The Lovasz-Softmax Loss: A Tractable Surrogate for the Optimization of the Intersection-Over-Union Measure in Neural Networks

摘要

The Jaccard loss, commonly referred to as the intersection-over-union loss,is commonly employed in the evaluation of segmentation quality due to itsbetter perceptual quality and scale invariance, which lends appropriaterelevance to small objects compared with per-pixel losses. We present a methodfor direct optimization of the per-image intersection-over-union loss in neuralnetworks, in the context of semantic image segmentation, based on a convexsurrogate: the Lov'asz hinge. The loss is shown to perform better with respectto the Jaccard index measure than other losses traditionally used in thecontext of semantic segmentation; such as cross-entropy. We develop aspecialized optimization method, based on an efficient computation of theproximal operator of the Lov'asz hinge, yielding reliably faster and morestable optimization than alternatives. We demonstrate the effectiveness of themethod by showing substantially improved intersection-overunion segmentationscores on the Pascal VOC dataset using a state-of-the-art deep learningsegmentation architecture.