How does the hue augmentation work?

[LukeAI]

What is the hue data augmentation unit?

What would be a good value for hue and saturation if the colour of the object of interest is basically irrelevant and I want the detector to be able to find objects of the same type but different colour to most of the training set.

context: I have a very large open source dataset on car license plates (CCPD 2019, 250000 images) but they are all chinese and so all have a blue background. I want to be able to detect license plates from anywhere in the world. My plan is to start by training a large network just on this very large database of chinese number plates and then using that to automatically annotate a large, more diverse dataset of cars and trucks with unmarked license-plates which I can manually check and then use to train a small network.

[LukeAI]

eg.

[AlexeyAB]

What is the hue data augmentation unit?

• What is the hue: https://en.wikipedia.org/wiki/HSL_and_HSV
• What is the hue data augmentation:
  • https://github.com/AlexeyAB/darknet/blob/eecc86116967b26dc9df86b0427e04e980cbdd83/src/data.c#L879-L881
  • https://github.com/AlexeyAB/darknet/blob/eecc86116967b26dc9df86b0427e04e980cbdd83/src/image_opencv.cpp#L1180-L1182

What would be a good value for hue and saturation if the colour of the object of interest is basically irrelevant and I want the detector to be able to find objects of the same type but different colour to most of the training set.

set hue=1.0

f.e.:

saturation = 3.0
exposure = 3.0
hue=1.0

Just check how does it work by training with flag -show_imgs