Fig. 3: Classification performance.

a Log-odds of being a match split by training set and true class membership for matrix − t distributions with 3, 5, 10, 15, 20, and 30 degrees of freedom. A log-odds ratio greater than 0 indicates greater odds of being a match than a non-match. The predictions for each training set are performed on all four sets of fracture surfaces. b Individual true match correlations for three repetitions of topographical imaging of the K-1 set of 9 knives and with 9 images per knife. The similarity among the three distributions demonstrates that similar results will be obtained upon re-imaging the same surface, which is important in forensic applications. The large dots indicate the means of the sets, and we display the covariance matrices through the 99% ellipses of concentration of their distributions.