Anchor-free Proposal Generation Network for Efficient Object Detection

Deep learning object detection methods are usually based on anchor-free or anchor-based scheme for extracting proposals and one-stage two-stage structure producing final predictions. As each has its own strength weakness, combining their in a unified framework is an interesting research topic. However, this topic not attracted much attention recent years. This paper presents method that utilizes generating the initial stage. For proposal generation, employs efficient network predicting corners assigns detected corners. prediction, designed to enhance both accuracy speed. The includes lightweight binary classification subnetwork removing most false positive candidates light-head Experimental results MS-COCO dataset demonstrate proposed outperforms baselines terms of performance.