Computer Vision and Image Understanding最新文献

{"title":"Occlusion-aware low-rank learning framework for facial expression recognition","authors":"Yanzhong Wang ,&nbsp;Daming Shi ,&nbsp;Muhammad Sadiq","doi":"10.1016/j.cviu.2026.104820","DOIUrl":"10.1016/j.cviu.2026.104820","url":null,"abstract":"<div><div>Facial expression recognition (FER) remains challenging when faces are partially occluded, as occlusions obscure expression-related regions and introduce irrelevant visual patterns, leading to substantial performance degradation in real-world scenarios. In this work, we propose an occlusion-aware low-rank learning framework for robust FER under occlusion. Unlike previous occlusion-robust FER methods that often require manually annotated occluded facial images, our method employs synthetically occluded images for training, thereby reducing the reliance on costly real-world occlusion annotations. The proposed framework integrates occlusion detection, masking, and feature-level low-rank recovery into a unified pipeline. Specifically, the model first detects occluded regions and replaces them with mask values to suppress the influence of diverse occlusion appearances, producing a cleaner but incomplete facial representation. Since this masked representation may still lose semantic information from occluded facial components, we introduce a low-rank learning module to recover missing expression-related features in the nonlinear feature space rather than reconstructing occluded pixels in the image space. Furthermore, because the corresponding non-occluded images in the synthetic training process provide complete facial semantics, we design a dual-branch guidance structure in which the non-occluded branch assists the occluded branch in feature recovery. By jointly optimizing these components, the proposed method achieves competitive performance on widely used in-the-wild FER benchmarks, including RAF-DB and AffectNet, and obtains state-of-the-art or highly competitive results on occluded FER datasets, including Occlusion-RAF-DB, Syn-RAF-DB, Syn-AffectNet, and FED-RO. These results demonstrate the robustness and generalization ability of the proposed framework under occluded conditions.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"269 ","pages":"Article 104820"},"PeriodicalIF":3.5,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148185465","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":"SGMP: Spatial-Semantic Guided Motion-Perception for infrared small targets","authors":"Wei Zhang,&nbsp;Tao Liu,&nbsp;Tianhang Guan,&nbsp;Jia Gao","doi":"10.1016/j.cviu.2026.104817","DOIUrl":"10.1016/j.cviu.2026.104817","url":null,"abstract":"<div><div>In temporal infrared small target detection, the effective exploitation of motion pattern discrepancies between targets and backgrounds is crucial for achieving accurate detection. Although existing convolutional neural network (CNN)-based architectures benefit from local feature perception, their limited receptive fields hinder the modeling of long-range temporal dependencies, especially in scenes with dynamic complexity, leading to a notable degradation in feature discrimination. Transformer-based architectures have been demonstrated to address the issue of long-sequence dependency through self-attention mechanisms, the uniform global interaction tends to dilute the representation of subtle local motion patterns, which are crucial for small target detection. To overcome these limitations, we propose a backbone network, called Spatial-Semantic Guided Motion Perception (SGMP). SGMP introduces a spatially semantic guided attention mechanism that (i) leverages spatial features to generate salient semantic query vectors and value maps for extracting potential semantic priors of targets, and (ii) incorporates a Transformer encoder to model long-range motion features, aligning global motion context with local semantic cues across modalities. The design enables it to focus on motion-sensitive spatiotemporal regions associated with small targets. Built upon SGMP, we design a lightweight and efficient detection framework called SGMP-UNet, which fully exploits motion-aware representations for end-to-end small target detection. Extensive experiments on the NUDT-MIRSDT, IRSDT and TSIRMT datasets demonstrate that SGMP-UNet consistently outperforms state-of-the-art (SOTA) methods across multiple evaluation metrics. Our code will be available at <span><span>https://github.com/Bethexone/SGMP.git</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"269 ","pages":"Article 104817"},"PeriodicalIF":3.5,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148185580","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":"MAS-Net: A medical image segmentation method based on memory augmentation and supplementation","authors":"Xiaoyan Zhang ,&nbsp;Zheng Zhao ,&nbsp;Weiqiang Sun ,&nbsp;Yongqin Zhang ,&nbsp;Chunlin Yu ,&nbsp;Xiangfu Meng","doi":"10.1016/j.cviu.2026.104816","DOIUrl":"10.1016/j.cviu.2026.104816","url":null,"abstract":"<div><div>Precise medical image segmentation plays a vital role in clinical applications, including computer-aided diagnosis and surgical procedures. However, challenges arise due to the blurred boundaries