FeatureNeRF

This repository is the official implementation of the paper:

FeatureNeRF: Learning Generalizable NeRFs by Distilling Foundation Models
Jianglong Ye, Naiyan Wang, Xiaolong Wang
ICCV, 2023

Project Page / ArXiv / Video

Environment Setup

(Our code has been tested with python 3.7, torch 1.12.1, CUDA 11.3 and RTX 3090)

To set up the environment, follow these steps:

conda create -n featnerf python=3.7 -y && conda activate featnerf
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch -y
pip install "git+https://github.com/facebookresearch/[email protected]"
pip install tqdm tensorboard wandb opencv-python dotmap imageio omegaconf pyhocon==0.3.59 imgviz seaborn scikit-image
# pyhocon 0.3.60 does not work

Note that the dependencies listed above are recommended, latest versions may result in sub-optimal performance.

Data Preparation

Processed Data

We render 6 categories (Chair, Car, Airplane, Table, Bottle and Motorcycle) from ShapeNet V1 dataset with Blender and collect annotations from ShapeNet Part Dataset, PartNet and KeypointNet.

Download our processed dataset from OneDrive or Google Drive and extract it to the data directory. The directory structure should look like this:

PROJECT_ROOT
├── ...
└── data
    ├── ...
    ├── img
    │   ├── 02691156
    │   ├── 02958343
    │   └── ...
    ├── camera
    ├── kp_2d_anno
    ├── kp_3d_anno
    ├── part_2d_anno
    └── part_3d_anno

Since the data processing envolves filtering, alignment, and rendering of multiple datasets, cleaning this part of the code is overcomplex. Please contact the author for any questions regarding the data processing.

Generate Teacher Features

python process_data/generate_dino_feat.py  # takes around 10 hours for 6 categories on one RTX 3090

Running

Training

wandb login  # if you want to use wandb
python train/train.py -n feature_car_dino_256_0.25_share_coord_exp -c conf/exp/feature_dino_256_0.25_share_coord.conf --synset 02958343 --epochs 500

Change different --synset for different categories. If do not want to use wandb, set --no_wandb in the command line.

Evaluation

We provide pre-trained models for evaluation. Download the pre-trained models from OneDrive or Google Drive and extract them to the outputs directory. Run the following commands for evaluation:

python eval/eval_2d_task.py -n feature_car_dino_256_0.25_share_coord_exp -c conf/exp/feature_dino_256_0.25_share_coord.conf -F feature_dino_256_2d_part_kp_anno --synset 02958343 --save_vis
python eval/eval_nv_2d_task.py -n feature_car_dino_256_0.25_share_coord_exp -c conf/exp/feature_dino_256_0.25_share_coord.conf -F feature_dino_256_2d_part_kp_anno --synset 02958343 --save_vis
python eval/eval_3d_task.py -n feature_car_dino_256_0.25_share_coord_exp -c conf/exp/feature_dino_256_0.25_share_coord.conf -F feature_dino_256_3d_part_kp_anno --synset 02958343 --save_vis

Evaluation results will be saved in the outputs directory. Change different --synset for different categories. If do not want to save visualization, remove --save_vis in the command line.

Citation

@article{ye2023featurenerf,
  title={FeatureNeRF: Learning Generalizable NeRFs by Distilling Foundation Models},
  author={Ye, Jianglong and Wang, Naiyan and Wang, Xiaolong},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={8962--8973},
  year={2023}
}

Acknowledgements

The code is heavily based on pixel-nerf, thanks for the great work!