Reproduce “A Three-Stage Self-Training Framework for Semi-Supervised Semantic Segmentation” on OpenResearch

About this paper

Rihuan Ke, Angelica Aviles-Rivero, Saurabh Pandey, Saikumar Reddy, Carola-Bibiane Schönlieb

A three-stage self-training framework for semi-supervised semantic segmentation progressively refines pseudo-masks and enhances model accuracy by integrating consistency regularization, strong data augmentations, and a multi-task learning approach. The framework outperforms existing semi-supervised methods on Cityscapes and PASCAL VOC, achieving a 54.85% mIoU on Cityscapes with only 100 labeled images, demonstrating improved performance with reduced annotation requirements.

Key insights

• Iteratively refining pseudo-masks across multiple training stages effectively reduces uncertainty and improves their quality for self-training.
• Consistency regularization combined with strong data augmentations is vital for enhancing the robustness and accuracy of semi-supervised semantic segmentation.
• Modeling pseudo-mask statistics with an auxiliary task, rather than direct pseudo-label supervision, provides a more stable and effective guidance signal for the segmentation network.

Summary by alphaXiv.