Dense range images from sparse point clouds using multi-scale processing

Multi-modal data processing based on visual and depth/range images has become relevant in computer vision for 3D reconstruction applications such as city modeling. robot navigation etc. In this paper, we generate high-accuracy dense range images from sparse point clouds to facilitate such applications. Our proposal addresses the problem of sparse data, mixed-pixels at the discontinuities and occlusions by combining multi-scale range images. The visual results show that our algorithm can create high-resolution dense range images with sharp discontinuities, while preserving the topology of objects even for environments that contain occlusions. To demonstrate the effectiveness of our approach, we propose an iterative perspective-to-point algorithm that aligns the edges between the color image and the range image from various viewpoints. The experimental results from 46 viewpoints show that the camera pose can be corrected when using high-accuracy dense range images, so that 3D reconstruction or 3D rendering can obtain a clearly higher quality.