Next: Noisy data
Up: Experiments
Previous: Experiments
Under ideal conditions, the test objects' surfaces are completely exposed to
the sensor and sensed data are free of noise.
Table 1 shows the achieved recognition rates and times for the six
criteria [cf. Section 4].
The measured times include all steps from drawing feature samples to the
output of the best matching object model.
Generating the surface mesh is not included.
Almost perfect classification has been achieved by the and
criteria.
Interestingly, the criterion performs dramatically weaker than the
criterion.
Apparently, the weighting of histogram differences by the reciprocal of the
trained histogram value alone is much more reliable than taking also the
estimate from the small test sample into account [cf. Equations (13),
(14)].
Correct classification and confusion rates between all pairs of objects are
shown in Figure 3.
All classifiers work well for simple shapes like cube or sphere.
Interestingly, the objects that are difficult to classify differ drastically
across the criteria. On the other hand, the and criteria
exhibit a strikingly similar pattern of classification performance.
This similarity will also be retained in all the other tests of the classifiers
we report below. The same similarity holds for the best, the and
criteria.
Next: Noisy data
Up: Experiments
Previous: Experiments
Eric Wahl
2003-11-06