search

mbanani / unsupervisedRR

[CVPR 2021 - Oral] UnsupervisedR&R: Unsupervised Point Cloud Registration via Differentiable Rendering
https://mbanani.github.io/unsupervisedrr
MIT License
136 stars 20 forks source link

Could you please give me training instructions #6

Closed linxcan closed 1 year ago

linxcan commented 1 year ago

Hello！Thank you for your work. I am very interested in your work. Could you please give me training instructions? I want to try to train. Thank you very much.

mbanani commented 1 year ago

Hi @linxcan,

Thank you for your interest and sorry I never updated the train docs. The config defaults should be the same one as the main experiment. So running the following command should work once you set up the environment and dataset appropriately. To train a base model on ScanNet, you can just use the debug.yaml config as follows:

python train.py train debug.yaml

You could define experiment groups in folder with modification configs, check this portion of the code: https://github.com/mbanani/unsupervisedRR/blob/4ba5b1cc76080f5c6fd99cec155453f43b547bcd/train.py#L36-L40

If you are interested in unsupervised registration, I would suggest checking my follow-up work byoc which builds on this work and simplifies it by removing the rendering component and achieving better performance. If you're interested in learning visual features, you should use the vis_byoc config. You should also check issue #5 for some tips and tricks to improve performance.

Please let me know if you have any other questions.

linxcan commented 1 year ago

Thank you for your reply. It helps me a lot.