Paper Info
Title
Real-time indoor scene reconstruction with Manhattan assumption.
Abstract
This paper presents a novel end-to-end system for real-time indoor scene reconstruction, which outperforms traditional image feature point-based method and dense geometry correspondence-based method in handling indoor scenes with less texture and geometry features. In our method, we fully explore the Manhattan assumption, i.e. scenes are majorly consisted with planar surfaces with orthogonal normal directions. Given an input depth frame, we first extract dominant axes coordinates via principle component analysis which involves the orthogonal prior and reduce the influence of noise. Then we calculate the coordinates of dominant planes (such as walls, floor and ceiling) in the coordinates using mean shift. Finally, we compute the camera orientation and reconstruct the scene by proposing a fast scheme based on matching the dominant axes and planes to the previous frame. We have tested our approach on several datasets and demonstrated that it outperforms some well known existing methods in these experiments. The performance of our method is also able to meet the requirement of real-time with an unoptimized CPU implementation.
Year
DOI
Venue
2019
10.1007/s11042-017-5519-7
Multimedia Tools Appl.
Keywords
Field
DocType
SLAM, Tracking, Depth sensor, Real-Time, AR
Computer vision,Computer science,Ceiling (aeronautics),Camera orientation,Planar,Artificial intelligence,Mean-shift,Principal component analysis
Journal
Volume
Issue
ISSN
78
1
1573-7721
Citations 
PageRank 
References 
0
0.34
24
Authors
7
Name
Order
Citations
PageRank
Zunjie Zhu102.03
Feng Xu219423.14
Chenggang Yan341032.87
Ning Li414548.40
Bingjian Gong501.01
Yongdong Zhang62544166.91
Qionghai Dai73904215.66