Input image format for Resnet-50

[kilickaya]

Hi Luck,

The post seems to be very useful. On the web I came across many Resnet implementations which is a combination of Keras+Tensorflow code (i.e. not pure Tensorflow).

One issue I can see is that the network does not apply any preprocessing to the images.

1. BGR not RGB. Keras models are mostly imported from Caffe, meaning that they probably used cv2 to read/feed images. In this case, the input format should be BGR and not RGB
2. Mean subtraction. Also, it is always advisable to subtract ImageNet mean from the images before feeding to the network (again, respecting to the BGR order). I can't see this either in your implementation.