A Semantic Relation Graph Reasoning Network for Object Detection

Abstract

Object detection is a basic task in computer vision, and it plays an important role in the fields of robotics, security, and autonomous driving. However, the object detection algorithms at present usually extract the features of a single region and then perform detection, ignoring the semantic context between objects and scenes, which will produce bad effect in detection. In order to use the semantic context between objects and scenes, this paper considers object detection as a graph reasoning problem. In this paper, we obtain the prior knowledge of the co-occurrence among the objects and between objects and scenes through statistics of the dataset, then we mainly use two modules to extract the semantic relationship between objects and scenes. The first one extracts the prior knowledge between objects through of graph convolutional networks(GCN), and introduces the graph attention networks(GAT) to learn hidden knowledge about the semantic context relationship between objects adaptively, and by concating these knowledge then use them for detection. The second one uses MLP to generate S-L coefficients and multiplies the scene features and the S-L coefficients to obtain scene-object related features for object detection. We have conducted experiments to verify our method on the PASCAL VOC dataset, and the experiments show that our method can effectively improve the accuracy of object detection.