Tilted Cross Entropy (TCE): Promoting Fairness in Semantic Segmentation

TL;DR

This paper introduces Tilted Cross Entropy (TCE), a novel loss function for semantic segmentation that reduces class performance disparities and enhances fairness across classes, outperforming traditional methods.

Contribution

The paper proposes TCE loss adapted for semantic segmentation, inspired by tilted ERM, to promote fairness and reduce class imbalance in model performance.

Findings

TCE improves performance of low-performing classes in Cityscapes and ADE20k datasets.

TCE enhances overall fairness in semantic segmentation models.

TCE achieves competitive or better results compared to standard cross-entropy.

Abstract

Traditional empirical risk minimization (ERM) for semantic segmentation can disproportionately advantage or disadvantage certain target classes in favor of an (unfair but) improved overall performance. Inspired by the recently introduced tilted ERM (TERM), we propose tilted cross-entropy (TCE) loss and adapt it to the semantic segmentation setting to minimize performance disparity among target classes and promote fairness. Through quantitative and qualitative performance analyses, we demonstrate that the proposed Stochastic TCE for semantic segmentation can efficiently improve the low-performing classes of Cityscapes and ADE20k datasets trained with multi-class cross-entropy (MCCE), and also results in improved overall fairness.