Computer Vision and Image Understanding最新文献

{"title":"ADuLTS: Appearance Descriptions under Long-Tailed Scenarios with diverse synthesized images","authors":"SeungJu Cha,&nbsp;Seunghee Choi,&nbsp;Kwanyoung Lee,&nbsp;Dong-Jin Kim","doi":"10.1016/j.cviu.2026.104684","DOIUrl":"10.1016/j.cviu.2026.104684","url":null,"abstract":"<div><div>Unlike class-balanced datasets such as ImageNet and CIFAR, real-world data often show an imbalanced distribution where the majority of image samples are concentrated in a few head classes while tail classes occupy a small portion of the data. If existing image classification models are trained using data with an imbalanced distribution, the results of the training can be biased towards head classes. Recently, supervised contrastive learning (SCL) methods using balanced datasets have shown effective results in the field of image classification. However, when applying SCL to image classification models with imbalanced datasets, the classification performance decreases because of the lack of tail class representations. With the recent success of synthesizing realistic images from generative models, to handle this representation shortage in tail classes, we propose Appearance Description under Long-Tailed Scenarios (ADuLTS) that utilize pre-trained large language models (LLM) for tail class sample generation. By asking LLM to describe the semantic appearance of each class, we synthesize images to alleviate the sample shortage in the tail classes. Furthermore, we propose a simple but effective joint training network to decrease the feature distribution difference while training with our balanced dataset with real and generated images. Our method shows noticeable classification performance enhancement on long-tailed scenarios, especially on the tail classes with fewer samples in CIFAR10-LT, CIFAR100-LT, and <em>mini</em>ImageNet-LT.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104684"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147422114","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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":"CurFed-CHARM: Curriculum sequential federated learning with a channel-hierarchical action recognition model for non-IID heterogeneous data in neuromorphic/event cameras","authors":"Muhammad Hamza Zafar,&nbsp;Syed Kumayl Raza Moosavi,&nbsp;Filippo Sanfilippo","doi":"10.1016/j.cviu.2026.104696","DOIUrl":"10.1016/j.cviu.2026.104696","url":null,"abstract":"<div><div>Human action recognition is a key technology in healthcare monitoring, security surveillance, human–computer interaction, and human–robot interaction. Traditional vision-based approaches, relying on RGB cameras, suffer from motion blur, high bandwidth demands, and poor performance in challenging lighting conditions. Neuromorphic event cameras offer promising solutions but still face challenges related to privacy-preserving model training and decentralized learning for non-IID heterogeneous data. To address these limitations, we propose Curriculum Federated Learning based Channel Hierarchical Action Recognition (<em>CurFed-CHARM</em>), a novel framework for neuromorphic vision-based action recognition. <em>CurFed-CHARM</em> leverages event-driven sensing and progressive knowledge distillation to enhance real-time recognition accuracy while preserving data privacy across distributed edge devices. Our approach implements a sequential federated learning strategy with attention-enhanced 3D CNNs that process spatio-temporal features from action sequences. The framework begins with a comprehensive client possessing all action classes and progressively integrates additional clients once performance thresholds are achieved. We introduce specialized metrics including Client Convergence Rate, Knowledge Transfer Efficiency, Catastrophic Forgetting Metric, and Communication Efficiency to evaluate the system’s performance. Our experimental results demonstrate that <em>CurFed-CHARM</em> achieves superior performance with 94.87% average accuracy across all clients on E-FedAct dataset, outperforming traditional federated approaches like FedAvg (92.99%), FedProx (92.42%), and FedCurv (92.09%). The ablation study confirms the critical importance of our Global Temporal Attention mechanism, which significantly enhances knowledge transfer efficiency by 60% compared to models without this component. Additionally, <em>CurFed-CHARM</em> reduces communication costs by 27.3% compared to conventional federated learning methods through its strategic client integration strategy.