Tiny convolution, decision tree, and binary neuronal networks for robust and real time pupil outline estimation

Wolfgang Fuhl

Abstract:

In this work, we compare the use of convolution, binary, and de-
cision tree layers in neural networks for the estimation of pupil
landmarks. These landmarks are used for the computation of the
pupil ellipse and have proven to be effective in previous research.
The evaluated structure of the neural networks is the same for all
layers and as small as possible to ensure a real-time application.
The evaluations include the accuracy of the ellipse determination
based on the Jaccard Index and the pupil center. Furthermore, the
CPU runtime is considered to make statements about the real-time
usability. The trained models are also optimized using pruning to
improve the runtime. These optimized nets are also evaluated with
respect to the Jaccard index and the accuracy of the pupil center
estimation. Link to the framework and models.