Convolutional neural network on neural compute stick for voxelized point-clouds classification

Abstract

2D Convolutional Neural Networks (CNNs) have enjoyed a surge in popularity over the last few years, mainly because they outperform traditional algorithms/methods in a myriad of computer vision (and other fields) tasks. On the other hand, the problem becomes more complex when dealing with 3D volumes. Lack of readily available training data, memory and computational requirements are just some of the factors hindering the progress of 3D CNNs. We propose a synthetic 3D voxelized point-clouds generation method containing object and scene in this paper. Furthermore, an efficient 3D volumetric representation called VOLA is applied. VOLA (Volumetric Accelerator) is a sexaquaternary (power-of-four subdivision) tree-based representation which aims to save significant memory for volumetric data. After training the model, it was deployed onto Movidius Neural Compute Stick which is a USB, containing a low-power processing unit as well as dedicated CNN hardware blocks. The trained model on NCS takes only ~ 90 frames per second to perform inference on each 3D volume, with an average power consumption of 1.2W.