In our technical demo, we constructed four separate facial recognition algorithms of varying merit: eigenfaces, fisherfaces, local binary pattern histograms, and convolutional neural networks. These four implementations have lots of existing documentation behind them regarding accuracy, and our general goal for this demo was to evaluate the relative accuracy of different facial recognition implementations on multiple different datasets. We also wanted to examine the extent of bias in facial recognition systems by evaluating the different implementations’ performance on a dataset that was divided by race and gender, allowing us to see how each implementations’ accuracy varies with respect to differences in demographic characteristics (which is a key research question for our project). Not only will our technical demo provide valuable information that could be cited in our own research project’s findings, but our technical demo will hopefully provide another useful benchmark for others to cite regarding the known accuracy differences between facial recognition implementations and the known biases of existing facial recognition systems.
Our full written technical report can be found here