Objects as Extreme Points

Object detection can be regarded as a pixel clustering task, and its boundary is determined by four extreme points (leftmost, top, rightmost, bottom). However, most studies focus on the center or corner of object, which are conditional results points. In this paper, we present an Extreme-Point-Prediction-Based object detector (EPP-Net), directly regresses relative displacement vector between each We also propose new metric to measure similarity two groups points, namely, Extreme Intersection over Union (EIoU), incorporate EIoU regression loss. Moreover, novel branch predict ground-truth prediction results, take it localization confidence filter out poor results. On MS-COCO dataset, our method achieves average precision (AP) 44.0% with ResNet-50 AP 50.3% ResNeXt-101-DCN. The proposed EPP-Net provides detect objects very competitive performance among state-of-the-art anchor-free detectors.