion understanding from clip-level labeling to temporally grounded, compositional recognition. Project page: https://ddz16.github.io/cammotion.github.io/.
Dipit Saha, Shah Mohammad Abdul Mannan, Mohammad Raihan Rashid +2cs.CV
Traffic surveillance cameras capture accidents continuously, yet converting raw CCTV footage into structured event records that pinpoint when, where, and what type of collision occurred remains unsolved at scale. The ACCIDENT @ CVPR benchmark evaluates exactly this joint prediction under a strict constraint: no labeled real-world training data is available. We introduce a training-free, two-pass coarse-to-fine pipeline that pairs a frozen Qwen3-VL-32B-Instruct vision-language model with YOLO11x object detection and BoT-SORT tracking. A first pass sparsely samples the full clip to anchor the collision moment in time; a second pass re-examines a tight window around that estimate using frames annotated with stable vehicle identities and normalized bounding-box coordinates, which gives the model both a visual overlay and an explicit numeric description of the same scene. On the official 2,027-clip real-CCTV test set, our system achieves a three-way harmonic mean score of 0.504, surpassing all organizer-published baselines including the best multi-model ensemble (0.412) by a 22% relative margin.
Masoumeh Sharafi, Muhammad Osama Zeeshan, Soufiane Belharbi +3cs.CV
Facial expression recognition (FER) in videos is challenging because models must identify subtle, temporally evolving affective states that vary across individuals. Although vision-language models provide transferable visual-semantic representations, models trained on subject-independent data often degrade under subject-specific distribution shifts at inference time. Existing test-time adaptation (TTA) methods commonly update model parameters during inference, increasing computational cost and latency. Cache-based methods avoid parameter updates, but they usually require enough target samples to form reliable class prototypes, which is difficult early in adaptation and for rarely observed classes. We introduce Energy-Based Cache Personalization (EB-CaP), a subject-based online TTA method for video FER that generates class-specific prototypes personalized to each target video. EB-CaP uses a lightweight energy-based model to sample prototypes from the current unlabeled video and populate a personalized cache online, without accumulating large amounts of target data or storing diverse source prototypes. Its energy function relies only on pretrained CLIP: similarities between the target video embedding and class text embeddings guide prototype sampling. In parallel, positive and negative caches store reliable and uncertain target embeddings. An adaptive entropy gate controls cache updates according to the evolving confidence distribution, while a diversity gate limits redundant samples. Final predictions combine cache-derived scores with the current CLIP scores. Experiments on BioVid, StressID, and BAH show that EB-CaP outperforms state-of-the-art TTA methods while maintaining low computational and memory overhead. Code is available at https://github.com/MasoumehSharafi/EB-CaP.
Optimal transport (OT) has emerged as an effective framework for unsupervised action segmentation. Yet, in existing OT-based methods, the latent action prototypes that define the OT costs are not re-estimated from the refined frame geometry. Instead, they evolve solely through gradients from the pseudo-label loss. We identify this \emph{representation--prototype inconsistency} as a central bottleneck, particularly around ambiguous transitions and for short or infrequent actions. To address this issue, we build on the recently introduced CLOT, which refines frame embeddings based on estimated segment embeddings, and further re-estimates the action prototypes from the refined frame embeddings. Specifically, we introduce a graph-constrained module that regularizes the OT-refined frame and segment representations by preserving the local neighborhood geometry of the encoder output. An action-embedding refinement step then periodically re-anchors the prototypes to this stabilized representation geometry. We study two instantiations that share the same backbone, graph module, and objective: D-CLOT updates the prototypes using $k$-means, whereas D-CLOT$_{B}$ updates them as OT barycenters weighted by the refined transport plan, yielding an assignment-aware prototype update consistent with the current transport geometry. Across five established benchmarks, both variants improve segment-level quality over CLOT, with per-video gains of up to $+12.7$ F1 and $+10.2$ mIoU (YTI) and activity-level gains of up to $+8.9$ F1 (FS-Eval). We further establish the first unsupervised action-segmentation baseline on Assembly101, a procedural and substantially more fine-grained benchmark than those commonly used in prior work. Extensive ablations and sensitivity analyses demonstrate that the two refinement mechanisms are complementary and robust.
