Figure S5. Detailed description of feature landmark points. 13 different
facial landmarks from 3D facial image are extracted by using the
OpenFace® software, which extracts 68 different facial
points and gaze direction from the face. a) Top eyebrow, lower eyelid,
top nose, mouth end, and bottom lip of each right and left face are
extracted, and 9 different Euclidean distances are calculated from both
3D depth map of the subject’s face. The 3D Euclidean distances measure
the reconstruction errors of 3D depth maps acquired from different LFC
depending on the various NIR filters. b) All 78 pairwise distance
features of each subject’s different facial expression image are
extracted (dp0-p1 to dp11-p12). The 78
features become a one input vector, and a total of 2,248 input vectors,
including 281 input vectors from 4 different expressions repeated 3
times of 32 subjects and augmented random noises for robustness, are
applied to the multi-layer perceptron (MLP) for reading facial
expression.