Relation-Aware Reasoning with Graph Convolutional Network.

Semantic dependencies among objects are crucial for the recognition system to enhance performance. However, utilizing object-object relationships is a non-trivial task as objects are of various scales and locations, leading to irregular relationships. In this paper, we present a novel visual reasoning framework that incorporates both semantic and spatial relationships to improve the recognition system. We at first construct a knowledge graph to represent the co-occurrence frequency and relative position among categories. Based on this knowledge graph, we are able to enhance the original regional features by a Graph Convolutional Network (GCN) that encodes the high-level semantic contexts. Experiments show that our framework manages to outperform the baselines and state-of-the-art on different backbones in terms of both per-instance and per-class classification accuracy.