search

huguesva / Face-Anti-Spoofing-Neural-Network

Implementation of "Learning Deep Models for Face Anti-Spoofing: Binary or Auxiliary Supervision" in PyTorch.
88 stars 22 forks source link

Depth Image Obtained From Spoof Images #4

Open leviethung2103 opened 4 years ago

leviethung2103 commented 4 years ago

Hello,

I would like to ask you about the depth images.

At the first glance, I've tested the depth images which is generated from this repo. https://github.com/cleardusk/3DDFA.

It's obvious that the model performs can generate the depth map for every case including genuine/spoof images.

Do we need do apply further post-processing to obtain the 3D-shape images for real and non 3D-shape images for spoof?

Thank you in advanced.

PhanMinhKhue commented 4 years ago

Hi Hung, I think groundtruthdepth image for spoof images is set zero (follow paper). I wonder about creating rPPG groundtruth signal. How do we create it?? I don't see this task in code. Can you explain me? Thank you

longnguyen2 commented 4 years ago

Hi Hung, I think groundtruthdepth image for spoof images is set zero (follow paper). I wonder about creating rPPG groundtruth signal. How do we create it?? I don't see this task in code. Can you explain me? Thank you

I'm struggling with these depth maps and rPPG signals for training input too. Just wondering that do we also need to generate dmap for every images while running this model like the demo here?

leviethung2103 commented 4 years ago

@Phan

Hi Hung, I think groundtruthdepth image for spoof images is set zero (follow paper). I wonder about creating rPPG groundtruth signal. How do we create it?? I don't see this task in code. Can you explain me? Thank you

Thank you for replying me.

Because I am not the author of the repo, please the author in order to get the correct answer. In my understanding, the author provide the code for rPPG in data_process2.py, the keyword is Anchors.

leviethung2103 commented 4 years ago

Hi Hung, I think groundtruthdepth image for spoof images is set zero (follow paper). I wonder about creating rPPG groundtruth signal. How do we create it?? I don't see this task in code. Can you explain me? Thank you

I'm struggling with these depth maps and rPPG signals for training input too. Just wondering that do we also need to generate dmap for every images while running this model like the demo here?

In inference stage, as we can see from model architecture, the depth map and feature map are fed into non-rigid registration layer (like intermediate layer) , and then forward into RNN part.

longnguyen2 commented 4 years ago

In inference stage, as we can see from model architecture, the depth map and feature map are fed into non-rigid registration layer (like intermediate layer) , and then forward into RNN part.

After reading the code carefully, i can understand how the model works now. Thing is, the training part did not work out as expected :( idk why the loss value keep growing up.

Did you get into that problem? Is this because of my customized dataset or something?

PhanMinhKhue commented 4 years ago

In inference stage, as we can see from model architecture, the depth map and feature map are fed into non-rigid registration layer (like intermediate layer) , and then forward into RNN part.

After reading the code carefully, i can understand how the model works now. Thing is, the training part did not work out as expected :( idk why the loss value keep growing up.

Did you get into that problem? Is this because of my customized dataset or something?

Hi Long, Did you resolve this problem? If you did, can you tell me the reason that caused the problem? Thank you

longnguyen2 commented 4 years ago

Hi Long, Did you resolve this problem? If you did, can you tell me the reason that caused the problem? Thank you

Not yet, I gave up on this :)

PhanMinhKhue commented 4 years ago

Hi Long, Did you resolve this problem? If you did, can you tell me the reason that caused the problem? Thank you

Not yet, I gave up on this :)

Oh :( I have a project about face anti spoofing. I am interesting in using face deep-map like this repo for doing this project. Can you recommend orther repo? or other solution for this project? I am confused. Thank you

longnguyen2 commented 4 years ago

Oh :( I have a project about face anti spoofing. I am interesting in using face deep-map like this repo for doing this project. Can you recommend orther repo? or other solution for this project? I am confused. Thank you

Check out this post: https://zhuanlan.zhihu.com/p/114313640 The writer mentioned some face antispoofing attack paper of CVPR 2020, source code included. I'm not sure if they provide the dataset or not, but worth trying.

longnguyen2 commented 4 years ago

@PhanMinhKhue You can download some public dataset that I have already collected here. FRAUD - Face Replay Attack UQ Datasets: http://staff.itee.uq.edu.au/lovell/FRAUD/ LCC FASD: https://drive.google.com/file/d/1NeyTFAwdJSjxA9ZtdviwdUjdptEVjM_i/view NUAA: https://drive.google.com/file/d/1fe80Vo366h4uKylFwsSN3apvLXZZm02L/view CASIA-SURF: https://drive.google.com/open?id=1BIIGhL02FlNNRYPGDCBTMgfC5PuLCf9U CASIA-FASD: https://drive.google.com/open?id=1n0zRgxzllNOzlrTrUSPs7BiWiuv4TEd5

PhanMinhKhue commented 4 years ago

In inference stage, as we can see from model architecture, the depth map and feature map are fed into non-rigid registration layer (like intermediate layer) , and then forward into RNN part.

After reading the code carefully, i can understand how the model works now. Thing is, the training part did not work out as expected :( idk why the loss value keep growing up.

Did you get into that problem? Is this because of my customized dataset or something?

Hi longnguyen2, Could you tell me How to create rPPG signal for sequence of consecutive frames as label for training. If you could, I hope you show me code about this function. I still want to explore this repo. Thanks you

nghiabka commented 3 years ago

@PhanMinhKhue You can download some public dataset that I have already collected here. FRAUD - Face Replay Attack UQ Datasets: http://staff.itee.uq.edu.au/lovell/FRAUD/ LCC FASD: https://drive.google.com/file/d/1NeyTFAwdJSjxA9ZtdviwdUjdptEVjM_i/view NUAA: https://drive.google.com/file/d/1fe80Vo366h4uKylFwsSN3apvLXZZm02L/view CASIA-SURF: https://drive.google.com/open?id=1BIIGhL02FlNNRYPGDCBTMgfC5PuLCf9U CASIA-FASD: https://drive.google.com/open?id=1n0zRgxzllNOzlrTrUSPs7BiWiuv4TEd5

thanks for sharing datasets , I downloaded casia-surf dataset , but validation set is not labeled, Can you help me ?

haiderasad commented 2 years ago

@nghiabka casia fasd link not updating kindly reshare