Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition

Existingdeepconvolutionalneuralnetworks(CNNs)requireafixed-size(e.g.,224224)inputimage.Thisrequirementis“artificial”andmayreducetherecognitionaccuracyfortheimagesorsub-imagesofanarbitrarysize/scale.Inthiswork,weequipthenetworkswithanotherpoolingstrategy,“spatialpyramidpooling”,toelimina