一、R-CNN:Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation
1、先通过图分割方法获取原始区域(Efficient Graph-Based Image Segmentation):
http://blog.csdn.net/surgewong/article/details/39008861
http://blog.csdn.net/huixingshao/article/details/42642973
2、再通过选择性搜索(Selective Search for Object Recoginition)的方法对上述原始区域进行合并操作:
http://blog.csdn.net/surgewong/article/details/39316931
通过颜色、文理、大小、吻合相似度约束对原始区域进行和合并,得到Buonding Box。常用的物体特征识别有HOG和bag-of-words,本文采用的是bag-of-words.
http://blog.csdn.net/charwing/article/details/27180421
3、R-CNN:Rich feature hierarchies for accurate object detection and semantic segmentation
http://yidianzixun.com/home?page=article&id=0EabZyXX&up=144
该博文很好的讲述了R-CNN的流程
4、R-CNN基础知识和算法细节分析讲解:
http://blog.csdn.net/u011534057/article/details/51218250
二、Fast R-CNN
1、SPP-Net网络主要是解决深度网络输入尺寸不灵活的问题,使用SPP进行检测,先用提候选proposals方法(selective search)选出候选框,不过不像RCNN把每个候选区域给深度网络提特征,而是整张图提一次特征,再把候选框映射到conv5上,因为候选框的大小尺度不同,映射到conv5后仍不同,所以需要再通过SPP层提取到相同维度的特征,再进行分类和回归,后面的思路和方法与RCNN一致。
http://blog.csdn.net/u011534057/article/details/51219959
2、Fast R-CNN:
http://blog.csdn.net/u011534057/article/details/51241831
2、Fast R-CNN:
http://blog.csdn.net/u011534057/article/details/51241831
三、Faster R-CNN
1、Towards Real-Time Object Detection with Region ProposalNetworks译文:
http://blog.csdn.net/u011326478/article/details/52872614
http://blog.csdn.net/u011534057/article/details/51259812
2、RPN网络解析:
http://blog.csdn.net/sloanqin/article/details/51545125
四、整体梳理:
http://blog.csdn.net/xyy19920105/article/details/50817725
http://blog.csdn.net/myarrow/article/details/51878004