search

korrawe / halo

74 stars 4 forks source link

Some questions about training and dataset. #6

Open jackyyang9 opened 1 year ago

jackyyang9 commented 1 year ago

Hello, Mr.Karunratanakul,

HALO is a nice work, especially the skeleton canonicalization part in paper.

It offers another feasible way to parameterize the pose.

In fact, I am reproducing HALO in FreiHAND dataset and I have some naive questions.

1): FreiHAND dataset doesn't has evaluation dataset, so I cannot observe the val loss in training.

My question is how can I know the training should be stopped from the real-time trainning loss?

2): Could you offer the Supplemently material you mentioned in paper?

I wish to know more details about the experiment.

3): I obtained YT3D dataset from official website.

However, they didn't provide joints and mano parameters.

And I notice that in line 32 of prepare_data_from_mano_param_keypoints.py , there is "default='../../data/youtuhand_raw'".

What is that? How can I get this?

Sincerely looking forward to your reply.

Best wishes.

korrawe commented 1 year ago

Hi,

  1. You can split a portion of the training data for validation.
  2. The supplementary is at the end of the paper on arxiv
  3. You can get key points using the J_regressor matrix in the MANO model. If you want MANO parameters, you can fit a MANO hand to the provided surface points using the fitting script in prepare_data_from_mano_param_keypoints.py. Other than that you can also use a method like SMPLify.

Best,

jackyyang9 commented 1 year ago

Thanks for your quick reply ! I already found the supplemental material in arxiv version and thansk for other answers. Best wishes to all your paper subbmision!

jackyyang9 commented 1 year ago

Thanks again. However I still have questions:

  1. I noticed that in prepare_data_from_mano_param_keypoints.py, you regarded the bones' translation matrixs of palm. If I want to get full transformation matrixs for each bone. How can I get it.
  2. If we have the transformation matrixs, why we still need the root_rotation_matrix?