This is the official implementation of the paper PDRF: Progressively Deblurring Radiance Field for Fast and Robust Scene Reconstruction from Blurry Images
Progressively Deblurring Radiance Field (PDRF) is a novel approach to efficiently reconstruct high quality radiance fields from blurry images. Compared to previous methods like NeRF and DeblurNeRF, PDRF is both much faster and more performant by utilizing radiance field features to model blur.
pip install -r requirements.txt
We evaluate our results based on the dataset captured by DeblurNeRF. You can download the data in here.
This codebase supports several configurations for radiance field modeling. Two parameters worth noting are args.mode={c2f,nerf} and args.kernel_type={none,DSK,PBE}. Specifically, args.mode=c2f leads to our coarse-to-fine rendering architectures (CRR+FVR), which leverages explicit representation (implemented based on TensoRF) and an improved importance sampling scheme; args.mode=PBE is our two-stage blur-modeling design. DeblurNeRF can be implemented with args.mode=nerf,args.kernel_type=DSK; you can also try any other combinations.
python3 run_nerf.py --config configs/defocustanabata/tx_defocustanabata_full.txt
To render a video on the learned scene:
python3 run_nerf.py --config configs/defocustanabata/tx_defocustanabata_full.txt --render_only
To render images on the learned scene:
python3 run_nerf.py --config configs/defocustanabata/tx_defocustanabata_full.txt --render_only --render_test
If you find this useful, please consider citing our paper:
@misc{peng2023pdrf,
title={PDRF: Progressively Deblurring Radiance Field for Fast and Robust Scene
Reconstruction from Blurry Images},
author={Cheng, Peng and Rama, Chellappa},
year={2023},
booktitle = {The 37th AAAI Conference on Artificial Intelligence}
}
This source code is derived from multiple sources, in particular: nerf-pytorch, Deblurnerf, TensoRF, and HashNeRF-pytorch. We thank the previous authors for their awesome and consistent implementations.