wentianzhang-ml / frt-pad Goto Github PK

how to use those dataset ？？

这个数据集怎么用啊，CASIA_faceAntisp/test_release ，CASIA_faceAntisp/train_release ，里面全是 .avi. 没有 jpg

In line 43 of data_find.py: path_list = glob.glob(dataset_path + '**/*.jpg', recursive=True) .

However, there is no .jpg in the MSU-MFSD dataset. Instead there are.faceand .mov files which causes program to crash. Should the jpg images be cropped face images with indices in the .face files or the whole frame from the video files?

FRT-PAD的作者您好，我是一名机器学习方向的博士生，我最近在做有关liveness detection的测试，看到您的论文水平非常高所以想自己来运行一下。而最近在用您的代码进行测试的时候我遇到了如下问题：在进行训练的时候，训练模型无法读取数据库，具体一点应该是get_dataset函数里的DataLoader无法正确读取输入的dataset。所以我想问下您在之前训练的时候是否有相似的问题，或者说您在data_find步骤，图片用的是怎样的格式，图片的名称是否对后来加载数据库有影响？
非常感谢您的理解和帮助，很期待跟您能有后续的交流与合作。我的联系方式是邮箱：[email protected]，期望您能抽出一点时间！提前感谢！

I wrote an inference script to try pretrained weights, but the model predicts everything around 0.5. Then, I recognized that there is no pretrained model loaded for GAT/Cross_Model_Adapter network. By any chance, is it possible for you to share GAT pretrained network?

Hi,

I'm testing your pretrained model on CASIA-FASD and I always get similar values to spoof and life faces.
I'm using as input 256*256 images of faces, Face_recognition and Step_by_Step_Graph.

Is there anything I'm missing?

Thank you

Hi, wonderful work!!!

I have a question about the method when testing.
It seems that the GraphAttentionLayer in GAT is related to the batch_size. Say I trained the model with batch_size 32, when testing, I have to have the same batch_size 32 to load the model weights? Otherwise the feature size is different with training when testing with batch_size 1.

Hope to hear from you soon. Thank you so much.