Improving mobile MR applications using a cloud-based image segmentation approach with synthetic training data
Marcel Klomann, Michael Englert, Kai Weber, Paul Grimm, Yvonne Jung · 2018
In this paper, we show how the quality of augmentation in mobile Mixed Reality applications can be improved using a cloud-based image segmentation approach with synthetic training data. Many modern Augmented Reality frameworks are based on visual inertial odometry on mobile devices and therefore have limited access to tracking hardware (e.g., depth sensor). Consequently, tracking still suffers from drift that makes it difficult to utilize in use cases that require a higher precision. To improve tracking quality, we propose a cloud tracking approach that uses machine learning based image segmentation to recognize known objects in a real scene, which allows us to estimate a precise camera pose. Augmented Reality applications that utilize our web service can use the resulting camera pose to correct drift from time to time, while still using local tracking between key frames. Moreover, the device's position in the real world, when starting the application, is usually used as reference coordinate system. Therefore, we simplify the authoring of MR applications significantly due to a well-defined coordinate system, which is context-based and not dependend on the starting position of a user. We present all steps from web-based initialization over the generation of synthetic training data up to usage in production. In addition, we describe the underlying algorithms in detail. Finally, we show a mobile Mixed Reality application, which is based on this novel approach and discuss its advantages.