International Journal of Computer Vision最新文献

{"title":"Convolutional Legendre Memory Unit (ConvLMU) for Stereo Neuromorphic Event-Based Depth Estimation","authors":"Avi Hazan, Hadar Cohen-Duwek, Elishai Ezra Tsur","doi":"10.1007/s11263-026-03032-6","DOIUrl":"https://doi.org/10.1007/s11263-026-03032-6","url":null,"abstract":"Event-based computer vision is distinguished by its high temporal resolution and energy efficiency, yet the asynchronous characteristics of event data impede its utilization in conventional embedded computing. In this work, we propose ConvLMU, a neuromorphic design that synergistically integrates Legendre Memory Units (LMUs) and convolutional networks for efficient spatiotemporal processing. Here, we used LMUs to capture long-range temporal dependencies and convolutional layers to provide robust spatial feature extraction for visual depth estimation in 3D from stereo event cameras. Evaluation on the DSEC dataset demonstrates that ConvLMU attains the lowest loss (0.2318) and pixel error rate (0.4904), surpassing the performance of both ConvLSTM (loss: 0.2701; pixel error: 0.7294) and ConvGRU (loss: 0.2818; pixel error: 0.7377). Furthermore, we validated ConvLMU’s enhanced spatial coherence and temporal smoothness using synthetic visual patterns. Our findings substantiate the potential of ConvLMUs as a promising foundation for energy-efficient, neuromorphic-compatible, event-based vision systems.","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"14 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149003643","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"SAM2TAP: Repurposing Pre-trained Video Object Segmentation Models for Point Tracking","authors":"Zhiyue Zhao, Wen Wang, Canyu Zhao, Zhoujie Fu, Zhouhang Zhu, Jianjun Zhou, Hao Chen, Tong He, Chunhua Shen","doi":"10.1007/s11263-026-03045-1","DOIUrl":"https://doi.org/10.1007/s11263-026-03045-1","url":null,"abstract":"Video object segmentation (VOS) and Tracking Any Point (TAP) share the common goal of tracking objects or points of interest in videos while predicting through potential occlusions, yet these two closely related tasks have largely evolved independently. Our research demonstrates that a well-pretrained VOS model, specifically SAM2, can be repurposed for TAP by adapting its entity-level tracking representations to point-level tracking, revealing a substantial connection between entity segmentation and point tracking. Building on this insight, we propose SAM2TAP, a novel framework that adapts SAM2 for the TAP task via a minimalist adaptation strategy, introducing only a lightweight token decoder and a specialized TAP head. To further tailor SAM2TAP for precise point estimation, we incorporate Long-Term Feature Memory and Spatial Context Attention mechanisms. SAM2TAP leverages the powerful spatiotemporal priors learned from VOS pretraining to enable online point tracking. Even though it is fine-tuned for point tracking exclusively on synthetic datasets, the model robustly handles complex occlusions and exhibits remarkable adaptability to diverse, high-resolution real-world scenes. Extensive experiments validate the effectiveness of our approach, confirming that repurposing foundation models offers a data-efficient pathway to advance point tracking and highlighting the potential of SAM2TAP as a unified framework for both point tracking and segmentation tasks. Our code and data will be made publicly available.","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"23 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149003639","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Constellation Dataset: Benchmarking High-Altitude Object Detection for an Urban Intersection","authors":"Mehmet Kerem Turkcan, Chengbo Zang, Sanjeev Narasimhan, Gyung Hyun Je, Bo Yu, Mahshid Ghasemi, Javad Ghaderi, Gil Zussman, Zoran Kostic","doi":"10.1007/s11263-026-03029-1","DOIUrl":"https://doi.org/10.1007/s11263-026-03029-1","url":null,"abstract":"As smart cities evolve, privacy-preserving edge processing at traffic intersections has become essential for real-time safety applications while reducing data transmission and centralized computation. High-altitude cameras with on-device inference provide an optimal solution that respects privacy while delivering low-latency results. We introduce Constellation, a dataset of 13K images for research on object detection in dense urban streetscapes from high-elevation cameras across varied temporal conditions. The dataset addresses challenges in small object detection, particularly for pedestrians observed from elevated positions with limited pixel footprints. Our evaluation of contemporary object