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104696"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147422116","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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":"Handscribe: A gloss-free framework for sign language translation and gloss sequence generation","authors":"Emanuele Colonna ,&nbsp;Ivan Rinaldi ,&nbsp;David Landi ,&nbsp;Gennaro Vessio ,&nbsp;Giovanna Castellano","doi":"10.1016/j.cviu.2026.104674","DOIUrl":"10.1016/j.cviu.2026.104674","url":null,"abstract":"<div><div>Sign language translation systems traditionally rely on intermediate gloss representations to bridge the gap between visual input and written language output. However, manual gloss annotation is costly, language-dependent, and often lossy, prompting growing interest in gloss-free alternatives. This paper introduces <span><math><mrow><mi>H</mi><mi>andscribe</mi></mrow></math></span>, a novel two-stage framework for gloss-free sign language translation and gloss sequence generation. <span><math><mrow><mi>H</mi><mi>andscribe</mi></mrow></math></span> first translates continuous sign language videos into written language sentences using a lightweight decoder built atop SlowFast-based spatiotemporal features and a frozen mBART model. Then, in the second stage, it generates gloss sequences from these sentences using a Large Language Model (LLaMa3.1-8B-Instruct) that has been fine-tuned with weak supervision. Our experiments on PHOENIX-2014-T and Wav2Gloss Fieldwork demonstrate strong translation performance and state-of-the-art multilingual gloss generation, even in zero-shot settings. The proposed framework reduces annotation bottlenecks while maintaining flexibility and interpretability, paving the way for scalable and inclusive sign language technologies. The code and fine-tuning scripts are available at <span><span>https://github.com/colonnaemanuele/Handscribe</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104674"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147422117","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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":"Statistical-driven adaptive data augmentation for single-domain generalized object detection","authors":"Chang Lu ,&nbsp;Shaochen Su ,&nbsp;Hao Li","doi":"10.1016/j.cviu.2026.104687","DOIUrl":"10.1016/j.cviu.2026.104687","url":null,"abstract":"<div><div>In the context of single-source domain generalized object detection, the diversification of the source domain is of paramount importance, as it significantly influences the model’s generalization ability across different target domains. Existing data augmentation strategies for single-source domain generalization often overlook the preservation of foreground features in the source domain during the augmentation process, inevitably compromising the effectiveness and integrity of source-domain foreground information. To address this issue, we propose an effective data augmentation method called statistical-driven adaptive piecewise mapping (SDAPM), which automatically adjusts mapping parameters based on image statistical features to maximally mitigate the detail loss in source-domain data caused by diversification operations during the augmentation stage. Additionally, we introduce a dynamic heterogeneous domain perturbation generator (DHDPG) built upon SDAPM, which simulates various types of domain shifts under natural weather conditions. Both SDAPM and DHDPG are designed to mimic domain shift phenomena while preserving source domain foreground information, enabling the detector to learn critical foreground domain-invariant factors and enhancing its performance robustness under harsh scenarios. On the multi-weather dataset and the large CityScapes-C dataset, the proposed method achieved significant improvements of 7.6% and 9.8%, respectively, compared with the baseline method—this highlights its stronger generalization ability in practical applications.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104687"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191649","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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":"OSTE: Omni-Scene Text Editing with Latent Decoupling","authors":"Tonghua Su ,&nbsp;Fuxiang Yang ,&nbsp;Lei Fan ,&nbsp;Donglin Di ,&nbsp;Zhongjie Wang ,&nbsp;Songze Li ,&nbsp;Xiangqian Wu ,&nbsp;Xiang Zhou","doi":"10.1016/j.cviu.2026.104673","DOIUrl":"10.1016/j.cviu.2026.104673","url":null,"abstract":"<div><div>In this study, we introduce a task, termed “Omni-Scene Text Editing”, which involves altering both the textual content and the text style of a source image while preserving the original text scene. Existing methods tend to overlook the fine-grained adjustment of foreground text style attributes, such as rotation angle, color, and font type. To address this challenge, we propose a framework for Omni-Scene Text Editing with Latent Decoupling (OSTE) that operates within the latent feature space to embed and refine foreground text styles. OSTE includes two modules: the background inpainting module and the foreground style editing module, which process the background texture and foreground text, respectively, facilitating independent editing of the text. The background inpainting module erases the foreground text content and recovers the background with highly fidelity textures. The foreground style editing module comprises a style encoder, a content encoder, and a fusion generator. The style encoder captures the style characteristics of the foreground text, while the content encoder provides representations of the target text within the latent feature space to facilitate content modifications. The fusion generator amalgamates information from these components and generates the final rendered text images. To better handle real-world scenes, we developed the shared weight approach and two model designs (“cut out text areas” and AdaIN) for training. To the best of our knowledge, our work stands as the first to enable fine-grained manipulation of foreground text content and style through deep semantic editing in the latent feature space. Extensive experiments validate OSTE’s ability to produce photorealistic foreground text and eliminate undesired text shadows in real-world scenes when altering text content.