Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution

By Y. Chen et al

Published on Aug. 18, 2019

Read the original document by opening this link in a new tab.

Table of Contents

1. Introduction
2. Related Work
3. Method

Summary

In this paper, the authors propose Octave Convolution as a method to reduce spatial redundancy in Convolutional Neural Networks. By factorizing feature maps into high and low frequencies, Octave Convolution efficiently processes information at different spatial resolutions. The approach is shown to improve accuracy for image and video recognition tasks while reducing memory and computational costs. The paper discusses the method, implementation details, and its integration with group and depth-wise convolutions.