Accurate Object Detection with Relation Module on Improved R-FCN
Jinglei Zhang, Shaoxing Chen, Yawei Hou · 2020
In object detection, modeling the relation of objects is the key to achieve performance improvement. In order to improve the detection accuracy of the R-FCN(Region-based Fully Convolutional Network), a new algorithm called accurate R-FCN is proposed. By introducing the relation module to make an interaction between their appearance and geometric features, the detection and recognition accuracy can be improved. In addition, aiming at the problem of position deviation in the region of interest(RoI) quantization, PSRoI Align(Position-Sensitive Region of Interest Align) is applied with the RPN(Region Proposal Network), which generates the accurate region proposal. Experiments on the MS COCO dataset show that the accurate R-FCN achieves state-of-the-art accuracy compared with the other seven classical algorithms.