runsong123/adafit

Code Release for ICCV 2021 (oral), "AdaFit: Rethinking Learning-based Normal Estimation on Point Clouds"

adafit's Issues

About surface reconstruction and denoising

Hello， your work is excellent!
Could you please tell me about the approaches to surface reconstruction and point cloud denoising in your Paper? This is because the unoriented normal vectors generated by AdaFit cannot be directly used for Poisson reconstruction.

about real dataset

many thanks for your great work! can you release sampleed point cloud and gt_normal from SceneNN dataset mentioned in the paper?

available trained model

thank you very much for your attention. Can you provide the trained model ?

How to visualize the error and estimated normals in Figure 7 and Figure 8 in the paper

Dear authors,
Thanks for your great work.
Can you share how to visualize the error and estimated normal in Figure 7 and Figure 8 in the paper.
Thank you very much for your reply.

Multi AdaFit Training： RuntimeError: inconsistent tensors dimensions input: 2 other: 3

Sorry for bothering you.
When I train a multi-model, the line 383 of train_n_est_multi_scale.py throws an error train_n_est_multi_scale.py.
The code is normal_loss = 0.5 * torch.norm(torch.cross(o_pred, o_target, dim=-1), p=2, dim=1).mean()
It seems that the shape of o_pred([128, 3]) and o_target([128, 175, 3) is not matched.
But I have no idea where to change, sorry for bothering you again.
Gratefully.

average time per frame

Dear authors,
Thank you for your great work.
I want to know the processing time per frame.
Thank you very much for for your reply.

question about code

thanks for your work.
I read the code ,have some questions, What does this line of code mean？