FCPN: Pruning redundant part-whole relations for more streamlined pattern parsing

Abstract

Most cropping-and-segmenting pattern parsers typically establish a single metric/scheme to reason diverse inner correlations, resulting in over-general and redundant representations. To make pattern parsing more streamlined and efficient, a fragile correlation pruner network (FCPN) with correlation-steered attention shifters (CSASs) and graph attention expectation-maximum routing agreement (GAEMRA) is proposed. CSASs prune fragile (weak and substitutable) part-to-whole correlations. They stipulate that only those primary entities (representing components) fulfilling the criteria of inter-part diversity and intra-object cohesiveness can update senior entities (representing the whole/intermediate composites). To further boost effects, GAEMRA is defined to shield the redundant voting signals of conventional routing agreement. With CSASs and GAEMRA, FCPN gradually parses objective semantic patterns by clustering highly associated secondary entities in a bottom-up "part backtracking" manner. Quantitative and ablation experiments surrounding face and human parsing demonstrate the superiority of FCPN over the state-of-the-arts, especially for the definition of fine-grained semantic boundaries.