SFP-NMS: Nonmaximum Suppression for Suppressing False Positives During Merging Patches of High-Resolution Image

Nonmaximum suppression (NMS) is a crucial postprocessing method for object detection. It aims to eliminate overlapping or redundant bounding boxes predicted by models. In tasks with high-resolution images like remote sensing object detection, images often undergo a process involving splitting, individual detection, and subsequent merging. During both detection and merging stages, NMS is used. However, due to the special position of split objects, some false positives (FPs) cannot be suppressed by NMS based on intersection-over-union (IoU) threshold. To address this issue, we analyze scenarios that might cause FPs during the merging process and propose SFP-NMS. This method has been validated on multiple models, yielding an improvement ranging from 0.19 to 1.38 on mAP.