Yuxiang Cai
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yuxiang Cai has a total of 37 co-authors. Here are the top co-authors by number of shared publications:
- Jianwei Yin
School of Software Technology, Zhejiang University, Ningbo, China
11 shared publications - Yongheng Shang
School of Computer Science and Technology, Zhejiang University, Hangzhou, China
7 shared publications Bin Chen
Zhejiang University
3 shared publications- Zhi Chen
University of Southern Queensland, Toowoomba, Australia
3 shared publications - Jingcai Guo
The Hong Kong Polytechnic University, Hong Kong SAR, China
2 shared publications Xu Jia
College of Computer Science and Technology, Zhejiang University, Zhejiang, China
2 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1129-1138
Continual Video-MLLM Adaptation over Evolving Domains
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 1129-1138
Open AccessContinual Video-MLLM Adaptation over Evolving Domains
- 17 September 2026
- Published by Springer Nature
- p. 221-239
Spectral Evolution-Guided Token Pruning in Multimodal Large Language Models
- 17 September 2026
- Published by Springer Nature
- p. 221-239
Spectral Evolution-Guided Token Pruning in Multimodal Large Language Models
- 14 September 2026
- Published by Springer Nature
- p. 213-233
Accelerating Multimodal Large Language Models with Prior-Corrected Token Reduction
- 14 September 2026
- Published by Springer Nature
- p. 213-233
Accelerating Multimodal Large Language Models with Prior-Corrected Token Reduction
- 23 July 2026
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 134 (8) , 361
Federated Learning Meets Test-Time Adaptation: Methods, Challenges, and Future Directions
- 23 July 2026
- Published by Springer Nature in International Journal of Computer Vision
- Vol. 134 (8) , 361
Federated Learning Meets Test-Time Adaptation: Methods, Challenges, and Future Directions
- 26 July 2025
- Published by Springer Nature in Service Oriented Computing and Applications
- Vol. 20 (2) , 215-223
TSAVD-CLIP: One visual encoder learn both temporal and spatial feat
- 26 July 2025
- Published by Springer Nature in Service Oriented Computing and Applications
- Vol. 20 (2) , 215-223
TSAVD-CLIP: One visual encoder learn both temporal and spatial feat
- 18 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 13 (3) , 369-394
Enhancing Ultrahigh Resolution Remote Sensing Imagery Analysis With ImageRAG: A new framework
Ultrahigh resolution (UHR) remote sensing imagery (RSI) (e.g. 10,000 × 10,000 pixels) poses a significant challenge for current RS vision-language models (RSVLMs). If one chooses to resize the UHR image to the standard input image size, the extensive spatial and contextual information that UHR images contain will be neglected. Otherwise, the original size of these images often exceeds the token limits of standard RSVLMs, making it difficult to process the entire image and capture long-range dependencies to answer the query based on the abundant visual context. In this article, we introduce ImageRAG for RS, a framework to address the complexities of analyzing UHR RSI with a little training requirement. By transforming the UHR RS image analysis task to the image’s long-context selection task, we design an innovative image contextual retrieval mechanism based on the retrieval-augmented generation (RAG) technique, denoted as ImageRAG. ImageRAG’s core innovation lies in its ability to selectively retrieve and focus on the most relevant portions of the UHR image as visual contexts that pertain to a given query. Fast path and slow path modes are proposed in this framework to handle this task efficiently and effectively. ImageRAG allows RSVLMs to manage extensive context and spatial information from UHR RSI, ensuring that the analysis is both accurate and efficient. The code will be released at https://github.com/om-ai-lab/ImageRAG.
