A equipe do Google Brain, compartilhou as filtragens realizadas por eles visando um incremento substâncial da qualidade do modelo. As operações são estatísticamente distribuídas, da seguinte forma:
def policy_v3 ():
""""Additional policy that performs well on object detection."""
# Each tuple is an augmentation operation of the form
# (operation, probability, magnitude). Each element in policy is a
# sub-policy that will be applied sequentially on the image.
policy = [
[('Posterize', 0.8, 2), ('TranslateX_BBox', 1.0, 8)],
[('BBox_Cutout', 0.2, 10), ('Sharpness', 1.0, 8)],
[('Rotate_BBox', 0.6, 8), ('Rotate_BBox', 0.8, 10)],
[('Equalize', 0.8, 10), ('AutoContrast', 0.2, 10)],
[('SolarizeAdd', 0.2, 2), ('TranslateY_BBox', 0.2, 8)],
[('Sharpness', 0.0, 2), ('Color', 0.4, 8)],
[('Equalize', 1.0, 8), ('TranslateY_BBox', 1.0, 8)],
[('Posterize', 0.6, 2), ('Rotate_BBox', 0.0, 10)],
[('AutoContrast', 0.6, 0), ('Rotate_BBox', 1.0, 6)],
[('Equalize', 0.0, 4), ('Cutout', 0.8, 10)],
[('Brightness', 1.0, 2), ('TranslateY_BBox', 1.0, 6)],
[('Contrast', 0.0, 2), ('ShearY_BBox', 0.8, 0)],
[('AutoContrast', 0.8, 10), ('Contrast', 0.2, 10)],
[('Rotate_BBox', 1.0, 10), ('Cutout', 1.0, 10)],
[('SolarizeAdd', 0.8, 6), ('Equalize', 0.8, 8)],
]
return policy
O Código fonte em TensorFlow pode ser obtida aqui: