A memory-efficient kinectfusion using octree

  • Authors:

  • Ming Zeng;Fukai Zhao;Jiaxiang Zheng;Xinguo Liu

  • Affiliations:

  • State Key Lab of CAD&CG, Zhejiang University, Hangzhou, China;State Key Lab of CAD&CG, Zhejiang University, Hangzhou, China;State Key Lab of CAD&CG, Zhejiang University, Hangzhou, China;State Key Lab of CAD&CG, Zhejiang University, Hangzhou, China

  • Venue:
  • CVM'12 Proceedings of the First international conference on Computational Visual Media
  • Year:
  • 2012

Quantified Score

Hi-index 0.00

Visualization

Abstract

KinectFusion is a real time 3D reconstruction system based on a low-cost moving depth camera and commodity graphics hardware. It represents the reconstructed surface as a signed distance function, and stores it in uniform volumetric grids. Though the uniform grid representation has advantages for parallel computation on GPU, it requires a huge amount of GPU memory. This paper presents a memory-efficient implementation of KinectFusion. The basic idea is to design an octree-based data structure on GPU, and store the signed distance function on data nodes. Based on the octree structure, we redesign reconstruction update and surface prediction to highly utilize parallelism of GPU. In the reconstruction update step, we first perform "add nodes" operations in a level-order manner, and then update the signed distance function. In the surface prediction step, we adopt a top-down ray tracing method to estimate the surface of the scene. In our experiments, our method costs less than 10% memory of KinectFusion while still being fast. Consequently, our method can reconstruct scenes 8 times larger than the original KinectFusion on the same hardware setup.