This is the implementation of From NeRFs to Gaussian Splats, and Back; An efficient procedure to convert back and forth between NeRF and GS, and thereby get the best of both approaches. New dataset in the paper can be downloaded from this google drive link.
Giannini_demo_compressed.mp4
This repository follows the nerfstudio method template
Please follow the Nerfstudio installation guide to create an environment and install dependencies.
Clone and navigate into this repository. Run the following commands:
pip install -e nerfsh
and
pip install -e nerfgs.
Finally, run ns-install-cli.
Run ns-train --help. You should be able to find two methods, nerfsh and nerfgs, in the list of methods.
You could download the Giannini-Hall and aspen datasets from this google drive link. Our new dataset (Wissahickon and Locust-Walk) can be downloaded from this google drive link.
Run the following command for training. Replace DATA_PATH with the data directory location.
ns-train nerfsh --data DATA_PATH --pipeline.model.camera-optimizer.mode off
To train on Wissahickon or Locust-Walk dataset, you need to add nerfstudio-data --eval-mode filename to properly split training and validation data, i.e.,
ns-train nerfsh --data DATA_PATH --pipeline.model.camera-optimizer.mode off nerfstudio-data --eval-mode filename
Replace CONFIG_LOCATION with the location of config file saved after training.
ns-export-nerfsh --load-config CONFIG_LOCATION --output-dir exports/nerfgs/ --num-points 2000000 --remove-outliers True --normal-method open3d --use_bounding_box False
Replace DATA_PATH with the data directory location. You also need to add nerfstudio-data --eval-mode filename if train on Wissahickon or Locust-Walk.
ns-train nerfgs --data DATA_PATH --max-num-iterations 1 --pipeline.model.ply-file-path exports/nerfgs/nerfgs.ply
We reduces the learning rate for finetuning. You also need to add nerfstudio-data --eval-mode filename if train on Wissahickon or Locust-Walk.
ns-train nerfgs --data DATA_PATH --pipeline.model.ply-file-path exports/nerfgs/nerfgs.ply --max-num-iterations 5000 --pipeline.model.sh-degree-interval 0 --pipeline.model.warmup-length 100 --optimizers.xyz.optimizer.lr 0.00001 --optimizers.xyz.scheduler.lr-pre-warmup 0.0000001 --optimizers.xyz.scheduler.lr-final 0.0000001 --optimizers.features-dc.optimizer.lr 0.01 --optimizers.features-rest.optimizer.lr 0.001 --optimizers.opacity.optimizer.lr 0.05 --optimizers.scaling.optimizer.lr 0.01 --optimizers.rotation.optimizer.lr 0.0000000001 --optimizers.camera-opt.optimizer.lr 0.0000000001 --optimizers.camera-opt.scheduler.lr-pre-warmup 0.0000000001 --optimizers.camera-opt.scheduler.lr-final 0.0000000001
Coming soon
In the new dataset, training images are rendered from splats. Replace CONFIG_LOCATION with the location of config file saved after training.
ns-nerfgs-render --load-config CONFIG_LOCATION --render-output-path exports/splatting_data --export-nerf-gs-data
ns-train nerfsh --data exports/splatting_data --pipeline.model.camera-optimizer.mode off nerfstudio-data --eval-mode filename
The conversion from NeRF to GS has inefficiency as mentioned at the discussion section of the paper. We welcome your efforts to reduce the inefficiency! The code for conversion is mainly in nerfsh/nerfsh/nerfsh_exporter.py.
@misc{he2024nerfs,
title={From NeRFs to Gaussian Splats, and Back},
author={Siming He and Zach Osman and Pratik Chaudhari},
year={2024},
eprint={2405.09717},
archivePrefix={arXiv},
primaryClass={cs.CV}
}