All experiments are based upon the 20 objects shown in Figure 2. The objects are initially given as surface meshes, which are, however, unrelated to the meshes we use as inputs to our algorithm. To ensure that classification cannot be dominated by object size, all objects are scaled to a common maximal diameter.
Models are trained by the following procedure. For each object, five sets of points, from 25,000 to 389,000 points per set, are drawn randomly from the surface and passed to a mesh generator. A training mesh consists of between 3,500 and 5,500 vertices. Features are built from pairs of surflets, which are in turn picked from each vertex. All features obtained from the five training meshes, that is, between 30,616,250 and 75,611,250 features, are collected into a histogram [cf. Section 3].
In the recognition phase, new meshes are generated from each object. Features are randomly subsampled from the vertices of these meshes. The number of features drawn is 0.005% of all available features. This arbitrary, low sampling rate turns out to be high enough for good recognition. Results presented on classification rate and timing are averaged over between 100 and 1000 meshes per test object.