rshaojimmy / rfmetafas Goto Github PK

View Code? Open in Web Editor NEW

140.0 12.0 29.0 814 KB

[AAAI 2020] Pytorch codes for Regularized Fine-grained Meta Face Anti-spoofing

Home Page: https://arxiv.org/pdf/1911.10771.pdf

Python 100.00%

face-antispoofing meta-learning domain-generalization

rfmetafas's Introduction

AAAI2020-RFMetaFAS

Pytorch codes for Regularized Fine-grained Meta Face Anti-spoofing (arxiv) in AAAI 2020

Idea of the proposed regularized fine-grained meta-learning framework. By incorporating domain knowledge as regularization, meta-learning is conducted in the feature space regularized by the domain knowledge supervision. Thus, generalized learning directions are more likely to be found for task of face anti-spoofing. Besides, the proposed framework adopts a fine-grained learning strategy that simultaneously conducts meta-learning in a variety of domain shift scenarios. Thus, more abundant domain shift information of face anti-spoofing task can be exploited.

Overview of proposed framework. We simulate domain shift by randomly dividing original N source domains in each iteration. Supervision of domain knowledge is incorporated via depth estimator to regularize the learning process of feature extractor. Thus, meta learner conducts the meta-learning in the feature space regularized by the auxiliary supervision of domain knowledge.

Setup

Prerequisites: Python3.6, pytorch=0.4.0, Numpy, TensorboardX, Pillow, SciPy, h5py
The source code folders:
1. "models": Contains the network architectures suitable for high-order derivatives calculation of network parameters. Please note that FeatExtractor, DepthEstmator and FeatEmbedder in the code are feature extractor, depth estimator and meta learner in the paper, respectively.
2. "core": Contains the training and testing files. Note that we generate score for each frame during the testing.
3. "datasets": Contains datasets loading
4. "misc": Contains initialization and some preprocessing functions

Training

To run the main file: python main.py --training_type Train

Testing

To run the main file: python main.py --training_type Test

It will generate a .h5 file that contains the score for each frame. Then, we use these scores to calculate the AUC and HTER.

Acknowledge

Please kindly cite this paper in your publications if it helps your research:

@InProceedings{Shao_2020_AAAI,
author = {Shao, Rui and Lan, Xiangyuan and Yuen, Pong C.},
title = {Regularized Fine-grained Meta Face Anti-spoofing},
booktitle = {Thirty-Fourth AAAI Conference on Artificial Intelligence (AAAI)},
year = {2020}
}

Contact: [email protected]

rfmetafas's People

Stargazers

Watchers

rfmetafas's Issues

Could you please provide optimal training model?

I want to compare the ROC curves of some methods for domain generalization.
Could you please provide optimal training model? Thank you.
My Email address:[email protected]

精度问题

您好，我按照您的要求送入网络rgb人脸图片，但是不管是real还是fake的样本，分数都在0.5左右，没有大的变化。请问您遇到过这种问题吗？ @ @rshaojimmy

Does the pretrained model will be released?

Wait for the pre-trained model?

Error encountered running "python main.py --training_type Test"

Hi,

Thanks for sharing your code.
I have installed and setup the prerequisite software.
However when I run the Testing mode using this command "python main.py --training_type Test", I encountered the following error. The program main.py seems to be asking for the command line argument parameter tstfile.
Appreciate your advice. Thanks

File "main.py", line 175, in
main(parser.parse_args())
File "main.py", line 22, in main
savefilename = osp.join(args.tstfile, args.tstdataset+'to'+args.dataset_target+args.snapshotnum)
AttributeError: 'Namespace' object has no attribute 'tstfile'

Best Regards,
Benjamin

Not able to run "python main.py --training_type Test"

I am trying to run the Test method but keep running into errors, can you please mention what are all the required files and settings that I would need to be able to run it.

Do you align the faces before feeding into the network? Or only crop the bounding box?

image_list_real.txt

这是数据集自带的吗

How did you get ~_depth images?

Hello, I'm Minha Kim researching anti-spoofing in korea.
Firstly, I appreciate you sharing this method with the public.

I have one question, how did you get ~_depth images?
As I know, all datasets don't contain depth images.
So, I wonder how did you make or get the dataset as a depth type of image.

Thank you :)
With best regards,
Minha Kim.

some questions about datasets partitioning

nice, it is a good job.but,when i run it ,some quetions occers.as shown below:

i do not know the composition of documents,please show me the hierarchy of data set files. looking forward to your replay.

Can you share the specific path format of the txt file?

I am confused about the path processing of the data txt file. Can you share the specific path format of the txt file? Thank you!

Questions about project details

Hi,
thanks for sharing the code.
I have some questions about your project that I would like to confirm.
1.What did your meta learner mainly learn?(FeatExtractor Weight or some hyperparameters?)
2.When testing, is it not necessary to generate a face depth map? because it is only used to auxiliary supervision, like an image label?
3.we need to generate a depth map through PRNet and add it to the input ? or your code contains this generate function?
4.Is it possible to achieve real-time detection?

Thanks.

Question about datasets

I have some questions about the code In datasets/DatasetLoader.py :

if getreal:
            filename = 'image_list_real.txt'
        else:
            filename = 'image_list_fake.txt'

What's the content in image_list_real.txt about and how do you get these files?
I cannot get these files after I download the datasets so I am curious about it.

Data processing problem

if getreal and name=='idiap':
depth_dir = os.path.join('depth', dirlist[0], dirlist[1], dirlist[2], imgname + '_depth.jpg')
elif getreal and name=='CASIA':
depth_dir = os.path.join('depth', dirlist[0], dirlist[1], dirlist[2], imgname + '_depth.jpg')
elif getreal and name=='MSU':
depth_dir = os.path.join('depth', dirlist[0], dirlist[1], imgname + '_depth.jpg')
elif getreal and name=='OULU':
depth_dir = os.path.join('depth', dirlist[0], imgname + '_depth.jpg')
else:
==> depth_dir = os.path.join('depth', 'fake_depth.jpg')

Why you let all name of fake_depth_image are 'fake_depth.jpg' ?
Did you not use them?

I find the model is too large, and when training, the results are very shocking. For example, when I use the epoch100 trained_model to test, the ACC in testing-set is 80%, but when I use the epoch101 trained_model to test, the ACC in testing-set is below 60%. Do you know why?
I cannot reproduce the results of the C&I&M to O scenario in your paper. The best AUC results that I achieved by using your code is 82%, which has a large gap compared to the results published in your paper 91.16%. How do you test your model on Oulu dataset? Do you use the all data (training and testing) or just the testing data? Why I cannot reproduce the results?

Thank you very much!!

How do you change the video data into image? How often should I get a screenshot? (Such as every 5 frames?) Could you tell me the whole num of images in every dataset after the pre-processing?
Should the input image been croped to face part or remain the whole scene?(In my experiment, PRNet can only generate depth for face image.)

Thanks for your great work and hope for relpy!