A model I have trained. Basic config is shown below.
T = 5
K = 540
landmark = 27
multi_scale = false
The model contains 5 stages and 540 carts in every stage. It uses 27 facial landmark points and doesn't use multi scale to extract the feature. This config.json is used to train this model.
This model is labeled luoyetx. I set a score threshold to remove most false positive result.
It depends on your optimization and parameter settings.
I have tested this model on a very simple video, see here.