help me

[nadongjin]

Hello
I have a question.
Personally, is there a way to detect faces in images I have, like glint360k or ms1m ?
mtcnn, retina, scrfd, etc. hold too tight when detecting faces.
thank you.

[leondgarse]

Ya, mostly detectors will not work well on those extracted training face images. I'm considering if I can hold a yolov5-face model here, but that may also not work well. May I know why you need that?

[nadongjin]

Hello
I am learning deepglint360k dataset.
Also, I have an additional data set.
I want to train with deepglint360k + additional dataset , but I don't know how to get a roi like deepglint360k.
One way I think of it is to get a face roi through face detection such as scrfd and retina, and then take a wider margin for both left and right, I don't know if this is the best way.
Thank you.

[leondgarse]

Uh, so you just wanna create your own dataset like glint360K one. I misunderstood you wanna do face detection in glint360K...

• Technically, faces extracted using scrfd / retinaface should be compatible with Glint360K. How your extracted yours?
• You can also refer to my efficientnetV2_basic_test.ipynb the last Keras insightface test part. It's an example of using scrfd and also some instances of extracted images. I think the margin just alright.

[leondgarse]

I've uploaded the YoloV5FaceDetector, which is a pure TF one, and also the above colab ipynb script. The usage is almost the same:

""" Previous """
# from eval_folder import Face_detection
# det = Face_detection()

""" Now """
import face_detector
det = face_detector.YoloV5FaceDetector()
# Or det = face_detector.SCRFD()

Other processes keep unchanged.

[nadongjin]

Thank you very much.
You're a great developer.