question in training

[muses0229]

No description provided.

[tiangexiang]

Hi, thanks for letting me know. I have now refactored the code to remove dependency on easy_kpconv, users are expected to install pykeops directly. I have also tried to clean the codes to be more organized. Please pull the newly updated repository and try again. Thank you!

[tiangexiang]

Yes! The training can be resumed from the last checkpoint (as in the same experiment), e.g. please run:
python train.py --cfg configs/occnerf/zju_mocap/387/occnerf.yaml resume True
to load the last checkpoint and resume training.

[tiangexiang]

Sorry for the problem. It has been fixed in the latest commit. Please pull again and give it a try!

[tiangexiang]

To evaluate on certain frame, please use render_allview.sh the script, it will render the 23 images wrt to the given camera configurations.

[tiangexiang]

Sorry for the late response and yes you are correct! Note that in the main paper we didn't report PSNR_body and SSIM_body, feel free to add those results if you find useful :)

[tiangexiang]

Have you recalculated the mask ratio and update the masking configures accordingly? The default masking configure is computed based on the default view and all frames. It is very likely the 100 images from view 4 may be mostly masked (leading to bad results) if you are still using the default masking configures. Please check the masking first!

[tiangexiang]

Yes, we basically decide occlusion_mid first based on the average width of visible human masks in all frames, then decide occlusion_width via binary search to make sure a certain portion of human body will be masked. This process has to be repeated for all NEW videos, include new subjects, new view etc...