Comparison of the Accuracy of Segmentation and Regression-Based Approaches for the Task of Eye Centre Localisation

Eye centre localisation is of great importance in eye movement analysis, which is useful in diagnosing neurodegenerative diseases. This project explores two distinct methodologies for eye centre localisation: segmentation- and regression-based approaches, using the EM-COGLOAD dataset9 , which contains 13,813 images of eyes with labelled coordinates of eye centres. We applied and evaluated regression models, known for their excellent precision, and segmentation models, noted for their flexibility in processing diverse images. The results demonstrated that while regression models, particularly one using MobileNetV23 , lead with over 99% accuracy in both eyes with a normalised error of 0.025, segmentation models, exemplified by the zero-shot Segment Anything Model (SAM), offered promising alternatives with significant potential for innovation and improvement. The strengths and limitations of each approach were also discussed, confirming the established effectiveness of the regression method and suggesting emerging capabilities of segmentation techniques in eye centre localisation tasks.