Dlafs cascade R-CNN: an object detector based on dynamic label assignment

Abstract

Object detection based on Deep Learning is the revolution of computer science in general and related problems of object detection in particular. In particular, recently, two-stage or multi-stage methods of the R-CNN family have shown outstanding results. These methods have two steps in common: Generating proposal boxes and object classification. In the step of the generating proposal, a Regional Proposal Network (RPN) will be learned to suggest high probability regions in the image, and the part of Label Assignment for RPN is of great interest. If the samples are obtained well, RPN will learn well and help the efficiency of the next stage increase sharply. In this study, we investigate and study to improve the object detection performance when applying Dynamic Label Assignment on the first stage of Cascade R-CNN called DLAFS Cascade R-CNN and perform some experiments to prove the effectiveness. Our DLAFS Cascade R-CNN outperform previous methods on three datasets: SeaShips (+0.2% AP), UIT-DODV (+5.7% AP), MS-COCO (+2.8% AP).