Hello, I am interested in your wonderful work on RCF. I am a postgraduate, and my search interest is edge detection. But I met some problems and want to discuss with someone who has worked on similar topics.
First, when I reproduce RCF using Pytorch, I find that the AP is relatively lower than other works like BDCN, CED, and recent BAN. Did you meet such a phenomenon during your implementation? And what's the reason do you think?
The second problem is about the BDCN. I notice that you asked the author of BDCN about not reproducible results. And when I implement using the released code, I met similar problems. Have you ever solve this problem? My retrained model has ODS 0.808 using BSDS and PASCAL, which is much lower than the reported 0.820. And I also met low AP problems.
If you can give some illustrations, I will much appreciate it. And I think such discussions are valuable.
Hello, I am interested in your wonderful work on RCF. I am a postgraduate, and my search interest is edge detection. But I met some problems and want to discuss with someone who has worked on similar topics.
First, when I reproduce RCF using Pytorch, I find that the AP is relatively lower than other works like BDCN, CED, and recent BAN. Did you meet such a phenomenon during your implementation? And what's the reason do you think?
The second problem is about the BDCN. I notice that you asked the author of BDCN about not reproducible results. And when I implement using the released code, I met similar problems. Have you ever solve this problem? My retrained model has ODS 0.808 using BSDS and PASCAL, which is much lower than the reported 0.820. And I also met low AP problems.
If you can give some illustrations, I will much appreciate it. And I think such discussions are valuable.