search

CZHQuality / Sal-CFS-GAN

TIP2019-GazeGAN Saliency Model
32 stars 5 forks source link

Data Augmentation and image padding parts #2

Open Sunshine-Ye opened 5 years ago

Sunshine-Ye commented 5 years ago

Dear author: I am very interesting in the data augmentation and image padding (as the results of the distorted dataset shown ) parts of this article. Although I can already run through your code, different input can affect the results. Can you share your code of the data augmentation and image padding parts? sincerely!

CZHQuality commented 5 years ago

Hi Sunshine Ye,

No problem.

Actually, we recommend to use 7 types of label-preserving transformations as data augmentation strategy for saliency detection task, which are: original reference image, mirrored version, inverted version, slight contrast change, slight shearing, slight JPEG compression, and slight Noise. I will neaten the data augmentation code (in Matlab) and share with you later. Could you send the failure cases ("different input can affect the results") to my e-mail "chezhaohuihy@gmail.com"? Thank you very much !!

Sincerely, Zhaohui

CZHQuality commented 5 years ago

Dear author: I am very interesting in the data augmentation and image padding (as the results of the distorted dataset shown ) parts of this article. Although I can already run through your code, different input can affect the results. Can you share your code of the data augmentation and image padding parts? sincerely!

Hi Sunshine Ye,

I upload the preprocessing code for unifying the aspect ratio of MIT1003 dataset by padding, please refer to : https://github.com/CZHQuality/Sal-CFS-GAN/blob/master/GazeGAN_using_CSC/img_padding_for_mit1003.py

Thanks a lot. Hope to get your precious feedbacks !!

Sincerely, Zhaohui

Sunshine-Ye commented 5 years ago

Dear author: I am very interesting in the data augmentation and image padding (as the results of the distorted dataset shown ) parts of this article. Although I can already run through your code, different input can affect the results. Can you share your code of the data augmentation and image padding parts? sincerely!

Hi Sunshine Ye,

I upload the preprocessing code for unifying the aspect ratio of MIT1003 dataset by padding, please refer to : https://github.com/CZHQuality/Sal-CFS-GAN/blob/master/GazeGAN_using_CSC/img_padding_for_mit1003.py

Thanks a lot. Hope to get your precious feedbacks !!

Sincerely, Zhaohui

Sunshine-Ye commented 5 years ago

Hi Zhaohui,

Thanks for your sharing! I have explained the failure cases caused by image resize by emailing to chezhaohuihy@gmail.com. I think this problem will be well handled by the uploaded preprocessing code.

Sincerely, Sunshine Ye.