Scene Inference Speedup

Hi, thanks for the great work!

Scene-level post-processing is extremely slow because each object must be processed in a for-loop, and generating one scene currently takes several minutes. I tried converting the object loop into threaded parallelism, but pytorch3d.transforms.Transform3D requires ordered execution, so threads end up blocking and no speedup is achieved.

Do you have any suggestions for speeding up scene composition?
For example: batching Transform3D, vectorized transforms, or any recommended approach to avoid the per-object loop?

Thanks!