PDF version of this entire document

Making a More Stable Classifier

With 3-D only (no texture) methods in mind, the pursuit for a stable classifier might be a combination of geodesic, geodesic-Euclidean, and purely Euclidean criteria (not photometric, but surface-based only). In addition, several PCA-based criteria are available, but they have not been combined yet as they measure range images or their derivatives very sparsely (the curse of dimensionality in PCA limits this considerably and makes it less pragmatic). We are combining these different criteria, excepting the weaker ones that do not help much in determining the outcome. But one that is explored today, following preliminary experiments that showed some merit, is one that gradually expands the surface around the eyes and the nose, then measuring Euclidean distances on these gradually-expanding boundaries. In a sense, this is the gradual measurement - in a 10-step process right about now - which accumulates distances by traversing the triangulated surface with the expectation that identical faces will give similar distance differences as the geodesic circles grow bigger and bigger (or conversely, smaller and smaller as it is currently implemented).

Results will be shown in terms of some ROC curves. This is slow enough to take hours for one ROC curve, especially because the code is a lot more inefficient than it can be (if optimised and polished a little).