Standard input image resolution and neural networks inconsistencies

01 introduced

If we want to verify the performance vgg in cifar10, the following problems may occur:

vgg standard input: 224 x 224 x3

Cifar10 resolution image: 32 x 32 x 3

02 Possible solutions

One view:

Modify the size of the last few fully connected layers! (Absolutely feasible)

Another view:

For ordinary image classification, resize to standard input the desired network, may have no effect.

But for our example is concerned, such an operation is clearly inappropriate! ! !

For targets in terms of, resize the image means a deformation occurs, is likely to affect the results.