search

huiyu8794 / LDCNet

Learnable Descriptive Convolutional Network for Face Anti-Spoofing (BMVC'22)
16 stars 4 forks source link

Pretrained LDCNet #1

Open angelabr opened 1 year ago

angelabr commented 1 year ago

Hi,

I see the pretrained model you provide is for visualization, could you provide the weights of your LDCNet model?

Thank you

huiyu8794 commented 1 year ago

Hello @angelabr , Which protocol of the pretrained model do you need? Which pretrained model do you need, FE and CF or the whole LDCNet?

angelabr commented 1 year ago

Thank you for your prompt reply. If possible, I would need the [O,C,I]->M protocol and the whole LDCNet.

huiyu8794 commented 1 year ago

Due to the limitation on the size of the uploaded file, I have placed it on google drive. https://drive.google.com/drive/folders/1aQ2J1hGgJg5tsIKV26xWiIDExcyPzkq-?usp=share_link (valid until Feb 28, 2023) I hope it will be helpful to you.

angelabr commented 1 year ago

Thank you very much. One last thing, is it necessary to perform some preprocessing on the images? I'm testing on a subset of Celeb a Spoof cropped with MTCNN and it seems to predict "1" all the time.

huiyu8794 commented 1 year ago

We use the MTCNN to preprocess the dataset, which should not pose a problem. I think there are two potential issues:

  1. We conduct cross-domain testing using only the O+C+I dataset, which consists of 7,740 images. This may not be sufficient for the larger CelebA-Spoof dataset, which includes 625,537 images. In addition, the CelebA-Spoof dataset contains a wider range of attack types, including print, replay, paper cut and 3D mask. In contrast, the O, C, I and M dataset only includes print and replay.
  2. After obtaining the predictions, we follow previous methods (SSAN, CDC) to use Youden Index Calculation for obtaining the threshold of binary classification.
zero-suger commented 9 months ago

Thank you very much. One last thing, is it necessary to perform some preprocessing on the images? I'm testing on a subset of Celeb a Spoof cropped with MTCNN and it seems to predict "1" all the time.

Thank you very much. One last thing, is it necessary to perform some preprocessing on the images? I'm testing on a subset of Celeb a Spoof cropped with MTCNN and it seems to predict "1" all the time.

First take dataset -> take frames/images -> detect face -> crop face -> make .npy for each cropped face (be careful to BGRtoRGB and while converting normilize each image npy value / 255 (0~1) and save) -> all saved npy files are lists so save all npy files into one .npy file (list) -> THEN just use as it is