NeRFuser

Official code release for "NeRFuser: Large-Scale Scene Representation by NeRF Fusion" [paper].

Original Videos

Raw NeRFs

NeRFuser Result

Installation

0. Create a conda environment and activate it

conda create -n nerfuser -y python=3.10 && conda activate nerfuser

1. Install dependencies

nerfstudio

pip install torch torchvision
pip install ninja git+https://github.com/NVlabs/tiny-cuda-nn/#subdirectory=bindings/torch
pip install nerfstudio

hloc

git clone --recurse-submodules git@github.com:cvg/Hierarchical-Localization.git && pip install -e Hierarchical-Localization

Misc

# due to a bug in open3d 0.17.0, we use the previous version
pip install imageio-ffmpeg open3d==0.16.0

2. Install NeRFuser

git clone git@github.com:ripl/nerfuser.git && cd nerfuser/
pip install .

Data Preparation

The data preparation assumes that you have several videos focusing on different yet overlapping portions of the same scene. Let ext denote the video file extension (e.g. mp4, mov, etc.), then one of the videos should be named test.ext, from which images will be extracted for blending evaluation. Others can be named whatever you like. Assume you have collected 3 more videos besides test.ext, whose names w/o the ext extension are stored as A, B and C. First put all the video files (including test.ext) in the directory DATASET_DIR, then run the following command to prepare data for training NeRFs:

python -m nerfuser.prep_data \
    --dataset-dir $DATASET_DIR \
    --vid-ids test $A $B $C \
    --downsample 8 \
    --extract-images \
    --run-sfm \
    --write-json \
    --vis

Please run python -m nerfuser.prep_data -h for more details. A sample dataset containing both videos and prepared data is provided here.

Training NeRFs

Let MODELS_DIR be the directory where you want to save the trained NeRF models. Run the following command to train a NeRF model corresponding to each video other than test:

for VID in $A $B $C; do
    ns-train nerfacto \
        --output-dir $MODELS_DIR \
        --data $DATASET_DIR/$VID \
        --viewer.quit-on-train-completion True \
        --pipeline.datamanager.camera-optimizer.mode off
done

Please run ns-train nerfacto -h for more details. Trained NeRF models on the sample dataset are provided here. Note that the provided model ckpts are trained with nerfstudio 0.3.2. If you encounter issues loading them, consider either installing the exact matching version of nerfstudio, or training your own as above.

NeRF Registration

Let TS_A, TS_B and TS_C be the timestamps of the trained NeRF models for videos A, B and C respectively. Run the following command to register the NeRF models:

python -m nerfuser.registration \
    --model-dirs $MODELS_DIR/$A/nerfacto/$TS_A/nerfstudio_models $MODELS_DIR/$B/nerfacto/$TS_B/nerfstudio_models $MODELS_DIR/$C/nerfacto/$TS_C/nerfstudio_models \
    --name my_scene \
    --model-names $A $B $C \
    --model-gt-trans I \
    --cam-info $DATASET_DIR/test/transforms.json \
    --render-views \
    --run-sfm \
    --compute-trans \
    --vis

Registration results are saved in outputs/registration by default. Please run python -m nerfuser.registration -h for more details.

NeRF Blending

Run the following command to query the NeRFs with test poses as in test and generate the blending results:

python -m nerfuser.blending \
    --model-dirs $MODELS_DIR/$A/nerfacto/$TS_A/nerfstudio_models $MODELS_DIR/$B/nerfacto/$TS_B/nerfstudio_models $MODELS_DIR/$C/nerfacto/$TS_C/nerfstudio_models \
    --name my_scene \
    --model-names $A $B $C \
    --cam-info $DATASET_DIR/test/transforms.json \
    --test-poses $DATASET_DIR/test/transforms.json \
    --test-frame world \
    --blend-views \
    --evaluate

Blending results are saved in outputs/blending by default. Please run python -m nerfuser.blending -h for more details.

NeRF Fusion

Alternative to the above two steps, you can run the following command to perform NeRF registration and blending in one go:

python -m nerfuser.fuser \
    --model-dirs $MODELS_DIR/$A/nerfacto/$TS_A/nerfstudio_models $MODELS_DIR/$B/nerfacto/$TS_B/nerfstudio_models $MODELS_DIR/$C/nerfacto/$TS_C/nerfstudio_models \
    --name my_scene \
    --model-names $A $B $C \
    --model-gt-trans I \
    --cam-info $DATASET_DIR/test/transforms.json \
    --render-views \
    --run-sfm \
    --compute-trans \
    --test-poses data/ttic/common_large/test/transforms.json \
    --test-frame world \
    --blend-views \
    --eval-blend

Please run python -m nerfuser.fuser -h for more details.

Citing NeRFuser

If you find our work useful in your research, please consider citing the paper as follows:

@article{fang23,
    Author  = {Jiading Fang and Shengjie Lin and Igor Vasiljevic and Vitor Guizilini and Rares Ambrus and Adrien Gaidon and Gregory Shakhnarovich and Matthew R. Walter},
    Title   = {{NeRFuser}: {L}arge-Scale Scene Representation by {NeRF} Fusion},
    Journal = {arXiv:2305.13307},
    Year    = {2023},
    Arxiv   = {2305.13307}
}