NVlabs / BundleSDF

[CVPR 2023] BundleSDF: Neural 6-DoF Tracking and 3D Reconstruction of Unknown Objects
https://bundlesdf.github.io/
Other
998 stars 110 forks source link

Using the reconstructed model for pose estimation #84

Closed fuzhao123232 closed 1 year ago

fuzhao123232 commented 1 year ago

请问作者我可以先把一个物体绕着相机拍一圈,然后建模,之后推理的时候直接用这个模型去做该物体的位姿估计吗?可以达到实时的效果吗

wenbowen123 commented 1 year ago

Yes, you can use the reconstructed CAD model similarly to how people normally deal with the instance-level objects (e.g. LINEMOD, YCB). BundleSDF enables to acquire such a model for unknown objects.

Once you have created the model, we have another repo for instance-level 6D pose tracking given the CAD models, which can run very fast (90 hz) https://github.com/wenbowen123/iros20-6d-pose-tracking

monajalal commented 10 months ago

@wenbowen123 do you have instructions on how to feed the data from milk output to iros20-6d-pose-trackig or any other custom object? Do you only need the CAD model or also the annotations? Any guidance would be really helpful.

VenkatSBitra commented 7 months ago

@wenbowen123 Could you please advice on the steps I should follow to get the model transferred to [iros20-6d-pose-tracking]?

wenbowen123 commented 7 months ago

@monajalal @VenkatSBitra We have a better version named FoundationPose, which will be seamless compatible with BundleSDF. Please stay tuned.

monajalal commented 7 months ago

@wenbowen123 Thanks for your updates. There is no code published yet in your linked repo. We also had to let go of BundleSDF in our exploration since it is not fast for 6D pose estimation unfortunately.

How fast is Foundation Pose?

VenkatSBitra commented 7 months ago

@monajalal Can you suggest some alternatives that you have explored for fast 6D pose estimation?