A Refined NMS Scheme Based on Center Distance and Corner Distribution

Abstract

Non-maximum suppression (NMS) is a significant post-process step in most object detection frameworks, the main task of which is to find the most optimal inferences. However, the low correlation between the classification confidence score and the location accuracy of bounding box severely affects the performance of object detector. Recently, several novel loss functions about the overlap rate of candidate bounding boxes have been proposed to reduce the negative impact of the low correlation problem, but there is no solution to pay attention to the influence of the background area in bounding box on NMS. In this paper, we present a refined NMS scheme based on the distribution of corner feature and the center distance of bounding box. The correlation between confidence score and location accuracy is optimized through reducing the impact of background area in bounding box. Specially, we first construct the corner feature and reorder the confidence score by its distribution. And then, a dynamically NMS threshold strategy is built by a center distance function. Extensive experiments are conducted on the Pascal VOC and the mean Average Precision (mAP) of FSSD is improved from 78.8% to 80.1%. What's more, the proposed method can be easily embedded into most of object detection frameworks without any extra training.