Generative Smoke Removal – efficient image desmoking using GAN

Supplementary code to the paper O Sidorov, C Wang, FA Cheikh. Generative Smoke Removal (Submitted to ICIP 2019).

Qualitative comparison. (a) Input smoke images and desmoked ones by: (b) DCP, (c) VAR, (d) EVID, (e) proposed method.

Get started

The implementation is based on pix2pix and PAN in PyTorch by DLHacks which is based on original pix2pix PyTorch implementation by Jun-Yan Zhu.

The framework is complemented by differentiable MS-SSIM loss in implementation of jorge-pessoa which borrows heavily from SSIM implementation by Po-Hsun-Su.

Requirements

• python 3.5 +
• pytorch 0.2.0 +

Optionally

• visdom and dominate for visualization pip install visdom dominate

Run the code

The easiest way to run training / testing:

1. Prepare your dataset as following

datasets/
        facades/
                train/
                       ...
                       ...
                test/
                       ...
                       ...

where /train and /test contains image pairs concatenated along a horizontal axis – the aim domain on the left, the initial domain on the right.

2. Run visdom to open training visualization (optional).
3. Run run.sh or runtest.sh correspondingly (you may also just copy the code to the command line).
4. Find the checkpoints and output in /checkpoints and /results.

---

Please kindly cite the paper if you use the code!

---

Benefit of MS-SSIM loss illustrated on other datasets