Improving rare relation inferring for scene graph generation using bipartite graph network

Jiale Lu, Lianggangxu Chen, Haoyue Guan, Shaohui Lin, Chunhua Gu, Changbo Wang, Gaoqi He

Abstract: Highlights•We propose a model-agnostic network, which makes better use of external texts.•We solve zero-shot and long-tail problem of rare relations in scene graph generation.•Our method performs better than baseline methods and other text-supervised methods.