the single_frame in mydata

[zj19921221]

hi thanks for you release the code;
I have run the single frame code on my own data of face_anti_spoofing;
in my opnion, the key point of the single frame method is the "Depthwise Spatial
Gradient Magnitude" but I found out that in my exprienment, it is not play a good role in my dataset;
how about your dataset;
on more question : Was the Depthwise Spatial Gradient Magnitude better than "short cut" in you expriment?

looking forward for your reply!
thanks

[clks-wzz]

1. "Depthwise Spatial Gradient Magnitude" indeed plays an important role in our experiment because it can grasp the detailed spoofing clues.
2. What's the setting in your experiments, such as backbone, learning rate, training iteration and the scale of your dataset?

[zj19921221]

thanks for your reply;
1、Have you compared Depthwise Spatial Gradient Magnitude with short_cut which keeped all parameters same;
2、in my experiment, backbone is mobilenet_v2 ; I replace the short_cut with the "Depthwise Spatial Gradient Magnitude" and keeped the parameters all same, I found out that "Depthwise Spatial Gradient Magnitude" is not as well as in my exp and data; how about you?

[clks-wzz]

The mobilenet_v2 may be too deep to converge for RSGB and depth supervised learning. Which loss function do you try in your experiment, binary or depth? A pretrained model initialization on IMAGENET might bring you a better result.

[zj19921221]

my loss: binary and depth;
thanks for your reply!
I have another question why you use architechture in the single_frame_net;
DO you use other models such as mobilenet_v2 shuffenet_v2 ?
you Net is better than those models?

thanks

[zj19921221]

The mobilenet_v2 may be too deep to converge for RSGB and depth supervised learning. Which loss function do you try in your experiment, binary or depth? A pretrained model initialization on IMAGENET might bring you a better result.

hi wzz
you mentioned that mobilenet_v2 is too deep; RSGB is more suitable used in the former layers?

[clks-wzz]

We have tried deeper network architectures, and found that the deeper models even may be not suitable to the depth supervision task.
You can try RSGB in the former layers, or directly use shallower network architectures.

[zj19921221]

ok thanks for your reply

[shahrzadesmat]

hi thanks for you release the code;
I have run the single frame code on my own data of face_anti_spoofing;
in my opnion, the key point of the single frame method is the "Depthwise Spatial
Gradient Magnitude" but I found out that in my exprienment, it is not play a good role in my dataset;
how about your dataset;
on more question : Was the Depthwise Spatial Gradient Magnitude better than "short cut" in you expriment?

looking forward for your reply!
thanks

could you please share your code and your data?thanksss

[punitha-valli]

@clks-wzz @zj19921221 @shahrzadesmat

Hi,

I would like to know about the test score, in the util_test_OULU_Protocol_1.py

the Testscor.txt, file

can you please tell me about that text file? how to generate those text files.

How did you calculating the map score?

it will be a great help for my studies

hope for a reply

thank you