Understanding complex surgical scenes requires recognizing multiple interdependent entities, such as instruments, actions, and targets, while maintaining their relational consistency across time. Existing surgical triplet recognition methods struggle to jointly model intra-frame label dependencies and inter-frame temporal semantics in a unified manner. To address these limitations, we propose a unified framework that integrates spatial, relational, and temporal cues for robust surgical triplet recognition. Specifically, class-specific spatial priors are first extracted through a multi-scale encoder. These priors are then refined by a Label Correlation Modeling module with multi-scale class activation map-guided relational extraction (MS-CAMRE), enabling the model to capture both static co-occurrence patterns and dynamic contextual dependencies among triplet components. Furthermore, a Bidirectional Temporal-Relational Fusion Attention (BTRFA) module harmonizes temporal and relational representations to achieve coherent temporal reasoning. We also introduce a new evaluation metric, the Triplet Consistency Error Rate (TCER), which quantitatively measures the model's ability to preserve causal and semantic consistency across triplets. Extensive experiments on the CholecT45 and ProstaTD datasets show that our method achieves state-of-the-art performance, improving AP_IVT by 5.1 percent and 7.8 percent, respectively. Moreover, according to TCER, our approach achieves relative reductions of more than 36 percent and 25 percent on the two datasets, respectively, demonstrating the effectiveness of our framework in temporal-relational co-reasoning.
Reliable physical reasoning from video requires understanding how objects move, interact, and respond to interventions. Existing vision-language models (VLMs) often struggle to interpret these dynamics and reason reliably about future and counterfactual outcomes. We introduce PhysMind, a training-free agentic framework that constructs one reusable, question-agnostic executable world per video. PhysMind recovers a temporally consistent dynamic scene through object segmentation, mesh reconstruction, and 6D pose tracking, then fits analytic continuous-time dynamics and latent physical parameters without unrolling a time-stepped simulator. Given a question, it inspects, continues, or edits the world and answers from the resulting trajectories and interactions. Relative to direct chain-of-thought (CoT) reasoning with the same VLM, PhysMind improves accuracy by 38.23 points on CLEVRER and 8.08 points on Physion++. On counterfactual questions, it exceeds the strongest evaluated VLM baseline, GPT-5.5, by 19.25 points.
Quynh Vo, Thong Nguyen, Vinh-Hien Do +2cs.CV cs.CL
We introduce Predictive State Retrieval (PSR), a task in which a model observes a short video prefix and a temporal question about an object's future state, then retrieves instances from other videos or images that depict that state. Unlike action anticipation, which predicts a label, moment retrieval, which localizes an observed event within a video, or video generation, which synthesizes pixels, PSR combines anticipation with cross-instance retrieval across multiple temporal horizons. We construct a benchmark from four datasets with graded, human-validated ground truth, difficulty tiers, and an oracle ceiling. We also propose LFTR, a lightweight retriever with frozen encoders that predicts a question- and horizon-conditioned future latent and matches it in complementary semantic and visual spaces. A ceiling decomposition reveals a clear bottleneck: the true future state is highly retrievable once specified, whereas every predictor we evaluate, including a large multimodal language model with access to the prefix frames, remains far below the oracle. Thus, forecasting rather than perception is the central learnable challenge. LFTR narrows this gap at substantially lower inference cost, and ablations attribute its gains to cross-space fusion and hard-negative training rather than latent rollout. We release the benchmark, code, and evaluation scripts.
Knowledge Distillation (KD) offers a promising yet underexplored path for compressing large action recognition models. However, existing KD methods suffer from two key limitations: 1) reliance on fixed input samples leads to suboptimal feature alignment between the frozen teacher (larger model) and the learnable student (smaller model), and 2) applying a uniform distillation strength for all channels fails to account for their varying importance in capturing distinct knowledge (e.g., motion tempo or magnitude) across training epochs. This motivates us to develop an Adaptive Sample-aware Channel-wise Dynamic (ASCD) KD approach, which operates in two stages. First, we use an adaptive sample generation module to create updated samples by incorporating semantics from sample gradients, which are derived by minimizing a feature loss weighted by channel centroid frequency differences at each layer. Meanwhile, crucial motion-related details are preserved by applying a Gaussian mask to frequency features. Second, we employ a channel-wise dynamic distillation module to train student on these generated samples, guided by sample gradients and feature frequencies. For efficiency, samples are updated periodically rather than per epoch. Extensive experiments on three video benchmarks (UCF101, Kinetics-400, Something-Something-v2) and two image datasets (CIFAR-100, ImageNet) demonstrate the state-of-the-art performance of our method. Code is available at https://github.com/mlvccn/ASCD_KD_Action.
