search

yaojieliu / ECCV2018-FaceDeSpoofing

http://cvlab.cse.msu.edu/project-face-anti.html
MIT License
161 stars 42 forks source link

cannot distinguish real or fake on Replay-Attack database #19

Open SE2AI opened 4 years ago

SE2AI commented 4 years ago

The code can be run successfully, but when taking both a real and a fake video from Replay-Attack database as inputs, the model generates both score sequence with value alomost equals to 0.4. So, it seems the model cannot distinguish real and fake videos with low resolution, right?

santo4ul commented 3 years ago

Hi @SE2AI, Any luck on this issue?

I see a similar problem. I get a score of about 0.4 for both live and replay-attack.

One thing that I found in facepad-test.py is that, by default opencv reads in BGR format, this needs to be converted to RGB before feeding to the network. However, this change doesn't help either.

@yaojieliu, any thoughts on this? Do we need a complete re-train for more number of epochs maybe?