Closed SatMa34 closed 2 years ago
Probably because i didn't use the imagenet pretrained model, doing the experiment right now
Yes, the pretrained model helps improve the results, which is also widely-adopted by lots of work in this area(e.g., HRNet, SimpleBasline, etc.). You may have a try.
Hi, thanks for your brilliant work! I have used provided yaml file to train SimBa-Res50-SimDR with input resolution set to 64*64, and the result is as follows:
The AP is only 35.4, much lower than 40.8 mentioned in the paper, did i miss something?