Haofan Cao, Zhichao You, Yunkai Yang +3cs.CV cs.LG cs.MM
Tracking links observations of the same object through visual change, yet cannot by itself determine when the object is empty or filled, intact or cut. We formulate identity-conditioned state-moment retrieval: given a tracked-object history and alternative state descriptions, localize an interval in which each described state holds. Absolute image-text similarity scores descriptions independently; because every visible frame depicts the same target, shared object compatibility can obscure the state evidence needed to identify the target interval. The alternatives provide the missing reference: evidence for one state should be measured against the others. We introduce Déjà Cue, a training-free framework that turns these alternatives into a vocabulary-relative coordinate system. It subtracts their state-balanced centroid from each description, calibrates frame scores, and scans multiple durations within contiguous visible runs using a frozen encoder. On 78 VOST histories, holding the temporal scan fixed and changing only the query reference nearly doubles R@1 at tIoU 0.5 from 10.3\% to 20.5\% and raises Top-1 tIoU from 16.0\% to 21.5\%. Candidate-rank analyses show that vocabulary-relative queries rank useful intervals higher within the same candidate set. Related state descriptions can therefore serve as an object-specific, query-time coordinate system for reading frozen visual representations.
Cheng-Yao Hong, Ting-Wei Lin, Yun-Chung Lai +3cs.CV
Long-horizon event-based action understanding remains underexplored because existing datasets largely comprise short, trimmed clips, while collecting native event streams with dense temporal annotations is costly. We introduce Event ActivityNet, a large-scale simulated-event benchmark derived from human-annotated, untrimmed ActivityNet videos. It comprises 3,263 videos, 200 action classes, and 106.94 hours, with matched 5-bin and 9-bin event-voxel representations, temporal action annotations, and timestamped captions. The benchmark supports annotated-segment action recognition, auxiliary event-language alignment, and causal online temporal action localization. We generate event voxels directly from non-interpolated source videos in decoded frame order, retain per-video rational nominal or average frame-rate metadata for approximate time mapping, and use action-center reconstruction LPIPS as a soft diagnostic of retained reconstructable content. We establish baselines for adaptive event framing, prompt-caption alignment, and event-only, RGB-only, and RGB-event localization. Under a progressive nested-scale training protocol, recognition Top-1 accuracy increases from 52.25 to 66.42, while online temporal localization average mAP improves from 21.7 to 29.0. Moreover, staged Event ActivityNet pretraining followed by native-event fine-tuning consistently outperforms target-only and joint-from-scratch training across multiple supervision budgets. Event ActivityNet provides a scalable benchmark for long-horizon event modeling, although native-camera evaluation remains essential for deployment-oriented conclusions.
Player-centric ball action spotting requires temporally precise event detection together with actor attribution in crowded, partially observed multi-agent sports videos. Existing Denoising Sequence Transduction (DST) baselines treat the player-role dimension as part of a flattened frame-level representation, which weakens the inductive bias for modeling player-specific temporal evolution and inter-player interactions. To address this limitation, we propose Multi-Entity Denoising Sequence Transduction (ME-DST). ME-DST keeps the role-slot dimension throughout encoding. It uses temporal attention to model the history of each role slot, and spatial attention to exchange information across role slots at each frame. This factorized design gives the model a direct structure for separating within-player evolution from inter-player context. We also add learnable role embeddings, tracking-derived tactical features, and fused visual predictions from X3D-L and Swin3D-S. Experiments on the FOOTPASS dataset show that ME-DST reaches a Micro F1 of 0.778. This improves the strongest official TAAD+DST baseline by 10.3 percentage points. Controlled ablations show that preserving the entity axis and encoding role identity are central to this gain. These results suggest that explicit entity modeling is an effective inductive bias for player-centric sports event understanding.
