search

minivision-ai / Silent-Face-Anti-Spoofing

静默活体检测（Silent-Face-Anti-Spoofing）
Apache License 2.0
1.36k stars 450 forks source link

Has anyone been able to create their own model? #100

Open tjkusnadi opened 2 years ago

tjkusnadi commented 2 years ago

How many datasets are used? How many images?

amilkcar commented 2 years ago

I will try, but i didn't understand how to resize the images to create the dataset.. it works for you or did you understand how to work with images or what kind of problems did you had ?

AlienX456 commented 2 years ago

did you were able to train the model?

tjkusnadi commented 2 years ago

nope.

tjkusnadi commented 2 years ago

how many images for each classes?

tjkusnadi commented 2 years ago

how do you get the datasets?

RayanAbdulnaser commented 1 year ago

300k live, 700k spoof

could please publish this project

Madina-S commented 1 year ago

How to prepare the data for retraining? I did not understand pre-processing stage

Nurmukhammeds commented 1 year ago

Try to search data gathering portals (Toloku)

Madina-S commented 1 year ago

@nurmukhammeds thank you for your response, but gathering is not problem. I have my data, but the accuracy of the model is not enough for me, so I want to add my data too, while keeping already computed weights (retrain). However, I could not even find good documentation for preparing data for training. The orginal image sizes must be the same? Or they can be different? How to split fake and real ones in the folder? And other questions. So, can you please provide some roadmap for retraining the existing models with my own dataset.

Nurmukhammeds commented 1 year ago

So, if you want to retrain existing models keeping weights, use patch size in the models name and resize your data to corresponding size with patch scale. Data directory can be structured as shown in Readme .

Madina-S commented 1 year ago

So, if you want to retrain existing models keeping weights, use patch size in the models name and resize your data to corresponding size with patch scale. Data directory can be structured as shown in Readme .

I understood, but how to split the traing data (real, fake faces)? In the code, there are 3 classes, why? What kind of classes?

Then, please, can you share the .py file for creating patches? generate_patches.py includes only CropImage class without usage example code.

pedromoraesh commented 1 year ago

Did someone understand why are there 3 classes and the kind of those classes? And finally, which size of the images and how to separate them?

eruslmu commented 1 year ago

folder 0 - 2D fake folder 1 - real folder 3 - 3D fake

If you have only two cases (real and fake photos), then you don't need to have folder 3 which holds 3D fake photos.

freedom9393 commented 10 months ago

folder 0 - 2D fake folder 1 - real folder 3 - 3D fake

If you have only two cases (real and fake photos), then you don't need to have folder 3 which holds 3D fake photos.

Are you sure? Cuz, they said directories 0, 1, 2 folders are for different scales of images (not for different classes). Please, correct me if I'm wrong