How to visualize the ILSVRC mean image

How to visualize the ILSVRC mean image.md

I had an issue with how to visualize the ILSVRC mean image. I just wanted to look at it and see how much does it differ from using pixel-wise mean subtraction instead of image-wise mean subtraction.

I assume that you have already downloaded the CaffeNet pretrained and model definition files.

The trick is to initialize two networks, one with mean file set (called net_mean) and the other one without mean file (called net). Then create a fake all 1 image. Use the net_mean to preprocess the fake image for data layer and save the result as fake_pre. Then use the net to deprocess fake_pre for data layer and save it as fake_re. If the two networks net and net_mean were the same then fake_re would be equal to fake, but since we have not set any mean file for net then we can visualize the mean image using 1 - fake_re. Take a look at the code.

The result looks like this:

imagenet_mean.png

the trick.py

import caffe
import numpy as np
from matplotlib import pylab as plt

net = caffe.Classifier('/path/to/caffe/models/bvlc_reference_caffenet/deploy.prototxt',
                       '/path/to/caffe/models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel',
                       channel_swap=(2, 1, 0), raw_scale=255)
net_mean = caffe.Classifier('/path/to/caffe/models/bvlc_reference_caffenet/deploy.prototxt',
                            '/path/to/caffe/models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel',
                            mean=np.load('/path/to/caffe/python/caffe/imagenet/ilsvrc_2012_mean.npy'),
                            channel_swap=(2, 1, 0), raw_scale=255)

fake = np.ones((227, 227, 3))

fake_pre = net_mean.preprocess('data', fake)
fake_re = net.deprocess('data', fake_pre)

mean_image = 1 - fake_re

plt.imshow(mean_image)

It seems that the image link is not corrupted.
And, how would it affect the prediction result of using the two ways of mean?