detection architectures reveals a 10% lower average precision (AP) for small pedestrians compared to vehicles. Pretraining models on structurally similar datasets increases mean AP by 1.8%. Domain-specific data augmentations and pseudo-labeled data from top-performing models further enhance performance. We evaluate deployment viability on resource-constrained edge devices including Jetson Orin, Raspberry Pi 5, and mobile platforms, demonstrating feasibility of privacy-preserving on-device processing. Comparing models trained on data collected across different time intervals reveals performance drift due to changing intersection conditions. The best-performing model achieves 92.0% pedestrian AP with 7.08 ms inference time on A100 machines, and 95.4% mAP. The best-performing edge model achieves a similar performance, with Jetson Orin Nano achieving 94.5% mAP and 27.5ms inference time using TensorRT.","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"35 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149003646","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"De-biasing Skeleton-Based Action Recognition with Convex Hull Adaptive Shift","authors":"Mengyuan Liu, Yuhang Wen, Yi Zhang, Songtao Wu, Hong Liu, Junsong Yuan, Beichen Ding","doi":"10.1007/s11263-026-03037-1","DOIUrl":"https://doi.org/10.1007/s11263-026-03037-1","url":null,"abstract":"Skeleton sequences can represent both individual actions and multi-entity interactions, encompassing human bodies, hands, objects, and robots. Existing approaches to recognize skeleton-based actions and interactions usually adopt a late fusion strategy, which expects individuals are independent and identically distributed to train a robust weight-shared entity encoder. However, observed entity bias in various skeletal data violates this assumption, leading to suboptimal optimization of backbone models that might produce wrong recognition results. This bias arises from the world coordinate system’s initial configuration, where the choice of origin often creates bias in representation. To this end, we propose a Convex Hull Adaptive Shift based normalization method to reduce Entity bias (CHASE), improving performance across a variety of skeleton-based action and interaction recognition tasks. To adaptively apply plausible shifts to the input skeletons, we formulate a plug-and-play parameterized network that ensures the relocated world origin lies within the skeleton convex hull, which avoids non-convergence by limiting the search space. To further minimize entity bias, we incorporate an auxiliary objective that leverages pair-wise distribution distances to guide network optimization. To support both single- and multi-entity actions, we propose a sub-entity strategy that offers a consistent formulation for both scenarios. Moreover, CHASE demonstrates compatibility with various intra-skeleton modalities, such as bones and velocities, highlighting its adaptability. Essentially, our method works as a normalization approach to reduce entity bias, enabling subsequent classifiers to achieve improved recognition performance across diverse settings. Extensive experiments on seven datasets, including NTU RGB+D, NTU RGB+D 120, H2O, Assembly101, Collective Activity, Volleyball, and HARPER, consistently verify our approach by seamlessly integrating with various backbones and significantly boosting their performance. Our code is publicly available at <ext-link ext-link-type=\"uri\" xlink:href=\"https://github.com/Necolizer/CHASE\">https://github.com/Necolizer/CHASE</ext-link>.","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"28 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149003635","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Spectral-Domain Anchor Learning for Scale-Robust Continuous Visual Fields","authors":"Xingran Guo, Tiaojie Xiao, Jie Liu, Keqin Li","doi":"10.1007/s11263-026-03038-0","DOIUrl":"https://doi.org/10.1007/s11263-026-03038-0","url":null,"abstract":"Continuous neural fields have advanced spatial-temporal signal parameterization, yet their efficacy is fundamentally limited by spectral bias and slow-convergence dynamics. These limitations severely restrict their capacity to resolve high-frequency stochastic textures and dense, highly oscillatory spatial fields. To address these limitations, we propose the Spectral-Domain Anchor Learning (SDAL) framework, which transitions from localized spatial gradient descent to explicit spectral-domain parameterization. SDAL constructs a frequency-domain “Spider-Web Topology” that functions as a geometrically-aware, multi-ring adaptive filter bank, enabling improved scale robustness and significantly reduced