Thanks for your implementation of R-CNN-family by PyTorch. It's very useful for me. But there always has a confusion for me while understanding the code, that is " I think this structure is Fast-RCNN rather than Faster-RCNN".
Based on Faster-RCNN paper, what I understood is Faster-RCNN is not simple "RPN + Fast-RCNN". The main difference is that bbox regression is inside RPN, what RPN output is delta_t<tx, ty, tw, th>, not real proposal<x,y,w,h>. So we can combine the output of RPN and Anchors to calculate the final result of bbox. It has no necessary to feed the result to fully_connected layer again. Here, Faster-RCNN paper said this is different with Fast-RCNN and RCNN-appendix-C, Faster RCNN has different bbox regression header for different scale/ration anchors.
I am looking forward your understanding....
Thanks for your implementation of R-CNN-family by PyTorch. It's very useful for me. But there always has a confusion for me while understanding the code, that is " I think this structure is Fast-RCNN rather than Faster-RCNN". Based on Faster-RCNN paper, what I understood is Faster-RCNN is not simple "RPN + Fast-RCNN". The main difference is that bbox regression is inside RPN, what RPN output is delta_t<tx, ty, tw, th>, not real proposal<x,y,w,h>. So we can combine the output of RPN and Anchors to calculate the final result of bbox. It has no necessary to feed the result to fully_connected layer again. Here, Faster-RCNN paper said this is different with Fast-RCNN and RCNN-appendix-C, Faster RCNN has different bbox regression header for different scale/ration anchors. I am looking forward your understanding....