Before the fix the demo_inference.py used batch_norm and it did the
normalization of input image implicitly. If at inference time the
batch_norm was disabled the inference produced incorrect results.
This fix does the proper input image normalization and disables the batch_norm
at inference time.