DyConfidMatch: Dynamic Thresholding and Re-sampling for 3D Semi-supervised Learning

TL;DR

This study investigates class-level confidence as an indicator of learning status in 3D SSL, proposing a novel method that utilizes dynamic thresholding to better use unlabeled data, particularly from underrepresented classes.

Abstract

Semi-supervised learning (SSL) leverages limited labeled and abundant unlabeled data but often faces challenges with data imbalance, especially in 3D contexts. This study investigates class-level confidence as an indicator of learning status in 3D SSL, proposing a novel method that utilizes dynamic thresholding to better use unlabeled data, particularly from underrepresented classes. A re-sampling strategy is also introduced to mitigate bias towards well-represented classes, ensuring equitable class representation. Through extensive experiments in 3D SSL, our method surpasses state-of-the-art counterparts in classification and detection tasks, highlighting its effectiveness in tackling data imbalance. This approach presents a significant advancement in SSL for 3D datasets, providing a robust solution for data imbalance issues.

Figures (6)

• Figure 1: The analysis of FixMatch, when trained on the ModelNet40 dataset using only 10% labeled data, reveals the following insights based on fixed threshold values of 0.9 and 0.3: (a) The number of pseudo-labels selected from the unlabeled data varies significantly. With a higher threshold of 0.9, only a small portion of the unlabeled data is incorporated, as the stricter criterion filters out many potential pseudo-labels. (b) Conversely, with a lower threshold of 0.3, a larger portion of the unlabeled data is utilized, as more pseudo-labels are accepted. Therefore, to fully leverage the unlabeled data, a relatively lower threshold should be considered.
• Figure 2: (a)The results comparing FixMatch with 0.3 and 0.9 thresholds, FlexMatch, and our method show that while a lower threshold (0.3) allows for greater utilization of unlabeled data, it also introduces more noise, which negatively impacts performance. In contrast, our method strikes a balance between pseudo-label quality and the effective use of unlabeled data, leading to superior overall performance. (b) There are notable differences in the test accuracy and the confidence level for each class in the unlabeled dataset. The learning challenge varies widely across different classes, and intriguingly, some less common classes register a better accuracy than the more prevalent ones. Furthermore, (1) The limited utilization of unlabeled data with a high threshold suggests a cautious approach to incorporating such data during training. (2) The observed differences in learning difficulty among classes highlight the importance of addressing class imbalance during training. (3) The strong correlation between class-level confidence and test accuracy underscores the reliability of the confidence scores in predicting class performance.
• Figure 3: The proposed method consists of three key components: (1) Leveraging class-level confidence derived from unlabeled data to determine the learning status for each class, (2) Utilizing the learning status to dynamically adjust the threshold for each class, and (3) Adjusting the re-sampling strategy based on the identified learning status.
• Figure 4: Analysis of results for 3DIoUMatch in SUN-RGBD Dataset. The result demonstrates that the class-level confidence for unlabeled data exhibits a strong correlation with the class test accuracy in the 3D SSL detection task.
• Figure 5: (a) Thresholds for each class in the last epoch of our method, applied to the ModelNet40 dataset with $10\%$ labeled data. (b) Each class's threshold of the FlexMatch in the last epoch. FlexMatch results in long-tail thresholds, introducing significant noise into pseudo labels, thereby diminishing performance. In contrast, the thresholds in our method are well-proportioned and balanced, enhancing the effective use of unlabeled data.