Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

De-Confusing Pseudo-Labels in Source-Free Domain Adaptation

Idit Diamant, Amir Rosenfeld, Idan Achituve, Jacob Goldberger, Arnon Netzer

TL;DR

A novel noise-learning approach tailored to address noise distribution in domain adaptation settings and learn to de-confuse the pseudo-labels to capture the label corruption of each class and learn the underlying true label distribution.

Abstract

Source-free domain adaptation aims to adapt a source-trained model to an unlabeled target domain without access to the source data. It has attracted growing attention in recent years, where existing approaches focus on self-training that usually includes pseudo-labeling techniques. In this paper, we introduce a novel noise-learning approach tailored to address noise distribution in domain adaptation settings and learn to de-confuse the pseudo-labels. More specifically, we learn a noise transition matrix of the pseudo-labels to capture the label corruption of each class and learn the underlying true label distribution. Estimating the noise transition matrix enables a better true class-posterior estimation, resulting in better prediction accuracy. We demonstrate the effectiveness of our approach when combined with several source-free domain adaptation methods: SHOT, SHOT++, and AaD. We obtain state-of-the-art results on three domain adaptation datasets: VisDA, DomainNet, and OfficeHome.

De-Confusing Pseudo-Labels in Source-Free Domain Adaptation

TL;DR

A novel noise-learning approach tailored to address noise distribution in domain adaptation settings and learn to de-confuse the pseudo-labels to capture the label corruption of each class and learn the underlying true label distribution.

Abstract

Source-free domain adaptation aims to adapt a source-trained model to an unlabeled target domain without access to the source data. It has attracted growing attention in recent years, where existing approaches focus on self-training that usually includes pseudo-labeling techniques. In this paper, we introduce a novel noise-learning approach tailored to address noise distribution in domain adaptation settings and learn to de-confuse the pseudo-labels. More specifically, we learn a noise transition matrix of the pseudo-labels to capture the label corruption of each class and learn the underlying true label distribution. Estimating the noise transition matrix enables a better true class-posterior estimation, resulting in better prediction accuracy. We demonstrate the effectiveness of our approach when combined with several source-free domain adaptation methods: SHOT, SHOT++, and AaD. We obtain state-of-the-art results on three domain adaptation datasets: VisDA, DomainNet, and OfficeHome.
Paper Structure (22 sections, 9 equations, 3 figures, 6 tables, 1 algorithm)

This paper contains 22 sections, 9 equations, 3 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Unsupervised Domain Adaptation.
  4. Source-Free Domain Adaptation.
  5. Pseudo Labels for SFDA.
  6. Learning with Label-Noise.
  7. SFDA Problem Setup
  8. De-Confusing Pseudo-Labels
  9. Pseudo-label generation
  10. Adaptation as learning with noisy pseudo-labels
  11. Domain adaptation regularization.
  12. Experiments
  13. Datasets
  14. Implementation Details
  15. Main Results
  16. ...and 7 more sections

Figures (3)

  • Figure 1: Schematic overview of our framework. DCPL learns both the target model and a transition matrix that adapts the model predictions to noisy pseudo-labels obtained by a general pre-trained network.
  • Figure 2: Real and learned noise transition matrices for VisDA.
  • Figure 3: t-SNE visualization for (a) Clipart to Sketch adaptation from the DomainNet dataset, (b) Real to Sketch adaptation from the DomainNet dataset, (c) synthetic to real adaptation for the VisDA dataset: (1) at the beginning of the target adaptation process, which demonstrates the source model's ability to separate the different classes, (2) at the end of target adaptation using DCPL with $\hat{T}=I$, (3) at the end of target adaptation using DCPL.