Single-camera, calibration-free gaze estimation using corneal reflections

Background

Mice make substantial eye movements during head-fixed visual stimulation, and uncorrected gaze shifts corrupt receptive field measurements and confound stimulus-response relationships. Corneal-reflection video oculography in rodents has provided the methodological foundation for calibrated angular gaze tracking since Stahl (2004) but the calibration procedures used by existing methods — physical camera rotation, motorized stages, behavioral tasks, or precisely co-aligned dual cameras — have limited their adoption in many mouse neuroscience laboratories. Most studies instead use uncalibrated pupil tracking, deep learning pose estimation that returns pixel coordinates without angular calibration, or learned shifter networks that lack independent validation.

New method

We present an open-source corneal-reflection eye tracking system for head-fixed mice with two methodological contributions. First, a geometric model recovers gaze in calibrated angular units from the pixel displacements of the pupil and corneal reflections, using the known 3D positions of multiple fiducial LEDs as the source of angular scale. The model requires no estimate of R _p , the per-animal eye-geometry parameter that earlier corneal-reflection methods determine through physical calibration. Second, a self-calibration procedure exploits the redundancy of multiple stationary fiducial LEDs: each LED produces an independent gaze estimate from the same geometric model, and a single residual calibration parameter is determined by minimizing the disagreement between per-LED estimates. This replaces the physical camera-rotation calibrations of earlier video oculography (Sakatani and Isa, 2004, 2007; van Alphen et al., 2013; Kretschmer et al., 2017), the motorized stages of Zoccolan et al. (2010), and the precision dual-camera alignment of Payne and Raymond (2017) with a software operation that requires no moving parts, no behavioral task, and no per-animal procedure. The system provides three interactive GUI stages: (1) pupil and LED detection via Difference-of-Gaussians filtering, (2) 3D geometry definition, and (3) gaze angle computation with blink detection, fiducial correction, and manual curation.

Results

Validation against a rotary-encoder-controlled artificial eye demonstrated mean absolute errors below 1° across all four fiducial LEDs over the ±20° working range of mouse eye movements, with Pearson correlations exceeding 0.998 between our method’s estimation and encoder ground truth. The self-calibration reduced inter-LED disagreement by a factor of 4–6 in mouse recordings. Gaze-corrected stimulus reconstruction applied to Neuropixels recordings from mouse V1 produced qualitatively sharper receptive field estimates with improved signal-to-noise ratios.

Comparison with existing methods

Our method is the first multi-LED, single-camera, fully software-calibrated corneal-reflection eye tracker for mice and includes an integrated open-source pipeline for detection, calibration, blink handling, and artifact correction. The multi-LED redundancy doubles as an internal consistency check — if two LEDs disagree on gaze direction, the calibration is wrong — providing a guarantee that learned approaches relying on neural-data-derived correction cannot offer.

Conclusions

Our method makes calibrated corneal-reflection eye tracking accessible to non-specialist mouse laboratories using consumer-grade hardware (∼ $2,000–2,700 USD), eliminates the per-animal calibration procedures of earlier methods, and is validated by two independent ground truths at both the absolute angular (artificial eye) and functional (V1 receptive fields) levels.

Highlights

• Open-source corneal-reflection eye tracking for head-fixed mice using a single camera and multiple stationary fiducial LEDs.

• Geometric gaze model derives angular scale from LED positions, eliminating per-animal eye-geometry calibration.

• Self-calibration via multi-LED redundancy replaces physical camera rotation, motorized stages, and dual-camera precision alignment.

• Validated to sub-degree accuracy against a rotary-encoder ground truth across the ± 20° range of mouse eye movements.

• Gaze correction produces sharper V1 receptive field estimates in Neuropixels recordings.