R2CNN++: Multi-Dimensional Attention Based Rotation Invariant Detector with Robust Anchor Strategy.

摘要: Object detection has been a building block in computer vision. Though considerable progress made, there still exist challenges for objects with small size, arbitrary direction, and dense distribution. Apart from natural images, such issues are especially pronounced aerial images of great importance. This paper presents novel multi-category rotation detector small, cluttered rotated objects, namely SCRDet. Specifically, sampling fusion network is devised which fuses multi-layer feature effective anchor sampling, to improve the sensitivity objects. Meanwhile, supervised pixel attention channel jointly explored object by suppressing noise highlighting feature. For more accurate estimation, IoU constant factor added smooth L1 loss address boundary problem rotating bounding box. Extensive experiments on two remote sensing public datasets DOTA, NWPU VHR-10 as well image COCO, VOC2007 scene text data ICDAR2015 show state-of-the-art performance our detector. The code models will be available at this https URL.