of lesions and the variability in the size of affected regions, making the task inherently difficult. Most existing methods enhance feature representations by combining convolutional neural networks and transformers. However, as model depth increases, networks tend to neglect lower-level details when extracting higher-order abstract features, leading to the loss of critical local features and spatial relationships. To address this issue, we propose MAS-Net, a novel image segmentation approach that leverages memory augmentation and memory supplementation techniques. Specifically, in the encoder, we design a Memory Augmentation (MA) module aimed at strengthening the expression of image details, thereby preventing the loss of key information during segmentation. In the decoder, we introduce a Memory Supplementation (MS) module to effectively recover missing features in ambiguous regions, especially where details are unclear or information is sparse. Experimental results on the Synapse, ACDC, ISIC2017, ISIC2018, and BUSI datasets confirm that MAS-Net outperforms existing state-of-the-art approaches, achieving significant performance improvements. Code available at: <span><span>https://github.com/zz0226zz/MAS-Net</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"269 ","pages":"Article 104816"},"PeriodicalIF":3.5,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148185578","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":"A Driving Oriented Objectness score for detecting known and unknown objects in road scenes","authors":"Corentin Bunel,&nbsp;Maxime Gueriau,&nbsp;Gilles Gasso,&nbsp;Samia Ainouz","doi":"10.1016/j.cviu.2026.104821","DOIUrl":"10.1016/j.cviu.2026.104821","url":null,"abstract":"<div><div>In complex and dynamic environments such as road scenes, autonomous driving systems must make fully informed decisions despite being trained on a limited set of known object classes. This challenge, known as Open-Set Object Detection (OSOD), requires perception systems to detect unknown objects while maintaining reliable detection of critical known classes. Most open-set detectors rely on an objectness score as a first step to distinguish foreground objects from the background. Since this score determines which regions are further processed, its relevance has a direct and significant impact on the overall detection performance. However, our study on road scenes shows that previous attempts to improve this objectness score for unknown object detection degrade known-class detection. This trade-off is paramount for autonomous driving, where reliability on known objects is critical. Thus, this paper aims to develop an objectness score that preserves known-class detection performance while improving unknown object detection in road scenes. We first introduce a unified formulation of previous objectness scores and then propose a training procedure that excludes unlabeled unknowns from the loss calculation. To achieve this, we categorize detected boxes into four types and assign tailored losses. This results in the Driving Oriented Objectness (DROO) score, which is learned through our proposed training procedure. Additionally, we propose an evaluation protocol that isolates the impact of objectness scoring. Our results on CODA dataset, which features challenging corner-case unknowns in road scenes, demonstrate that the approach enhances unknown detection while maintaining accuracy on known classes.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"269 ","pages":"Article 104821"},"PeriodicalIF":3.5,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148185581","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":"Towards robust 3D human reconstruction with uncertainty-aware low-rank adaptation","authors":"Inho Chang,&nbsp;Jumi Kang,&nbsp;Yong-Hoon Kwon,&nbsp;Ju Hong Yoon,&nbsp;Min-Gyu Park","doi":"10.1016/j.cviu.2026.104813","DOIUrl":"10.1016/j.cviu.2026.104813","url":null,"abstract":"<div><div>We propose an uncertainty-aware 3D human reconstruction framework that enhances the robustness of reconstructing clothed humans from a single image, even under extreme conditions. To address this problem, we use Low-Rank Adapters (LoRA) to the depth prediction network, <em>i.e</em>., the encoder module, to adapt the pre-trained depth network for predicting depth maps for noisy input. Furthermore, we train a lightweight network that predicts a pixel-wise uncertainty map for each predicted depth map. With LoRA and the uncertainty prediction network, it is possible to predict a depth map and a corresponding uncertainty map that can be fused when multiple depth maps are given. We refer to this as the Uncertainty-aware Low-Rank Adapter (U-LoRA). Here, we employ multiple U-LoRAs with different noise distributions and fuse the outputs of U-LoRAs with the aid of predicted uncertainty maps. Comprehensive experiments demonstrate that our method can reveal underlying human shapes even in challenging scenarios. For reproducibility, we will release the source code to the public.