Aleksandrs Koselevs¹, Veronica Pisu¹, Ömer Yıldıran^2,3, Chloe Lam¹, Saivydas Villani¹, Pascal Mamassian², Dominik Straub^4,5, Constantin Rothkopf⁴, Guido Maiello¹; ¹University of Southampton, ²Laboratoire des Systèmes Perceptifs, DEC, ENS, PSL University, Paris, France, ³Department of Psychology, New York University, New York, USA, ⁴Institute of Psychology, Centre for Cognitive Science, Technical University of Darmstadt, Darmstadt, Germany, ⁵Department of Engineering, Cambridge University, Cambridge, UK

Continuous psychophysics relies on tracking behaviour to estimate visual functions (Bonnen et al., 2015). Poorer tracking reflects greater perceptual uncertainty: harder-to-see targets are tracked less faithfully. But which is the best way to quantify tracking? Performance can be quantified from the peak, lag, and width of the cross-correlation between target and tracking velocities. As performance worsens, peak cross-correlation decreases, occurs at longer lags, and the cross-correlation broadens. However, these measures are not purely perceptual: they also reflect motor and cognitive factors, and are strongly intercorrelated, limiting their usefulness as independent descriptors. Fitting a Kalman filter to the tracking data instead yields a single observation-noise parameter, but its relationship to cross-correlation measures remains unclear. Here, we compared cross-correlation measures to the Kalman-derived parameter across five datasets: the Bonnen et al. dataset and four in-house datasets (pooled N=27). Participants tracked Gaussian-blob targets via mouse or eye tracking, and we manipulated target visibility and hand performance (dominant vs non-dominant hand). As expected, cross-correlation peak, lag, and variability were strongly intercorrelated, although interestingly, both the magnitude and direction of intercorrelations varied across datasets (range of correlations: r=[ -.81, .70]). Nevertheless, across all datasets pooled together, the Kalman-derived observation-noise parameter consistently correlated with each cross-correlation metric (peak–noise: r= -0.37; lag–noise: r=0.20; variability–noise: r=0.17, all ps< 0.05), suggesting that it could serve as a single summary index of variation across these metrics. Together, these findings show that the Kalman-derived observation-noise estimate provides a unified descriptor of tracking performance across datasets and tracking modalities, although it cannot by itself disentangle visual, motor, and cognitive contributions. The ultimate goal of this project is to apply inverse optimal control (Straub & Rothkopf, 2022) to decompose these components and provide more interpretable, mechanism-level estimates of the processes underlying tracking behaviour.

Acknowledgements: This research was funded by Fight for Sight (Grant Reference: RESSGA2302, awarded to GM) and Wessex Medical Research (Grant Reference: AF04, awarded to GM).