Expected Confusion as a Method of Evaluating Recognition Techniques

We derive an expected confusion metric, as opposed to reporting percent correct with a limited database, as a method to evaluate recognition techniques. This metric allows us to predict how well a given feature vector will filter identity in a large population. Our expected confusion is the ratio of the average individual variation of a feature vector to that of the population variation of the feature vector. We evaluate our gait-recognition technique [2] that recovers static body and stride parameters of walking subjects with the expected confusion metric to demonstrate its use.