AN ANCHOR-BASED OBJECT DETECTION METHOD FOR EFFECTIVELY REDUCING THE NUMBER OF ANCHORS

The method of object detection based on anchor has the disadvantage of huge computation. In this work, we propose a two-stage solution to effectively reduce the number of anchors, which is pre-position estimation network (PPENet). We estimated the target location and category information on the Gaussian heatmap, generated only a small number of anchors based on the target location information, and estimated the target shape deviation. This method putted the target position estimation in front, and used the position estimation information to effectively reduce the number of anchors and the amount of computation. This solution was end-to-end. On the MSCOCO public dataset, the maximum AP value reached 50.2, and the optimal FPS was 33.1. The scheme reached the state-of-the-art level. The experimental results prove the effectiveness of the proposed method.