Self-Supervised Linear Motion Deblurring

This repository contains the code and datasets to reproduce the results from the paper Self-Supervised Linear Motion Deblurring by Peidong Liu, Joel Janai, Marc Pollefeys, Torsten Sattler and Andreas Geiger. Additional supplementary materials can be found from arXiv. The project website is also available at blog.

We present a differentiable reblur model for self-supervised motion deblurring. We are able to train the networks with two consecutive blurry images and do not require any ground truth sharp image for supervision. During inference, our network takes a single blurry image as an input and produce the corresponding sharp estimate, as in the following examples:

You can find detailed usage instructions for training your own models and using pretrained models below.

If you find our code, datatsets or paper useful, please consider citing:

@article{Liu2020RAL,
  author = {Peidong Liu and Joel Janai and Marc Pollefeys and Torsten Sattler and Andreas Geiger},
  title = {Self-Supervised Linear Motion Deblurring},
  journal = {Robotics and Automation Letters},
  year = {2020}
}

Dependencies installation

To train or test the model, you need to install the dependent packages via

pip install -r requirements.txt

The code is tested with PyTorch 0.4.0 and 1.1.0 with CUDA 9.0.

Install correlation package

cd ./correlation_package
python setup.py install

Install reblur_package

cd ./reblur_package
python setup.py install

Demo with our pretrained model

You can now test our code with the provided images in the demo folder. To do this, simply run

bash download_pretrained_models.sh
bash demo.sh

Datasets

If you want to evaluate your/our algorithm with our proposed datasets, you can download them as follows.

• Fastec training data: the training data (~6G) from the synthetic Fastec dataset.

• Fastec test data: the test data (~2G) from the synthetic Fastec dataset.

• Real dataset: the real dataset (~379MB) which contains both the training data and test data.

Evaluation with our pretrained model

If you want to re-produce the same experimental results as what our paper demonstrates, please download the Fastect test data and the real dataset to your local computer.

Then you can run following commands to get the results.

bash download_pretrained_models.sh
# !! Please update the path to test data in 'evaluate_pretrained.sh'
# !! with your own local path, before run following command!!
bash evaluate_pretrained.sh

Training with our datasets

If you want to re-train the network with our proposed datasets, please download the Fastec training data and the real dataset to your local computer.

Then you can run following commands to re-train the networks.

# !! Please update the corresponding paths in 'train.sh' with  #
# !! your own local paths, before run following command!!      #

bash train.py

Training with your own dataset

If you want to train our network with your own dataset, please prepare a paired_image_list.log file in both your training data root folder and test data root folder. The format is defined as follows. Assuming you have two sequences of images, i.e., Sequence_01 and Sequence_02, example file entries can be:

#--------------------------------------------------------------------------------------#
#                          File 'paired_image_list.log'                                #
#--------------------------------------------------------------------------------------#
# im_blurry_path, im_sharp_path, exposure_time and timestamp (in the same unit         #
# as exposure time)                                                                    #
#--------------------------------------------------------------------------------------#
# The image path is the relative path with respect to the 'paired_image_list.log' file #
# Please use '*' to seperate two different sequences.                                  #
#--------------------------------------------------------------------------------------#
* Sequence_01
seq01_im_blur_0.png seq01_im_sharp_0.png 0.02 0.05
seq01_im_blur_1.png seq01_im_sharp_1.png 0.05 0.10
seq01_im_blur_2.png seq01_im_sharp_2.png 0.03 0.15
.
.
.
* Sequence_02
seq02_im_blur_0.png seq02_im_sharp_0.png 0.04 0.05
seq02_im_blur_1.png seq02_im_sharp_1.png 0.03 0.10
seq02_im_blur_2.png seq02_im_sharp_2.png 0.04 0.15
.
.
.

If there is no corresponding sharp images (e.g., a real dataset), you can simply duplicate the path to the blurry image for the sharp image. More details can be found from the provided paired_image_list.log file from one of our datasets.