Class RotatedFrontalFace_CDataGenerator

      ml.ml.DataGenerator --+    
                            |    
ml.cla.dataset.CDataGenerator --+
                                |
                               RotatedFrontalFace_CDataGenerator

Generate face and non-face image patches of size 32-by-32.

Frontal faces can have:

• a : rotation angle in multiples of 45 degree, clockwise

  a = (M_PI / 4) * (i_a - 4), where i_a in {0..7}

• s : size 19x19, 21x21, 23x23, 25x25

  s = 19 + 2 * i_s, where i_s in {0..3}

• x : shift in x in {-3.5, -2.5, ..., 2.5, 3.5} pixels

  x = i_x - 3.5, where i_x in {0..7}

• y : shift in y in {-3.5, -2.5, ..., 2.5, 3.5} pixels

  y = i_y - 3.5, where i_y in {0..7}

We have in total 8*4*8*8 = 2048 face locations (a,s,x,y). Use static method class_to_location() to convert from class to location.

Neighboring patches:

• Translation: move by 8 pixels.
• Scaling: scale down with scaling factor 1.5

Compared to traditional methods:

Shifting: traditional 1.5 pixel, ours 1.0 pixel Scaling: traditional 1.2-1.5, ours 1.02-1.12 Overall: we consider 2-4 times more locations than traditional methods