i dont know the mcs cascades too well, but most probably they were trained on a closed mouth pose, where both lips are visible, –
that might explain, why it cannot find the mouth here
(and i’d also give up on finding the mouth using cascades here. simply skip that part, and infer on the “lower half” of the detected face)
unfortunately, there is no simple machinelearning (like the SVM) in opencv.js but you could still try to either:
- train a small cnn on toungue out / in images (later use the dnn module to infer it)
- train a HogDescriptor
- try a simple
sum(absdiff)and find a threshold
again, dont expect a “ready-made” solution here !