When training R-Net, should the input image minus 'pixel_mean'? about deep3dfacereconstruction HOT 4 CLOSED

We did not do any preprocessing for input images. All input image used to train R-Net have intensities range from 0-255, in BGR order.

from deep3dfacereconstruction.

Ok, but in many image process cases, the image will minus the 'pixel_mean' before sending to CNN, so have you tried this kind of process?

from deep3dfacereconstruction.

We did not try this preprocessing step in our experiments. Actually, I think minus it with pixel_mean is not necessary because network can learn this bias by itself. I agree that not doing so may increase the network's burden, but it could be neglect in the reconstruction task in my opinion.

from deep3dfacereconstruction.

Ok, well, I would like to have a try. Thank you.

from deep3dfacereconstruction.