Non-Maximum Suppression for Rotated Object Detection During Merging Slices of High-Resolution Images

In object detection on remote sensing images or aerial images, high-resolution images and low relative area ratio of objects need to be solved. Usually, a high-resolution image should be split and detected separately. Then, the prediction results would be merged as a result of the complete image. Due to the overlap between adjacent image slices, non-maximum suppression (NMS) is used to suppress redundant prediction boxes during the merging process. However, redundant boxes from splitting are different from those generated by detection on a single image. Some special cases cannot be effectively solved by NMS. We have studied these cases and summarized judgment conditions to identify them. Based on NMS, we propose a new method named RODM-NMS for rotated object detection during merging slices of high-resolution images. Cases challenging for NMS are classified into two categories and addressed with specific handling methods. Compared to the traditional NMS, RODM-NMS could achieve a 3.9 improvement in AP50 during merging detection results on $640\times 640$ slices of DOTA by FCOSR. It has low requirements for the overlap gap of slices and shows a significant improvement in performance for weak models. Hence, it is more suitable for computing-constrained mobile platforms such as drones to detect objects on high-resolution images.