An accurate method for 3D object reconstruction from unordered sparse views

Abstract

3D object reconstruction from depth scans has been long investigated problem, but complete solutions are still missing. Recent advances of consumer depth sensors resulted in a number of excellent methods for 3D object reconstruction, with the most impressive results obtained by KinectFusion and Visual Odometry. Both methods reconstruct an object from a dense set of RGB-D video frames, with the important restrictions that the object has to be static and camera transitions has to be smooth during the reconstruction. As an alternative, we propose a method which is more accurate, while being able to reconstruct the object from a sparse set of unordered depth scans. Our method starts by pairwise view matching which relies on robust point pair features. It is followed by the efficient graph-based view consistency check to produce an initial 3D model reconstruction. To improve the accuracy, we perform a simultaneous optimization of the camera positions and the object model, using depth and color information. A quantitative evaluation demonstrates that we can reconstruct objects with equal or better accuracy than the KinectFusion and Visual Odometry, but using significantly lower number of unordered input scans.