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"269 ","pages":"Article 104813"},"PeriodicalIF":3.5,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148185582","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":"Are Candidate Models Really Needed for Active Learning?","authors":"Harshini Mridula Mohan ,&nbsp;Maanya Manjunath ,&nbsp;Vipul Arya ,&nbsp;Shabbeer Basha S.H. ,&nbsp;Nitin Cheekatla","doi":"10.1016/j.cviu.2026.104809","DOIUrl":"10.1016/j.cviu.2026.104809","url":null,"abstract":"<div><div>Deep learning has profoundly impacted domains such as computer vision and natural language processing by uncovering complex patterns in vast datasets. However, the reliance on extensive labeled data poses significant challenges, including resource constraints and annotation errors, particularly in training Convolutional Neural Networks (CNNs) and transformers due to a larger number of parameters. Active learning offers a promising solution to reduce labeling burdens by strategically selecting the most informative samples for annotation. However, the current active learning frameworks are time-intensive; select the samples iteratively with the help of initial candidate models. This study investigates the feasibility of using CNNs and transformers with randomly initialized weights, eliminating the need for initial candidate models while achieving results comparable to active learning frameworks that depend on such candidate models. We evaluate three confidence-based sampling strategies—high confidence (HC), low confidence (LC), and a combination of high confidence in the early stages of training and low confidence at later stages of training (HCLC). Among these, mostly LC demonstrated the best performance in our experiments, showcasing its effectiveness as an active learning strategy without the need for candidate models. Further, extensive experiments verify the robustness of the proposed active learning methods. By challenging traditional frameworks, the proposed work introduces a streamlined approach to active learning, advancing efficiency and flexibility across diverse datasets and domains. The code of this research is available here: <span><span>Source Code</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"269 ","pages":"Article 104809"},"PeriodicalIF":3.5,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148185576","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":"Text-guided context learning for multi-source domain adaptation","authors":"Bo Zhou,&nbsp;Long Liu,&nbsp;Chenyue Fan,&nbsp;Zhipeng Zhao","doi":"10.1016/j.cviu.2026.104818","DOIUrl":"10.1016/j.cviu.2026.104818","url":null,"abstract":"<div><div>Multi-source Unsupervised Domain Adaptation (MUDA) aims to address the domain transfer challenge between multiple labeled source domains and an unlabeled target domain by reducing domain discrepancies. Existing methods typically learn domain-invariant features in a shared feature space between the source and target domains. However, during multi-source distribution alignment, cross-domain distribution adjustment may shift the model’s attention away from category-discriminative regions. Moreover, most approaches focus primarily on global distribution alignment while neglecting the applicability of source domain supervision to the target domain. To address these issues, we propose a multi-source domain adaptation method called Text-guided Context Learning (TCL), which reformulates multi-source transfer as a cross-domain textual context optimization problem within a pretrained vision–language model. Specifically, we first develop a target domain style context-guided feature augmentation strategy. This strategy utilizes textual modal features that incorporate the style of the target domain to optimize the source domain features, improving the applicability of the source domain supervisory information to the target domain while preserving the category-discriminative information of the source domain visual features. Furthermore, TCL proposes active boundary optimization to ensure the class distinguishability of cross-domain context. By designing a boundary loss and a boundary sampling strategy, we guide the model to focus on difficult samples near decision boundaries during learning, ensuring the learned cross-domain context is more sensitive to inter-class differences. Experimental results demonstrate that our method effectively mitigates the shortcomings of prompt-tuning approaches in domain shift tasks and achieves competitive performance across multiple benchmark datasets. Additionally, our method can be effectively generalized to semi-supervised and partial-set domain adaptation tasks, further validating its robustness and versatility.