arnaldog12 / icaonet Goto Github PK

Hi,

I tried to download AR, FRGC and PUT dataset but not successfully.
Could you please help to share it via google drive or some where that I can download it.

Thank you alot

I can't seem to find your model, please is it publicly available?

I would like to be able to test the model but I don't know what the order of execution of the notebooks is. You will have the step by step to do tests

It seems you used TensorFlow V1 for the face detection and the ICAONET. Is there a V2 model for each of them especially for face detection? Also is possible to load the models with V2 in both python and C++.

Or is there a way to convert it to ONNx and then load it with OpenCV dnn

I noticed the preprocessing method you implemented does not crop the face. I am curious to know if that was intentional and if it was, I would like to know why you decided not to use the cropped face for training

Hello,

I tried running the model using a couple of images and using the weights shared in release 1.0, but I'm not sure how to interpret them. Does a result closer to 1.0 imply compliance with the requirement or does 1.0 mean non-compliance?

These are the image and the requirement scores I obtained:

[02] Blurred: 0.5959023833274841
[03] Looking away: 0.9834322333335876
[04] Ink marked/creased: 0.9980897307395935
[05] Unnatural skin tone: 0.3168644905090332
[06] Too dark/light: 0.9694647789001465
[07] Washed out: 0.9932628273963928
[08] Pixelation: 0.9621853232383728
[09] Hair across eyes: 0.9957687854766846
[10] Eyes closed: 0.9981269240379333
[11] Varied background: 0.5940201878547668
[12] Roll/pitch/yaw rotations greater than a predefined thresholds: 0.5451245903968811
[13] Flash reflection on skin: 0.5639101266860962
[14] Red eyes: 0.9995801448822021
[15] Shadows behind head: 0.6668246388435364
[16] Shadows across face: 0.8134157061576843
[17] Dark tinted lenses: 0.9999613165855408
[18] Flash reflection on lenses: 0.9192411303520203
[19] Frames too heavy: 0.9310649633407593
[20] Frame covering eyes: 0.9684894680976868
[21] Hat/cap: 0.0000643364832243
[22] Veil over face: 0.9917857646942139
[23] Mouth open: 0.972870409488678
[24] Presence of other faces or toys too close to face: 0.9891334176063538

As you can see, most values are very close to 1.0, with notable excepions being Hat/cap, flash reflection on skin, Varied background, Blurred and Unnaturak skin tone. This is peculiar as I believe the image I used as input should be compliant with all of the requirements, particularly those with the lower scores.