aliasing artifacts. Crucially, we introduce an Alternating Optimization Strategy equipped with a closed-form analytical solver to decouple the highly non-convex spatial learning problem. This architectural innovation yields a zero-iteration cold-start capability: by resolving the projection weights <italic>W</italic> via a closed-form analytical solver at network initialization, SDAL achieves high reconstruction fidelity prior to any gradient update. Comprehensive evaluations across highly oscillatory continuous spatial fields, extreme cross-domain sensing modalities (e.g., subsurface radar topologies), and severely degraded visual signals support our theoretical assertions. SDAL consistently achieves effective spectral disentanglement, enabling the sharp recovery of faint structural signatures in noisy empirical environments. Furthermore, SDAL achieves equivalent reconstruction fidelity with zero gradient iterations, at the cost of a one-time <inline-formula><alternatives><mml:math><mml:mrow><mml:mi mathvariant=\"script\">O</mml:mi><mml:mo stretchy=\"false\">(</mml:mo><mml:msup><mml:mi>K</mml:mi><mml:mn>2</mml:mn></mml:msup><mml:mo stretchy=\"false\">)</mml:mo></mml:mrow></mml:math><tex-math>documentclass[12pt]{minimal} usepackage{amsmath} usepackage{wasysym} usepackage{amsfonts} usepackage{amssymb} usepackage{amsbsy} usepackage{mathrsfs} usepackage{upgreek} setlength{oddsidemargin}{-69pt} begin{document}$$mathcal {O}(K^{2})$$end{document}</tex-math></alternatives></inline-formula> Cholesky factorization. This approach trades iterative gradient search for a single analytical matrix inversion, which scales favorably for small-to-moderate anchor counts compared to conventional coordinate networks. Ultimately, this work advances implicit representations into robust, geometrically regularized solvers, providing a scalable foundation for large-scale computer vision and continuous spatial learning.","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"191 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149003637","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Towards Understanding Token Evolution in Continual Learning of MLLMs","authors":"Xusheng Cao, Yifan Meng, Jiaze Li, Jianzhong Ju, Zhenbo Luo, Jian Luan, Fei Yang, Ming-Ming Cheng, Xialei Liu","doi":"10.1007/s11263-026-03035-3","DOIUrl":"https://doi.org/10.1007/s11263-026-03035-3","url":null,"abstract":"Multimodal large language models (MLLMs) have recently exhibited impressive generalization and reasoning capabilities across diverse vision–language tasks. Despite their success, continual learning on domain-specific data often triggers catastrophic forgetting, severely degrading previously acquired knowledge and hindering their scalability to real-world multidomain and multitask settings. Existing attempts typically extend conventional continual learning paradigms such as replay buffers, modular expansion, or heuristic training schedules combining data rebalancing and hyperparameter tuning. However, these strategies overlook the unique architectural and representational properties of MLLMs, limiting their effectiveness. In this work, we revisit catastrophic forgetting from an interpretability perspective. By examining layer-wise token evolution across different training stages, we uncover semantic shifts in how MLLMs internalize and reinterpret tokens during continual learning. Motivated by this observation, we introduce correctness-guided layer alignment (CLA), which allows the new model to inherit multi-layer semantic consistency from its predecessor while learning from ground-truth supervision. Experiments on the UCIT and DCL benchmarks demonstrate that our approach effectively mitigates forgetting and promotes more stable continual learning across domains and tasks.","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"118 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149003630","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"C-Flat++: Towards a More Efficient and Powerful Framework for Continual Learning","authors":"Wei Li, Hangjie Yuan, Zixiang Zhao, Yifan Zhu, Aojun Lu, Guodong Du, Yanan Sun, Tao Feng","doi":"10.1007/s11263-026-03009-5","DOIUrl":"https://doi.org/10.1007/s11263-026-03009-5","url":null,"abstract":"Balancing sensitivity to new tasks and stability for retaining past knowledge is crucial in continual learning (CL). Recently, sharpness-aware minimization has proven effective in transfer learning and has also been adopted in continual learning (CL) to improve memory retention and learning efficiency. However, relying on zeroth-order sharpness alone may favor sharper minima over flatter ones in certain settings, leading