Reproduce “Semi-Supervised Semantic Segmentation Using Unreliable Pseudo-Labels” 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
Yuchao Wang, Haochen Wang, Yujun Shen, Jingjing Fei, Wei Li, Guoqiang Jin, Liwei Wu, Rui Zhao and 1 more
U²PL is a semi-supervised semantic segmentation framework that effectively leverages low-confidence pixel predictions by treating them as negative supervisory signals. This approach yields stronger discriminative features, particularly in ambiguous regions, achieving state-of-the-art mean IoU on PASCAL VOC 2012 and competitive results on Cityscapes.
Key insights
- Unreliable pixel-wise predictions, traditionally discarded, contain valuable negative supervisory information that can be effectively leveraged for training.
- Framing unreliable predictions as negative samples within a contrastive learning framework forces the model to learn more discriminative features, particularly beneficial for ambiguous regions and object boundaries.
- Maximizing data utilization by making 'every pixel matter' in semi-supervised training, rather than relying solely on high-confidence pseudo-labels, leads to improved performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.