Train RGB and Test BGR performs better than Train RGB and Test RGB
Guys I've got a weird issue.
I'm trying to solve a Salient Object Detection (SOD) problem, using Encoder+Decoder architecture.
The module's encoder is VGG16, which accepts RGB as input. Also, I'm using a pre-trained model, that's why I need to normalize the input by subtracting the mean through channels.
R_MEAN = 123.68
G_MEAN = 116.78
B_MEAN = 103.94
CHANNEL_MEANS = [R_MEAN, G_MEAN, B_MEAN]
Everything looks fine, but when I test my trained model with BGR input (instead of RGB), it performs better (~1%-3%) by almost all metrics and visual comparison.
Do you have any explanations for this or any idea?