• PDF / 2,290,138 Bytes
• 17 Pages / 439.37 x 666.142 pts Page_size
• 40 Downloads / 196 Views

DOWNLOAD

REPORT

---

Technische Universit¨ at M¨ unchen, Munich, Germany {mira.slavcheva,nassir.navab}@tum.de, {kehl,slobodan.ilic}@in.tum.de 2 Siemens AG, Munich, Germany Abstract. This paper addresses the problem of 3D object reconstruction using RGB-D sensors. Our main contribution is a novel implicit-to-implicit surface registration scheme between signed distance fields (SDFs), utilized both for the real-time frame-to-frame camera tracking and for the subsequent global optimization. SDF-2-SDF registration circumvents expensive correspondence search and allows for incorporation of multiple geometric constraints without any dependence on texture, yielding highly accurate 3D models. An extensive quantitative evaluation on real and synthetic data demonstrates improved tracking and higher fidelity reconstructions than a variety of state-of-the-art methods. We make our data publicly available, creating the first object reconstruction dataset to include ground-truth CAD models and RGB-D sequences from sensors of various quality. Keywords: Object reconstruction · Signed distance field · RGB-D sensors

1

Introduction

The persistent progress in RGB-D sensor technology has prompted exceptional focus on 3D object reconstruction. A key research goal is recovering the geometry of a static target from a moving depth camera. This entails estimating the device motion and fusing the acquired range images into consistent 3D models. Depending on the particular task, methods differ in their speed, accuracy and generality. Most existing solutions are SLAM-like, thus their applications lie in the field of robotic navigation where precise reconstructions are of secondary importance. In contrast, the growing markets of 3D printing, reverse engineering, industrial design, and object inspection require rapid prototyping of high quality models, which is the aim of our system. One of the most influential works capable of real-time reconstruction is KinectFusion [12,24]. It conveniently stores the recovered geometry in an incrementally built signed distance field (SDF). However, its frame-to-model camera tracking via iterative closest points (ICP [1,6]) limits it to objects with distinct Electronic supplementary material The online version of this chapter (doi:10. 1007/978-3-319-46448-0 41) contains supplementary material, which is available to authorized users. c Springer International Publishing AG 2016  B. Leibe et al. (Eds.): ECCV 2016, Part I, LNCS 9905, pp. 680–696, 2016. DOI: 10.1007/978-3-319-46448-0 41

SDF-2-SDF: Highly Accurate 3D Object Reconstruction

681

Synthetic data

Industrial sensor

Kinect

Fig. 1. SDF-2-SDF reconstructions of the proposed dataset objects. Colors vary due to difference between synthetic rendering and 3D-printed models, and camera radiometrics (Color figure online)

appearance and to uniform scanning trajectories. Other techniques use a pointto-implicit scheme [3,5] that avoids explicit correspondence search by directly aligning the point clouds of incoming depth frames with the growing SDF. Such registration has proven to be more