search

codeslake / DMENet

[CVPR 2019] Official TensorFlow Implementation for "Deep Defocus Map Estimation using Domain Adaptation"
GNU Affero General Public License v3.0
122 stars 24 forks source link

The max_coc of datasets is not 28. #9

Closed chanwental closed 4 years ago

chanwental commented 4 years ago

Hi， I read a blur_map in your dataset as follows： image = (np.float32(cv2.imread(file_name, cv2.IMREAD_UNCHANGED))/10.)[:, :, 1]. Then the maximum of image is 15. #15=(61-1)/4 It implys max_coc=61, which is not consistent with 28 in your paper. What is the problem?

codeslake commented 4 years ago

Hi, Wentao

Regarding with your first answer, I've looked at the initial code and I found the accurate answer.

In the initial SYNDOF, the sigma was computed as:

sigma = (max_coc - 1) / 2. with initial max_coc = 15.

But after generation of SYNDOF, we found out that the largest blur size was acutally around 28~30. So, we phisically analyze the blur kernel, and found out;

sigma = (max_coc - 1) / 4

was acuatlly making a blur kernel to have the diameter similar to max_coc.

However, instead of regenerating SYNDOF and retrain DMENet (as we ran out of time for CVPR2019), we wrote "sigma = (max_coc) / 4" in the paper, and defined our max_coc = 28. Then, we reflected the equation and refined the SYNDOF generation code for the final release, in which sigma is defined as sigma = (max_coc - 1) / 4.

Please not that that doesn't make any difference in SYNDOF generation process. With that said, if random seeds are the same, initial_SYNDOF(max_coc=15) generates the same images as the refined_SYNDOF(max_coc = 29).

I hope these answer solves your question.

Best, Junyong