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104673"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191651","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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":"FusGaze: Full range gaze estimation with multi-scale fusion","authors":"Dayeon Yoo,&nbsp;Jaehun Cho,&nbsp;Kwangho Song","doi":"10.1016/j.cviu.2026.104666","DOIUrl":"10.1016/j.cviu.2026.104666","url":null,"abstract":"<div><div>Gaze estimation plays a crucial role in a wide range of applications, including human–computer interaction (HCI), augmented and virtual reality (AR/VR), driver monitoring and robotic intelligence. Most of the conventional approaches rely on external detectors to extract head or face regions prior to gaze estimation. It causes a fundamental issue by inheriting the detector errors, which leads to incorrect gaze predictions based on inaccurate input data. To overcome these limitations, we propose FusGaze, a novel end-to-end gaze estimation framework that seamlessly integrates detection and estimation within a single architecture. FusGaze eliminates external detector dependencies by appending an object detection head and a gaze estimation head to a single detection backbone, thereby enabling more stable and robust gaze estimation in practical environments. To enable effective multi-scale fusion, we design an Adaptive Weighted Late Fusion (AWLF) block that data-drivenly fuses multi-scale gaze vectors. Moreover, a 2D Gaussian heatmap-based auxiliary loss function is introduced to improve estimation stability, even under challenging scenarios such as side or rear views where visual cues are limited. Experiments conducted on the Gaze360 benchmark demonstrate that FusGaze achieves state-of-the-art performance with mean angular error of 12.85°, among single image-based gaze estimation models. In particular, the model shows significant improvements under front-facing conditions while maintaining competitive performance under backward-facing conditions, thus experimentally validating its effectiveness in full 360° gaze estimation.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104666"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191752","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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":"HGAvatar: One-shot high-quality 3D Gaussian head avatar","authors":"Xueping Wang,&nbsp;Jun Xu,&nbsp;Feihu Yan,&nbsp;Guangzhe Zhao","doi":"10.1016/j.cviu.2026.104660","DOIUrl":"10.1016/j.cviu.2026.104660","url":null,"abstract":"<div><div>3D Gaussian Splatting (3DGS) has sparked growing interest among researchers in 3D head avatars, which is an important topic in the fields of digital human and artificial intelligence. Due to the limitations of monocular videos and the complexity of facial textures and motions, creating high-fidelity 3D head avatars still poses significant challenges. In this paper, we propose a High-Quality 3D Gaussian Head Avatar (HGAvatar) that achieves modeling head avatars with realistic textures and accurate poses within the limitations of monocular video. Specifically, we design a Detail Optimization Module (DOM) to enhance the capability in capturing intricate facial texture details. Additionally, we introduce a Geometric Deformation Module (GDM), which decouples head pose from facial expression and utilizes them as guiding factors to further adjust the properties of the Gaussian ellipsoid, allowing the model to learn more accurate fine details of facial motions. Both qualitative and quantitative experiments demonstrate that the proposed method can better preserve the fine facial texture details and accurately reproduce subtle differences in facial dynamics compared to the state-of-the-art methods.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104660"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146081623","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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":"GAN inversion via cross-domain feature fusion and invertibility decomposition","authors":"Jinyong Sun ,&nbsp;Ruiheng Ling ,&nbsp;Zhigang Sun ,&nbsp;Xiang Zhao ,&nbsp;Guoyong Cai ,&nbsp;Zhiliang Shang ,&nbsp;Chuang Lei","doi":"10.1016/j.cviu.2026.104700","DOIUrl":"10.1016/j.cviu.2026.104700","url":null,"abstract":"<div><div>Generative Adversarial Networks (GANs) have significantly enhanced the capability of unconditional image synthesis by mapping latent codes to high-fidelity synthetic images. Editing real images with GANs requires finding a latent code that can faithfully reconstruct the input images, a process commonly referred to as GAN inversion. However, existing methods often struggle to effectively reconstruct Out-Of-Domain (OOD) features and suffer from a trade-off between