Recognition phase

The goal of the recognition phase is to obtain feature distributions from sensed objects, to compare them with the model database, and to find the closest match. In order to avoid excessive computation time during recognition, we draw only a tiny subset (0.005%) of all available features $\cal {S}$.

Not the whole range of feature parameters $S=(\alpha, \beta, \gamma, \delta)$ is necessarily covered by every subsample, or even full sample of features from an object. Hence, some bins of a histogram may remain zero. This leads to numerical problems when computing divisions or logarithms. In such cases, all zero bins of a histogram are set to a common value, lower than the lowest non-zero value occurring in all histograms. This value has the effect of a penalty term.

In this section, $H_O$ denotes the histogram of an object model $O$ from the database. We define six different criteria that we will evaluate for their classification performance. Five of them are based on comparison of $H_O$ to a histogram $H_{O'}$ that is built at recognition time from a test object $O'$. One implements the maximum-likelihood classifier.