Semantic boundary refinement by joint inference from edges and regions
Chao Yang · 2017
We study the problem of detecting boundaries for specific classes of objects. Our approach leverages recent advances in semantic segmentation and bottom-up boundary detection. We propose a mechanism for combining multiple sources of information: predicted segmentation masks, bottom-up contours, and a novel local class-specific boundary detector. These are jointly mapped to final category-specific boundary strength estimated by a trained classifier. In experiments on VOC2012 [7] and Microsoft COCO [15], our method dramatically outperforms recent prior work, for some classes doubling the accuracy of boundary prediction.