M.S. Candidate: Mehmet Can Baytekin
Program: Information Systems
Date: 19.01.2023 / 13:30
Place: -

Abstract: In this work, we propose a dynamic anchor, dynamic assignment region proposal network named DA2RPN to improve the traditional Region Proposal Network (RPN). Classical Region Proposal Network places a set of pre-determined anchor boxes to generate object proposals which are later consumed by detection heads to produce final results. Problems of the Region Proposal Network are for each feature map point same anchor boxes are placed and static threshold values based on intersection-over-union scores are used to label anchor boxes. However, anchors mostly residing out of the image plane are not that useful to detect objects as objects only reside in the image plane. On the other hand, anchor box-ground truth box pairs having the same intersection-over-union values may not be equally useful for the detection of different objects. One anchor can be the best candidate for that object but another anchor can be a poor choice for another object even if they have the same iou values. To mitigate these problems, we generate a different number of anchors per feature map point and use a dynamic thresholding mechanism adaptive to the quality of selected anchors per ground-truth box. Additionally, to ease the training and to prevent sparsity caused by dynamic anchor generation we decouple anchor dimensions. Our results experimented on the COCO dataset show improvements over the Region Proposal Network.