search

zc-alexfan / hold

[CVPR 2024✨Highlight] Official repository for HOLD, the first method that jointly reconstructs articulated hands and objects from monocular videos without assuming a pre-scanned object template and 3D hand-object training data.
https://zc-alexfan.github.io/hold
MIT License
271 stars 7 forks source link

Inquiry Evaluation On Hampali et al.'s Method #3

Closed HaixinShi closed 4 months ago

HaixinShi commented 5 months ago

Hello Zicong,

It is a very great work! Looking forward to the code! I am also working on HO3D dataset and want to compare my result with Hampali et al. I wonder how did you compare your results with their Method? I found the shape in their website is actually Point Cloud and there are tons of points inside the object. Our methods are SDF based so I think they are not fully comparable? image

Thanks a lot in advance!

Best, Haixin

zc-alexfan commented 5 months ago

We ran marching cube to convert SDF to meshes and then we can use the points in the vertices to compare.

HaixinShi commented 5 months ago

Thanks a lot for the reply!!! My concern was that: the GT mesh should only contain thin object surface and our SDF method with marching cube would also generate a thin object surface; however, the point cloud downloaded from Hampali et al contains a lot of points inside the object surface. If we sample(or consider) a lot of points that are too far away from the object surface then its performance would be pretty poor?

zc-alexfan commented 5 months ago

Hmm. I didn't noticed that but we checked qualitatively first before getting any numbers

HaixinShi commented 5 months ago

Same here! Previously I also found my method was better than theirs qualitatively. I just found this problem recently. Do you plan to publish meshes from your method on HO3D? Looking forward to doing some comparison :D

zc-alexfan commented 4 months ago

yes, we will publish data, models, and weights.

zc-alexfan commented 4 months ago

You can sign up for an HOLD account on our README. We will notify the code release there.

anilesec commented 2 months ago

yes, we will publish data, models, and weights.

I could not find the released meshes on HO3D sequences. Am I missing something? THanks in advance!