Gaze-Aware Graph Convolutional Network for Social Relation Recognition

Social relation, as the basic relation in our daily life, is vital for social action analysis. However, how to learn feature between people still not tackled. In this work, we propose a gaze-aware graph convolutional network (GA-GCN) recognition, which targets discovering context-aware inference with attention. To predict gaze direction, apply trained direction loss. Then, build module, two-stream both attention and distance-aware The can pick up relevant context objects representation. We further introduce additional scene features construct multiple fusion adaptively representation from feature. Extensive experiments on PISC PIPA datasets demonstrate that GA-GCN find interesting contextual achieves state-of-the-art performances.