Skip to content

Learning Dense and Continuous Optical Flow from an Event Camera (TIP 2022)

License

Notifications You must be signed in to change notification settings

danqu130/DCEIFlow

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

16 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

DCEIFlow

This repository contains the source code for our paper:

Learning Dense and Continuous Optical Flow from an Event Camera

IEEE Transactions on Image Processing (TIP 2022)

Zhexiong Wan, Yuchao Dai, Yuxin Mao

Project Page, arXiv, IEEE, Supp.

If you have any questions, please do not hesitate to raise the issue or contact my email [email protected]

Requirements

The code has been tested with PyTorch 1.12.1 and Cuda 11.7.

Pretrained Weights

Pretrained weights can be downloaded from Google Drive.

Please put them into the checkpoint folder.

Evaluation

To evaluate our model, you need first download the HDF5 files version of MVSEC datasets.

data/MVSEC_HDF5
├── indoor_flying
│   ├── indoor_flying1_data.hdf5
│   ├── indoor_flying1_gt.hdf5
│   ├── indoor_flying2_data.hdf5
│   ├── indoor_flying2_gt.hdf5
│   ├── indoor_flying3_data.hdf5
│   ├── indoor_flying3_gt.hdf5
├── outdoor_day
│   ├── outdoor_day1_data.hdf5
│   ├── outdoor_day1_gt.hdf5
│   ├── outdoor_day2_data.hdf5
│   ├── outdoor_day2_gt.hdf5

After the environment is configured and the pretrained weights is downloaded, run the following command to get the consistent results as reported in the paper.

python main.py --task test --stage mvsec --checkpoint ./checkpoint/DCEIFlow_paper.pth

The results reported in our paper are simulated with esim_py. We also provide another pretrained model using a new simulator, DVS-Voltmeter, thanks for this open source project. It improves the performance on Chairs and Sintel, while MVSEC is basically unchanged. We recommend using the updated model because of better generalization performance.

python main.py --task test --stage mvsec --checkpoint ./checkpoint/DCEIFlow.pth

Training

To train our model, you need to download the FlyingChairs2 datasets and simulate the events corresponding to every two frames.

data/FlyingChairs2
├── train
├── val
├── events_train
├── events_val

The simulated events are stored in HDF5 format with name ******-event.hdf5. Please refer to the read_event_h5() function in utils/file_io.py.

After completing the simulation, you can run the following command to start the training.

CUDA_VISIBLE_DEVICES="0," python main.py --task train --stage chairs2 --isbi --model DCEIFlow --batch_size 4 --epoch 200 --lr 0.0004 --weight_decay 0.0001 --loss_gamma=0.80 --name DCEIFlow

Citation

If our work or code helps you, please cite our paper.

If our code is very useful for your new research, I hope you can also open source your code including training.

@article{wan2022DCEIFlow,
    author={Wan, Zhexiong and Dai, Yuchao and Mao, Yuxin},
    title={Learning Dense and Continuous Optical Flow From an Event Camera}, 
    journal={IEEE Transactions on Image Processing}, 
    year={2022},
    volume={31},
    pages={7237-7251},
    doi={10.1109/TIP.2022.3220938}
}

Acknowledgments

This research was sponsored by Zhejiang Lab.

Thanks the assiciate editor and the reviewers for their comments, which is very helpful to improve our paper.

Thanks for the following helpful open source projects:

RAFT, event_utils, EV-FlowNet, mvsec_eval, esim_py, DVS-Voltmeter, Spike-FlowNet.

About

Learning Dense and Continuous Optical Flow from an Event Camera (TIP 2022)

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages