Mixup damages color nuances of input images ? Difficult to classify the objects with same design but different color.

[gouldberg]

I would like to discuss or get advice on discriminating objects with same design/size but different color.

During training applying mixup, 1 image (foreground + background) is mixed up with other 1 image (foreground + background).
In current yolox, the blended ratio is 0.5 and 0.5.

Foreground image is blended with, sometimes other foreground object in other 1 image or sometimes background in other 1 image.
If the foreground object is middle color such as pink, orange, brown .., (not red, green, blue), the foreground color is changed depending on mixed-up image during mixup training.
For example, pink + black is almost gray, pink + orange is almost red ? ...

So I think by applying mixup,
objects with same design/size but different color (including middle colors) are not properly learned, leading to annotation confusion during training.

Better NOT to apply mixup in this case or
any other idea for 'object detection' to discriminate objects with same design/size but different colors.
(I am really afraid that if not applying mixup, the detection performance is really dropped even if small model such as yolox-s)