Introduction
Eye tracking technology is increasingly vital across diverse fields, from immersive gaming to clinical diagnostics. However, stationary eye trackers suffer from a persistent "drift" problem. Even with perfect initial calibration, accuracy degrades as users shift their seating position or change their head pose. Michael Xuelin Huang and Andreas Bulling address this challenge in their paper, "SacCalib: Reducing Calibration Distortion for Stationary Eye Trackers Using Saccadic Eye Movements," by introducing a background correction method based on the intrinsic physics of eye movements.
The Problem: Calibration Drift
Stationary eye trackers utilize a "gaze projection plane"—a mathematical mapping that translates camera-captured eye positions into screen coordinates. This mapping is highly sensitive; minor head movements or tracker shifts distort the projection plane, leading to significant gaze estimation errors.
Traditional self-calibration methods often rely on external "ground truth" events, such as mouse clicks or visual saliency maps. However, these are often sparse, computationally intensive, or inaccurate. SacCalib offers a non-intrusive, content-agnostic alternative that maintains accuracy without requiring user interaction or knowledge of on-screen content.
Methodology
The fundamental insight of SacCalib is geometric: saccades (rapid eye movements between fixation points) are physically straight. When a miscalibrated eye tracker records these movements as curves, it indicates systematic distortion in the projection plane.
The SacCalib pipeline consists of three primary stages:
- Segmentation: Saccades are identified using a velocity-threshold algorithm (I-VT).
- Saccade Selection: Since some saccades (e.g., vertical or oblique) are naturally curved due to physiology, a Random Forest classifier filters for "ideal" saccades that should be straight.
- Warping for Undistortion: The system employs Moving Least Squares (MLS) to warp the gaze projection plane, effectively "straightening" the identified curved saccades.
Warping Strategies
The authors explored different methods for projecting curved trajectories back into straight lines. They compared three strategies: projecting toward the "Bottom" (chord), the "Peak" of the curve, or a "Middle" point between them.
Empirical results showed that the "Middle" strategy yielded the highest accuracy in restoring the projection plane.
Experimental Setup and Results
The authors evaluated SacCalib using a new dataset of approximately 3,000 on-screen saccades from 10 participants. To simulate real-world drift, sessions included "Small" and "Large" head pose variations.
Key findings include:
- Accuracy Improvement: SacCalib reduced average error from 1.79° to 1.75°, a 2.5% overall improvement.
- Drift Mitigation: The method was most effective in scenarios with significant head pose variation, where traditional calibration typically fails.
- The Importance of Selection: Using a Random Forest to select saccades significantly outperformed using all available saccades, highlighting the need to filter out physiological noise.
Conclusion and Future Work
SacCalib demonstrates that biological constraints can be leveraged as mathematical priors for self-correcting systems. While the current implementation shows promise, future work will focus on real-time integration, accounting for ocular dominance in binocular tracking, and exploring the use of microsaccades for even finer calibration adjustments.
Summary updated