Reproduce “InfoSeg: Unsupervised Semantic Image Segmentation with Mutual Information Maximization” 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
Robert Harb, Patrick Knöbelreiter
Researchers at Graz University of Technology developed InfoSeg, an unsupervised semantic image segmentation method that learns pixel-wise class assignments by maximizing mutual information between local image features and global class-specific features. The approach achieved a Pixel Accuracy of 38.8% on COCO-Stuff, representing a 26% relative improvement over previous methods, and 69.6% on the new COCO-Persons benchmark.
Key insights
- Decoupling feature learning and segmentation into distinct iterative steps provides a more robust training process, effectively preventing the model from converging to degenerate, low-level feature-based segmentations.
- Employing soft assignments for global feature integration, rather than hard class assignments, significantly improves performance by accounting for uncertainty and enabling smoother learning.
- Maximizing mutual information using a Jensen-Shannon Divergence (JSD) based estimator between local and soft global features is more effective for unsupervised segmentation than alternative estimators or direct application of existing MI objectives like those in IIC.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.