Closed Sixkplus closed 4 years ago
@Sixkplus We tried it and the performance was really bad. The conv. layers with fixed weights have no way to know which objects they should segment because both the input feature maps and weights are the same (except for the relative coordinates). I still encourage you to do the experiment by yourself, though. Maybe I am wrong.
@tianzhi0549 Thank you for your reply!
Hi, thank you for your nice work!
I'm wondering whether you have contucted the experiments where the normal conv layers are used instead of conditional convolution in the mask head, just like Table 3 of your paper.
I think this can help to better verify the effectiveness of conditional(dynamic) filters. Can you provide some of the related experimental results?