Matlab version about humbi HOT 13 CLOSED

One more question, did you used only netural model for all subjects?
I noticed there are only a few frames provided (seems every 8 frames), do we have the rest frames or they just remain unopened?

from humbi.

Same problem with hand reconstruction.

from humbi.

@halcyon370 Hi what do you mean by same problem?
The matlab version or you also found that the reconstruction meshes are inaccurate in some cases ?

from humbi.

@MengXinChengXuYuan The reconstruction meshes of hand are also inaccurate. I use matlab of version 2018b.

from humbi.

@halcyon370 I see!

from humbi.

@MengXinChengXuYuan I think the visualization code itself is good. I will ping Jae Shin to try to address your issue since he was working on SMPL body mesh fitting. But as far as I know, we admit that it is not perfect.

from humbi.

@halcyon370 Similarly, I think the visualization code for hand itself is good. But as you are aware of, for some cases the fitting is not perfect and we will try to correct those cases in next round of refinement.

from humbi.

@zhixuany Just curious about your hand fitting procedure,In your paper,you mentioned that you use 2D-3D reprojection error term to reconstruct the MANO hand model,But in your annotation, you provide us 3d keypoints GroundTruth,why you did not use the 3D-3D Euclidean distance error to fit mano, Is that more accurate than your current method?

from humbi.

@zhixuany Actually I think no one can generate perfect smpl params for now, in most cases :p
Thank you for sharing your data, and looking forward to the refinement!

from humbi.

@MengXinChengXuYuan Thanks for having interest in our dataset. The ground-truth that we can perfectly ensure is the "Multiview images". For mesh reconstruction, due to the fitting error (e.g., the error from the 3D keypoint reconstruction propagates to the mesh fitting), it could be inaccurate for some cases. - Jae Shin

from humbi.

@quyanqiu Actually I did "use the 3D-3D Euclidean distance error to fit mano" as you said. In "section 3.3. Hand" of paper we mentioned that "We align the hand model to multiview images by minimizing the Euclidean distance between hand keypoints". Is this sentence somehow misleading?

from humbi.

@zhixuany sorry for misunderstanding,When I check the formular in your appendix its truly 3D-3D loss

from humbi.

@quyanqiu No problem. Thanks for being interested in HUMBI~

from humbi.