background
Solve image classification ConvNet method to detect problems become a craze, but these methods need to first resize images to a fixed w * h, and then thrown into the network, after resize images may lose some information. The authors invented the SPP pooling (space pyramid pooling) layer, so that the network can accept input of arbitrary size.
method
First, think about a problem, why ConvNet need a fixed size of the input picture as we know, Conv layer need only channel fixed (picture 3 color, grayscale 1), but can take any input w * h, of course, the output of w * h will follow changes; however, the back layer of FC it takes as input a fixed length vector, image size change -> size change in the output layer conv -> vector length changes FC input layer, which generates an error.
How to solve this problem? The method presented in the back of the last layer plus a layer Conv SPP pooling layer, SPP pooling input conversion layer may be a different size of the received output be fixed to ensure a fixed length of the input layer FC.
As shown, the principles of the SPP pooling layer is very simple example is as follows: Given a feature of FIG w * h, respectively, into the bin 4 * 4,2 * 2,1 * 1 in each bin for pooling the above operation (as used herein is max pooling), the last to give 16 * 256-d (256-d is the output channel number of the last conv layer), 4 * 256-d, 1 * 256-d of the feature vector, and finally connected together, get that feature vector 21 * 256-d's.
We can see, no matter how much the beginning of the w and h values, and finally can get a fixed length feature vector as input FC layers, so, ConvNet will be able to accept different size of images as input.
to sum up
Authors SPP pooling by adding a layer in front of the FC layer, an effective solution to ConvNet must accept the fixed size of the picture.