Real-time visual object classification for augmented reality

Abstract

Augmented Reality (AR) is becoming one of the most interesting and valuable technologies in the digital space. AR applications integrated with image classification functionality are able to identify objects and provide meaningful interactions for users. The study on the deep CNN models has experienced great success in terms of image classification tasks. Considering the users’ requirements on object of interest may keep changing, the image classification models for an AR application are required to keep updating and retraining. The project focused on the automation of the neural network model training utilising a promising and solid technology called Google AutoML. The key contribution of the project was to design and implement an image classification tool which allows the user to build their own models with growable database. Furthermore, the project also researched on implementation of a novel approach of hierarchical image classification. A tree-based multi-model system was developed with a special prediction mechanism. The results of the prediction accuracy were analysed and compared with a flat n-way classifier. Comparison among different approaches exploiting label relations was conducted as well.