Pix2NeRF: Unsupervised Conditional π-GAN for Single Image to Neural Radiance Fields Translation (CVPR 2022)
Pix2NeRF: Unsupervised Conditional π-GAN for Single Image to Neural Radiance Fields Translation
Shengqu Cai, Anton Obukhov, Dengxin Dai, Luc Van Gool
Abstract: We propose a pipeline to generate Neural Radiance Fields (NeRF) of an object or a scene of a specific class, conditioned on a single input image. This is a challenging task, as training NeRF requires multiple views of the same scene, coupled with corresponding poses, which are hard to obtain. Our method is based on π-GAN, a generative model for unconditional 3D-aware image synthesis, which maps random latent codes to radiance fields of a class of objects. We jointly optimize (1) the π-GAN objective to utilize its high-fidelity 3D-aware generation and (2) a carefully designed reconstruction objective. The latter includes an encoder coupled with π-GAN generator to form an auto-encoder. Unlike previous few-shot NeRF approaches, our pipeline is unsupervised, capable of being trained with independent images without 3D, multi-view, or pose supervision. Applications of our pipeline include 3d avatar generation, object-centric novel view synthesis with a single input image, and 3d-aware super-resolution, to name a few.
We use pytorch 1.7.0 with CUDA 10.1. To build the environment, run:
pip install -r requirements.txt
For CelebA, download from https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html and extract the img_align_celeba split.
For Carla, download from https://github.com/autonomousvision/graf.
For ShapeNet-SRN, download from https://github.com/sxyu/pixel-nerf and remove the additional layer, so that there are 3 folders chairs_train, chairs_val and chairs_test within srn_chairs. Instances should be directly within these three folders.
Copy img_csv/CelebA_pos.csv to /PATH_TO/img_align_celeba/.
Copy srn_chairs_train.csv, srn_chairs_train_filted.csv, srn_chairs_val.csv, srn_chairs_val_filted.csv, srn_chairs_test.csv and srn_chairs_test_filted.csv under /PATH_TO/srn_chairs.
CELEBA
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train_con.py --curriculum=celeba --output_dir='/PATH_TO_OUTPUT/' --dataset_dir='/PATH_TO/img_align_celeba' --encoder_type='CCS' --recon_lambda=5 --ssim_lambda=1 --vgg_lambda=1 --pos_lambda_gen=15 --lambda_e_latent=1 --lambda_e_pos=1 --cond_lambda=1 --load_encoder=1
CARLA
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train_con.py --curriculum=carla --output_dir='/PATH_TO_OUTPUT/' --dataset_dir='/PATH_TO/carla/*.png' --encoder_type='CCS' --recon_lambda=5 --ssim_lambda=1 --vgg_lambda=1 --pos_lambda_gen=15 --lambda_e_latent=1 --lambda_e_pos=1 --cond_lambda=1 --load_encoder=1
SHAPENET CHAIRS
CUDA_VISIBLE_DEVICES=0,1,2,3 python3 train_con.py --curriculum=srnchairs --output_dir='/PATH_TO_OUTPUT/' --dataset_dir='/PATH_TO/srn_chairs' --encoder_type='CCS' --recon_lambda=5 --ssim_lambda=1 --vgg_lambda=1 --pos_lambda_gen=15 --lambda_e_latent=1 --lambda_e_pos=1 --cond_lambda=1 --load_encoder=1
Note that the training script has been refactored and has not been fully validated yet. It may not reproduce exactly the results from the paper. Please let the authors know if results are not at reasonable levels!
Render novel views of the given image:
python render_video_from_img.py --path=/PATH_TO/checkpoint_train.pth --output_dir=/PATH_TO_WRITE_TO/ --img_path=/PATH_TO_IMAGE/ --curriculum="celeba" or "carla" or "srnchairs"
Render videos and create gifs for the three datasets:
CELEBA
python render_video_from_dataset.py --path PRETRAINED_MODEL_PATH --output_dir OUTPUT_DIRECTORY --curriculum "celeba" --dataset_path "/PATH/TO/img_align_celeba/" --trajectory "front"
CARLA
python render_video_from_dataset.py --path PRETRAINED_MODEL_PATH --output_dir OUTPUT_DIRECTORY --curriculum "carla" --dataset_path "/PATH/TO/carla/*.png" --trajectory "orbit"
SRNCHAIRS
python render_video_from_dataset.py --path PRETRAINED_MODEL_PATH --output_dir OUTPUT_DIRECTORY --curriculum "srnchairs" --dataset_path "/PATH/TO/srn_chairs/" --trajectory "orbit"
Render images and a video interpolating between 2 images.
python linear_interpolation --path=/PATH_TO/checkpoint_train.pth --output_dir=/PATH_TO_WRITE_TO/
Since our model is feed-forward and uses a relatively compact latent codes, it most likely will not perform that well on yourself/very familiar faces---the details are very challenging to be fully captured by a single pass. Therefore, we provide a script performing hybrid optimization: predict a latent code using our model, then perform latent optimization as introduced in pi-GAN. The command to use is:
python --path PRETRAINED_MODEL_PATH --output_dir OUTPUT_DIRECTORY --curriculum ["celeba" or "carla" or "srnchairs"] --img_path /PATH_TO_IMAGE_TO_OPTIMIZE/
Note that compare with vanilla pi-GAN inversion, we need significantly less iterations.
We provide pretrained model checkpoint files for the three datasets. Download from https://www.dropbox.com/s/lcko0wl8rs4k5qq/pretrained_models.zip?dl=0 and unzip to use.
@inproceedings{cai2022pix2nerf,
title={Pix2NeRF: Unsupervised Conditional p-GAN for Single Image to Neural Radiance Fields Translation},
author={Cai, Shengqu and Obukhov, Anton and Dai, Dengxin and Van Gool, Luc},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={3981--3990},
year={2022}
}
The code repo is built upon https://github.com/marcoamonteiro/pi-GAN. We thank the authors for releasing the code and providing support throughout the development of this project.