14 hours ago
- Eye tracking is a vital tool for many domains but is not yet pervasive.
- The goal is to enable eye tracking on commodity hardware like mobile phones without extra sensors.
- GazeCapture is the first large-scale eye tracking dataset with over 1,450 people and 2.5 million frames.
- iTracker, a convolutional neural network, achieves significant error reduction and runs in real-time (10–15 fps) on mobile devices.
- Prediction error is 1.7 cm (phones) and 2.5 cm (tablets) without calibration; improved to 1.3 cm and 2.1 cm with calibration.
- iTracker's learned features generalize to other datasets, achieving state-of-the-art results.