Darknet: Image Modifications during training

[golars497]

Hi,

Wanted to ask if these parameters in the .cfg file are all the available configs that can be used to "modify" the training images when training. Also are these values fixed? for example if I set the exposure to say, 4.0, do all images gets trained on a higher level of exposure? or is the value more like a "range" where the different level of exposures will be applied to the images rather than just a fixed amount.
angle=0
saturation = 1.5
exposure = 1.5
hue=.1

I hope my question makes sense. Many Thanks :D

[AlexeyAB]

Hi, You can change these values. Different levels of exposures will be applied to the images.

Will be calculated temporary variables:
float dhue = [-hue, hue]
float dsat = [1/saturation, saturation]
float dexp = [1/exposure, exposure]

Then these values will be applied in a such way:

 hsv[1] = hsv[1] * dsat; 
 hsv[2] = hsv[2] * dexp; 
 hsv[0] = hsv[0] + 179 * dhue; 

https://github.com/AlexeyAB/darknet/blob/701f4fab63b3f6826ae6095ce32b9b99b3ece203/src/http_stream.cpp#L270-L272

About HSV: https://en.wikipedia.org/wiki/HSL_and_HSV

[golars497]

Thanks for the swift reply! Sorry if this sounds silly, so basically the higher the values I set for saturation, exposure and hue the greater the range of changes are applied to the images? Also, was wondering what angle does, my guess is that it rotates the image (and if so if I set it to 360, will it rotate the image in all possible angles when training)?

Many Thanks

[AlexeyAB]

Thanks for the swift reply! Sorry if this sounds silly, so basically the higher the values I set for saturation, exposure and hue the greater the range of changes are applied to the images?

Yes.

Also, was wondering what angle does, my guess is that it rotates the image (and if so if I set it to 360, will it rotate the image in all possible angles when training)?

Angle isn't used for Yolo Detector at this moment. Angle is used for Classification only: https://github.com/AlexeyAB/darknet/blob/701f4fab63b3f6826ae6095ce32b9b99b3ece203/src/image.c#L1230-L1239

[Pattorio]

Hi Alexey, one more question. You said that

Angle isn't used for Yolo Detector at this moment. Angle is used for Classification only

Does it means there is no difference between angle=0 and angle=360 when I use YOLO for object detection?

[AlexeyAB]

@Pattorio

Does it means there is no difference between angle=0 and angle=360 when I use YOLO for object detection?

Yes.

[joel-simon]

@AlexeyAB
It would be very helpful to others to document the fact that angle has no effect on detection (I had assumed it did)

Thank you for your work on this great library,
Joel

[Pattorio]

Hi @AlexeyAB, Thank you for your answer.
If I want to detect human head with different angles (some of heads may be upside down), is there any good way to do the data augmentation under condition that angle is not used in detection?
What about rotating original images and adding them into training set for training?