About the BBox Detection for "masked-Face"

[wjwang0816]

Thanks for this master work!
About the "masked face", I have some question, hope to give me some advice please. Thanks a lot !
(1) Compare to the normal face (non-masked face), Why the BBox of the masked face is smaller than normal face ?
The bbox of masked face seems like more approximating "square", but the bbox of normal face like "rectangle".
(2) Is it possible to train a model to predict "two classes" - "non face" & "face (masked & normal face)"?
and get high accuracy bbox for masked face as same as normal face?
In the other word, I just hope to predict "two" classes :
- class one : background (non face)
- class two : face (including "normal face" & "masked face")
In my application, I want to detect normal face & masked-face, regardless of masked or non-masked.
So I hope the BBox accuracy of masked face can be near to normal face
In the "maked_face" Folder of the repo, the prediction result is seperated to three classes, how can I modify to "two classes" - "non face" & "face (masked & normal face)"?
Thanks a lot !