Reproduce “Adaptive Affinity Loss and Erroneous Pseudo-Label Refinement for Weakly Supervised Semantic Segmentation” on OpenResearch

About this paper

Xiangrong Zhang, Zelin Peng, Peng Zhu, Tianyang Zhang, Chen Li, Huiyu Zhou, Licheng Jiao

This work enhances single-stage weakly supervised semantic segmentation by introducing an Adaptive Affinity (AA) loss and a Label Reassign (LR) loss. The method achieves 63.9% mIoU on the PASCAL VOC 2012 validation set, surpassing prior single-stage methods while maintaining computational efficiency comparable to its baseline.

Key insights

• Weighting pixel affinities based on the maximum confidence of connected pixels effectively suppresses the propagation of errors from ambiguous boundary regions while preserving clear semantic properties.
• Refining semantic similarity in the embedding space through a metric-based loss, specifically targeting samples with high ambiguity, helps the network discern and avoid reinforcing erroneous pseudo-labels.
• Combining adaptive affinity learning and erroneous label refinement within a single-stage framework can close the performance gap between efficient single-stage methods and more complex multi-stage approaches without significant computational overhead.

Summary by alphaXiv.