Research Team Led by Professor Cho Seong-in of the Department of Artificial Intelligence Has Paper Accepted at IEEE/CVF CVPR 2026
A collaborative research paper titled “Measure The Feature Universe: Topology-based Pseudo Labeling and Gravity Consistency for Source-Free Domain Adaptation,” authored by Professor Cho Seong-in’s research team (Master’s students Lee Jae-yoon and Nam Hyuk) from the Department of Artificial Intelligence, has been accepted for publication at the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026, the world’s most prestigious conference in the field of computer vision.
This paper addresses the performance degradation inherent in Source-Free Domain Adaptation (SFDA) environments. In particular, it alleviates the difficulty of achieving robust adaptation performance by relying solely on target domain data when source data is unavailable.
To address these SFDA challenges, the research team proposes “Topology-based Pseudo Labeling,” which enables high-confidence pseudo-labeling by reflecting the model's feature space structure, along with “Gravity Consistency” to enhance training stability.
The research team’s method consists of two main stages, as illustrated in the figure below: ▲ generating high-confidence pseudo labels, and ▲ training the model stably based on those labels. In the pseudo-labeling stage, to minimize mislabeling risks near boundaries or in sparse regions of the target feature space, the method combines virtual feature generation, feature traversal, and reliable area definition. This ensures more stable identification of samples near class centers and enhances the reliability of the pseudo-labels.
Subsequently, during the training phase, the model enforces consistency to ensure that its features and logits do not vary significantly between weak and strong augmentation inputs. In particular, the proposed “Gravity Consistency” adjusts the learning signal according to prediction reliability, thereby mitigating training instability caused by uncertain samples.
Finally, the study conducted experiments on representative SFDA benchmark datasets, such as Office-Home, VisDA-C, and DomainNet-126, and validated the effectiveness of the proposed method.
(Top) Topology-based pseudo-labeling via Gaussian-based virtual feature generation, confidence region scheduling, and manifold-aware distance measurement
(Bottom) Target training based on weak/strong augmentation, and the design of feature/logit and gravity consistency signals
According to the CVPR 2026 Program Committee announcement, 16,092 papers underwent the review process this year (excluding withdrawals and desk rejections), of which 4,090 were accepted, resulting in an acceptance rate of 25.42%. Meanwhile, the Poster / Highlights / Oral categories will be announced separately at a later date.
CVPR 2026 is scheduled to be held from June 3 to 7, 2026, in Denver, Colorado, USA.
Title: Measure The Feature Universe: Topology-based Pseudo Labeling and Gravity Consistency for Source-Free Domain Adaptation
Author Information: Lee Jae-yoon (Co-first author, Sogang University), Nam Hyuk (Co-first author, Sogang University), Cho Seong-in (Corresponding author, Sogang University)
▶IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 Website: https://sites.google.com/view/csi2267svm/
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 Website: https://sites.google.com/view/csi2267svm/
[SEO 키워드]
CVPR 2026, Source-Free Domain Adaptation, Pseudo Labeling
[Summary]
The research team secured a paper acceptance at CVPR 2026 for their study on Source-Free Domain Adaptation, introducing a new Pseudo Labeling method to enhance model stability and accuracy.
This paper addresses the performance degradation inherent in Source-Free Domain Adaptation (SFDA) environments. In particular, it alleviates the difficulty of achieving robust adaptation performance by relying solely on target domain data when source data is unavailable.
To address these SFDA challenges, the research team proposes “Topology-based Pseudo Labeling,” which enables high-confidence pseudo-labeling by reflecting the model's feature space structure, along with “Gravity Consistency” to enhance training stability.
The research team’s method consists of two main stages, as illustrated in the figure below: ▲ generating high-confidence pseudo labels, and ▲ training the model stably based on those labels. In the pseudo-labeling stage, to minimize mislabeling risks near boundaries or in sparse regions of the target feature space, the method combines virtual feature generation, feature traversal, and reliable area definition. This ensures more stable identification of samples near class centers and enhances the reliability of the pseudo-labels.
Subsequently, during the training phase, the model enforces consistency to ensure that its features and logits do not vary significantly between weak and strong augmentation inputs. In particular, the proposed “Gravity Consistency” adjusts the learning signal according to prediction reliability, thereby mitigating training instability caused by uncertain samples.
Finally, the study conducted experiments on representative SFDA benchmark datasets, such as Office-Home, VisDA-C, and DomainNet-126, and validated the effectiveness of the proposed method.
(Top) Topology-based pseudo-labeling via Gaussian-based virtual feature generation, confidence region scheduling, and manifold-aware distance measurement
(Bottom) Target training based on weak/strong augmentation, and the design of feature/logit and gravity consistency signals
According to the CVPR 2026 Program Committee announcement, 16,092 papers underwent the review process this year (excluding withdrawals and desk rejections), of which 4,090 were accepted, resulting in an acceptance rate of 25.42%. Meanwhile, the Poster / Highlights / Oral categories will be announced separately at a later date.
CVPR 2026 is scheduled to be held from June 3 to 7, 2026, in Denver, Colorado, USA.
Title: Measure The Feature Universe: Topology-based Pseudo Labeling and Gravity Consistency for Source-Free Domain Adaptation
Author Information: Lee Jae-yoon (Co-first author, Sogang University), Nam Hyuk (Co-first author, Sogang University), Cho Seong-in (Corresponding author, Sogang University)
▶IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 Website: https://sites.google.com/view/csi2267svm/
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 Website: https://sites.google.com/view/csi2267svm/
[SEO 키워드]
CVPR 2026, Source-Free Domain Adaptation, Pseudo Labeling
[Summary]
The research team secured a paper acceptance at CVPR 2026 for their study on Source-Free Domain Adaptation, introducing a new Pseudo Labeling method to enhance model stability and accuracy.