Question about d (depth)

Jaraxxus-Me / VoxDet

Official code for NeurIPS 2023 SpotLight: VoxDet: Voxel Learning for Novel Instance Detection

24 stars 4 forks source link

Open junjun1023 opened 6 months ago

junjun1023 commented 6 months ago

Good work!

I'm confused to the following questions:

Is there any strategy to define d (depth) to reshape 2D features into 3D voxels?
How does pre-defined depth work to all scenarios and instances?
Why does 'Voxel Relation' output 6D continuous representation instead of a rotation matrix directly?

Thank you for the open source code and data!

Jaraxxus-Me commented 6 months ago

We didn't use depth, we just used the rotation matrix to rotate the 3D voxels. This strategy is inspired by VideoAutoEncoder.
I think we didn't predefine any depth hyperparameter. VoxDet only needs to rotate the voxel (we assume object-centric feature).
I tried both rotation matrix and 6D continuous vector and found the vector estimation is slightly better.