hyper-parameters about CRF?
pigcv opened this issue · 2 comments
Hi! Thanks for you code! This repo doesn't provided detailed hyper-parameters (i.e, crf_t, alpha) when applying dCRF to the pseudo CAM label, can you provide them?
Hello, @pigcv
Thanks for your interest.
We conducted an experiment for comparing the performance of pseudo-masks. We also observed the performance when applying dCRF. We tested various CRF hyper-parameters based on the previous papers and chose them with the best performance. We referred to the official code/paper of DeepLab, PSA(https://github.com/jiwoon-ahn/psa), and OAA.
The difference in the performance according to the hyper-parameters among the papers is not that big, but not negligible.
Finally, we chose the hyper-parameters of DeepLab-V2 (ResNet101-based) for dCRF in the experiments. Please see the website of DeepLabV2 for them. By the way, the hyper-parameters are different according to the backbone network in the official DeepLab code.
If you have any question, please leave an issue here :)
Thanks!
Thanks for your reply~