EventHDR
EventHDR copied to clipboard
This is the implementation and dataset for Learning To Reconstruct High Speed and High Dynamic Range Videos From Events, CVPR 2021, by Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani and Ying Fu
EventHDR
This is the implementation and dataset for Learning To Reconstruct High Speed and High Dynamic Range Videos From Events, CVPR 2021, by Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani and Ying Fu.
News
- 2023/09/19: Both the training and testing datasets are avalable at OneDrive
- 2023/09/27: Release the evaluation code and pre-trained model
Introduction
In this work, we present a convolutional recurrent neural network which takes a sequence of neighboring event frames to reconstruct high speed HDR videos. To facilitate the process of network learning, we design a novel optical system and collect a real-world dataset with paired high speed HDR videos and event streams
Highlights
- We propose a convolutional recurrent neural network for the reconstruction of high speed HDR videos from events. Our architecture carefully considers the alignment and temporal correlation for events
![](https://github.com/jackzou233/EventHDR/raw/main/figs/overview.png)
- To bridge the gap between simulated and real HDR videos, we design an elaborate system to synchronously capture paired high speed HDR video and the corresponding event stream
![](https://github.com/jackzou233/EventHDR/raw/main/figs/system.png)
- We collect a real-world dataset with paired high speed HDR videos (high-bit) and event streams. Each frame of our HDR videos are merged from two precisely synchronized LDR frames
![](https://github.com/jackzou233/EventHDR/raw/main/figs/dataset.png)
Dataset
- Download our EventHDR dataset at OneDrive
- Both the training and testing data are stored in
hdf5
format, you can useh5py
tool to read the data. To create a pytorch dataset, please refer to DynamicH5Dataset - We build our dataset based on the data structure of events_contrast_maximization, and the structure tree is shown below
EventHDR Dataset Structure Tree
+-- '/'
| +-- attribute "duration"
| +-- attribute "num_events"
| +-- attribute "num_flow"
| +-- attribute "num_neg"
| +-- attribute "num_pos"
| +-- attribute "sensor_resolution"
| +-- attribute "t0"
| +-- group "events"
| | +-- dataset "ps"
| | +-- dataset "ts"
| | +-- dataset "xs"
| | +-- dataset "ys"
| |
| +-- group "images"
| | +-- attribute "num_images"
| | +-- dataset "image000000000"
| | | +-- attribute "event_idx"
| | | +-- attribute "size"
| | | +-- attribute "timestamp"
| | | +-- attribute "type"
| | |
| ...
| +-- group "images"
| | +-- attribute "num_images"
| | +-- dataset "image000002827"
| | | +-- attribute "event_idx"
| | | +-- attribute "size"
| | | +-- attribute "timestamp"
| | | +-- attribute "type"
| | |
| |
|
Usage
- Create
conda
environment and download our repository
conda create -n eventhdr python=3.6
conda activate eventhdr
git clone https://github.com/jackzou233/EventHDR
cd EventHDR
pip install -r requirements.txt
- Compile dependencies for deformable convolution with
BASICSR_EXT=True python setup.py develop
- Download our
hdf5
foramt evaluation data from OneDrive, and put it in folder./eval_data
- Download the pre-trained model at pretrained, and move it to
./pretrained
- Run the following script to reconstruct HDR videos from event streams!
bash test.sh
- To make a comparing video of the reconstruction, please run
python mk_video.py
Then, you will obtain results like
Results
The quantitative results are shown below
Method | PSNR | SSIM | LPIPS | TC |
---|---|---|---|---|
HF | 10.99 | 0.2708 | 0.4434 | 1.2051 |
E2VID | 12.78 | 0.5753 | 0.3541 | 1.0305 |
EF | 13.23 | 0.5914 | 0.3030 | 0.9729 |
Ours | 15.31 | 0.7084 | 0.2424 | 0.5198 |
Citation
If you find this work useful for your research, please cite:
@inproceedings{zou2021learning,
title={Learning To Reconstruct High Speed and High Dynamic Range Videos From Events},
author={Zou, Yunhao and Zheng, Yinqiang and Takatani, Tsuyoshi and Fu, Ying},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={2024--2033},
year={2021}
}
contact
If you have any problems, please feel free to contact me at [email protected]
Acknowlegment
The code borrows from event_cnn_minimal, EDVR and E2VID, please also cite their great work