BV-CNNs: Binary volumetric convolutional networks for 3D object recognition

Abstract

Though 3D convolutional neural networks (CNNs) have achieved impressive performance for object recognition, they still face challenges in computational and memory cost. In this paper, we propose binary volumetric convolutional neural networks (namely, BV-CNNs) for efficient 3D object recognition. Specially, it transforms the inputs and weights in the network to binary values through binary transformation, then the floating-point arithmetic convolutions are replaced with bitwise operations to reduce the computational and memory cost. Three binary volumetric CNNs are introduced from the traditional CNNs using our BV-CNN approach. Experimental results on the ModelNet and Sydney Urban Objects datasets demonstrated that computational and memory cost can be significantly reduced while achieving a high recognition accuracy comparable to the networks with floating point weights.