Next: Generalization across mesh resolution
Up: Experiments
Previous: Noisy data
In real applications, objects to be recognized are often just partially
visible.
Reasons are self-occlusion in single-view data or occlusions by other objects.
Partial objects yield incomplete surface meshes.
Therefore, in this set of experiments, each test object is meshed and
classified with varying fraction of visible surface.
Visible parts are determined by intersecting point clouds by a random plane.
Subsequently, data on one side of the plane are processed by the mesh generator.
Visibility is defined as the sum of remaining triangle areas in percent of the
complete surface area.
Figure 4(c) gives an example of a partially visible mesh (33%).
Results on recognition rates for various visibilities are plotted
in Figure 5(b).
Performance can be seen to drop off more gradually with occlusion than with data
corruption by noise [cf. Figure 5(a)].
Correct classification by the and criteria
remains above 80% down to roughly 65% visibility.
We note that recognition with partial visibility depends in fact heavily on the
particular section of the object that remains visible.
Eric Wahl
2003-11-06