search

NVlabs / BundleSDF

[CVPR 2023] BundleSDF: Neural 6-DoF Tracking and 3D Reconstruction of Unknown Objects
https://bundlesdf.github.io/
Other
992 stars 109 forks source link

Resulting mesh after step 1 is very weird and wrong #123

Closed monajalal closed 7 months ago

monajalal commented 9 months ago

I used a tripod and set the camera 1 meter away (z-far i run_custom.py : 1) from the object of interest and did use a rotating table with controller.

At the end the resulting mesh ended up being so weird Screenshot from 2023-11-22 09-57-50

Pose visualization after step 1 sounds reasonable for the most part.

Screenshot from 2023-11-22 09-59-24 Screenshot from 2023-11-22 09-59-35

Could you please provide any suggestion as to why the reconstructed mesh is damaged?

here is the log of step 1 https://drive.google.com/file/d/1o1wq_Yw9TZDCedkBBMzGFZXqNJTFlVyT/view?usp=sharing

monajalal commented 9 months ago

also here is the entire input and output for blue cup if you wanna reproduce the results or check it

https://drive.google.com/drive/folders/1FtfCd1MaE7z2U3OMCMDuv-ps951dFJEx?usp=sharing

fedona commented 9 months ago

How are the masks and depth frames?

I believe that it might be better if the object is centered in the picture, as at the edges depth information might not be present. Moreover the mug should be better if is not reflecting light.

Are you sure the pose is correct? Since your object lacks of a texture, could happen that as soon as the handle is hidden, the object will look standing still when is actually rotating.