An enhanced N-point interpolation method to eliminate average precision distortion

Abstract

• Existing N -point interpolation method to calculate average precision raises three anomalies. • The average precision distortion problem is commonly present in the N -point interpolation method. • The proposed enhanced N -point interpolation method reduces more than 90% of the average precision distortion. Existing N -point interpolation methods generate large errors in the average precision calculation for object detection. These errors lead to average precision distortion, which makes it impossible to accurately evaluate the performance of the model. We investigate the reason for the average precision distortion and propose an enhanced N -point interpolation method. These improvements are based on the N -point interpolation method and can be summarized in two parts: (1) The interpolation point position is changed to the middle interpolation. (2) Dynamic selection of parameters for calculating the area of the interpolation interval. Experiments verify the existence of severe average precision distortion in the N -point interpolation method. Furthermore, the proposed enhanced N -point interpolation method reduces the average precision distortion by more than 90% to only 0.04%. In this way, the enhanced N -point interpolation method is able to replace the all-point interpolation method for fast and accurate evaluation of object detection model.