Boosting speed- and accuracy of gradient based dark pupil tracking using vectorization and differential evolution

Abstract

Gradient based dark pupil tracking [Timm and Barth 2011] is a simple and robust algorithm for pupil center estimation. The algorithm's time complexity of O(n4) can be tackled by applying a two-stage process (coarse center estimation followed by a windowed refinement), as well as by optimizing and parallelizing code using cache-friendly data structures, vector-extensions of modern CPU's and GPU acceleration. We could achieve a substantial speed up compared to a non-optimized implementation: 12x using vector extensions and 65x using a GPU. Further, the two-stage process combined with parameter optimization using differential evolution considerably increased the accuracy of the algorithm. We evaluated our implementation using the "Labelled pupils the wild" data set. The percentage of frames with a pixel error below 15px increased from 28% to 72%, surpassing algorithmically more complex algorithms like ExCuse (64%) and catching up with recent algorithms like PuRe (87%).