search

3D-FRONT-FUTURE / NeuDA

NeuDA: Neural Deformable Anchor for High-Fidelity Implicit Surface Reconstruction (CVPR 2023)
MIT License
88 stars 2 forks source link

About Training time #4

Closed yec22 closed 1 year ago

yec22 commented 1 year ago

Hi, thanks for your great job!

I run the code on an NVIDIA A100 GPU (following neuda_wmask.conf), and the complete training time cost ~5 hours, which is similar to NeuS. This confuses me. As NeuDA leverages hash-table as instant-NGP, shouldn't NeuDA converges faster than NeuS?

Looking forward to you reply!!!

BwCai commented 1 year ago

Thanks for your interest. We tested NeuDA on the Nvidia V100 GPU, NeuDA is a bit faster than NeuS. NeuDA has a less time cost on each iteration since it employs a shallower MLP (4 vs. 8 for NeuS). Besides, NeuDA leverages deformable anchor grids instead of hash grids to achieve better convergence. But NeuDA is slower than Instant-NGP in this implementation because there are some accelerated strategies (such as occupancy grids and fully fused MLPs) used in Instant-NGP not implemented in this repo.