FOCM: Faster Octave Convolution Using Mix-scaling

Abstract

Octave convolution that separates the feature maps for different resolutions is an effective method to reduce the spatial redundancy in Convolution Neural Networks (CNN). In this paper, we propose a faster version of octave convolution, FOCM, which can further reduce the computation cost of CNNs. Similar to the octave convolution, FOCM divides the input and output feature maps into the domains of different resolutions, but without explicit information exchange among them. In addition, FOCM utilizes the mix-scaled convolution kernels to learn different sized spatial features. Experiments on various depth ResNet with ImageNet data-set have shown that FOCM can reduce 33.9% to 46.4% operations of the original models, and save 11.1% to 21.7% FLOPS of the models using octave convolutions, with similar top-1 and top-5 accuracy.