method: DAL(multi-scale)2020-08-14

Authors: Qi Ming

Affiliation: Beijing Institute of Technology


Description: Based on RetiaNet without extra preset anchors for theta(only a few horizontal anchors are set for each pix on feature map), we try to use the less horizontal anchor to achieve better performance than oriented anchors. In order to match oriented objects, we adopt a novel label assignment method named Dynamic Anchor Learning(DAL) to achieve a better match result between ground truth boxes and anchors.