A Multi-View Stereo Algorithm Based on Homogeneous Direct Spatial Expansion with Improved Reconstruction Accuracy and Completeness

Yalan Li,Zhiyang Li

doi:10.3390/app7050446

Abstract

Reconstruction of 3D structures from multiple 2D images has wide applications in such fields as computer vision, cultural heritage preservation, etc. This paper presents a novel multi-view stereo algorithm based on homogeneous direct spatial expansion (MVS-HDSE) with high reconstruction accuracy and completeness. It adopts many unique measures in each step of reconstruction, including initial seed point extraction using the DAISY descriptor to increase the number of initial sparse seed points, homogeneous direct spatial expansion to enhance efficiency, initial value modification via a conditional-double-surface-fitting method before optimization and adaptive consistency filtering after optimization to ensure high accuracy, processing using a multi-level image pyramid to further improve completeness and efficiency, etc. As demonstrated by experiments, owing to above measures the proposed algorithm attained much improved reconstruction completeness and accuracy.

Highlights

Multi-view 3D reconstruction can be traced back to binocular stereopsis based on window matching and triangulation
When it was extended to multi-view situations, it was adapted to feature-matching and triangulation, giving rise to Structure From Motion (SFM) [4], which is widely used today for extraction of sparse seed points and camera calibration
We extended the framework of feature expansion algorithm by an additional step for initial value modification and made important improvements to other existing steps, which is why our algorithm outperforms Patch-based Multi-View Stereopsis (PMVS)

Summary

Introduction

Multi-view 3D reconstruction can be traced back to binocular stereopsis based on window matching and triangulation. This window-based matching produces reliable results only for texture-rich regions. When it was extended to multi-view situations, it was adapted to feature-matching and triangulation, giving rise to Structure From Motion (SFM) [4], which is widely used today for extraction of sparse seed points and camera calibration. To reconstruct the remaining regions and cope with difficulties like occlusion, non-Lambert illumination, photo noise, etc., many kinds of algorithm have been developed. They could be roughly classified into volume-based algorithms, depth-map fusion algorithms and feature-expansion algorithms.

Methods

Results

Conclusion