next up previous contents
Next: Two Indoor Scenes Up: Experimental Results Previous: Using the Edge Distribution

A Larger Scene

  figure969
Figure 6.9: Scene II.

In this section we consider a slightly more complicated scene, a larger configuration-space and a lower sampling density (or higher sample spacing). Scene II is depicted in Figure 6.9. As in Scene I, the camera is mounted on the end-effector of a gantry-mounted robot arm. The camera faces the scene at a fixed orientation from a distance of about 1.0m, and 256 training images are collected at 2.0cm intervals (tex2html_wrap_inline4758=2.0cm) over a 30cm by 30cm grid. 100 images are taken from random poses as test subjects. As in Scene I, the x-axis lies in the image plane of the camera pointing to the right, and is parallel to the horizon. The y-axis is perpendicular to the image plane of the camera, pointed in the direction that the camera faces.

Figure 6.10 demonstrates the accuracy of the method for tex2html_wrap_inline4890 and tex2html_wrap_inline4834. The mean error in position is 0.38cm or 19% of tex2html_wrap_inline4758.

  figure978
Figure 6.11: Scene II pose estimates for 100 test cases, tex2html_wrap_inline4890 and tex2html_wrap_inline4876. The mean estimation error is 0.38cm

In the case of appearance-based estimation, Figure 6.11 demonstrates the accuracy of the method for tex2html_wrap_inline4862 and tex2html_wrap_inline4876. The mean error in position is 0.8cm, or 40% of tex2html_wrap_inline4758.

  figure988
Figure: Scene II pose estimates for 100 test cases, tex2html_wrap_inline4862 and tex2html_wrap_inline4876. The mean estimation error is 0.8cm



Robert Sim
Tue Jul 21 10:30:54 EDT 1998