Large-scale, real-time 3D scene reconstruction on a mobile device

Abstract

Google’s Project Tango has made integrated depth sensing and onboard visual-intertial odometry available to mobile devices such as phones and tablets. In this work, we explore the problem of large-scale, real-time 3D reconstruction on a mobile devices of this type. Solving this problem is a necessary prerequisite for many indoor applications, including navigation, augmented reality and building scanning. The main challenges include dealing with noisy and low-frequency depth data and managing limited computational and memory resources. State of the art approaches in large-scale dense reconstruction require large amounts of memory and high-performance GPU computing. Other existing 3D reconstruction approaches on mobile devices either only build a sparse reconstruction, offload their computation to other devices, or require long post-processing to extract the geometric mesh. In contrast, we can reconstruct and render a global mesh on the fly, using only the mobile device’s CPU, in very large (300 m$^2$) scenes, at a resolutions of 2–3 cm. To achieve this, we divide the scene into spatial volumes indexed by a hash map. Each volume contains the truncated signed distance function for that area of space, as well as the mesh segment derived from the distance function. This approach allows us to focus computational and memory resources only in areas of the scene which are currently observed, as well as leverage parallelization techniques for multi-core processing. Furthermore, we describe an on-device post-processing method for fusing datasets from multiple, independent trials, in order to improve the quality and coverage of the reconstruction. We discuss how the particularities of the devices impact our algorithm and implementation decisions. Finally, we provide both qualitative and quantitative results on publicly available RGB-D datasets, and on datasets collected in real-time from two devices.