Bayes' Rays is built on top of Nerfstudio. After cloning Bayes' Rays repository, install Nerfstudio as a package by following the installation guide on Nerfstudio installation page
Specifically, perform the following steps
- Create environment
- Dependencies
- Installing nerfstudio. Follow the From pip guidelines, no need to clone Nerfstudio.
Then install Bayes' Rays as a package using pip install -e .
This will allow you to run uncertainty-related commands in the terminal like ns-uncertainty.
The code is tested to run with Nerfstudio version 0.3.1 (commit hash e2e5637d05fc281a28abe7c4f9c86a93e130a085) and version 0.3.3.
For experiments on clean-up task and comparing with Nerfbusters the nerfbusters-changes branch of Nerfstudio is used, to ensure fair comparison with Nerfbusters code. Therefore, if you wish to compare directly to Nerfbusters, install nerfbusters-changes branch of Nerfstudio instead.
First train a NeRF model using ns-train command in Nerfstudio. Currently our code only has support for Nerfacto, instant-NGP and Mip-NeRF models. Here is the demo example on the 'Loader Truck' scene. Download the dataset from here.
- Train a Nerfacto model normally:
ns-train nerfacto --vis viewer --data {PATH_TO_DATA} --experiment-name {EXP_NAME} --output-dir {OUTPUT_DIR} --timestamp main --relative-model-dir=nerfstudio_models --max-num-iterations=30000 nb-dataparser --eval-mode eval-frame-index --train-frame-indices 0 --eval-frame-indices 1 --downscale-factor 2 --center_method focus
Note in this example we use nb-dataparser (instead of the usual nerfstudio-data), which is implemented based on nerfbusters-changes branch, and enables using a small part of the dataset as train set and the rest as test.
- Then extract uncertainty:
ns-uncertainty --load-config {PATH_TO_CONFIG} --output-path {PATH_TO_UNCERTAINTY} --lod 8
Note {PATH_TO_UNCERTAINTY} must be a path to .npy file and 2^lod denotes the uncertainty grid length (i.e. the grid length in the example above is 256).
- Render or view the uncertainty!
To run viewer in the browser and visualize uncertainty:
ns-viewer-u --load-config={PATH_TO_CONFIG} --unc_path {PATH_TO_UNCERTAINTY}
Note that ns-viewer-u is a modified version of ns-viewer command in Nerfstudio and works with the same options. However it additionally takes --unc_path option as the path to the extracted uncertainty .npy file.
In the browser, toggle Output Render to uncertainty and optionally set colormap to inferno. For performing clean-up task interactively, use the Filter Threshold slider to gradually filter out uncertain parts!
To render uncertainty use ns-render-u. :
ns-render-u camera-path --load-config={PATH_TO_CONFIG} --unc_path {PATH_TO_UNCERTAINTY} --output-path {PATH_TO_VIDEO} --downscale-factor 2 --rendered_output_names rgb depth uncertainty --filter-out True --filter-thresh 1. --white-bg True --black-bg False --camera_path_filename {PATH_TO_CAMERA_PATH}
For the 'Loader Truck' scene, {PATH_TO_CAMERA_PATH} is inclued as camera_path.json in the downloaded dataset.
Note that ns-render-u is a modified version of ns-render command in Nerfstudio and works with the same options. However it additionally takes --unc_path option as the path to the extracted uncertainty .npy file. If the option --filter-out True is set and the threshold --filter-thresh is set to a value less than 1, then the clean-up task is performed with the given threshold. With --white-bg and --black-bg control the color of rendered empty parts after clean-up.
The uncertainty evaluation is done on the Light Field (LF) dataset and ScanNet Scenes by correlating depth error with uncertainty, computing via the AUSE metric.
For Light Field (LF) dataset download the data here and for ScanNet scenes use the data here. Note that these folders contain the ground truth depth files for test views, alongside the approximate scale in scale_parameters.txt. The scale parameter is computed using utils/scale_solver.py and is used for the purpose of evaluation to solve the scale ambiguity of NeRF for depth comparison with GT depth. To run scale_solver.py on another scene or to verify the provided scales, you need to have the sparse COLMAP generated point cloud and the ground truth depthmaps then you can run the script as:
python /path/to/scale_solver.py --dataset [scannet, LF] --scene [scene's name]
Train Nerfacto models using the provided settings, and then evaluate by:
ns-eval-u --load-config {PATH_TO_CONFIG} --output-path {PATH_TO_METRICS} --unc_path {PATH_TO_UNCERTAINTY} --dataset_path {PATH_TO_DATA}
where {PATH_TO_METRICS} is a json file.
