从faster-rcnn谈起:
faster-rcnnhttps://zhuanlan.zhihu.com/p/31426458具体ROI-pooling的解读:
ROIpoolinghttps://blog.csdn.net/u011436429/article/details/80279536RFCN到底怎么改进了ROI-pooling
RFCNhttps://www.cnblogs.com/shouhuxianjian/p/7710707.htmlDCN1继续改进了ROI-pooling
https://arxiv.org/abs/1703.06211https://arxiv.org/abs/1703.06211DCNv2改进DCNv1的Deformabel Convolutions为 Modulated Deformabel Convolutions,同时还要加入 RCNN feature mimicking (网络模仿的训练策略)。
这里的主分支和子分支的规则是类似的,子分支的Whole Image Regions对应了主分支的原图,子分支的输入是原图的crop,所以子分支里不需要经过主分支的Region Proposals。
DCNv2https://blog.csdn.net/u014380165/article/details/88072737Reppoint的代码解读
RepPointhttps://zhuanlan.zhihu.com/p/326058238