New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Train/val on MS-COCO with fully (80 things + 91 stuff) categories #13
Comments
Sorry for my mistake, I didn't modify the setting of 'EvalCOCO' on lines 287 and 296 in 'train_picie.py' to generate 171 full categories labels for evaluation. I have fixed this issue and got the following results:
Do these results make sense? |
Hi zyy-cn, Thank you for your interest in our work! In our work, we strictly limit to the coarse set of classes. Our loss is under the premise that classes are visually distinguishable under geometric and photometric variations. With the complete set of labels in MS-COCO, this will not be true as there are many classes that are semantically different but visually very similar. Unfortunately, I think this is still a limitation in unsupervised learning in general. Hence, regarding your experimental results, I have not tried this on my end, but I would not be surprised if it performs poorly. |
OK, I got your point. Many thanks for your reply! |
Hi:
Thanks for sharing this great work! Recently I try to adapt this work to the task of training PiCIE on MS-COCO with full class categories without merging them into superclasses (i.e. 12 things + 15 stuff), by doing the following modifications:
here are my train logs:
I got only 2.7377 mIoU after training. Was this result as expected? Or do you have any suggestions for this full-class setting? Thanks.
The text was updated successfully, but these errors were encountered: