How darknet (YoloV4) manages color channels?

[CdAB63]

Good evening.

Yesterday I was wondering (but without really going into the code): do darknet features (filters) for RGB images (thus 3 channel features) have RGB layers all equal? like:

1 0 0 - 1 0 0 - 1 0 0
1 0 1 - 1 0 1 - 1 0 1
0 1 1 - 0 1 1 - 0 1 1

Or do I have all possible combinations?

1 0 0 - 0 1 0 - 0 0 1
0 1 0 - 1 0 0 - 0 0 1
0 1 0 - 1 0 0 - 1 1 0
...

For in the first case I have 512 (8 * 8 * 8) possible features and in the second case possible values or 512^3 (512 * 512 * 512) possible features.

In the first case it´s about bringing the convoluted image to an unweighted grayscale and doing that is about getting a 3 channel image and converting it into a 1 channel grayscale and then applying the 2D features (filters).

But in this case it´s supposed that objects will be illuminated by white light, which is not always true because under certain conditions (night lights with non uniform spectrum) things can get quite weird.

On the other hand, if I use all possible 3 channel features (512^3) that will result in 136590875 possible filters and it is necessary some criterion to keep things small and healthy