Defocus Map Estimation and Deblurring from a Single Dual-Pixel Image

This repository is an implementation of the method described in the following paper:

Shumian Xin, Neal Wadhwa, Tianfan Xue, Jonathan T. Barron, Pratul P. Srinivasan, Jiawen Chen, Ioannis Gkioulekas, and Rahul Garg. "Defocus Map Estimation and Deblurring from a Single Dual-Pixel Image", ICCV 2021.

If you use our code or dataset, please cite our paper:

@article{Xin_2021_ICCV_dual_pixel,
    author    = {Xin, Shumian and Wadhwa, Neal and Xue, Tianfan and Barron, Jonathan T. and Srinivasan, Pratul P. and Chen, Jiawen and Gkioulekas, Ioannis and Garg, Rahul},
    title     = {Defocus Map Estimation and Deblurring From a Single Dual-Pixel Image},
    journal   = {IEEE/CVF International Conference on Computer Vision (ICCV)},
    year      = {2021}
}

Dataset

We captured a new dataset of 17 indoor and outdoor scenes using a Google Pixel 4 smartphone camera. Data can be found in ./DP_data_pixel_4. Google Pixel 4 camera provides dual-pixel (DP) images in the green channel. These DP images are 14-bit, with a black level of 1024. Please refer to this GitHub Repo for more details about Google Pixel's DP data.

We also provide calibrated blur kernels and vignetting patterns of our device in ./DP_data_pixel_4/calibration.

Code

Code implementation is in ./code. It is written in Python, with autograd package Jax. Note: When installing Jax, make sure to install with GPU support.

To reproduce results in the paper (in the paper, results are postprocessed for better visualization), run:

cd ./code; python ./run.py

Each optimization runs for 10,000 iterations with an Adam optimizer, and takes about 2 hours on an Nvidia Titan RTX GPU.

Code has been tested with:

• Python 3.7.8
• Jax 0.2.19
• OpenCV 4.4.0

Note: To run the code on your own Google Pixel 4 data, please adjust the preprocessing step in ./code/util.py/load_data_and_calibration if needed, such that the input dual pixel images are normalized to the range of [0, 1].