How Many Samples Does Convincible Performance Evaluation of a Biometric System Need?

Rongfeng Li, Darun Tang, Beining Huang, Wenxin Li

Abstract: Before a biometrics system is implied, we evaluated its performance by testing data. While a test dataset, instead of the universal dataset, is used, the performance indices cannot be exactly accurate. Previous researchers have attempted, based on the estimation of the standard deviation, to determine the size of test dataset. Apparently, the larger test sample size is, the more convincible performance estimation will come out. However the test data cannot be as many as we want, considering of the costs of collecting data. In this paper, we firstly defined a cost function of collecting test data, and secondly a new second-level partition bootstrap was implied aiming to predict the standard deviation of universal set using test datasets. Then we determined the test sample size using the cost function. Lastly, the experiment validated that the bootstrap prediction was accurate enough. For a 10150-finger finger-vein database, we need 2954 test samples.