Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CNG-SFDA:Clean-and-Noisy Region Guided Online-Offline Source-Free Domain Adaptation

Hyeonwoo Cho, Chanmin Park, Dong-Hee Kim, Jinyoung Kim, Won Hwa Kim

TL;DR

This work proposes to update cluster prototypes and their structure in the target domain formulated by the source model in online manners and shows that the method, CNG-SFDA, achieves state-of-the-art for most cases.

Abstract

Domain shift occurs when training (source) and test (target) data diverge in their distribution. Source-Free Domain Adaptation (SFDA) addresses this domain shift problem, aiming to adopt a trained model on the source domain to the target domain in a scenario where only a well-trained source model and unlabeled target data are available. In this scenario, handling false labels in the target domain is crucial because they negatively impact the model performance. To deal with this problem, we propose to update cluster prototypes (i.e., centroid of each sample cluster) and their structure in the target domain formulated by the source model in online manners. In the feature space, samples in different regions have different pseudo-label distribution characteristics affected by the cluster prototypes, and we adopt distinct training strategies for these samples by defining clean and noisy regions: we selectively train the target with clean pseudo-labels in the clean region, whereas we introduce mix-up inputs representing intermediate features between clean and noisy regions to increase the compactness of the cluster. We conducted extensive experiments on multiple datasets in online/offline SFDA settings, whose results demonstrate that our method, CNG-SFDA, achieves state-of-the-art for most cases. Code is available at https://github.com/hyeonwoocho7/CNG-SFDA.

CNG-SFDA:Clean-and-Noisy Region Guided Online-Offline Source-Free Domain Adaptation

TL;DR

This work proposes to update cluster prototypes and their structure in the target domain formulated by the source model in online manners and shows that the method, CNG-SFDA, achieves state-of-the-art for most cases.

Abstract

Domain shift occurs when training (source) and test (target) data diverge in their distribution. Source-Free Domain Adaptation (SFDA) addresses this domain shift problem, aiming to adopt a trained model on the source domain to the target domain in a scenario where only a well-trained source model and unlabeled target data are available. In this scenario, handling false labels in the target domain is crucial because they negatively impact the model performance. To deal with this problem, we propose to update cluster prototypes (i.e., centroid of each sample cluster) and their structure in the target domain formulated by the source model in online manners. In the feature space, samples in different regions have different pseudo-label distribution characteristics affected by the cluster prototypes, and we adopt distinct training strategies for these samples by defining clean and noisy regions: we selectively train the target with clean pseudo-labels in the clean region, whereas we introduce mix-up inputs representing intermediate features between clean and noisy regions to increase the compactness of the cluster. We conducted extensive experiments on multiple datasets in online/offline SFDA settings, whose results demonstrate that our method, CNG-SFDA, achieves state-of-the-art for most cases. Code is available at https://github.com/hyeonwoocho7/CNG-SFDA.
Paper Structure (13 sections, 17 equations, 5 figures, 9 tables)

This paper contains 13 sections, 17 equations, 5 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Domain Adaptation Scenarios
  4. Online-Offline Source-Free Domain Adaptation
  5. The Proposed Method
  6. Distinguishing Clean and Noisy Regions in Clusters
  7. Training Strategy for Clean and Noisy Regions
  8. Instance and Prototype-aware Contrastive Learning
  9. Experiments
  10. Experimental Setup
  11. Experimental Results
  12. Analysis and Discussion
  13. Conclusion

Figures (5)

  • Figure 1: The upper left presents the feature distribution of clean samples on the source domain. The upper right shows the previous method that utilizes fixed class prototypes in offline manner and learn nearest neighborhoods of class prototypes. The bottom depicts the process of CNG-SFDA that updates class prototypes in online manners and learns intermediate features in gray zone with progressive weight.
  • Figure 2: Framework of CNG-SFDA (a): partition the cluster into 'Clean' (close to the cluster prototype) and 'Noisy' (far from the cluster prototype) region based on the clean probability distribution. (b): training samples in clean regions with pseudo-labels generated from prediction of closest features. (c): reducing a gap between clean and noisy region with mix-up weight. (d): prototype/instance-aware contrastive learning.
  • Figure 3: Feature distribution of clean and noisy samples in target domain. Here, 'clean' and 'noisy' refer to cases where the Ground-Truth (GT) and the Pseudo-Label (PL) match and mismatch. Circle and triangular denote clean and noisy samples predicted by CNG-SFDA. For clean samples (circle), each color represent the GT class. For noisy samples (triangular), pink and black triangular represent cases where CNG-SFDA correctly (i.e., Prediction: Noisy, GT: Noisy) and incorrectly (i.e., Prediction: Noisy, GT: Clean) identifies noisy data.
  • Figure 4: Qualitative evaluation of samples in predicted clean and noisy regions by CNG-SFDA at each target domain (each row). PL and CP are Pseudo-Label and Clean Probability estimated by CNG-SFDA, respectively.
  • Figure 5: Robustness analysis of different hyper-parameters.