Suppression and Enhancement of Overlapping Bounding Boxes Scores in Object Detection
Ye Pan, Feng Dong · 2019
Non-Maximum Suppression (NMS) and Soft-NMS are integral parts of the object detection pipeline. Based on the analysis of the above two algorithms, we propose a new algorithm named Suppression and Enhancement (SE) in this paper. We aim at solving the problem that crowd between objects make the detection scores low in object detection. The detection scores will be decreased while being partially crowded by other objects. If we remove the objects' bounding boxes with smaller score by a threshold after NMS or Soft-NMS, it is very likely to remove the crowded object, which will reduce the precision. Increasing the scores of the crowded objects with our algorithm, we can detect the crowded objects with the algorithm more accurately.