Edge-Aware Graph Matching Network for Part-Based Semantic Segmentation

Abstract Semantic segmentation of parts objects is a marginally explored and challenging task in which multiple instances within those must be recognized an image. We introduce novel approach (GMENet) for this combining object-level context conditioning, part-level spatial relationships, shape contour information. The first target achieved by introducing class-conditioning module that enforces class-level semantics when learning the ones. Thus, intermediate-level features carry prior to decoding stage. To tackle ambiguity relationships among we exploit adjacency graph-based aims at matching between ground truth predicted maps. Last, additional further leverage edges localization. Besides testing our framework on already used Pascal-Part-58 Pascal-Person-Part benchmarks, two benchmarks large-scale part parsing, i.e., more version Pascal-Part with 108 classes ADE20K-Part benchmark 544 parts. GMENet achieves state-of-the-art results all considered tasks furthermore allows improve accuracy.