Improving RGB-D-based 3D reconstruction by combining voxels and points.

We propose a flexible 3D reconstruction method based on the RGB-D data stream. Compared to previous methods using pure voxels or pure points as representations, our works propose a new representation combining voxels and points to improve the reconstruction accuracy. A key insight is that points can store additional depth data that are not sampled by regular voxels. Thus, by integrating points and voxels, the 3D reconstruction process can be accelerated due to higher data utilization. Furthermore, depth information stored in points is used to refine the noisy depth image through a depth image refinement method, consequently improving the reconstructed shape quality. Extensive comparative experiments are performed including different representations (pure voxels/points) and various methods (fusion-based/learning-based and online/offline) to illustrate the effectiveness of our work. Experimental results demonstrate that our method can achieve real-time performance, effectively avoid artifacts, and reach state-of-the-art accuracy levels. More importantly, we provide a novel idea to balance the conflict between memory overhead and reconstruction accuracy. [ABSTRACT FROM AUTHOR]