Reproduce “Bootstrapping Semantic Segmentation with Regional Contrast” 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
Shikun Liu, Shuaifeng Zhi, Edward Johns, Andrew J. Davison
Researchers from the Dyson Robotics Lab and Robot Learning Lab at Imperial College London developed ReCo, a pixel-level contrastive learning framework that significantly improves semantic segmentation accuracy, particularly in low-label semi-supervised settings. ReCo achieved relative mean IoU improvements of 5-10% in these scenarios while producing sharper object boundaries with minimal additional memory overhead.
Key insights
- Actively sampling 'hard' pixels and focusing on semantically confusing negative classes significantly enhances the efficiency and effectiveness of pixel-level contrastive learning.
- Regional contrastive learning can effectively disentangle ambiguous features and improve discriminative power, leading to sharper object boundaries and more accurate predictions.
- A batch-wise, on-the-fly approach to class distribution approximation in contrastive learning can be as effective and more efficient than large memory bank methods.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.