Novel Analytical Models of Face Recognition Accuracy in Terms of Video Capturing and Encoding Parameters

To fit the tight resource constraints, including network bandwidth, the video streams in Computer Vision systems are adapted dynamically by changing the video capturing and encoding parameters. We propose two novel analytical models that characterize the face recognition accuracy in terms of these parameters, specifically resolution, quantization, and actual bitrate. We find that the accuracy is a logistic function of the video quantization parameter, with the value of the Sigmoid's midpoint being a function of the resolution. Alternatively, we find that the accuracy is equal to the sum of two exponentials of the actual video bitrate, with the resolution as a multiplicative factor with one exponential. We develop an evaluation framework to validate the models using two distinct video datasets with 99 videos and the widely used Labeled Faces in the Wild (LFW) dataset with 13, 233 images. We conduct 1,668 experiments that involve varying combinations of encoding parameters. We show that both models hold true for the deep-learning and statistical-based face recognition. The developed models achieve an average coefficient of determination (R
²
) of 98.7% to 99.8%.