Classify 3D voxel based point-cloud using convolutional neural network on a neural compute stick

TLDR

CNNs have achieved high performance in many tasks, but deploying them on low‑power mobile devices is challenged by limited training data, memory demands, and the need for efficient volumetric representations such as VOLA. The study proposes generating synthetic 3D point‑clouds from CAD scene models and leveraging a Vision Processing Unit to train volumetric CNNs for mobile deployment. The authors train several volumetric CNNs on synthetic data using VOLA, then port the best model to the Fathom Neural Compute Stick via its VPU. The best model achieved 91.3 % test accuracy and, when deployed on the Fathom NCS, performed inference in 11 ms (≈90 fps) at 1.2 W, yielding 75.75 inferences per second per watt.

Abstract

With the recent surge in popularity of Convolutional Neural Networks (CNNs), motivated by their significant performance in many classification and related tasks, a new challenge now needs to be addressed: how to accommodate CNNs in mobile devices, such as drones, smartphones, and similar low-power devices? In order to tackle this challenge we exploit the Vision Processing Unit (VPU) that combines dedicated CNN hardware blocks and very high power efficiency. The lack of readily available training data and memory requirements are two of the factors hindering the training and accuracy performance of 3D CNNs. In this paper, we propose a method for generating synthetic 3D point-clouds from realistic CAD scene models (based on the ModelNet10 dataset), in order to enrich the training process for volumetric CNNs. Furthermore, an efficient 3D volumetric object representation (VOLA) is employed. VOLA (Volumetric Accelerator) is a sexaquaternary (power-of-four subdivision) tree-based representation which allows for significant memory saving for volumetric data. Multiple CNN models were trained and the top performing model was ported to the Fathom Neural Compute Stick (NCS). Among the trained CNN models, the maximum test accuracy achieved is 91.3%. After deployment on the Fathom NCS, it takes 11ms (~ 90 frames per second) to perform inference on each input volume, with a reported power requirement of 1.2W which leads to 75.75 inference per second per Watt.