Sure About That Line? Approaching Confidence-Based, Real-Time Line Assignment in Reading Gaze Data
Low-latency eye tracking algorithm works even when readers skip back lines.
A team led by Franziska Kaltenberger (TU Munich) has published CONF-LA (Confidence-score-based Online Fixation-to-Line Assignment), a novel algorithm for real-time line tracking in multi-line reading from remote or webcam-based eye trackers. Traditional approaches either process gaze data after the fact or restrict natural reading behaviors like regressions (re-reading previous lines). CONF-LA integrates knowledge of reading behavior with Gaussian line likelihoods over fixations to compute a posterior-line-score, deferring assignment when uncertainty is high.
Evaluated on open-source datasets, CONF-LA achieves a mean per-fixation latency of just 0.348 ms—fast enough for real-time applications—while closing the online-offline accuracy gap to only 1-2%. Notably, it shows particular invariance to regressions, achieving ad hoc median accuracies of approximately 95% on children's reading data, significantly outperforming existing algorithms. Accepted at ETRA 2026, this work opens the door to interactive reading aids that work with basic webcams, potentially helping children and adults with reading difficulties.
- CONF-LA assigns gaze fixations to lines in multi-line reading with 0.348 ms latency.
- Closes the online-offline accuracy gap to just 1-2% using confidence-based deferral.
- Achieves ~95% median accuracy on children's data even with regressions (re-reading).
Why It Matters
Enables real-time reading support from webcams, aiding children and those with reading difficulties.