From NeRFs to Gaussian Splats, and Back#
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. The official code can be found here.
Installation#
This repository follows the nerfstudio method template
0. Install Nerfstudio dependencies#
Please follow the Nerfstudio installation guide to create an environment and install dependencies.
1. Install the repository#
Clone and navigate into this repository. Run the following commands:
pip install -e nerfsh
and
pip install -e nerfgs.
Finally, run ns-install-cli.
2. Check installation#
Run ns-train --help. You should be able to find two methods, nerfsh and nerfgs, in the list of methods.
Downloading data#
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.
NeRFGS: NeRFs to Gaussian Splats#
Training NeRF-SH#
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
NeRFGS: Converting NeRF-SH to Guassian splats#
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
Visualize converted Gaussian splats#
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
Fintuned NeRFGS#
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
GSNeRF: Gaussian Splats to NeRFs#
Scene modification#
Coming soon
Rendering new training images#
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
GSNeRF: Training on new training images#
ns-train nerfsh --data exports/splatting_data --pipeline.model.camera-optimizer.mode off nerfstudio-data --eval-mode filename
Extending the method#
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.
Method#
NeRF-SH#
The NeRF-SH field structure is shown above in the overview figure. NeRF-SH is modified from Nerfacto to predict spherical harmonics (degree 3 by default) for each rgb channel. The volumetric rendering process remains unchanges: at each point along a ray, we predict the spherical harmonics and calculate color based on the view direction.
NeRFGS#
Given a trained NeRF-SH, we extract pointcloud based on rendered depth, following the pointcloud export pipeline in Nerfstudio. In addition to exporting the location of each point, we export spherical harmonic coefficients and density predicted by NeRF-SH. We exclude rays with low opacity or corresponding to the sky. We initialize each Gaussian as isotropic where the scale depends on the sparsity of points in the neighborhood. Specifically, the scale of each Gaussian is half of the average distance between each point and its three nearest neighbors. To avoid large Gaussians, the scale is clipped between 0 and 0.8-th quantile of the scales in the scene. The exported Gaussian splats already captures the geometric and photometric properties of the scene. To obtain fine-grained splats and remove outliers during exportation, we finetune the splats using training views.
GSNeRF#
After editing the Gaussian splats, rendered training views from edited Gaussian splats can be used to create a new dataset. The new dataset can be used to train/update other models (especially implicit representations that are difficult to edited directly).