work support

[picli3]

hi Luca, I would like to know if you could make me a budget for a more accurate trained model of mouth and eyes. You can write to my email.

maykol.rey@services.mss.pe

[picli3]

[Luca96]

Hi, you can try training the model by following this script.

I don't know your application requirements and constraints, but you can use the following training options:

options = dlib.shape_predictor_training_options()
options.tree_depth = 4
options.nu = 0.1
options.cascade_depth = 15
options.feature_pool_size = 1000
options.num_test_splits = 300

With which you should achieve the best accuracy, but at the cost of a slower inference and a bigger model.

I've seen the screenshot you post. Maybe the problem is about the face location. How do you get the face bounding box? The trained models, that I provide (like the Dlib's original one), assume that the face is located by the Dlib HOG face detector.

Hope it helps, Regards.

[picli3]

hello Luca, thanks for your answer.

the face detection I'm doing with haar cascade. for now I have little time to do a training and try because I have more activities in the company.

I was wondering if you have a more accurate model, the hardware is a Jetson TX2.

[Luca96]

What set of landmarks should be located by the model? Only the eyes (points from 36 to 47) and mouth (points from 48 to 67) landmarks?

[picli3]

eyes from 36 to 47 and mouth from 48 to 59 and point 30 for the tip of the nose.

[Luca96]

I've trained this model with the landmarks you need, and with a more accurate set of parameters. Let me know how it works.

[picli3]

thank you I await your answer, thank you really.