Enhancing Smartphone Eye Tracking with Cursor-Based Interactive Implicit Calibration

2024
Chang Liu, Xiangyang Wang, Chun Yu, Yingtian Shi, Chongyang Wang, Ziqi Liu, Chen Liang, Yuanchun Shi

Limited accuracy of eye tracking on smartphones restricts its use. Existing RGB-camera-based eye tracking rely on extensive datasets, which could be enhanced by continuous fine-tuning using calibration data implicitly collected from interaction. In this context, we propose COMETIC (Cursor Operation Mediated Eye-Tracking Implicit Calibration), which introduces a cursor-based interaction and utilizes the inherent correlation between cursor and eye movement.

By filtering valid cursor coordinates as proxies for the ground truth of gaze and fine-tuning the eye-tracking model with corresponding images, COMETIC enhances accuracy during interaction. Both filtering and fine-tuning use pre-trained models and could be facilitated using personalized, dynamically updated data. Results show COMETIC achieves an average eye tracking error of 208.04 px (1.2 cm), representing a 49.64% improvement compared to that without fine-tuning. We found that filtering cursor points whose actual distance to gaze falls within 250 to 300 px (1.44 to 1.73 cm) yields the best eye tracking results.

Publication:

• Enhancing Smartphone Eye Tracking with Cursor-Based Interactive Implicit Calibration(CHI ’25)

Eye Tracking on Smartphone

Current smartphone eye tracking faces two main challenges: limited personalization for individual users and difficulty adapting to frequent posture changes. To address these issues, we developed a system that dynamically calibrates eye tracking through implicit user interactions, improving both accuracy and personalization.

COMETIC

By filtering valid cursor coordinates as proxies for gaze ground truth and fine-tuning the eye-tracking model with corresponding images, COMETIC enhances accuracy during interaction. Both filtering and fine-tuning utilize pre-trained models and can be further improved with personalized, dynamically updated data.

Data Collection Experiment

Evaluation

Results show that COMETIC achieves an average eye-tracking error of 208.04 px (1.2 cm), representing a 49.64% improvement compared to the system without fine-tuning. The analysis reveals that filtering cursor points with an actual gaze distance between 250 and 300 px (1.44 to 1.73 cm) yields the best eye-tracking results.