[target detection] r-cnn, fast r-cnn, fast r-cnn learning

Thunderbolt boss learns trilogy ：

1. First read others' explanations on the Internet , Leave a general impression

2. Read the original .（ It's very important , details , Find more details ）

3. Read code .github Search a large number of , Pick a code warehouse that is updated frequently , mark star More ,clone. according to readme Run through code , Analyze author code （ First analyze the network construction part , Re analyze data preprocessing and loss calculation , Basic mastery , Combination of code and paper ）

---

R-CNN：SS The algorithm generates candidate boxes , For each candidate box, use CNN The extracted features , Send in SVM classifier , Then use the regressor to fine correct the position of the candidate box .

Fast R-CNN： Optimize ： Do not extract features from each candidate box , Instead, feature map is extracted from the whole image , then SS The candidate frame generated by the algorithm is projected onto the feature map to obtain the corresponding feature matrix .

Faster R-CNN： Optimize ： The method of generating candidate boxes is determined by SS The algorithm is changed to RPN The Internet

Reference material

1.1Faster RCNN A collection of theories

R-CNN

Fast R-CNN

Loss calculation

Faster R-CNN

review Fast R-CNN


RPN