IEEE Transactions on Circuits and Systems for Video Technology最新文献_第4页

A Masked Pretraining Framework With Multi-Scale Token Embedding and Multi-Head Conv-Attention for Hyperspectral Image Classification 高光谱图像分类的多尺度标记嵌入和多头反关注掩膜预训练框架

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-07-30 DOI: 10.1109/TCSVT.2026.3718651

Yu Xiang;Xiaodong Yu;Hongbin Dong;Shuying Zang

{"title":"A Masked Pretraining Framework With Multi-Scale Token Embedding and Multi-Head Conv-Attention for Hyperspectral Image Classification","authors":"Yu Xiang;Xiaodong Yu;Hongbin Dong;Shuying Zang","doi":"10.1109/TCSVT.2026.3718651","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3718651","url":null,"abstract":"Hyperspectral image (HSI) classification benefits from rich spectral-spatial information but suffers from limited labeled samples. Recent masked image modeling (MIM) methods provide a promising self-supervised solution. However, existing HSI MIM methods typically inherit Transformer-based architectures originally designed for natural images, which fail to fully exploit the intrinsic spectral continuity and spatial structures of hyperspectral data. Consequently, these approaches often suffer from limited efficiency and suboptimal representation learning under low-label regimes. To address these challenges, this paper proposes M3Net, an efficient masked pretraining framework specifically designed for hyperspectral image classification under limited labeled samples. M3Net constructs multi-scale spectral-spatial representations at the token embedding stage without introducing additional branches, replaces global self-attention with a convolution-based multi-head attention encoder to filter features, and adopts a lightweight reconstruction head for efficient masked pretraining. Extensive experiments conducted on three benchmark datasets, including Indian Pines, Pavia University, and WHU-Hi-HanChuan, demonstrate the effectiveness and robustness of the proposed framework. Under the challenging setting with only five labeled samples per class, M3Net achieves overall accuracies of 78.19%, 87.81%, and 79.87%, respectively, consistently outperforming existing supervised and masked pretraining methods. Moreover, M3Net significantly reduces pretraining time while maintaining competitive classification performance without requiring dataset-specific hyperparameter tuning, highlighting its efficiency and generalization capability across different hyperspectral scenes.","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"15996-16009"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149097700","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Spatial-Frequency Awareness for Low-Light Instance Segmentation 基于空间频率感知的弱光实例分割

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-07-06 DOI: 10.1109/TCSVT.2026.3710188

Chang Liu;Junbao Zhuo;Bochao Zou;Jiansheng Chen;Qianchuan Zhao;Huimin Ma

{"title":"Spatial-Frequency Awareness for Low-Light Instance Segmentation","authors":"Chang Liu;Junbao Zhuo;Bochao Zou;Jiansheng Chen;Qianchuan Zhao;Huimin Ma","doi":"10.1109/TCSVT.2026.3710188","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3710188","url":null,"abstract":"The low-light instance segmentation task tends to suffer from degraded performance due to contrast deficiency and noise interference. Most existing methods primarily rely on image enhancement preprocessing or spatial-domain architectures, which suffer from semantic representation deficiency and cross-frequency noise issues. To address these issues, we propose SFNet, which innovatively integrates spatial-frequency awareness cues into the low-light instance segmentation task. Specifically, for the semantic representation deficiency issue, we introduce a cross-frequency fusion unit (CFU) that decomposes spatial features into high- and low-frequency components via the stationary wavelet transform (SWT) and then fuses them with frequency-aware attention to enrich the semantic representation. Furthermore, we design a multi-domain mixer (MDM) for complementary spatial-frequency feature fusion and enhancement. To tackle the cross-frequency noise issue, we design a learnable adaptive frequency-domain filtering mechanism after the SWT in the CFU, which effectively suppresses noise in both low- and high-frequency domains. In addition, we design a novel frequency consistency constraint (FCC) to reduce noise interference across different frequency bands. Our method effectively suppresses noise interference while enhancing the complementarity between spatial and frequency features. Extensive experiments on the LIS dataset demonstrate substantial accuracy improvements, and ablation studies further verify the effectiveness of each proposed component. The code is made publicly available at: <uri>https://github.com/LuCky-cv-14/SFNet</uri>.","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"15275-15290"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149097750","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Multi-Modal Fine-Grained Discriminative Cues Transfer for Efficient Person Re-Identification 多模态细粒度鉴别线索转移对人的有效再识别

