We found that there are more than 1~2% of the data are dirty data (without face / crop failed).
In this rate the dirty data issue will seriously affect the result of the contest.
Thanks for your comments. Compared to other face anti-spoofing datasets, HiFiMask is one of the large-scale face anti-spoofing datasets. Therefore, some noise labels maybe exist, because we cannot check all videos manually. The noise labels may be two reasons:
1) false face detection by face detection algorithm, such as background-image, black image...
2) the capture environment setting is too complex, which may cause some black images. But we note that the black images also include the real/fake faces.
But we think it is also fair for all participants. And you don't care about these noises.
PS: some famous datasets also include noise, such as Imagenet, Webface... We believe the proposed HiFiMask dataset would also push the cutting-edge research for face anti-spoofing.
Lastly, please enjoy this contest.
Organizing Team
Posted by: gesture_challenge @ June 13, 2021, 4:13 p.m.We just want to point out that in the first condition, what the model do is pure guessing, which might cause the following condition:
If participant A's model guess real and B's model guess fake, while the label of the blank/dirty image is real, than it won't be fair.
Of course maybe this particular scenario won't happen. If the organizer thinks that's fine than we have no other concern.
We appreciate the hard work. Thanks.