- 18 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Magazine
- Vol. 13 (3) , 369-394
Enhancing Ultrahigh Resolution Remote Sensing Imagery Analysis With ImageRAG: A new framework
Ultrahigh resolution (UHR) remote sensing imagery (RSI) (e.g. 10,000 × 10,000 pixels) poses a significant challenge for current RS vision-language models (RSVLMs). If one chooses to resize the UHR image to the standard input image size, the extensive spatial and contextual information that UHR images contain will be neglected. Otherwise, the original size of these images often exceeds the token limits of standard RSVLMs, making it difficult to process the entire image and capture long-range dependencies to answer the query based on the abundant visual context. In this article, we introduce ImageRAG for RS, a framework to address the complexities of analyzing UHR RSI with a little training requirement. By transforming the UHR RS image analysis task to the image’s long-context selection task, we design an innovative image contextual retrieval mechanism based on the retrieval-augmented generation (RAG) technique, denoted as ImageRAG. ImageRAG’s core innovation lies in its ability to selectively retrieve and focus on the most relevant portions of the UHR image as visual contexts that pertain to a given query. Fast path and slow path modes are proposed in this framework to handle this task efficiently and effectively. ImageRAG allows RSVLMs to manage extensive context and spatial information from UHR RSI, ensuring that the analysis is both accurate and efficient. The code will be released at https://github.com/om-ai-lab/ImageRAG.
- Open Access
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1168-1177
MultiDAN: Unsupervised, Multistage, Multisource and Multitarget Domain Adaptation for Semantic Segmentation of Remote Sensing Images
Unsupervised domain adaptation (UDA) has been a crucial way for cross-domain semantic segmentation of remote sensing images and reached apparent advents. However, most existing efforts focus on single source single target domain adaptation, which don't explicitly consider the serious domain shift between multiple source and target domains in real applications, especially inter-domain shift between various target domains and intra-domain shift within each target domain. In this paper, to address simultaneous inter-domain shift and intra-domain shift for multiple target domains, we propose a novel unsupervised, multistage, multisource and multitarget domain adaptation network (MultiDAN), which involves multisource and multitarget domain adaptation (MSMTDA), entropy-based clustering (EC) and multistage domain adaptation (MDA). Specifically, MSMTDA learns feature-level multiple adversarial strategies to alleviate complex domain shift between multiple target and source domains. Then, EC clusters the various target domains into multiple subdomains based on entropy of target predictions of MSMTDA. Besides, we propose a new pseudo label update strategy (PLUS) to dynamically produce more accurate pseudo labels for MDA. Finally, MDA aligns the clean subdomains, including pseudo labels generated by PLUS, with other noisy subdomains in the output space via the proposed multistage adaptation algorithm (MAA). The extensive experiments on the benchmark remote sensing datasets highlight the superiority of our MultiDAN against recent state-of-the-art UDA methods.
- 28 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 1168-1177
Open AccessMultiDAN: Unsupervised, Multistage, Multisource and Multitarget Domain Adaptation for Semantic Segmentation of Remote Sensing Images
Unsupervised domain adaptation (UDA) has been a crucial way for cross-domain semantic segmentation of remote sensing images and reached apparent advents. However, most existing efforts focus on single source single target domain adaptation, which don't explicitly consider the serious domain shift between multiple source and target domains in real applications, especially inter-domain shift between various target domains and intra-domain shift within each target domain. In this paper, to address simultaneous inter-domain shift and intra-domain shift for multiple target domains, we propose a novel unsupervised, multistage, multisource and multitarget domain adaptation network (MultiDAN), which involves multisource and multitarget domain adaptation (MSMTDA), entropy-based clustering (EC) and multistage domain adaptation (MDA). Specifically, MSMTDA learns feature-level multiple adversarial strategies to alleviate complex domain shift between multiple target and source domains. Then, EC clusters the various target domains into multiple subdomains based on entropy of target predictions of MSMTDA. Besides, we propose a new pseudo label update strategy (PLUS) to dynamically produce more accurate pseudo labels for MDA. Finally, MDA aligns the clean subdomains, including pseudo labels generated by PLUS, with other noisy subdomains in the output space via the proposed multistage adaptation algorithm (MAA). The extensive experiments on the benchmark remote sensing datasets highlight the superiority of our MultiDAN against recent state-of-the-art UDA methods.
