Automated image-based 3D reconstruction system for precise plant modeling

Xu, Yiwen

Permalink

https://hdl.handle.net/2142/97704 Copy

Description

Title

Automated image-based 3D reconstruction system for precise plant modeling

Author(s)

Xu, Yiwen

Issue Date

2017-04-18

Director of Research (if dissertation) or Advisor (if thesis)

Hart, John C.

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• Plant modeling
• Object segmentation
• Image-based modeling
• Image reconstruction

Abstract

Plant modeling has been an interesting topic in computer graphics for decades because of the growing demand for realistic plants in video games or movies. It also attracts researchers in crop science to digitize real plants for scientific research. However, a plant is a hard object to model due to its irregularity in shapes and complexity in structure. In this thesis, we present an automated non-parametric system, which reconstructs a high-quality 3D dense point cloud for plants from video inputs that are taken in a casual setting. The input data can be easily collected using an ordinary hand-held camera or phone camera in a daily environmental setting, by slowly walking around the plant. And the resulting dense point cloud is comparable to the original image when reprojected back. The system can run without any additional inputs while also providing users with flexibility for advanced settings. To build such a system, existing systems for plant modeling lead us to image-based modeling methods. Under the context of image-based modeling, related components such as Structure from Motion(SFM), Multi-View Stereo(MVS) or 3D Object Segmentation, are explicitly analyzed with preliminary experiments that were performed individually on the plant objects. The results show that simple combination of existing work does not generate satisfying results when modeling a plant. To solve this problem, we proposed a non-parametric gentle segmentation algorithm based on a novel combination of these approaches that can preserve the detailed structure of plants as much as possible while cleaning the noisy point cloud that ordinarily results, yielding a result tending towards a photo-consistent level when compared to the original images.

Graduation Semester

2017-05

Type of Resource

text

Permalink

http://hdl.handle.net/2142/97704

Copyright and License Information

Copyright 2017 Yiwen Xu

Owning Collections