Object Detection Made Simpler by Eliminating Heuristic NMS

It is valuable and promising to remove post-processing non-maximum suppression (NMS) for object detectors, making detectors simpler purely end-to-end. Removing NMS possible if the detector can identify only one positive sample prediction each ground-truth instance in an image. In this work, we propose a compact plug-in head, named PSS which be attached any one-stage make them NMS-free. Specifically, head works by automatically selecting detected, so that with our directly NMS. The success of lies three aspects, namely one-to-one label assignment, stop-gradient operation eliminating optimization conflicts, pss loss ranking specifically designed head. Experiments on COCO dataset demonstrate effectiveness method. particular, when compared stage-of-the-art NMS-free methods, (attaching VFNET) achieves 44.0% mAP, exceeds 41.5% mAP DeFCN large margin. When taking Res2Net-101-DCN as backbone network, 50.3% test set, performance even among NMS-based methods.