- Open Access
- 26 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 2148-2158
Exploring High-Correlation Source Domain Information for Multi-Source Domain Adaptation in Semantic Segmentation
Multi-source domain adaptation (MSDA) aims to transfer knowledge from multiple source domains to one target domain. Although multi-source domains contain more complementary information than single source domain, MSDA involves some disturbed source samples, which will degrade the adaptation performance. To solve this problem, we propose a novel MSDA method for semantic segmentation. Specifically, to fully explore the optimal source samples for target domain, we propose a novel correlation measurement mechanism, weighing domain-level source-target correlation (DSC) and pixel-level source-target correlation (PSC). For each pair of source and target domains, DSC and PSC estimate the source-target correlations via the distances between target class prototypes and source class prototypes, and between target class prototypes and every pixel of source features, respectively. Built upon PSC, we propose a novel mix-up strategy, which pastes high-correlation source pixels to target images, to construct augmented mixing images for adaptation. Then we train the segmentor on the mixed images with pseudo labels and labeled source images, with DSC and PSC to suppress the negative effects of the low-correlation source domains and pixels. Furthermore, an attentive prototype alignment loss, based on DSC, is proposed to align target and multi-source domains, which attaches more importance to high-correlation source domains. The experimental results on the representative benchmark datasets (i.e., GTA5 and SYNTHIA → Cityscapes) highlight that our method substantially outperforms the state-of-the-art single-source domain adaptation and MSDA methods.
- 26 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 2148-2158
Open AccessExploring High-Correlation Source Domain Information for Multi-Source Domain Adaptation in Semantic Segmentation
Multi-source domain adaptation (MSDA) aims to transfer knowledge from multiple source domains to one target domain. Although multi-source domains contain more complementary information than single source domain, MSDA involves some disturbed source samples, which will degrade the adaptation performance. To solve this problem, we propose a novel MSDA method for semantic segmentation. Specifically, to fully explore the optimal source samples for target domain, we propose a novel correlation measurement mechanism, weighing domain-level source-target correlation (DSC) and pixel-level source-target correlation (PSC). For each pair of source and target domains, DSC and PSC estimate the source-target correlations via the distances between target class prototypes and source class prototypes, and between target class prototypes and every pixel of source features, respectively. Built upon PSC, we propose a novel mix-up strategy, which pastes high-correlation source pixels to target images, to construct augmented mixing images for adaptation. Then we train the segmentor on the mixed images with pseudo labels and labeled source images, with DSC and PSC to suppress the negative effects of the low-correlation source domains and pixels. Furthermore, an attentive prototype alignment loss, based on DSC, is proposed to align target and multi-source domains, which attaches more importance to high-correlation source domains. The experimental results on the representative benchmark datasets (i.e., GTA5 and SYNTHIA → Cityscapes) highlight that our method substantially outperforms the state-of-the-art single-source domain adaptation and MSDA methods.
- 26 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 61 (01962892) , 1-18
DASRSNet: Multitask Domain Adaptation for Super-Resolution-Aided Semantic Segmentation of Remote Sensing Images
Unsupervised domain adaptation (UDA) has become an important technique for cross-domain semantic segmentation (SS) in remote sensing community and obtained remarkable results. However, when transferring from high-resolution (HR) remote sensing images to low-resolution (LR) images, the existing UDA frameworks always fail to segment the LR target images, especially for small objects (e.g., cars), due to the severe spatial resolution shift problem. In this paper, to improve the segmentation ability of UDA models for LR target images and small objects, we propose a novel multi-task domain adaptation network (DASRSNet) for SS of remote sensing images with the aid of super-resolution (SR). The proposed DASRSNet contains domain adaptation for semantic segmentation (DASS) branch, domain adaptation for super-resolution (DASR) branch and feature affinity (FA) module. Specifically, the DASS and DASR branches share the same encoder to extract the domain-invariant features for the target and source domains, and these two branches utilize different decoders and discriminators to conduct cross-domain SS task and SR task, which align the domain shift in output space and image space respectively. Finally, FA module, which involves the proposed FA loss, is applied to enhance the affinity of SS features and SR features for both source and target domains. The experimental results on the cross-city aerial datasets demonstrate the effectiveness and superiority of our DASRSNet against the recent UDA models.
