A refined prior-box generator for anchor-based object detector

This paper proposes a refined prior-box generator for anchor-based object detectors, which allows detectors to be trained on a more balanced candidate set. Concretely, we present two schemes to alleviate the sample imbalance caused by prior-box mechanism. First, a more elaborated presupposed dimension calculation scheme is introduced to obtain more positive samples. Then, a location weighting mask originating from the edge map is constructed to guide the sampling of prior-box, which has the ability to avoid sampling in the background as much as possible. The widely used Feature Fusion Single Shot multi-box Detector (FSSD) is selected as baseline. Extensive experiments are conducted on the Pascal VOC (07+12) dataset and demonstrate that FSSD with the refined prior-box generator achieves 1.3% higher mean Average Precision (mAP) with 70% fewer prior candidate bounding boxes. Under the same input resolution, our method receives the best detection performance.