open mouth about clm-framework HOT 3 CLOSED

CLM framework is quite out of date now, you should have a look at OpenFace project - https://github.com/TadasBaltrusaitis/OpenFace, which is the continuation of CLM work but with greater functionality and more accurate models, which should address the mouth opening issue (at least partally). The cause is both the training and regularization: 1) the training data does not contain many examples of widely open mouths, 2) open mouth is an unlikely face so regularization penalizes its likelihood.

For a more detail of the terms of the regularization you can have a look at my thesis - https://www.dropbox.com/s/gk6nlgn4y9vcm4d/TechReport.pdf?dl=0, specifically section 4.5.2 which explains where all the terms come from (this is actually what encourages the model to stay close to "neutral" faces)

from clm-framework.

You are right that it uses the same PDM, however, as the patch experts are more reliable, the regularization on the PDM can be reduced somewhat, allowing for more "unusual" expressions.

from clm-framework.

thanks a lot for your answers.

from clm-framework.