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-06-09 DOI: 10.1109/TCSVT.2026.3701704

Shenyu Xu;Boqian Liu;Hong Zhang;Haojie Li;Yi Wei

{"title":"Multi-Modal Fine-Grained Discriminative Cues Transfer for Efficient Person Re-Identification","authors":"Shenyu Xu;Boqian Liu;Hong Zhang;Haojie Li;Yi Wei","doi":"10.1109/TCSVT.2026.3701704","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3701704","url":null,"abstract":"In Person Re-identification (ReID), multi-modal features such as text can complement image features with discriminative semantic cues. Existing methods use CLIP to optimize learnable textual prompts, which are then utilized to enhance the features for ReID. However, these textual prompts are often susceptible to background noise and struggle to capture fine-grained discriminative attributes. Moreover, during the inference phase, networks of the multi-modal, such as the text encoder, must remain operational alongside the image encoder, thereby increasing computational overhead. In this paper, we design a Multi-Modal Fine-grained Discriminative Cues Transfer model for efficient person ReID. To address the limitation of existing text modal in capturing fine-grained discriminative information, we propose a Multi-Modal Fine-grained Knowledge Extraction (MMFE) module. This module leverages pre-trained foundation models to generate detailed person attribute descriptions and anatomical structure characterizations, and encodes both modals into multi-modal representations for enhancing the visual features. For improved inference efficiency, we design an Explicit-Implicit combined Knowledge Injection (EIKI) module. Through explicit feature alignment and implicit multi-modal attention mechanism, EIKI guides the image encoder to focus on fine-grained discriminative semantics and structural information within images. This design enables the method to effectively utilize multi-modal knowledge while discarding auxiliary modalities-related networks during inference. Experimental results on multiple public datasets demonstrate that our approach not only enhances ReID accuracy but also maintains computational efficiency with reduced resource consumption, making it particularly suitable for edge devices.","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"15804-15817"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149097847","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Human-Like Single Image Reflection Removal via Deep Unfolding With Large Kernels 基于大核深度展开的类人单图像反射去除

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-06-17 DOI: 10.1109/TCSVT.2026.3704623

Haoyang Zhang;Junkang Zhang;Faming Fang;Tingting Wang;Guixu Zhang;Haichuan Song

{"title":"Human-Like Single Image Reflection Removal via Deep Unfolding With Large Kernels","authors":"Haoyang Zhang;Junkang Zhang;Faming Fang;Tingting Wang;Guixu Zhang;Haichuan Song","doi":"10.1109/TCSVT.2026.3704623","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3704623","url":null,"abstract":"Single image reflection removal aims to remove the reflection part from a reflection image. Existing approaches predominantly rely on either mathematical modeling or deep learning techniques. However, their reflection removal methods often neglect the global understanding inherent to human visual perception when observing reflected scenes. The human visual system employs a “global-to-local” hierarchical processing: first rapidly comprehending global scene information, then progressively focusing on local details to effectively distinguish real scenes from reflection images. Inspired by this observation, we developed LKCDU-Net, a deep unfolding model that innovatively apply large kernel convolution. This allows our model to effectively capture and utilize the global context of reflection information, while combining flexibility and interpretability. Specifically, we have meticulously crafted an initialization module for our model. This module employs large kernel convolutions to emulate the global perception capabilities of the human visual system, thereby capturing a more holistic view of the image. We then build an optimization model for local details refinement, improving the performance of our model through the deep unfolding technology. Benefiting from the initialization module, our deep unfolding model is capable of providing powerful performance with a small number of parameters. Extensive experiments on real-world reflection images show that LKCDU-Net outperforms SOTA models, effectively removing reflections while preserving details, which demonstrates its superior performance. The source code and pre-trained model will be released soon at <uri>https://github.com/ZhangHaoyang493/HumanLikeSIRR/tree/main</uri>","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"14974-14990"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149097918","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

Post-Training Quantization for Hybrid Iterative Generative Models 混合迭代生成模型的训练后量化

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-07-07 DOI: 10.1109/TCSVT.2026.3710967