- 26 December 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 61 (01962892) , 1-18
DASRSNet: Multitask Domain Adaptation for Super-Resolution-Aided Semantic Segmentation of Remote Sensing Images
Unsupervised domain adaptation (UDA) has become an important technique for cross-domain semantic segmentation (SS) in remote sensing community and obtained remarkable results. However, when transferring from high-resolution (HR) remote sensing images to low-resolution (LR) images, the existing UDA frameworks always fail to segment the LR target images, especially for small objects (e.g., cars), due to the severe spatial resolution shift problem. In this paper, to improve the segmentation ability of UDA models for LR target images and small objects, we propose a novel multi-task domain adaptation network (DASRSNet) for SS of remote sensing images with the aid of super-resolution (SR). The proposed DASRSNet contains domain adaptation for semantic segmentation (DASS) branch, domain adaptation for super-resolution (DASR) branch and feature affinity (FA) module. Specifically, the DASS and DASR branches share the same encoder to extract the domain-invariant features for the target and source domains, and these two branches utilize different decoders and discriminators to conduct cross-domain SS task and SR task, which align the domain shift in output space and image space respectively. Finally, FA module, which involves the proposed FA loss, is applied to enhance the affinity of SS features and SR features for both source and target domains. The experimental results on the cross-city aerial datasets demonstrate the effectiveness and superiority of our DASRSNet against the recent UDA models.
- 31 August 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 60 (01962892) , 1-17
IterDANet: Iterative Intra-Domain Adaptation for Semantic Segmentation of Remote Sensing Images
When segmenting the continuous proliferation of unlabeled remotely sensed images, unsupervised domain adaptation (UDA) has become one of the most critical techniques and achieved significant performance. But, in fact, there still exists a large performance gap between the existing UDA frameworks and supervised learning methods, for the majority of UDA frameworks do not consider the intra-domain gap in the target domain. In this article, to further minimize the complex intra-domain shift within the target domain in remote sensing, we propose a novel iterative intra-domain adaptation framework (IterDANet), which conducts inter-domain adaptation (InterDA), entropy-based ranking (ER), and iterative intra-domain adaptation (IntraDA). Specifically, first, to enhance the performance of InterDA built upon generative adversarial network (GAN)-based image-to-image (I2I) translation, we propose a new generator selection strategy (GSS) to assess and choose a well-trained generator for the inter-domain classifier. Then, to produce more accurate pseudolabels for IntraDA, we propose a new pseudolabel generation strategy (PLGS) to remove both high-entropy and low-confident pixels in predicted maps of inter-domain classifier. Finally, to better reduce the intra-domain gap, we propose to cluster all the target images into multiple subdomains using ER and iteratively align the cleanest subdomain with other noisy subdomains. The extensive experiments on the benchmark dataset, which includes cross-city aerial images, highlight the superiority and effectiveness of our IterDANet against the state-of-the-art UDA frameworks.
- 31 August 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 60 (01962892) , 1-17
IterDANet: Iterative Intra-Domain Adaptation for Semantic Segmentation of Remote Sensing Images
When segmenting the continuous proliferation of unlabeled remotely sensed images, unsupervised domain adaptation (UDA) has become one of the most critical techniques and achieved significant performance. But, in fact, there still exists a large performance gap between the existing UDA frameworks and supervised learning methods, for the majority of UDA frameworks do not consider the intra-domain gap in the target domain. In this article, to further minimize the complex intra-domain shift within the target domain in remote sensing, we propose a novel iterative intra-domain adaptation framework (IterDANet), which conducts inter-domain adaptation (InterDA), entropy-based ranking (ER), and iterative intra-domain adaptation (IntraDA). Specifically, first, to enhance the performance of InterDA built upon generative adversarial network (GAN)-based image-to-image (I2I) translation, we propose a new generator selection strategy (GSS) to assess and choose a well-trained generator for the inter-domain classifier. Then, to produce more accurate pseudolabels for IntraDA, we propose a new pseudolabel generation strategy (PLGS) to remove both high-entropy and low-confident pixels in predicted maps of inter-domain classifier. Finally, to better reduce the intra-domain gap, we propose to cluster all the target images into multiple subdomains using ER and iteratively align the cleanest subdomain with other noisy subdomains. The extensive experiments on the benchmark dataset, which includes cross-city aerial images, highlight the superiority and effectiveness of our IterDANet against the state-of-the-art UDA frameworks.