Pre-NeRF 360: Enriching Unbounded Appearances for Neural Radiance Fields

The repository contains the code release for paper: Pre-NeRF

How it run?

We highly recommend to use the docker image in /docker. Please make sure that your installation is a GPU friendly.

DATA_DIR=path/to/the/scene/directory
docker run -v $(pwd):/workspace \
           -v /path/to/360_v2_nk:/workspace/360_v2_nk \
           --gpus all --shm-size 24G --name prenerf --rm -it \
           --entrypoint bash -d prenerf/prenerf:latest \
           single_scene_processing.sh $DATA_DIR

Do it yourself?

If you want to run the scripts instead of downloading the data. You can do the following:

bash rsync.sh 

Then, you need to run our multi_scene_processing.sh

bash multi_scene_processing.sh n5k360

Or for a single scene

bash single_scene_processing.sh n5k360/dish_1550705786

Run on your custom data?

Each scene should be as one or more videos in a directory. For example

n5k360l/
├── dish_1550705786 (scene 1)
│   ├── camera_A.h264
│   ├── camera_B.h264
│   ├── camera_C.h264
│   └── camera_D.h264
├── dish_1550705888 (scene 2)
│   ├── camera_A.h264
│   ├── camera_B.h264
│   ├── camera_C.h264
│   └── camera_D.h264
└── dish_1550705939 (scene 3)
    ├── camera_A.h264
    ├── camera_B.h264
    ├── camera_C.h264
    └── camera_D.h264

Or if you have the scene as images, each scene folder should have images folder

n5k360/dish_1550704903/
└── images
   ├── 0001.png
   ├── 0002.png
   ├── 0003.png
   ├── 0004.png
   ├── ...
   ├── 0063.png
   └── 0064.png

How fast is it?

[We] used GNU Parallel, it also shows a better utilisation for the multiprocessing on the GPU. However, the extending it takes 1-2hrs of preparing data.

Licence

Code: GNU General Public License v3.0 (see LICENSE), except for the third-party code below, which remains under the terms of its authors:

• colmap2nerf.py, from NVIDIA's instant-ngp (scripts/colmap2nerf.py), which keeps its NVIDIA copyright notice and is governed by the NVIDIA Source Code License (non-commercial use). Helper functions taken from the same script (for example closest_point_2_lines) also appear in llff2nerf.py.
• blur_fft/ follows the FFT blur detector of the PyImageSearch tutorial "OpenCV Fast Fourier Transform (FFT) for blur detection in images and video streams" (A. Rosebrock, 2020).
• The git submodules (LLFF, COLMAP, Nutrition5k, pixel-perfect-sfm, Hierarchical-Localization, detectron2) are separate projects under their own licences.

Data and models: the N5k360 dataset, the evaluation dataset and the pre-trained models are released under the Creative Commons Attribution-NonCommercial 4.0 International licence (CC BY-NC 4.0). You may share and adapt them for any non-commercial purpose, including research and teaching, provided you give appropriate credit. N5k360 is derived from Nutrition5k (Thames et al., CVPR 2021), released by Google under CC BY 4.0; please credit Nutrition5k as well when you use N5k360.

Citation

If you use this code or the N5k360 dataset, please cite the paper:

@inproceedings{almughrabi2025enriching,
  title     = {Enriching Unbounded Appearances for Neural Radiance Fields},
  author    = {AlMughrabi, Ahmad and Haroon, Umair and Marques, Ricardo and Radeva, Petia},
  booktitle = {Pattern Recognition and Image Analysis (IbPRIA 2025)},
  series    = {Lecture Notes in Computer Science},
  volume    = {15937},
  pages     = {80--96},
  publisher = {Springer},
  year      = {2026},
  doi       = {10.1007/978-3-031-99565-1_7}
}

An earlier version is available as a preprint:

@misc{almughrabi2023prenerf,
      title={Pre-NeRF 360: Enriching Unbounded Appearances for Neural Radiance Fields},
      author={Ahmad AlMughrabi and Umair Haroon and Ricardo Marques and Petia Radeva},
      year={2023},
      eprint={2303.12234},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Contact

If you have any questions about the code, the N5k360 dataset or the paper, please email the authors or open an issue.