3-Dimensional object recognition using 1-dimensional capsule neural networks

Currently, indoor home object recognition systems lack the degree of accuracy required for reliable automated operations. In this paper, a 3-Dimensional (3D) object recognition deep neural network system, capable of recognizing indoor objects from 3D images with a view to assisting indoor robotic devices in performing home tasks, is presented. Almost invariably such systems drastically increase the number of total parameters which must be trained leading to very time-consuming training processes. Furthermore, the lack of suitably large annotated datasets for indoor objects adds to the training difficulties. To address these challenges a 3D dataset arranged in a 1D array format along with new architectures of 1D CapsNet is proposed. This effectively reduces the total number of parameters, resulting in faster and more accurate training of a neural network. Using the proposed architecture also appears to improve the training accuracy even when the datasets are relatively small. For this work, ModelNet-10 and the ModelNet-40 datasets are used. They have indoor home object images in 3D, which are few when compared to other datasets.