Training Settings
For ScanNet dataset (setting {SCENE_NAME} to scene_001, scene_079, scene_316 or scene_158):ns-train nerfacto --vis viewer --data {PATH_TO_DATA} --experiment-name {SCENE_NAME} --output-dir {OUTPUT_DIR} --timestamp main --relative-model-dir=nerfstudio_models/ --steps-per-save=2000 --max-num-iterations=30000 --logging.local-writer.enable=False --pipeline.datamanager.camera-optimizer.mode off --pipeline.model.proposal-initial-sampler uniform --pipeline.model.use-average-appearance-embedding True --pipeline.model.background-color random --pipeline.model.disable-scene-contraction True --pipeline.model.distortion-loss-mult 0.001 --pipeline.model.max-res 4096 sparse-nerfstudio --dataset-name {SCENE_NAME}
For LF dataset (setting {SCENE_NAME} to statue, torch, basket or africa):
ns-train nerfacto --vis viewer --data {PATH_TO_DATA} --experiment-name {SCENE_NAME} --output-dir {OUTPUT_DIR} --timestamp main --relative-model-dir=nerfstudio_models/ --steps-per-save=2000 --max-num-iterations=30000 --logging.local-writer.enable=False --pipeline.datamanager.camera-optimizer.mode off --pipeline.model.disable-scene-contraction True --pipeline.model.distortion-loss-mult 0.0 --pipeline.model.near-plane 1 --pipeline.model.far-plane 100. --pipeline.model.use-average-appearance-embedding True --pipeline.model.proposal-initial-sampler uniform --pipeline.model.background-color random --pipeline.model.max-res 4096 sparse-nerfstudio --dataset-name {SCENE_NAME}
The NeRF clean up task is performed on the Nerfbusters dataset, which can be downloaded here.
First make sure that you have nerfbuster-changes branch of nerfstudio installed.
Train Nerfacto models using the provided settings, and then evaluate by:
ns-eval-u --load-config {PATH_TO_CONFIG} --output-path {PATH_TO_METRICS} --unc_path {PATH_TO_UNCERTAINTY} --filter-out
Optionally, you can use --nb-mask option to have the exact same metric definition as Nerfbusters (i.e. with the pseudo-ground-truth visibility masks applied.). The visibility masks can be downloaded here.
If running a comparison on scene x against Nerfbusters, you should run the uncertainty commands (extraction and evaluation) on the x---nerfacto model in the Nerfbusters outputs (i.e Nerfbusters "baseline" output model that does not use the Nerfbusters postprocessing techniques). Evaluation can be run as:
ns-eval-u --load-config {PATH_TO_CONFIG} --output-path {PATH_TO_METRICS} --unc_path {PATH_TO_UNCERTAINTY} --filter-out --nb-mask True --visibility-path {PATH_TO_VISIBILITY_MASKS}
where {PATH_TO_VISIBILITY_MASKS} are the paths to the scene specific visibility masks.
Training Settings
For Nerfbusters initial training model (using `nerfbusters-changes` branch of Nerfstudio):ns-train nerfacto --vis viewer --data {PATH_TO_DATA} --experiment-name nerfbusters --output-dir {OUTPUT_DIR} --timestamp base --relative-model-dir=nerfstudio_models/ --max-num-iterations=30000 nerfstudio-data --eval-mode eval-frame-index --train-frame-indices 0 --eval-frame-indices 1
which is then passed to Nerfbusters pipeline to get baseline and Nerfbusters postprocessed models (Baseline model is just the same Nerfacto model trained for 5K longer).
@article{goli2023,
title={{Bayes' Rays}: Uncertainty Quantification in Neural Radiance Fields},
author={Goli, Lily and Reading, Cody and Sellán, Silvia and Jacobson, Alec and Tagliasacchi, Andrea},
journal={arXiv preprint arXiv:2309.03185},
year={2023}
}