Mild-NMS: A Suppression Algorithm for Enhanced Object Detection and Classification for High Intra-Class Overlap Ranges

Greedy Non-Maximum Suppression (NMS) is a necessity in the object detection pipelines of state-of-the-art models. Because this greedy heuristic can often lead to false negatives when more overlap or crowdedness is introduced into an image, Soft-NMS was created to handle the suppression less aggressively. Soft-NMS, however, can have the same effect as Greedy NMS when more crowdedness is introduced due to its repeated decaying of confidence scores. To this end, Mild-NMS is proposed as a means to handle higher overlap scenarios where Soft-NMS might perform like Greedy NMS and produce false negatives. Its idea is to limit the number of times a prediction's confidence score can be decayed by Soft-NMS to only one. Improvements in the performance of a stock YOLOV8x model pre-trained on the Common Objects in Context dataset for detecting only the person class were explored. It was found that, in this person class detection problem, Mild-NMS consistently outperformed both NMS and Soft-NMS in the subsets with the highest degree of intra-class overlap for all four datasets tested. In particular, it performed best on the three highest overlap subsets for the first dataset, the two highest overlap subsets for the second dataset, and the highest overlap subsets in the last two datasets. When considering only the subsets with the highest rate of overlap for each dataset, Mild-NMS outperformed the next best-performing suppression algorithm by an average of 0.02773 in terms of Average Precision. This was done while incurring no additional computational cost, making this suppression algorithm particularly advantageous for object detection tasks where consistently high overlap between objects is expected.