Generalization across mesh resolution

Since we have relied upon surface meshes as the input representation, it is interesting to ask how recognition performance is affected by changes to the mesh procedure. The most demanding scenario is generalization across mesh procedures, that is, being confronted at recognition time with a mesh of a type essentially different from what training has been based on.

In a final set of experiments, we thus have investigated the effect of varying the mesh resolution for the test objects. Figure 5(c) shows plots of correct-classification rates under such conditions, where mesh resolution is given in percent of the (constant) resolution in the training phase. Apparently, recognition performance does not critically depend on test-mesh resolution. Only below 50% of the training resolution, recognition performance drops off. In part, this can be ascribed to the low absolute number of feature samples drawn. In particular, the ${\cal {K}}$ and ${\cal {L}}$ criteria exhibit a high degree of generalization across meshes.