Will the code for inference be released?

[hermosayhl]

Your work is great! I'm of much interest. However, I'm having trouble how to run the code for a single image or a dir with many pictures. The code written with Chainer is too hard to understand all procedures for me and the demo.py is not useful for me, because no direct image produced is saved. I need the produced images to do more experiments.
I'm looking forward to the inference code.

[satoshi-kosugi]

I am not planning to release the inference code.
If you want to save the result of the demo code, please replace line 249 in chainer_spiral/environments/env.py

                edit_demo(self.original_original_images[i] * 255, clipped_action[i])

with

                editted_image = self.photo_editor(self.original_original_images[i].copy(), clipped_action[i])
                cv2.imwrite(os.path.basename(self.file_names[i]), (editted_image * 255).astype(np.uint8))

Then, please run demo.py.

[hermosayhl]

I am not planning to release the inference code.
If you want to save the result of the demo code, please replace line 249 in chainer_spiral/environments/env.py

                edit_demo(self.original_original_images[i] * 255, clipped_action[i])

with

                editted_image = self.photo_editor(self.original_original_images[i].copy(), clipped_action[i])
                cv2.imwrite(os.path.basename(self.file_names[i]), (editted_image * 255).astype(np.uint8))

Then, please run demo.py.

At first, thanks for your help!

I managed to generate enhanced results and saved them.
However, as mentioned in 《Global and Local Enhancement Networks for Paired and Unpaired Image Enhancement》，the enhanced images is somehow underexposed. "For instance, FRL fails to increase brightness sufficiently" is True. Please look up the literaure at 12/16 of pages.

Listed below are some results:

Compared the satisfying results produced by your code while training:

However, low resolution images produced look pleasing.

Compared to the results of 《DeepLPF: Deep Local Parametric Filters for Image Enhancement》

[satoshi-kosugi]

Reason for underexposed results

Thank you for your information about "Global and Local Enhancement Networks for Paired and Unpaired Image Enhancement".
As mentioned in the paper, our enhanced images are somewhat underexposed. I think this is the limitation of our method, not implementation error.

Low-resolution images generated during training may be different from results by demo.py.
This is because our method generates slightly different images even if the same images are used for training.

Comparison with DeepLPF

DeepLPF is a method for paired photo enhancement, but our method is for unpaired photo enhancement.
Naturally, DeepLPF will give better results.