to less robust and potentially suboptimal solutions. In this paper, we propose <bold>C</bold>ontinual <bold>Flat</bold>ness (<bold>C-Flat</bold>), a method that promotes flatter loss landscapes tailored for CL. C-Flat offers plug-and-play compatibility, enabling easy integration with minimal modifications to the code pipeline. Besides, we present a general framework that integrates C-Flat into all major CL paradigms and conduct comprehensive comparisons with loss-minima optimizers and flat-minima-based CL methods. Our results show that C-Flat consistently improves performance across a wide range of settings. In addition, we introduce <bold>C-Flat++</bold>, an efficient yet effective framework that leverages selective flatness-driven promotion, significantly reducing the update cost required by C-Flat. Extensive experiments across multiple CL methods, datasets, and scenarios demonstrate the effectiveness and efficiency of our proposed approaches. Code is available at <ext-link ext-link-type=\"uri\" xlink:href=\"https://github.com/WanNaa/C-Flat_PP.\">https://github.com/WanNaa/C-Flat_PP.</ext-link>","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"16 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-09","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149003638","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Dual Adaptive Visual-Semantic Prompt Collaboration for Generalized Zero-Shot Learning","authors":"Huajie Jiang, Zhengxian Li, Yuankai Qi, Yongli Hu, Baocai Yin, Amin Beheshti, Jian Yang, Ming-Hsuan Yang","doi":"10.1007/s11263-026-02998-7","DOIUrl":"https://doi.org/10.1007/s11263-026-02998-7","url":null,"abstract":"Generalized zero-shot learning (GZSL) addresses the challenging task of recognizing both seen and unseen classes by leveraging shared semantic knowledge. A core challenge in this domain is achieving robust visual-semantic alignment to transfer knowledge from seen classes to novel classes. Current state-of-the-art methods typically fine-tune large-scale visual backbones on scarce training data. However, this approach frequently leads to severe overfitting to seen classes, which significantly degrades performance on novel categories. To mitigate this issue, we propose the Dual Adaptive Visual-Semantic Prompt Collaboration Network (VSPCN+), a novel framework that utilizes prompt-tuning for effective feature adaptation. Our method introduces a dual-prompt mechanism comprising both visual and semantic prompts. The semantic prompts guide the visual encoder to learn visual features that are more semantically consistent with class attributes, while the visual prompts steer the semantic encoder to generate semantic representations that are more visually grounded. This collaborative process enhances the overall visual-semantic consistency. A key innovation of our work is the dynamic generation of instance-adaptive prompts, which contrasts with existing prompt-learning methods that rely on static, global prompts. By tailoring prompts to individual instances, our approach enhances the model’s robustness and generalization capabilities across diverse visual inputs. This collaborative adaptation, guided by our dual-prompt mechanism, allows the visual and semantic encoders to produce consistent representations for effective visual-semantic alignment. Extensive experiments on standard GZSL benchmarks demonstrate that our proposed VSPCN+ performs favorably against several state-of-the-art methods.","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"27 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884071","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"3D Scene Generation: A Survey","authors":"Haozhe Xie, Beichen Wen, Zhaoxi Chen, Fangzhou Hong, Ziwei Liu","doi":"10.1007/s11263-026-02999-6","DOIUrl":"https://doi.org/10.1007/s11263-026-02999-6","url":null,"abstract":"","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"133 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-03","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884072","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"OpenVid++: A Large-scale High-quality Dataset for Text-to-video Generation","authors":"Kepan Nan, Tiehan Fan, Rui Xie, Penghao Zhou, Yilin Zhang, Xiang Li, Jian Yang, Zhenheng Yang, Ying Tai","doi":"10.1007/s11263-026-02989-8","DOIUrl":"https://doi.org/10.1007/s11263-026-02989-8","url":null,"abstract":"","PeriodicalId":13752,"journal":{"name":"International Journal of Computer Vision","volume":"70 1","pages":""},"PeriodicalIF":19.5,"publicationDate":"2026-09-02","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148884074","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}