ARM-NMS: Shape Based Non-Maximum Suppression For Instance Segmentation in Large Scale Imagery
Detecting objects in aerial scenes is a fundamental and critical task in remote sensing. However, state-of-the-art object detectors are susceptible to producing correlated scores in neighboring detections resulting in increased false positives. In addition, detection on large-scale images requires a tiling scheme with usually overlapping windows, consequently creating more double detections. Therefore, a non-maximum suppression (NMS) approach can be exploited as integral to the detection pipeline. NMS suppresses overlapping detections in regards to their scores. Current NMS algorithms filter detections by utilizing their corresponding bounding boxes. However, one can assume that comparing bounding boxes to determine the overlap of non-rectangular objects involves a certain degree of inaccuracy. Therefore, we propose Area Rescoring Mask-NMS (ARM-NMS), which uses object shapes for filtering. ARM-NMS exploits instance masks instead of the conventional boxes to eliminate detections and does not require retraining for instance segmentation pipelines. To exhibit the effectiveness of our approach, we evaluate our method on the large-scale aerial instance segmentation dataset iSaid. Our approach leads to considerable improvements for the COCO-style mAP metric of 3.3 points for segmentations and 3.5 points for boxes.