NTIRE 2025 the First Challenge on Event-Based Deblurring @ CVPR 2025

May 7, 2025 · View on GitHub

This is a simple introduction to the dataset and basic codes.

Downloading Testset

Blurry images and Raw events

Testset input:

CodaLab_downloading_link; Kaggle_downloading_link

(Optional) Voxel input for testset: CodaLab_downloading_link; Kaggle_downloading_link

Downloading Trainset and Valset

Voxel grids

If you find the data loading too slow, we also provide processed voxel grid (bin=6) for convenience, which is optional.

Processed voxel grid of events

The structure of the HighREV dataset with raw events is as following:

    --HighREV
    |----train
    |    |----blur
    |    |    |----SEQNAME_%5d.png
    |    |    |----...
    |    |----event
    |    |    |----SEQNAME_%5d_%2d.npz
    |    |    |----...
    |    |----sharp
    |    |    |----SEQNAME_%5d.png
    |    |    |----...
    |----val
    ...

For each blurry image, there are several NPZ files containing events. By concatenating them, the events for the entire exposure time can be obtained. More details please refer to ./basicsr/data/npz_image_dataset.py

Converting events to voxel

By using ./basicsr/utils/npz2voxel.py you can convert raw events to voxel grids by you own offline.

Dataset codes:

./basicsr/data/npz_image_dataset.py for processing raw events. ./basicsr/data/voxelnpz_image_dataset.py for processing voxel grids.

How to start training?

We provide a simple codebase here:

git clone https://github.com/AHupuJR/NTIRE2025_EventDeblur_challenge
cd NTIRE2025_EventDeblur_challenge
pip install -r requirements.txt
python setup.py develop --no_cuda_ext

Single GPU training:

python ./basicsr/train.py -opt options/train/HighREV/EFNet_HighREV_Deblur.yml

Multi-GPU training:

python -m torch.distributed.launch --nproc_per_node=4 --master_port=4321 basicsr/train.py -opt options/train/HighREV/EFNet_HighREV_Deblur.yml --launcher pytorch

How to start testing?

Example:

python3 basicsr/test.py -opt options/test/HighREV/EFNet_HighREV_Deblur.yml

Calculating flops: set print_flops to true and set your input shapes in flops_input_shape in the test yml file. Example:

print_flops: true 
flops_input_shape: 
  - [3, 256, 256] # image shape
  - [6, 256, 256] # event shape

Be sure to modify the path configurations in yml file.

Develop your own model

We recommand to used basicsr (already used here, tutorial) for developing. It is easy to change the models in ./basicsr/models.

Use your own codes

The dataset-related code is in ./basicsr/data/npz_image_dataset.py. If you are not using the code from this repository, please integrate it into your own code for convenience.

Others

The aim is to obtain a network design / solution that fusing events and images and produce high quality results with the best performance (i.e., PSNR). We suggest using HighREV dataset for training.

For the sake of fairness, please do not train your model with the (HighREV) validation GT images.

The top ranked participants will be awarded and invited to follow the CVPR submission guide for workshops to describe their solution and to submit to the associated NTIRE workshop at CVPR 2025.

Solution codebases from the attending teams

All team solutions codebases are provided in a single Google Drive archive. Download it here: Google Drive link. After extracting the archive, you'll find ./codebases/<team_name> folder structure for each team.

One folder per team
Example: ./codebases/IVISLAB/
Self-contained instructions
Every team folder includes its own README.md that lists:
- prerequisites
- installation or environment setup
- commands to run the code and reproduce results

Competition results download

Each team has also submitted pre‑computed results (e.g., model outputs, prediction files) so you can evaluate or compare without re‑running experiments:

Citations

@inproceedings{sun2023event,
  title={Event-based frame interpolation with ad-hoc deblurring},
  author={Sun, Lei and Sakaridis, Christos and Liang, Jingyun and Sun, Peng and Cao, Jiezhang and Zhang, Kai and Jiang, Qi and Wang, Kaiwei and Van Gool, Luc},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={18043--18052},
  year={2023}
}

@inproceedings{sun2022event,
  title={Event-Based Fusion for Motion Deblurring with Cross-modal Attention},
  author={Sun, Lei and Sakaridis, Christos and Liang, Jingyun and Jiang, Qi and Yang, Kailun and Sun, Peng and Ye, Yaozu and Wang, Kaiwei and Gool, Luc Van},
  booktitle={European Conference on Computer Vision},
  pages={412--428},
  year={2022},
  organization={Springer}
}