Efficient piecewise training of deep structured models for semantic segmentation

摘要: Recent advances in semantic image segmentation have mostly been achieved by training deep convolutional neural networks (CNNs). We show how to improve through the use of contextual information; specifically, we explore `patch-patch' context between regions, and `patch-background' context. For learning from patch-patch context, formulate Conditional Random Fields (CRFs) with CNN-based pairwise potential functions capture correlations neighboring patches. Efficient piecewise proposed structured model is then applied avoid repeated expensive CRF inference for back propagation. capturing patch-background that a network design traditional multi-scale input sliding pyramid pooling effective improving performance. Our experimental results set new state-of-the-art performance on number popular datasets, including NYUDv2, PASCAL VOC 2012, PASCAL-Context, SIFT-flow. In particular, achieve an intersection-over-union score 78.0 challenging 2012 dataset.