Jing Gao;Junyi Wu;Wei Wang;Yan Yan;Yao Zhao

{"title":"Post-Training Quantization for Hybrid Iterative Generative Models","authors":"Jing Gao;Junyi Wu;Wei Wang;Yan Yan;Yao Zhao","doi":"10.1109/TCSVT.2026.3710967","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3710967","url":null,"abstract":"Iterative Generative Models (IGMs) span autoregressive and diffusion paradigms, and hybrid variants that couple them can achieve remarkable image-generation fidelity. However, their iterative inference incurs substantial computational overhead, making Post-training Quantization (PTQ) appealing for acceleration, while directly applying vanilla PTQ to hybrid IGMs can trigger model collapse. By analyzing these failures, we identify two critical challenges: Excessive Outliers (EOs) in the activations create an irreconcilable trade-off between preserving normal precision and covering EOs, resulting in severe degradation in generation quality; Amplified Anomalies (AAs) arising unpredictably from minor quantization errors, create a mismatch between calibration and inference, thus iteratively triggering model collapse. To address these challenges, we introduce HyGenQ, a PTQ framework for hybrid IGMs. HyGenQ comprises Hierarchical Cluster Decoupling (HCD) and Scaling Recalibration (SR). HCD identifies and decouples outlier channels via a multi-stage clustering process, effectively isolating EOs while maintaining normal value precision, thereby alleviating performance degradation. SR scales AAs beyond Gaussian Bound, thereby avoiding model collapse caused by aggressive truncation. Extensive experiments demonstrate that HyGenQ successfully quantizes representative hybrid IGMs to 8-bit precision (W8A8), significantly outperforming existing baselines and validating its robustness across different model families.","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"15332-15346"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149097882","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

ART-VOD: Adaptive Relational Distillation and Temporal Modeling for Open-Vocabulary Video Object Detection ART-VOD：开放词汇视频目标检测的自适应关系蒸馏和时间建模

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-06-30 DOI: 10.1109/TCSVT.2026.3708773

Yang Cao;Junlong Gao;Yan Yan;Hanzi Wang

{"title":"ART-VOD: Adaptive Relational Distillation and Temporal Modeling for Open-Vocabulary Video Object Detection","authors":"Yang Cao;Junlong Gao;Yan Yan;Hanzi Wang","doi":"10.1109/TCSVT.2026.3708773","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3708773","url":null,"abstract":"Open-vocabulary video object detection aims to recognize and localize objects in videos from an unrestricted set of categories. Despite recent advances driven by knowledge transfer from large-scale vision–language models, existing methods still suffer from insufficient relational modeling, unreliable similarity matching, and inaccurate feature aggregation, which collectively hinder discriminative semantic learning and temporal coherence. To address these challenges systematically, we propose ART-VOD, an adaptive relational distillation and temporal modeling method designed to strengthen representation learning, association robustness, and temporal consistency. Specifically, we introduce an Adaptive Relational Contrastive Distillation (ARCD) module, which integrates information entropy weighting, curricular negative sample modulation, and relational contrastive loss to capture rich contextual dependencies during feature distillation, thereby improving knowledge transfer and enhancing discriminative capacity. In addition, we develop a Current Distribution-Aligned Matching (CDAM) module that establishes cross-frame correspondences by aligning the matching process with the embedding distribution of the current frame, effectively mitigating distributional drift and boosting association accuracy. Finally, we present a Semantic Relevance-Based Feature Aggregation (SRFA) module that dynamically aggregates the matched region embeddings by jointly considering frame-level reliability and region-level quality. This selective aggregation suppresses irrelevant information and enriches temporal representations. Extensive experiments on multiple datasets validate the effectiveness of our method. Notably, the proposed ART-VOD surpasses OV-VOD by 3.9% in <inline-formula> <tex-math>$text{AP}_{n}$ </tex-math></inline-formula> on the LV-VIS dataset, demonstrating its superiority over existing methods.","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"14535-14548"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149096966","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

FSUS-Seg: A Fusion-Spectral-Driven Universal Medical Ultrasound Image Segmentation Network FSUS-Seg：一种融合光谱驱动的通用医学超声图像分割网络

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-07-06 DOI: 10.1109/TCSVT.2026.3707106

