Estimating Gaze Depth Using Multi-Layer Perceptron

Y. Lee , C. Shin , A. Plopski , Y. Itoh , T. Piumsomboon , A. Dey , G. A. Lee , S. Kim , M. Billinghurst

ISUVR

Head-mounted eye trackers measure gaze direction well but struggle with depth. This paper feeds the gaze vector from a binocular, two-camera tracker into a multi-layer perceptron neural network to estimate how far away a user is looking. Trained on data from 13 people fixating targets 1–5 m away, it achieved roughly 10% classification error and an average error of about 0.42 m — enough for some AR applications, though estimating gaze in continuous space needs further work.

@inproceedings{lee2017,
  author    = {Y. Lee and C. Shin and A. Plopski and Y. Itoh and T. Piumsomboon and A. Dey and G. A. Lee and S. Kim and M. Billinghurst},
  title     = {Estimating Gaze Depth Using Multi-Layer Perceptron},
  booktitle = {ISUVR},
  pages     = {26-29},
  doi       = {10.1109/ISUVR.2017.13},
  year      = {2017}
}