Confusing Pair Correction Based on Category Prototype for Domain Adaptation under Noisy Environments

TL;DR

This paper proposes a new method to detect and correct confusing class pair and applies label correction to the noisy samples within the confusing pair to train the model with more accurate labels.

Abstract

In this paper, we address unsupervised domain adaptation under noisy environments, which is more challenging and practical than traditional domain adaptation. In this scenario, the model is prone to overfitting noisy labels, resulting in a more pronounced domain shift and a notable decline in the overall model performance. Previous methods employed prototype methods for domain adaptation on robust feature spaces. However, these approaches struggle to effectively classify classes with similar features under noisy environments. To address this issue, we propose a new method to detect and correct confusing class pair. We first divide classes into easy and hard classes based on the small loss criterion. We then leverage the top-2 predictions for each sample after aligning the source and target domain to find the confusing pair in the hard classes. We apply label correction to the noisy samples within the confusing pair. With the proposed label correction method, we can train our model with more accurate labels. Extensive experiments confirm the effectiveness of our method and demonstrate its favorable performance compared with existing state-of-the-art methods. Our codes are publicly available at https://github.com/Hehxcf/CPC/.

Figures (6)

• Figure 1: Demonstration of our confusing pair correction strategy. (b) Noisy samples of the confusing class pair frequently switch their predicted labels. (c) Class A becomes dominant, class B becomes weak. Samples of B will be misclassified as A. (d) Label correction between the confusing class pair. (a) The classification boundary between the two classes becomes clearer.
• Figure 2: The framework of our method. It contains two components: target domain refinement and confusing pair correction. Firstly, we construct a shared prototype to align the source and target domains and obtain pseudo labels for the target samples. Next, pair labels are generated from the prototype by using top-2 prediction. Secondly, the confusing pair correction method finds the most confusing pair and performs label correction for the noisy samples of the pair. Once the labels are corrected, the prototypes and all pseudo labels are recalculated.
• Figure 3: The classification accuracy of target domain refinement under different noisy rates in Office-31.
• Figure 4: Examples of confusing pairs in domain Dslr. (a) and (b) is a confusing pair. (c) and (d) is another pair.
• Figure 5: Demonstration of how confusing pair correction contributes to improving the accuracy of the confusing pair and overall accuracy. The numbers in parentheses represent the number of times the confusing pair has been corrected.