Overview:

The Instant Neural Surface Reconstruction repository contains an implementation of NeRF (Neural Radiance Fields) and NeuS (Neural Surface) for neural surface reconstruction. It is based on the Instant-NGP and Pytorch-Lightning frameworks. The training time for NeRF on a NeRF-Synthetic scene is approximately 5 minutes, while it is around 10 minutes for NeuS on a single RTX3090 GPU.

Features:

Rendering capabilities for both NeRF (rendering-nerf) and NeuS (rendering-neus)
Mesh generation for NeRF (mesh-nerf) and NeuS (mesh-neus)
Acceleration techniques such as multiresolution hash encoding and fully fused networks by tiny-cuda-nn
Occupancy grid pruning and rendering by nerfaccout-of-the-box multi-GPU and mixed precision training by PyTorch-Lightning
Hierarchical project layout designed for customization and extension
Flexible experiment configuration with OmegaConf

Summary:

The Instant Neural Surface Reconstruction repository provides an efficient and customizable implementation of NeRF and NeuS for neural surface reconstruction. It offers various features such as rendering, mesh generation, acceleration techniques, multi-GPU training, and flexibility in experiment configuration. The installation process requires installing PyTorch, the tiny-cuda-nn extension, and the necessary dependencies.

Bennyguo

Instant Nsr Pl