Mitigating Motion Blur in Neural Radiance Fields with Events and Frames

This is the official Pytorch implementation of the CVPR 2024 paper Mitigating Motion Blur in Neural Radiance Fields with Events and Frames.

Method Overview

We propose a novel approach for recovering a sharp NeRF in the presence of motion blur, incorporating both model-based priors and learning-based modules, and exploiting the complementary information provided by events and frames. We exploit the event double integral as an additional model-based prior and model the event-pixel response using an end-to-end learnable response function, allowing our method to adapt to non-idealities in the real event-camera sensor.

Project Overview

1. Install conda environment

We highly recommend using Mambaforge to reduce the installation time. In such case, replace conda with mamba in the following command:

conda env create -f environment.yml

Dependencies (click to expand)

- python=3.8
- pytorch-cuda=11.6
- pytorch==1.13.1
- torchvision==0.14.1
- configargparse=1.5.3
- einops=0.7.0
- imageio=2.22.0
- kornia=0.6.9
- numba=0.56.4
- numpy=1.23.1
- pandas=2.0.3
- plotly=5.18.0
- scikit-image=0.19.2
- scipy=1.9.1
- tqdm=4.65.0
- h5py=3.8.0
- pillow=9.2.0
- pyyaml=6.0
- open3d=0.15.1
- imageio-ffmpeg>=0.4.9
- matplotlib>=3.7.3
- opencv-python==4.6.0.66
- tensorboardx>=2.5.1

2. Download datasets

Download the Ev-DeblurNeRF-CDAVIS and Ev-DeblurNeRF-Blender datasets and place them in the datasets/ folder.

The dataset is provided in LLFF format. Please refer to the original LLFF codebase for a more detailed description of camera poses format.

Folder structure (click to expand)

datasets/
├─ evdeblurnerf_cdavis/
│  ├─ blurbatteries/
│  │  ├─ images_1/            ◁─┬ Interleaved blur and sharp images, see llffhold and 
│  │  │  ├─ 00.png              └ llffhold_end args. Images are already undistorted.
│  │  │  ├─ 01.png
│  │  │  ├─ ...
│  │  │  └─ timestamps.npz    ◁─┬ Image timestamps in npz format with keys (timestamps, 
│  │  │                         │ timestamps_start, timestamps_end, start, end). Timestamps
│  │  │                         │ are either in us or ns, see events_tms_files_unit arg. Sharp
│  │  │                         └ images are indicated with timestamps_start = timestamps_end.
│  │  ├─ poses_bounds.npy     ◁── Camera poses in LLFF format.
│  │  ├─ events.h5            ◁─┬ Events saved in HDF5 file format with (p, t, x, y) keys. 
│  │  │                         └ Timestamps are either in us or ns, see events_tms_unit arg.
│  │  ├─ ev_map.npz           ◁─┬ Mapping between original image and undistorted space (used 
│  │  │                         └ with color events only, not available in Blender dataset).
│  │  ├─ all_poses_bounds.npy ◁── All available poses for interpolation in LLFF format.
│  │  └─ all_timestamps.npy   ◁── All available timestamps for interpolation.
│  └─ ... 
└─ evdeblurnerf_blender/
   └─ ...

3. Setting parameters

We provide ConfigArgparse config files for the main experiments in the paper in the configs/ folder. You might want to modify the datadir parameter in the config file if you decided to store the datasets in a custom folder, and the basedir and tbdir parameters to change where the model will be saved and the tensorboard logs will be stored, respectively.

Note: Tensorboard logs are disabled by default in place of W&B logging. You can enable tensorboard logs by setting --use_tensorboard, and disable W&B logging by using the --no_wandb parameter.

4. Execute

python run_nerf.py --config configs/evdeblurnerf_blender/tx_blurfactory_evdeblurnerf_ediprior_evcrf.txt

This command will train a model using the parameters specified in the config file. Video and test set renderings will be produced at the end of training, or every i_testset and i_video iterations, respectively.

Citation

If you find this useful, please consider citing our paper:

@InProceedings{Cannici_2024_CVPR,
  author  = {Marco Cannici and Davide Scaramuzza},
  title   = {Mitigating Motion Blur in Neural Radiance Fields with Events and Frames},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year    = {2024},
}

Acknowledgments

This source code is derived from multiple sources, in particular: PDRF, DP-NeRF, and TensoRF. We thank the authors for releasing their code.