SacCalib: Straightening Out Eye Tracking Drift
Eye tracking is a powerful tool for everything from gaming to medical diagnostics, but it has a frustrating "drift" problem. Even if you calibrate a stationary tracker perfectly at the start of a session, your accuracy will inevitably degrade as you shift in your seat or change your head pose.
In their paper "SacCalib: Reducing Calibration Distortion for Stationary Eye Trackers Using Saccadic Eye Movements," Michael Xuelin Huang and Andreas Bulling introduce a clever way to fix this drift in the background, using nothing but the physics of the eye itself.
TL;DR
SacCalib is the first "eye-only" method to reduce calibration distortion in stationary eye trackers without requiring mouse clicks or knowledge of on-screen content. By exploiting the fact that saccades (fast eye jumps) are naturally straight but appear curved when a tracker is miscalibrated, the system uses image warping to "straighten" these trajectories and restore accuracy.
What Problem Does This Paper Address?
Stationary eye trackers rely on a "gaze projection plane"—a mathematical mapping that translates where the camera sees your eye looking into a coordinate on the screen. This mapping is fragile. If you move your head slightly or the tracker shifts, the projection plane becomes distorted.
Existing "self-calibration" methods try to fix this by looking for "ground truth" events, like a mouse click (assuming the user is looking at the cursor) or visual saliency maps (assuming the user is looking at the most interesting thing on screen). However, mouse clicks are sparse, and saliency maps are computationally expensive and often wrong. SacCalib addresses the need for a non-intrusive, content-agnostic way to maintain accuracy during long-term use.
Methods
The core insight of SacCalib is geometric: Saccades are physically straight. When you jump your gaze from point A to point B, your eye moves in a nearly straight line. If an eye tracker records that movement as a curve, it’s a "smoking gun" for calibration distortion.
The SacCalib pipeline works in three main stages:
- Segmentation: It identifies saccades using a velocity-threshold algorithm (I-VT).
- Saccade Selection: Not all saccades are created equal. Some are naturally curved due to human physiology (especially vertical or oblique ones). The authors trained a Random Forest classifier using 10 attributes (like amplitude, velocity, and duration) to pick the "ideal" saccades that should have been straight.
- Warping for Undistortion: Once the "bad" curves are identified, the system uses a technique called Moving Least Squares (MLS). It treats the gaze projection plane like a flexible sheet and warps it until the curved saccades become straight lines. Specifically, they found that projecting the curve toward a "middle" point between the chord and the peak of the curve yielded the best results.
Datasets
The authors collected a new dataset specifically for this task:
- Participants: 10 individuals with diverse backgrounds.
- Data: Approximately 3,000 on-screen saccades.
- Setup: Users followed a "shrinking circle" on a 5x5 grid. To simulate real-world drift, the authors recorded three sessions per user, manually moving the chin rest between sessions to create "Small" and "Large" head pose variations.
Key Insights and Ideas
- The "Straightness" Prior: The paper turns a biological constraint into a mathematical advantage. By assuming the "intent" of a saccade is a straight line, they create a self-correcting mechanism that doesn't need to know what the user is looking at.
- Selection is Everything: You can't just straighten every line. The authors show that using random saccades actually hurts accuracy. The Random Forest filter is essential because it weeds out saccades that are naturally curved by the brain's motor control.
- Systematic vs. Random Error: The curving effect caused by miscalibration isn't random noise; it’s a systematic distortion of the projection plane. This allows a global warping fix to improve accuracy across the entire screen.
Key Results
- Accuracy Boost: SacCalib achieved a 2.5% overall improvement in accuracy, pushing the average error from 1.79° down to 1.75°.
- Handling Drift: The method was most effective in the "Small" and "Large" head pose variation scenarios—exactly where traditional calibration fails.
- Spatial Success: The improvement was particularly strong on the right side of the screen, suggesting the method effectively handles regional distortions.
- Selection Matters: Using the Random Forest to select saccades outperformed using "all" available saccades, proving that physiological noise must be filtered out before warping.
Limitations and Future Work
- Preparation Time: In the experiment, users had one second to "prepare" for a saccade, which helps ensure straightness. In wild, rapid-fire UI interaction, saccades might be naturally messier.
- Ocular Dominance: The study only looked at the left eye. Because of ocular dominance, the distortion patterns might differ between the left and right eyes, requiring a binocular approach.
- Real-Time Integration: While the math is efficient, the authors suggest future work is needed to integrate this into a truly seamless, real-time "background" service for consumer eye trackers.
- Microsaccades: The authors suggest that even tiny eye movements (microsaccades) could potentially be used for even finer-grained calibration in the future.
Summary updated