Skip to content

Latest commit

 

History

History
143 lines (110 loc) · 4.55 KB

README.md

File metadata and controls

143 lines (110 loc) · 4.55 KB
Raw

UP-NeRF: Unconstrained Pose-Prior-Free Neural Radiance Fields

Project Page | Paper

Injae Kim*, Minhyuk Choi*, Hyunwoo J. Kim†.

This repository is an official implementation of the NeurIPS 2023 paper UP-NeRF (Unconstrained Pose-Prior-Free Neural Radiance Fields) using pytorch-lightning.

🏗️ Installation

git clone https://github.com/mlvlab/UP-NeRF.git
cd UP-NeRF

We recommend using Anaconda to set up the environment.

conda create -n upnerf python=3.8 -y
conda activate upnerf

conda install pytorch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 cudatoolkit=11.3 -c pytorch -y
pip install -r requirements.txt

💻 Quick Start

Data download

1. Manual download

Download the Phototourism data of the scene that you want to train from here and train/val split file of the data from here.

And place them as shown below.

UP-NeRF
├── data
│   └── phototourism
│       ├── brandenburg_gate
│       │   ├── brandenburg_gate.tsv
│       │   └── dense
│       ├── taj_mahal
│       │   ├── taj_mahal.tsv
│       │   └── dense
│       ...
...

2. Script

Or you can simply use our automated download script.

# Example
sh scripts/download_phototourism.sh brandenburg_gate

Scenes provided are {brandenburg_gate, british_museum, lincoln_memorial_statue, pantheon_exterior, sacre_coeur, st_pauls_cathedral, taj_mahal, trevi_fountain}

3. Custom Dataset

To run with your own dataset, please check the format of metadata data/example/metadata.json and configuration file configs/custom.yaml (You can omit c2w fields in metadata.json if pose evaluation is not necessary. In addition, c2w matrices must be right up back format) You must put images in dense/images (mandatory for compatability).

Data Preprocessing

Before training you need to save DINO feature maps and DPT mono-depth maps.

Initialize the external submodule. (Last sanity check on Jan 2st, 2024)

git submodule update --init --recursive

Run the script for preprocessing. (example of Brandenburg Gate scene)

sh ./preprocess/preprocess_all.sh brandenburg_gate

❗ Our script includes downloading checkpoint of DPT. If download fails due to some reasons, you can do it manually by downloading it from here, and move the weight to "./DPT/weights/".

Additionally, we highly recommend saving the cache data.

Run (example)

python prepare_phototourism.py --config configs/brandenburg_gate.yaml

Caching Custom Dataset

Script is slightly different, so we separate the script file.

sh ./preprocess/preprocess_all_custom.sh data/example # you have to specify root directory of dataset
python prepare_phototourism.py --config configs/example.yaml

Training

# If you saved the cache data.
python train.py --config configs/brandenburg_gate.yaml

# If you did not save the cache data.
python train.py --config configs/brandenburg_gate.yaml phototourism.use_cache False

You can change the yaml file to change the scene. Check the config files in ./configs

🔎 Evaluation

Test time optimization

Use tto.py for test time optimization. It optimizes camera poses and appearance embeddings for test images.

Run (example)

python tto.py \
  --result_dir ./outputs/brandenburg_gate/UP-NeRF \
  --ckpt last \
  --optimize_num -1 \
  --wandb

Print results

eval.py prints results (PSNR, SSIM, LPIPS, rotation & translation errors).

Run (example)

python eval.py \
  --result_dir ./outputs/brandenburg_gate/UP-NeRF \
  --ckpt last

📂 Weights

You can download pretrained weights from here. (brandenburg_gate, sacre_coeur, taj_mahal, trevi_fountain)

Cite

@inproceedings{kim2023upnerf,
  title={UP-NeRF: Unconstrained Pose-Prior-Free Neural Radiance Fields},
  author={Kim, Injae and Choi, Minhyuk and Kim, Hyunwoo J},
  booktitle={Advances in Neural Information Processing Systems},
  year={2023}
}

Acknowledge

Our code is based on the implementation of NeRF in the Wild (NeRF-W) and BARF (BARF).