lincolnhard / face-point-shape-model-tracking

ancient code with deprecated function, refactoring required

face-point-shape-model-tracking

Modified and inspired from https://github.com/MasteringOpenCV/code/tree/master/Chapter6_NonRigidFaceTracking

The method is similar to ASM, but it use template matching for landmark tracking instead of active contour.

I use 24 landmarks to build shape model, which are label 27-36, 38-44, 48, 49, 51, 53, 54, 56, 58 in MUCT annotation. http://www.milbo.org/muct/muct-landmarks.html

There are 8 bases in shape model, 4 rigids and 4 non-rigids.

The 4 rigid bases are representing scale, in-plane rotation, translation in x direction, and translation in y direction.

The 4 non-rigid bases are head pitch, certain facial expression, head yaw, certian facial expression:

By projecting landmarks to the subspace formed by 4 rigid and 4 non-rigid bases, then project back, we can have a standard enough face shape. Furthoemore, limit the weight of each non-rigid basis using its standard deviation makes every face shape common.

Tracking code is re-written in C-style, and scaled down patch matching search area for faster performance.