load_balanced_frontal_face_fitter

menpofit.aam.load_balanced_frontal_face_fitter()[source]

Loads a frontal face patch-based AAM fitter that is a good compromise between model size, fitting time and fitting performance. The model returns 68 facial landmark points (the standard IBUG68 markup).

Note that the first time you invoke this function, menpofit will download the fitter from Menpo’s server. The fitter will then be stored locally for future use.

The model is a PatchAAM trained using the following parameters:

Parameter

Value

diagonal

110

scales

(0.5, 1.0)

patch_shape

[(13, 13), (13, 13)]

holistic_features

menpo.feature.fast_dsift()

n_shape

[5, 20]

n_appearance

[30, 150]

lk_algorithm_cls

WibergInverseCompositional

It is also using the following sampling grid:

import numpy as np

patch_shape = (13, 13)
sampling_step = 4

sampling_grid = np.zeros(patch_shape, dtype=np.bool)
sampling_grid[::sampling_step, ::sampling_step] = True
sampling = [sampling_grid, sampling_grid]

Additionally, it is trained on LFPW trainset, HELEN trainset, IBUG and AFW datasets (3283 images in total), which are hosted in http://ibug.doc.ic.ac.uk/resources/facial-point-annotations/.

Returns

fitter (LucasKanadeAAMFitter) – A pre-trained LucasKanadeAAMFitter based on a PatchAAM that performs facial landmark localization returning 68 points (iBUG68).