The joint research paper 'Measure The Feature Universe: Topology-based Pseudo Labeling and Gravity Consistency for Source-Free Domain Adaptation' by Lee Jae-yoon, Nam Hyuk and Cho Sung-in has been accepted as a regular paper at CVPR 2026 (IEEE/CVF Conference on Computer Vision and Pattern Recognition), the most prestigious international conference in computer vision.

The paper addresses the performance degradation that arises in source-free domain adaptation (SFDA). In particular, it alleviates the difficulty of securing stable adaptation performance from target domain data alone when source data is unavailable.

To solve this SFDA problem, the team proposes topology-based pseudo labeling, which reflects the structure of the feature space as the model understands it so that reliable pseudo labeling is possible, together with gravity consistency to improve training stability.

As shown in Figure 1, the method consists broadly of (1) a stage that produces reliable pseudo labels and (2) a stage that trains the model stably on that basis. In the pseudo labelling stage, virtual feature generation, feature traversal and reliable area setting are used together to reduce the risk of mislabelling where the target feature space is empty (sparse) or near boundaries. This finds samples around class centres more stably and raises the reliability of the pseudo labels.

In the training stage, the model then learns consistency so that its features and final predictions (logits) do not differ greatly between weakly augmented and strongly augmented inputs. The proposed gravity consistency in particular modulates the training signal according to prediction confidence, alleviating the training instability caused by uncertain samples.

Finally, the paper reports experiments on representative SFDA benchmark datasets including Office-Home, VisDA-C and DomainNet-126, verifying the effectiveness of the proposed method.

Figure 1. Overview of the Measure The Feature Universe framework. (Top) Topology-based pseudo labeling through Gaussian-based virtual feature generation, reliable area scheduling and manifold-aware distance measurement. (Bottom) Target training based on weak/strong augmentation, with the design of feature/logit consistency and the gravity consistency signal.

According to the announcement of the CVPR 2026 programme committee, 16,092 papers went through the review process this year (excluding withdrawals and desk rejects), of which 4,090 were accepted, an acceptance rate of 25.42%. The poster / highlight / oral designations will be announced separately at a later date.

CVPR 2026 will be held in Denver, Colorado, United States, from Wednesday 3 June to Sunday 7 June 2026.

[References]

Paper title: Measure The Feature Universe: Topology-based Pseudo Labeling and Gravity Consistency for Source-Free Domain Adaptation

Authors: Lee Jae-yoon (joint first author, Sogang University), Nam Hyuk (joint first author, Sogang University), Cho Sung-in (corresponding author, Sogang University)

IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) 2026 Website

Website : https://sites.google.com/view/csi2267svm/