Object detection by crossing relational reasoning based on graph neural network

Utilizing relational representations to facilitate object detection has attracted growing research attention in recent years. However, previous studies mainly focus on relationships within the region proposals or within the label embeddings and pay less attention to the relationships between them. To fill this gap, we propose a novel object detection framework that fully explores the relationships across visual feature space and label embedding space to facilitate the proposal classification in object detection. Specifically, we model the region proposals and class labels into a uniform relation graph, where the extracted proposals and labels are regarded as nodes and each pair of them is associated by an assignment edge, and convert the problem of classifying proposals to the problem of selecting reliable edges from the constructed relation graph. Furthermore, a graph convolutional module is developed to perform relational reasoning on the graph, which finally predicts a label for each assignment edge to indicate whether the classification is reliable or not. The updated relational representations for proposals are used for bounding box regression. Embedding our framework into state-of-the-art baselines, we perform extensive comparison experiments on two public benchmarks, i.e., Pascal VOC and COCO2017. And the experimental results demonstrate the flexibility and effectiveness of the proposed framework.