Neighbouring Relationship Exploration Based on Graph Convolutional Network for Object Detection

Current object detection methods are mostly based on the Faster R-CNN which is composed of two stages: 1) the region proposal network which roughly sifts anchors and produces object proposals and 2) the detection network which inputs the generated region proposals and makes predictions of category and bounding box for each region. However, the object proposals are always treated individually in the detection network without taking their relationship into consideration. Specifically, most researches utilize the non-maximum suppression algorithm to reduce the number of proposals by subtracting the neighbouring proposals for either training or testing process. Though a single proposal may be not accurate enough, a union of its neighbouring proposals and itself would represent this object region more comprehensively. Since the neighbouring relationship can be expressed by an edge between two proposals (nodes), the set of region proposals can be modelled in a graph structure. In this paper, we propose a novel neighbouring relationship exploration model (NREM) to improve the object detection by aggregating the neighbouring proposal graph (NPG) based on the graph convolutional network (GCN). Owing to the effective exploration of complementary relationship among neighbouring proposals, our method can improve the detection results significantly. Experiments on the PASCAL VOC 2007 dataset demonstrate the superiority of our proposed method.