search

JizhiziLi / P3M

[ACM MM 2021] Privacy-Preserving Portrait Matting
MIT License
289 stars 33 forks source link

there are no 'train/fg/' & 'train/bg/' in the release dataset #3

Open yjymickey opened 2 years ago

yjymickey commented 2 years ago

Thanks for your efforts!But there are no 'train/fg/' & 'train/bg/' in the release dataset! Please help! If there are not these files,we can not train for your model!

yjymickey commented 2 years ago

I read from config.py But i can not find fg and fg floders in the dataset P3M-10k.zip DATASET_PATHS_DICT={ 'P3M10K':{ 'TRAIN':{ 'ROOT_PATH':P3M_DATASET_ROOT_PATH+'train/', 'ORIGINAL_PATH':P3M_DATASET_ROOT_PATH+'train/blurred_image/', 'MASK_PATH':P3M_DATASET_ROOT_PATH+'train/mask/', 'FG_PATH':P3M_DATASET_ROOT_PATH+'train/fg/', 'BG_PATH':P3M_DATASET_ROOT_PATH+'train/bg/',

JizhiziLi commented 2 years ago

Hi there,

We do not provide the foregrounds and backgrounds directly, you need to generate them following the closed form method as in the paper Levin, Anat, Dani Lischinski, and Yair Weiss. "A closed-form solution to natural image matting." IEEE transactions on pattern analysis and machine intelligence, 2007. Please refer to this code-base page: Prepare Datasets (2) for the implementation details. Thanks.

yjymickey commented 2 years ago

thank for you reply! But the alpha inference reuslt depends on I=αF+(1−α)B; the closed-form result does not real. So if we used the wrong foregroud alpha png,we could not train real network!

huangyangke commented 2 years ago

only composition loss need fg and bg, you can note it.

JizhiziLi commented 2 years ago

Hi @yjymickey,

Since matting is a highly ill-posed problem, the accurent foreground cannot be calculated. Foregrounds calculated following the closed-form paper is an approximate solution to this problem and have been proved performing better than the ones generated by alpha blending solution, i.e., F = I*α, B = I*(1-α) in many previous matting papers.

Besides, like what @924726976 has pointed out, fg and bg are only used in composition loss in training stage. The advantage of using composition loss while training can be refered to the paper Ning Xu, Brian Price, Scott Cohen, Thomas Huang. "Deep Image Matting" CVPR 2017. Thanks.