Modern autonomous-driving fleets record far more video than human reviewers can inspect. This motivates the need for an automatic clip triage mechanism, to surface rare and review-worthy clips, so that driving models can be fine-tuned to better handle unideal circumstances. We test a label-free approach that scores clips by the prediction-error "novelty" of a self-supervised joint-embedding predictive architecture (JEPA); a frozen V-JEPA video encoder is paired with a lightweight predictor head to reconstruct masked clip embeddings, and clips whose embeddings are hard to predict are flagged as interesting. Evaluated under a realistic protocol that trains on one dataset and tests against footage from others, this approach appears highly effective. We show that this apparent success is actually a domain-shift consequence: on a fair benchmark drawn from a single dataset, this mechanism collapses to chance and is on par with simple no-training baselines. A lightly supervised probe on the same frozen embeddings results in almost double the average precision, indicating that the bottleneck is indeed the self-supervised objective, rather than the representation. We present this as a study for evaluating the effectiveness of self-supervised learning, where cross-dataset protocols can silently reward domain separation over novelty.
Recent advances in large vision-language models (LVLMs) have enabled long-video understanding and analysis. However, processing the large number of frames in a video incurs substantial computational overhead. Existing methods reduce LVLM inference costs by scoring frame-query relevance before inference and selecting keyframes accordingly. Nevertheless, the distribution of relevant frames varies across queries, and these methods often need to score hundreds or thousands of frames. To address this limitation, we propose CSES, a training-free semantic keyframe selector that adaptively determines the numbers of frames to score and keyframes to select. CSES estimates the prominence of the frame-query relevance profile to guide active acquisition and adapt the temporal coverage of each input. It then formulates keyframe selection as a coverage problem that jointly accounts for semantic relevance, temporal redundancy, and visual redundancy. Active acquisition and keyframe selection terminate based on coverage saturation. The selection objective is monotone and submodular, enabling greedy optimization with a standard approximation guarantee. Experiments with four LVLMs on two benchmarks show that our method preserves accuracy while scoring $4$-$13\times$ fewer frames and selecting $18.4\%$-$20.5\%$ fewer input keyframes than existing baselines. CSES further achieves a $3.1$-$5.4\times$ speedup in frame selection over baselines.
Video Scene Procedure Planning (VSPP) supplies the target start-goal observations in advance, leaving open how a planner should act when the evidence must itself be retrieved. We introduce Cross-Video Scene Procedure Planning (CVSPP): given an answer-redacted start-goal query and K candidate videos, a model must retrieve the supporting video, localize the relevant window, and predict the action sequence. Two obstacles couple here. Same-task demonstrations share stages and windows, and an early hard selection passes the wrong scene chain to the planner. We build an eleven-source benchmark with typed negative roles, a fail-closed answer-leakage gate, and separate Evidence- and Plan-axis metrics. On its 14 source-horizon cells we adapt nine planner families against a majority-sequence floor. We then present One-Step Evidence Fusion (OSEF), which scores a query-conditioned cell-and-span lattice over all candidates and feeds the full soft lattice to the planner through a token-global adapter, cropping no window beforehand. OSEF ranks first on all six cells the benchmark certifies as method-rankable. On four matched same-task COIN and CrossTask cells it improves exact-video-and-plan success by 2.9-10.7 points over an enhanced hard-selection SOTA, and a component study assigns the largest single increment to the token-global interface. Five converted-source cells sit at or near the majority-sequence floor, the benchmark's remaining headroom. The supplementary package includes model constructors and evaluation code.
Gait recognition is shaped by its input representation. Silhouettes encode projected body shape, skeletons encode sparse joint coordinates, and 3D meshes encode dense surface geometry. In each case, identity-bearing articulation is observed through geometric carriers that also vary with clothing, skeletal scale, or body shape. We investigate whether gait can instead be recognized from compact articulated controls. We introduce Momentum Human Rig (MHR) pose as a gait representation, describing each frame using 184 semantically organized body and hand parameters estimated from monocular video. MHRGait groups these heterogeneous controls by anatomy, models their intra-frame coordination and temporal evolution, and produces compact body and hand descriptors. We further introduce MHRGait++, which combines MHR pose with silhouettes through modality-balanced distance fusion, preventing descriptor count from determining modality importance. Experiments on four benchmarks show that MHRGait attains the best overall performance among compared model-based methods on CCPG and SUSTech1K and transfers effectively across datasets, while its recognition network requires only 2.76M parameters and 0.69 GFLOPs for a 30-frame input. MHRGait++ consistently improves silhouette recognizers with a favorable accuracy-efficiency trade-off. These results establish rig-space articulation as an effective standalone gait representation and a complementary cue to projected body shape. Our code is available at https://github.com/duanhuiran/MHRGait.