Data agumentation—what are reasonable values for tf.image.random_[...]?

General Discussion
,
#1

I am using TensorFlow functions to do image data augmentation for CNNs that are trained to recognize tree species on aerial images of forest. However, I am not 100%-ly sure about the parameters of the functions I use, since there is little documentation available.

I use the following code:

input_image = tf.image.random_brightness(input_image, max_delta = 0.1)
input_image = tf.image.random_contrast(input_image, lower = 0.75, upper = 1.25)
input_image = tf.image.random_saturation(input_image, lower = 0.75, upper = 1.25)

What is max_delta? Is it a fraction? I expect a value of 0.1 to cause some change in brightness in the range of -10 % to +10 %, is that the case?
Are there any units to contrast and saturation?
Are the values of a reasonable or would these functions expect some input in a whole other magnitude, e.g., values between 0 and 100 (some percent value)?

#3

Hi @Manuel_Popp, tf.image.random_brightness adjusts the brightness of images by a random factor by taking the value between [-max_delta,max_delta]. The values you pass to the max_delta, lower and upper bound depends on your choice. Thank You.