Zero-Shot Scene Graph Generation with Knowledge Graph Completion

Abstract

Limited by the incomprehensive training samples, existing scene graph generation (SGG) methods perform poorly on predicting zero-shot (i.e., unseen) subject-predicate-object triples. To address this problem, we propose a general SGG framework to improve their zero-shot performance. The main idea of our method is to generate the information of zero-shot triples before the training of the predicate classifier and thus make the original zero-shot triples non-zero-shot. Specifically, the missing information of zero-shot triples is generated by our proposed knowledge graph completion strategy and then integrated with visual features of images. Therefore, the predicate classification of zero-shot triples is no longer just regarded as a single visual classification task but also transformed into a prediction task of missing links in a knowledge graph. The experiments on the dataset Visual Genome demonstrate that our proposed method outperforms the state-of-the-art methods in popular zero-shot metrics (i.e., zR@N, ng-zR@N) for all popular SGG tasks.