Improving Initial Estimations for Structure from Motion Methods

In proceedings of The 13th Central European Seminar on Computer Graphics (CESCG 2009), Apr. 2009
Präsentiert: The 13th Central European Seminar on Computer Graphics (CESCG 2009)
 

Abstract

In Computer Graphics as well as in Computer Vision and Autonomous Navigation, Structure from Motion is a common method to register cameras. Usually several steps are involved with bundle-adjustment as the final one. A good initial estimation of camera positions is of crucial importance for the success of the bundle-adjustment and is the core component of any Structure from Motion system. Yet there are some limitations to current Structure from Motion tools regarding the quality of the initial estimation. With our proposed method of merging different connected components resulting from a lack of good input images, we aim to overcome the fact that at first glance no global initial estimation could be found. We will show that in many of these situations our method is applicable and may even be used to speed up the Structure from Motion process and limit its memory consumption in general.

Stichwörter: 3D scene analysis, Reconstruction, structure from motion

Bilder

Paper herunterladen

Paper herunterladen

Zusätzliches Material

Bibtex

@INPROCEEDINGS{schwartz-2009-improvesfm,
      author = {Schwartz, Christopher and Klein, Reinhard},
       title = {Improving Initial Estimations for Structure from Motion Methods},
   booktitle = {The 13th Central European Seminar on Computer Graphics (CESCG 2009)},
        year = {2009},
       month = apr,
    keywords = {3D scene analysis, Reconstruction, structure from motion},
    abstract = {In Computer Graphics as well as in Computer Vision and Autonomous Navigation,
                Structure from Motion is a common method to register cameras. Usually several
                steps are involved with bundle-adjustment as the final one.
                A good initial estimation of camera positions is of crucial importance for the success
                of the bundle-adjustment and is the core component of any Structure from Motion
                system. Yet there are some limitations to current Structure from Motion tools
                regarding the quality of the initial estimation. With our proposed method
                of merging different connected components resulting from a lack of good input images,
                we aim to overcome the fact that at first glance no global initial estimation could be found.
                We will show that in many of these situations our method is applicable and may even
                be used to speed up the Structure from Motion process and limit its memory consumption
                in general.},
  conference = {The 13th Central European Seminar on Computer Graphics (CESCG 2009)}
}