Real-time object detection on Android devices - A step towards automatic GeoPose estimation on mobile AR platforms

Abstract

Augmented Reality (AR) technologies are becoming more available than ever. What the technologies lack, is a system making it possible for users to share and communicate across all the different AR platforms. This is the motivation behind the AR-Cloud and the open source AR-Cloud community Open AR-Cloud. A research and development program aims to find optimal ways to automatically estimate the geographical position and orientation of AR-devices GeoPose using object detection on a phone and a continuously updated model of the world. There is a rising interest in building small and efficient Neural Networks for mobile deployment. This master thesis has investigated real-time object detection on an Android device by looking into how a combination of a dataset and an object detecting method will solve the task of identifying stationary objects found in an everyday scene as a step towards automatic GeoPose estimation. The chosen method achieves an acceptable inference speed of five frames per second on a Samsung s9 and a mAP of 33,46%.