search

korrawe / grasping_field

GNU General Public License v3.0
124 stars 13 forks source link

Pretrained models for RGB image to 3d reconstruction #3

Open anilesec opened 3 years ago

anilesec commented 3 years ago

Dear Authors,

Really nice work! Thank you for making the code available. Would you be releasing the pre-trained model for RGB to 3D reconstruction network?

Thanks!

korrawe commented 3 years ago

Hi,

We are working on making the reconstruction model and the training code available. It should be ready in the next few days or next week.

As a side note for the reconstruction task, our aim is only to show that we could use the GraspingField for RGB reconstruction but the numbers might not be directly comparable to previous works. As mentioned in appendix C3, because we do not have the keypoints from the GraspingField prediction, we instead align the predicted mesh to the ground truth mesh (scale and translation) for evaluation. But, because our predicted mesh has more vertices than MANO mesh, this alignment and evaluation tend to favor our mesh more than the MANO mesh from other method.

Best, Korrawe

anilesec commented 3 years ago

Hi Korrawe,

Thanks for the quick response and the note. I will look forward to the models.

Best, Anil