Fast FFT-Based Inference in 3D Convolutional Neural Networks

Recognizing real world objects based on their 3D shapes is an important problem in robotics, computational medicine, and the internet of things (IoT) applications. In the recent years, deep learning has emerged as the foremost tool for a wide range of recognition and classification problems. However, the main problem of convolutional neural networks, which are the primary deep learning systems for such tasks, lies in the high computational cost required to train and use them, even for 2D problems. For 3D problems, the problem becomes even more pressing, and requires new methods to keep up with the further increase in computational cost. One such method is the use of Fast Fourier Transforms to reduce the computational cost by performing convolution operations in the Fourier domain. Recently, this method has seen widespread use for 2D problems. In this paper, we implement and test the method for 2D and 3D object recognition problems and compare it to the traditional convolution methods. We test our network on the ShapeNet 3D object library, achieving superior performance without any loss in accuracy compared to conventional methods.