Graph Attention Network Based Object Detection and Classification in Crowded Scenario

Non-maximum Suppression (NMS) is an essential post-processor for most detectors, responsible for merging excessive detections from detector. The standard NMS is a crude greedy algorithm that is not accurate enough sometimes and even fail entirely in crowded scenario, which results from making inadequate use of information of detection. We regard the detection set as the data of the graph structure, and adopt the Graph Attention Network (GAT), which fits the data structure very well, as the backbone to perform NMS. In addition, our algorithm is aware of the density of objects and therefore can handle partly crowded scenario. The result of our algorithm is better than the traditional algorithm, which shows the superiority of performing NMS based on global information.