Optimization of the Jaccard index for image segmentation with the Lovász hinge

The Jaccard loss, commonly referred to as the intersection-over-union loss, is commonly employed in the evaluation of segmentation quality due to its better perceptual quality and scale invariance, which lends appropriate relevance to small objects compared with per-pixel losses. We present a method for direct optimization of the per-image intersection-over-union loss in neural networks, in the context of semantic image segmentation, based on a convex surrogate: the Lovász hinge. The loss is shown to perform better with respect to the Jaccard index measure than other losses traditionally used in the context of semantic segmentation; such as cross-entropy. We develop a specialized optimization method, based on an efficient computation of the proximal operator of the Lovász hinge, yielding reliably faster and more stable optimization than alternatives. We demonstrate the effectiveness of the method by showing substantially improved intersection-overunion segmentation scores on the Pascal VOC dataset using a state-of-the-art deep learning segmentation architecture.