search

liamheng / Restoration-of-Cataract-Images-via-Domain-Adaptation

A restoration of cataract images using damain adaptation
6 stars 2 forks source link

pre-trained model and instructions #1

Open AugustLee93 opened 2 years ago

AugustLee93 commented 2 years ago

Hi, there, I was very impressed by your work ! Can you provide a pretrain model and give some instrucions of your work ? I would like to test how good the restoration and prediction will be. Thank you very much!

HeverLaw commented 2 years ago

Thank you for your interest in our study. We have uploaded a trained model as a demo for the algorithm. README.md has been updated for instructions. Please connect us if there is any further question.

AugustLee93 commented 2 years ago

Thank you for your interest in our study. We have uploaded a trained model as a demo for the algorithm. README.md has been updated for instructions. Please connect us if there is any further question.

Thank you for your reply! I tried your pre-trained model, but I'm not sure what your format of the dataset as input is. Isn't it supposed to be a single pic as input, or i need to format the inputs data like one pic but concatenate a lable, like {cataract pic, clear cataract pic(lable)}?

HeverLaw commented 2 years ago

From your understanding, I think the former one it's right for the target dataset and the latter one it's right for the source dataset. A dataset contains two parts, including the source dataset and the target dataset. The source dataset contains the image pairs where a pair is constructed like {cataract image, clear fundus image(label)}, and the target dataset contains real cataract images. See the file structure in https://github.com/liamheng/Restoration-of-Cataract-Images-via-Domain-Adaptation/tree/master/images/cataract_dataset to get a more intuitive understanding.

AugustLee93 commented 2 years ago

From your understanding, I think the former one it's right for the target dataset and the latter one it's right for the source dataset. A dataset contains two parts, including the source dataset and the target dataset. The source dataset contains the image pairs where a pair is constructed like {cataract image, clear fundus image(label)}, and the target dataset contains real cataract images. See the file structure in https://github.com/liamheng/Restoration-of-Cataract-Images-via-Domain-Adaptation/tree/master/images/cataract_dataset to get a more intuitive understanding.

Thanks for your useful reply! But I still have a few question. Firstly, if in train stage, dataset contains source file like {cataract image, clear fundus image(label)}, I get it. But within a pre-trained model, if source file is empty ( but target file has a image to test), then I got None result output. I was wondering how it works. By the way, I tried if source and target file contain images, then the test image in target file has a good result, good work!