search

VIS-VAR / LGSC-for-FAS

Learning Generalized Spoof Cues for FaceAnti-spoofing
MIT License
226 stars 56 forks source link

Experiment setting, experiment results #28

Open weidandan1997 opened 3 years ago

weidandan1997 commented 3 years ago

Hello,author. I have several problems to ask you for help. First,I want to know what you mean in the paper " We resample the training examples to keep the live-spoof ratio to 1:1. ".Do you mean the number of live and spoof images in each batch is 1:1? Or is 1:1 only for training set ,not for each batch? Second,I want to know how many images do you resample for one video. Third, I have implemented some experiments, but the result is not good. The result in the protocol1,2 of OULU is not good.The results in bad in the protocol3 of SIW. The results is worse in the Replay-to-Casia.

Here are some spoof cues. I dont know why the spoof cues are so different in diffent dataset. And I dont know why there are always some Highlights in each spoof cues. oulu_protocol1 image image oulu_protocol2 image image

siw_protocol2 image image

Help!

adamhtoolwin commented 3 years ago

It seems your images are in a different color format to RGB. Or are they normalized?

weidandan1997 commented 3 years ago

It seems your images are in a different color format to RGB. Or are they normalized?

Yes,the images are in a BGR color format.

adamhtoolwin commented 3 years ago

The input to the model has to be RGB if I am not mistaken. Or is this only for visual purposes?

Just to share some of my experiences, I got an ACER of around 0.1 on the SiW-M dataset and similar on protocol 4 of the OULU dataset. However, I am not using paddle but the pytorch version (heavily modified version of this repo credit to Poddiving). But the performance drops when tested on a new dataset (or even different environment actually).

CHNxindong commented 3 years ago

The input to the model has to be RGB if I am not mistaken. Or is this only for visual purposes?

Just to share some of my experiences, I got an ACER of around 0.1 on the SiW-M dataset and similar on protocol 4 of the OULU dataset. However, I am not using paddle but the pytorch version (heavily modified version of this repo credit to Poddiving). But the performance drops when tested on a new dataset (or even different environment actually).

@AdamHtooLwin hello, I am also training on Oulu on the pytorch version code, But I can not get a good results as you, I get ACER 1.0 on Protocol1 and ACER 3.3 on Protocol2, can you share how to train it on this code, do you modify some code? Or do you have some good advice for me? Thank you! Thank you! Thank you!!!

adamhtoolwin commented 3 years ago

I suggest you take a look at this repo. I based my code off of this and it's Pytorch based too. And maybe give him a star too :)

PangziZhang523 commented 3 years ago

Hello, have you solved this Highlights problem in each spoof cues?

weidandan1997 commented 2 years ago

Hello, have you solved this Highlights problem in each spoof cues? Sorry,I haven't been following this problem.

Aluooooo commented 2 years ago

Have you solved this problem?