distortion and editability, primarily due to limited semantic coverage of GAN latent spaces. In this paper, we propose a novel GAN inversion method based on cross-domain feature fusion and invertibility decomposition, which enhances OOD feature modeling while constraining the editability of In-Domain (ID) features. To achieve high-fidelity inversion, we propose a residual-structured Feature Pyramid Network to extract OOD features, which are integrated into the encoder-generator pipeline via residual connections to enhance the modeling capability. To enable high-quality attribute manipulation, we introduce a multi-scale feature masking mechanism that balances the influence of ID and OOD features. Extensive experiments on real-world datasets demonstrate the superiority of our method in reconstructing OOD features while preserving latent editability of ID features.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104700"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147422110","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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":"GASP: Gaussian Splatting for physics-based simulations","authors":"Piotr Borycki ,&nbsp;Weronika Smolak-Dyżewska ,&nbsp;Joanna Waczyńska ,&nbsp;Marcin Mazur ,&nbsp;Sławomir Tadeja ,&nbsp;Przemysław Spurek","doi":"10.1016/j.cviu.2026.104699","DOIUrl":"10.1016/j.cviu.2026.104699","url":null,"abstract":"<div><div>Physics simulation is paramount for modeling and utilizing 3D scenes in various real-world applications. However, integrating with state-of-the-art 3D scene rendering techniques such as Gaussian Splatting (GS) remains challenging. Existing models use additional meshing mechanisms, including triangle or tetrahedron meshing, marching cubes, or cage meshes. Alternatively, we can modify the physics-grounded Newtonian dynamics to align with 3D Gaussian components. Current models take the first-order approximation of a deformation map, which locally approximates the dynamics by linear transformations. In contrast, our GS for Physics-Based Simulations (GASP) pipeline uses parametrized flat Gaussian distributions. Consequently, the problem of modeling Gaussian components using the physics engine is reduced to working with 3D points. In our work, we present additional rules for manipulating Gaussians, demonstrating how to adapt the pipeline to incorporate meshes, control Gaussian sizes during simulations, and enhance simulation efficiency. This is achieved through the Gaussian grouping strategy, which implements hierarchical structuring and enables simulations to be performed exclusively on selected Gaussians. The resulting solution can be integrated into any physics engine that can be treated as a black box. As demonstrated in our studies, the proposed pipeline exhibits superior performance on a diverse range of benchmark datasets designed for 3D object rendering. The project webpage, which includes additional visualizations, can be found at <span><span>https://waczjoan.github.io/GASP</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104699"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191648","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"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":"Deep learning based empty shelf detection based on autonomous mobile robot","authors":"Giuseppe De Simone,&nbsp;Alessia Saggese,&nbsp;Pasquale Foggia,&nbsp;Mario Vento","doi":"10.1016/j.cviu.2026.104697","DOIUrl":"10.1016/j.cviu.2026.104697","url":null,"abstract":"<div><div>The issue of out-of-stock (OOS) represents a substantial challenge for retailers, often resulting in significant sales losses. To address this problem, this paper introduces an autonomous mobile robotic platform built on the Robot Operating System (ROS) framework, designed to accelerate the restocking process in supermarkets. The platform autonomously detects empty shelves and notifies human operators, streamlining inventory management. Equipped with advanced navigation capabilities, the proposed system employs a deep learning-based, two-stage architecture that identifies shelving areas and subsequently detects empty shelves. To validate the performance of the proposed two-stage artificial vision algorithm, two datasets were used: the first comprises approximately 2000 images (900 of them collected by our team from three different supermarkets), while the second dataset consists of around 5600 manually annotated images extracted from videos recorded in a supermarket by the robotic platform itself. Additionally, in order to validate the entire robotic system, an extensive experimental evaluation was conducted in a supermarket during regular business hours. The results demonstrate that the proposed platform substantially outperforms human operators, identifying OOS items eight times faster than traditional human operator based methods. This advancement provides valuable assistance to supermarket staff, significantly enhancing operational efficiency.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"265 ","pages":"Article 104697"},"PeriodicalIF":3.5,"publicationDate":"2026-03-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"146191650","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}