ScaNeRF: Scalable Bundle-Adjusting Neural Radiance Fields for Large-Scale Scene Rendering

This research combines tile-based hybrid neural fields with parallel distributive optimization to improve bundle-adjusting neural radiance fields.

Environment

• Operating system: Ubuntu 18.04

• g++/gcc: 9.4.0

• GPU/NVCC: V100/11.2

• Python: 3.8.0

• PyTorch/NumPy: 1.9.0/1.24.2

Compile

Follow the steps below to compile the code:

cd cuda 
bash make.sh

Compile the code of ray-mesh intersection calculation.

cd fastMesh
bash make.sh

Compile the code of hash grid:

cd hashgrid 
bash make.sh

Directory Structure

Root directory of a dataset

images: the directory of captured images with .png format

mesh: the directory of proxy mesh

logs: the directory for training logs

camera.log: camera paramaters file

Preprocess

Allocate tiles for a given scene:

python preprocess/build_tiles.py config/[SCENE].yaml [GPU IDX]

Training

Training in parallel:

python admm_trainer.py config/[SCENE].yaml

Rendering

python render.py config/[SCENE].yaml [GPU IDX] [SCENE]

Here [SCENE] file is inside the log file after training.

License

Copyright (C) 2007 Free Software Foundation, Inc. https://fsf.org/ Everyone is permitted to copy and distribute verbatim copies of this license document, but changing it is not allowed.