Mean subtraction after cropping

[mgarbade]

There is a bug in the image transformation. The mean subtraction must happen before the zero padding performed by tf.image.resize_image_with_crop_or_pad.

I'm trying to squeeze the functions
img = tf.random_crop(img, [h, w, 3])
label = tf.random_crop(label, [h, w, 1])
in there as well to get a random instead of a center crop, but I'm not sure if they take the same crop from image and label if I do it like this... So I keep looking into that.

[arslan-chaudhry]

Yes, this looks like a bug.
As for random crop, you can change the code to check if image dimensions are greater than '321', then do a random crop using tf.random_crop , else pad the inputs accordingly. I am also doing the same thing, as always doing a centre-crop effectively reduces the amount of training data.

[DrSleep]

@mgarbade, nice catch! Fixed!

The tf.random_crop functions has a seed argument that you can provide to both instantiations to guarantee that the crop is the same for the image and the mask.