Boheng Zhang;Haorui Huang;Jialin Li;Cailing Lin;Rui Fan;Mingjian Sun

{"title":"FSUS-Seg: A Fusion-Spectral-Driven Universal Medical Ultrasound Image Segmentation Network","authors":"Boheng Zhang;Haorui Huang;Jialin Li;Cailing Lin;Rui Fan;Mingjian Sun","doi":"10.1109/TCSVT.2026.3707106","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3707106","url":null,"abstract":"Medical ultrasound images are often degraded by severe noise, low contrast and ambiguous anatomical boundaries. These characteristics limit segmentation methods that rely solely on spatial features, as they struggle to capture global context, local detail and precise boundary information. Spectral-domain modeling complements spatial features by capturing frequency-distribution patterns and multi-scale structural details. However, accurate segmentation of complex ultrasound images requires coordinated modeling of global frequency context, localized structural cues and long-range dependencies across network stages. To this end, we propose FSUS-Seg, a fusion-spectral driven universal network for medical ultrasound image segmentation with stage-specific spatial-spectral modeling. In the encoder, we design a Fast Spatial-Spectral Convolution (FSSC) block to extract and fuse local and global features through joint convolutional and Fourier-domain modeling, enabling coupled representations in the spatial and spectral domains. At the skip connections, we introduce a Wavelet-Mamba Cross-Scale Fusion (WMCSF) module that integrates multi-resolution global semantics with fine-grained details in the spectral domain. By exploiting Mamba’s capacity for efficient long-range dependency modeling, WMCSF helps bridge the semantic gap between the encoder and decoder. At the bottleneck, a Wavelet-Mamba module further strengthens the joint modeling of global semantics and boundary precision through discrete wavelet transform-based decomposition and reconstruction. Extensive experiments on four public ultrasound datasets and one external dataset show that FSUS-Seg consistently outperforms state-of-the-art methods. Its superior performance on an in-house photoacoustic multi-organ segmentation dataset further demonstrates strong cross-scenario generalization ability.","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"15900-15916"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149097136","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

BPA-Net: Bidirectional Prototype Alignment for Image Captioning BPA-Net：图像标题的双向原型对齐

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-06-10 DOI: 10.1109/TCSVT.2026.3701826

Tianci Li;Cheng Xu;Yizhe Song;Jiazheng Yuan;Songyin Dai;Jiancheng Zhang

{"title":"BPA-Net: Bidirectional Prototype Alignment for Image Captioning","authors":"Tianci Li;Cheng Xu;Yizhe Song;Jiazheng Yuan;Songyin Dai;Jiancheng Zhang","doi":"10.1109/TCSVT.2026.3701826","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3701826","url":null,"abstract":"Image captioning aims to generate accurate textual descriptions for visual inputs. A major challenge is the semantic gap between continuous visual features and discrete linguistic concepts. Existing methods often rely on the unidirectional injection of external textual priors, which inevitably introduces semantic noise. Furthermore, CLIP-based image captioning methods lack direct supervision from ground-truth captions, leading to weak cross-modal alignment. To address these issues, we propose BPA-Net, a framework for bidirectional prototype alignment. Specifically, the Dual Prototype Extraction module independently projects visual features and ground-truth text tokens into a shared semantic space to form structured prototypes. Then, the Cross-Modal Prototype Alignment module enables bidirectional interaction and alignment between visual and textual prototypes guided by symmetric matching objectives. This approach effectively mitigates the noise inherent in external proxy texts and establishes robust semantic alignment. Extensive experiments on the MS-COCO, Flickr8k, and Flickr30k benchmarks demonstrate that BPA-Net achieves state-of-the-art performance. Furthermore, evaluations on the Nocaps dataset confirm its strong generalization capability for describing novel objects. The code is available at <uri>https://github.com/CarasuLee/BPA-Net</uri>","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"14710-14725"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149097280","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

TLight: A Triple-Modality Benchmark for Extreme Low-Light Enhancement TLight：超低光增强的三模态基准

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-06-08 DOI: 10.1109/TCSVT.2026.3701028

