Increase the sensitivity of moderate examples for semantic image segmentation

Quan Tang, Fagui Liu, Dengke Zhang, Jun Jiang, Xuhao Tang, C. L. Philip Chen

Abstract: Highlights•Reconsidering pixels from a difficulty imbalance perspective in semantic segmentation.•Fitting extremely hard pixels during training leads to model over-fitting.•A dynamically scaled sensitive loss that focuses model learning on moderate pixels.•Sensitive loss notably increases the generalization capacity of existing models.