Closed s0rrymaker77777 closed 1 year ago
Hi, lab4d runs a single-stage optimization as opposed to 3 stages in banmo.
The design goal of lab4d is to provide a simplified/unified codebase of BANMo, RAC, Total-Recon, PPR, etc, but keep the method the same.
I would say better engineering (or implementation details) plays a major role in terms of the speed-up. Here are some key differences:
Lab4d | BANMo | Type of improvement | |
---|---|---|---|
Sampling unit | Pixels | Lines/images | Fewer iterations |
Camera Extrinsics | MLP + pre-training | Explicit+MLP | Fewer iterations |
Camera Intrinsics | MLP + pre-training | Explicit | Fewer iterations |
Rotation/Skinning Ops | (Dual) Quaternion | Matrix | Per eval time |
Shape/Skinning MLP Size | Small | Large | Per eval time |
Depth samples | 64 | 128 | Per eval time |
2D-3D Features matching | Reuse sampled rays | New samples | Per eval time |
Thanks for your explanation! By the way, have you already implemented the "comp" (fg+bg) situation in Lab4d?
Yes, please check out this: https://lab4d-org.github.io/lab4d/tutorials/single_video_cat.html#reconstruct-the-total-scene
Thanks for your help!
Hi, thanks for your great work. I also tried BANMo before. I want to ask why lab4d takes so much less time than banmo and if lab4d contains three optimization stages in BANMo if I run lab4d 120 rounds. Thanks.