The Lov\'asz-Softmax loss: A tractable surrogate for the optimization of the intersection-over-union measure in neural networks

TL;DR

This paper introduces a new loss function based on the Lovász extension to directly optimize the intersection-over-union measure in neural network training, leading to improved segmentation performance.

Contribution

It proposes a tractable surrogate loss for IoU optimization in neural networks, outperforming traditional pixel-wise losses like cross-entropy.

Findings

Significant improvement in IoU scores on Pascal VOC and Cityscapes datasets.

Better alignment of training loss with evaluation metric (IoU).

Enhanced segmentation quality with state-of-the-art architectures.

Abstract

The Jaccard index, also referred to as the intersection-over-union score, is commonly employed in the evaluation of image segmentation results given its perceptual qualities, scale invariance - which lends appropriate relevance to small objects, and appropriate counting of false negatives, in comparison to per-pixel losses. We present a method for direct optimization of the mean intersection-over-union loss in neural networks, in the context of semantic image segmentation, based on the convex Lov\'asz extension of submodular losses. The loss is shown to perform better with respect to the Jaccard index measure than the traditionally used cross-entropy loss. We show quantitative and qualitative differences between optimizing the Jaccard index per image versus optimizing the Jaccard index taken over an entire dataset. We evaluate the impact of our method in a semantic segmentation pipeline and show substantially improved intersection-over-union segmentation scores on the Pascal VOC and Cityscapes datasets using state-of-the-art deep learning segmentation architectures.