（附论文&代码）不再是方方正正的目标检测输出
点击左上方蓝字关注我们
论文：
https://openaccess.thecvf.com/content/CVPR2021/papers/Han_ReDet_A_Rotation-Equivariant_Detector_for_Aerial_Object_Detection_CVPR_2021_paper.pdf
源代码：
https: //github.com/csuhan/ReDet
1
简要
2
背景
3
新框架
(a) Overall architecture of the proposed Rotation-equivariant Detector. We first adopt the rotation-equivariant backbone to extract rotation-equivariant features, followed by an RPN and RoI Transformer (RT) to generate RRoIs. Then we use a novel Rotation-invariant RoI Align (RiRoI Align) to produce rotation-invariant features for RoI-wise classification and bounding box (bbox) regression.
(b) Rotation-equivariant feature maps. Under the cyclic group CN , the rotation-equivariant feature maps with the size (K, N, H, W ) have N orientation channels, and each orientation channel is corresponding to an element in CN .
(c)RiRoI Align. The proposed RiRoI Align consists of two parts: spatial alignment and orientation alignment. For an RRoI (x, y, w, h, θ), spatial alignment warps the RRoI from the spatial dimension, while orientation alignment circularly switches orientation channels and interpolates features to produce completely rotation-invariant features.
4
实验结果
END
整理不易，点赞三连↓