Subtract mean value

[hytseng0509]

Hi,

Do we have to subtract the mean before the predicting?
Thanks for the sharing. =)

[tracer9]

Hi, @hytseng0509 , I checked the caffe implementation. Since caffe use OpenCV, so the image is :

• [0, 255]
• BGR mode
• mean sub-straction.

The implementation here use a skimage library. Input image is:

• [0, 255]
• RGB mode
• non-substraction

So you need first substraction the mean value according to each channel; then transfer RGB to BGR mode.

A sample code for reference:

X = get_sport_clip('roger')
X = Variable(X)
X.data[:, 0, :, :, :] -= 122.6 # R channel
X.data[:, 1, :, :, :] -= 116.6 # G channel
X.data[:, 2, :, :, :] -= 104.0 # B channel
X = X[:, [2,1,0], :, :, :] # channel swap
X = X.cuda()

According to experiment, the difference is really small.

[hytseng0509]

Thanks for the information!

[apple2373]

I think this is ImageNet mean. It's true that we should subtract mean but we should use the mean from Sports1M.