Darknet: My training set is a grayscale graph. In addition to the channels in the cfg file changed to 1, do I need to change anything else?

[ss199302]

@AlexeyAB My training set is a grayscale graph. In addition to the channels in the cfg file changed to 1, do I need to change anything else? thanks

[AlexeyAB]

@ss199302 You even shouldn't change channels to 1 in the cfg-file. Just leave channels=3: https://github.com/AlexeyAB/darknet/blob/57fce97fd5c2486997a6ee9ac67cccf2a5f35c81/cfg/yolov3.cfg#L10

Using channel=1 instead of 3 affects performance less than 1%.

---

If all of your training and detection images/videos will be gray-scale, then you can disable color data augmentation, leave only exposure=1.5: https://github.com/AlexeyAB/darknet/blob/57fce97fd5c2486997a6ee9ac67cccf2a5f35c81/cfg/yolov3.cfg#L14-L16
Use these:

saturation = 1
exposure = 1.5
hue=0

[AlexeyAB]

But I didn't test it, also there is as issue realted to this case: https://github.com/AlexeyAB/darknet/issues/359

Try to train with channel=3 about 2000 iteration, and test it, can it detect anything with -thresh 0.05?
And what mAP can you get?

[ss199302]

I train with channel=3, i can detect object,and i don't have map function.

[AlexeyAB]

@ss199302 Once training and detection work fine, then use channel=3

[ido-ran]

@AlexeyAB,
I also have grayscale training images. I've read this and other threads but I still didn't understand a clear answer: is it better to train on color images than grayscale images?

[AlexeyAB]

@ido-ran @ss199302
Currently you should change channels=1 in the cfg-file to use 1-channel Images or Video: https://github.com/AlexeyAB/darknet/pull/936

[dharinik]

what do the following lines in the cfg actually do?

saturation = 1
exposure = 1.5
hue=0