Reproduce “Re-distributing Biased Pseudo Labels for Semi-supervised Semantic Segmentation: A Baseline Investigation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ruifei He, Jihan Yang, Xiaojuan Qi
The DARS method effectively tackles biased pseudo-label distributions in semi-supervised semantic segmentation, achieving an 8.89% mIoU gain over the baseline on the Cityscapes dataset with 1/8 labeled data. It integrates distribution alignment, random sampling, and a progressive training strategy to improve learning, especially for underrepresented classes.
Key insights
- The bias in pseudo-label distributions is a primary bottleneck in semi-supervised semantic segmentation, and explicit distribution alignment significantly improves learning, particularly for tail classes.
- "Confidence overlapping" in DCNN predictions prevents simple confidence thresholding from precisely controlling pseudo-label counts per class, requiring a combination with random sampling for effective alignment.
- Progressive increases in both the pseudo-labeling ratio and data augmentation strength across iterative training rounds are essential to prevent premature overfitting and ensure sustained performance gains.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.