Shaonan Liu;Guoshun Nan;Leisen Niu;Ke Xu;Xiaofeng Tao;Xudong Jiang

{"title":"TLight: A Triple-Modality Benchmark for Extreme Low-Light Enhancement","authors":"Shaonan Liu;Guoshun Nan;Leisen Niu;Ke Xu;Xiaofeng Tao;Xudong Jiang","doi":"10.1109/TCSVT.2026.3701028","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3701028","url":null,"abstract":"Extreme Low-Light Enhancement (ELLE) aims to improve the visibility and quality of images captured under very low-illuminance conditions, typically below 0.1 lux. Existing learning-based methods commonly learn a direct mapping from a low-light image to a clean image. However, these methods struggle to recover visual details when they are overwhelmed by noise. Intuitively, the near-infrared image is robust to illumination variations such that is able to provide more structural cues under extremely low-illuminance conditions. Motivated by this, we propose TLight, a triple-modality (infrared-RAW-RGB) benchmark for RAW-domain ELLE. Our TLight consists of 300 indoor and outdoor scenarios, with each scenario comprising a triplet of three images: 1) a low-light RAW image as the input, 2) a normal-light RGB as the groundtruth and 3) an infrared image as additional input that is invariant to lighting conditions to provide supplementary structural cues. On top of our TLight, we develop a dual-fusion network (DFN) with a dual-branch encoder to exploit the complementary information in RAW and infrared images. Following DFN, a trainable ISP module is constructed to mimic the traditional ISP pipeline to obtain the final result. Extensive experimental results validate the effectiveness of our method, which achieves state-of-the-art accuracy on diverse benchmark datasets while maintaining competitive efficiency.","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"15382-15395"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149097321","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0

M-Mask: A Unified Parameter-Free Multi-Dimensional Masking Strategy for Universal Video Segmentation M-Mask：用于通用视频分割的统一无参数多维掩蔽策略

IF 10.8 1区工程技术

IEEE Transactions on Circuits and Systems for Video Technology Pub Date : 2026-10-01 Epub Date: 2026-07-08 DOI: 10.1109/TCSVT.2026.3711186

Ting Liu;Jiamin Cao;Chen Liang;Yuyan Li;Mingxiu Zhang;Luoqi Liu;Xiang Li

{"title":"M-Mask: A Unified Parameter-Free Multi-Dimensional Masking Strategy for Universal Video Segmentation","authors":"Ting Liu;Jiamin Cao;Chen Liang;Yuyan Li;Mingxiu Zhang;Luoqi Liu;Xiang Li","doi":"10.1109/TCSVT.2026.3711186","DOIUrl":"https://doi.org/10.1109/TCSVT.2026.3711186","url":null,"abstract":"State-of-the-art (SOTA) approaches in universal video segmentation often rely on a decoupled “segment, track, then refine” paradigm. Performance gains are typically sought by adding extra modules or auxiliary losses, which increases architectural complexity. In this work, we shift the focus from model optimization to the training strategy and propose M-Mask, a unified and trainable parameter-free <bold>M</b>ulti-dimensional <bold>Mask</b>ing strategy that systematically improves segmentation across three key dimensions, spatial, temporal, and object. The core principle of M-Mask is to train the model on masked inputs while supervising it with the original, complete ground-truth. This discrepancy compels the model to infer the missing information by leveraging the surrounding context, which inherently enhances its contextual awareness and discriminative feature extraction capabilities. M-Mask applies this principle across three dimensions: First, Spatial Patch Masking (SPM) randomly occludes image patches. Second, Temporal Consistency Masking (TCM) removes tubelet regions across frames, guiding the model to exploit motion continuity for stable instance association. Finally, Object Confidence Masking (OCM) dynamically prunes background-related queries before cross-attention, purifying spatiotemporal context and improving foreground precision. M-Mask is model-agnostic, parameter-free, and adds zero inference overhead. Extensive experiments demonstrate its generality and effectiveness. Notably, M-Mask delivers substantial gains on the challenging OVIS dataset, improving AP by 5.0%, 1.5%, and 0.4% under different architectures, and establishes new state-of-the-art results on both OVIS and VSPW datasets. These results underscore that improving the training process through multi-dimensional masking offers an efficient and effective toward achieving better universal video segmentation.","PeriodicalId":13082,"journal":{"name":"IEEE Transactions on Circuits and Systems for Video Technology","volume":"36 10","pages":"15917-15930"},"PeriodicalIF":10.8,"publicationDate":"2026-10-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"149097438","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":1,"RegionCategory":"工程技术","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

引用次数: 0