A Survey of Scene Graph: Generation and Application

EasyChair Preprint no. 3385

Abstract

Scene Graph is a data structure, which is mainly used to describe the objects, attributes and object relationships in a scene. Scene Graph is a deep representation of a scene, and is very conducive to many visual tasks, such as image retrieval, image/video captions, VQA, and even to image generation and specific relationship detection. At present, numbers of research works about scene graph are proposed, including the scene graph generation methods and the related applications. These proposed methods based on scene graph have great improvements in relative performances compared with the corresponding traditional methods, which also proves the effectiveness of scene graph in the visual understanding of a scene. Therefore, In this paper, we provide a systematic review of the existing techniques of scene graph generation and application, including not only the state-of-the arts but also those with latest trends. Particularly, we discuss the scene graph generation methods according to the inference models for visual relationship detection, and the applications of scene graph are stated according to the specific visual tasks. Finally, we point out several problems in the current scene graph generation methods, related applications and the future research directions of scene graph.

Keyphrases: object detection, prior information, Scene Graph, visual feature extraction, visual relationship recognition

@Booklet{EasyChair:3385,
  author = {Pengfei Xu and Xiaojun Chang and Ling Guo and Po-Yao Huang and Xiaojiang Chen and Alex Hauptmann},
  title = {A Survey of Scene Graph: Generation and Application},
  howpublished = {EasyChair Preprint no. 3385},

  year = {EasyChair, 2020}}