This may be the reason why the segmentation result is lower than the paper value.

lhoyer / MIC

[CVPR23] Official Implementation of MIC: Masked Image Consistency for Context-Enhanced Domain Adaptation

261 stars 40 forks source link

This may be the reason why the segmentation result is lower than the paper value. #62

Open QChhh123 opened 1 year ago

QChhh123 commented 1 year ago

I followed the recommends of issue #8 and #9 to set the cuda 11.0 and random seed 2, but still got a lower result of 74.7 mIoU. So I compare my log file with yours and find a difference. Can you provide some information about the losed parameters? Looking forward to your early reply. diff

lhoyer commented 1 year ago

These parameters are for a feature that I have discarded later. In the provided config, this feature is disabled anyway. After removing the related code, I ran MIC again and was able to reproduce the results in my environment. Have you tried the simplified environment suggested in #8? This seems to solve reproducibility issues for others.

QChhh123 commented 1 year ago

Yes, I have tried it. There is no solution for me at the moment. Could you check the log file for me, thank you very much. 20230811_221438.log