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

CVPR 2018 Maxim BermanAmal Rannen TrikiMatthew B. Blaschko

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 Lovász extension of submodular losses... (read more)

PDF Abstract

Evaluation results from the paper


Task Dataset Model Metric name Metric value Global rank Compare
Real-Time Semantic Segmentation Cityscapes ENet + Lovász-Softmax mIoU 63.1% # 9
Real-Time Semantic Segmentation Cityscapes ENet + Lovász-Softmax Time (ms) 13 # 1
Real-Time Semantic Segmentation Cityscapes ENet + Lovász-Softmax Frame (fps) 76.9 # 1
Semantic Segmentation Cityscapes ENet + Lovász-Softmax Mean IoU (class) 63.06% # 27
Semantic Segmentation PASCAL VOC 2012 Deeplab-v2 + Lovász-Softmax Mean IoU 79.0% # 14