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"269 ","pages":"Article 104818"},"PeriodicalIF":3.5,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148185579","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":"DiTalker: A unified DiT-based framework for high-quality and style-controllable portrait animation","authors":"He Feng ,&nbsp;Yongjia Ma ,&nbsp;Lei Fan ,&nbsp;Donglin Di ,&nbsp;Tonghua Su","doi":"10.1016/j.cviu.2026.104819","DOIUrl":"10.1016/j.cviu.2026.104819","url":null,"abstract":"<div><div>Portrait animation aims to synthesize talking videos from a static reference face, conditioned on audio and style frames, while ensuring precise lip synchronization and faithful reproduction of speaking styles. Existing diffusion-based portrait animation methods primarily focus on lip synchronization or static emotion transformation, often overlooking dynamic styles such as head movements. Moreover, most of these methods rely on a dual U-Net architecture, which preserves identity consistency but incurs additional architectural complexity. To this end, we propose DiTalker, a unified DiT-based framework for style-controllable portrait animation. We design a Style-Expression Encoding Module that employs two separate branches: a style branch extracting identity-specific facial dynamics, and an expression branch extracting global expression semantics. We further introduce an Audio-Style Fusion Module that processes audio and style features in parallel via two cross-attention layers, whose outputs are scaled and fused through element-wise addition to guide the animation process. To ensure the quality of animation results, we adopt and modify two optimization constraints: one to improve lip synchronization and the other to preserve fine-grained identity and background details. Extensive experiments demonstrate the superiority of DiTalker in terms of lip synchronization and speaking style controllability. Code is available at <span><span>https://github.com/fenghe12/DiTalker-CVIU/</span><svg><path></path></svg></span>.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"269 ","pages":"Article 104819"},"PeriodicalIF":3.5,"publicationDate":"2026-06-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"148185577","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":"GoT-CQA: Graph-of-Thought guided compositional reasoning for chart question answering","authors":"Lingling Zhang ,&nbsp;Muye Huang ,&nbsp;Qianying Wang ,&nbsp;Yaxian Wang ,&nbsp;Wenjun Wu ,&nbsp;Ziqi He ,&nbsp;Jun Liu","doi":"10.1016/j.cviu.2026.104745","DOIUrl":"10.1016/j.cviu.2026.104745","url":null,"abstract":"<div><div>Chart Question Answering (CQA) aims at answering questions based on the visual chart content, which plays an important role in chart summarization, business data analysis, and data report generation. CQA is a challenging multi-modal task because of the strong context dependence and complex reasoning requirement. The former refers to answering this question strictly based on the analysis of the visual content or internal data of the given chart, while the latter emphasizes the various logical and numerical reasoning involved in answer prediction process. In this paper, we pay more attention on the complex reasoning in CQA task, and propose a novel Graph-of-Thought (GoT) guided compositional reasoning model called GoT-CQA to overcome this problem. At first, we transform the chart-oriented question into a directed acyclic GoT composed of multiple operator nodes, including localization, numerical and logical operator. It reflects the human brain’s solution process to this question intuitively. After that, we design an efficient auto-compositional reasoning framework guided by the GoT, to execute the multi-step reasoning operations in various types of questions. Comprehensive experiments on ChartQA and PlotQA-D datasets show that GoT-CQA achieves outstanding performance, especially in complex human-written and reasoning questions, comparing with the latest popular baselines.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"268 ","pages":"Article 104745"},"PeriodicalIF":3.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147803114","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":"SAFIR: Spatially-aware Activation Function for Implicit Neural Representations","authors":"Ehsan Zeraatkar,&nbsp;Jelena Tešić","doi":"10.1016/j.cviu.2026.104754","DOIUrl":"10.1016/j.cviu.2026.104754","url":null,"abstract":"<div><div>Neural Implicit Representations (INRs) have reshaped a range of vision tasks by modeling signals as continuous functions, yet existing architectures, such as SIREN, suffer from spectral bias due to their reliance on spatially uniform frequency parameters. This work introduces SAFIR, a novel spatially aware adaptive frequency modulation framework that learns coordinate-dependent frequency parameters via a convolutional omega-prediction network. SAFIR enables adaptive allocation of high- and low-frequency components across spatial regions, thereby addressing the persistent problem of over-smoothing or overfitting in conventional INRs. Extensive evaluations on five benchmark datasets demonstrate that SAFIR not only offers substantial improvements in PSNR and SSIM for 2D image reconstruction, super-resolution, and generative vision tasks, but also achieves superior parameter efficiency and faster convergence than state-of-the-art alternatives. By effectively bridging local signal complexity with adaptive neural modulation, SAFIR represents a practical step toward high-fidelity, efficient signal representation in modern computer vision applications.</div></div>","PeriodicalId":50633,"journal":{"name":"Computer Vision and Image Understanding","volume":"268 ","pages":"Article 104754"},"PeriodicalIF":3.5,"publicationDate":"2026-05-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"147803106","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}