Gaze event detection is a central step in eye-tracking analysis, as raw gaze samples are typically transformed into fixation and saccade sequences before they are used in further analyses. However, different detection principles can segment the same gaze recording into different event sequences, which can affect downstream analyses that use event-level features. This thesis implements and descriptively compares four selected fixation and saccade detection approaches: Velocity- Threshold Identification (I-VT), Dispersion-Threshold Identification (I-DT), Bayesian Decision Theory Identification (I-BDT), and Bayesian Mixture Model Identification (I-BMM). The goal is not to propose a new detection algorithm or to evaluate a user-profiling model, but to examine how the selected approaches transform the same gaze recordings under shared processing assumptions.

The evaluation is based on recordings from the Hollywood eye-tracking database. All gaze coordinates are converted to degrees of visual angle before algorithm execution, and reduced- sampling conditions are generated by nearest-timestamp resampling. Because no manual fixation and saccade annotations are used as a ground-truth reference, the evaluation reports descriptive output metrics rather than accuracy, precision, recall, F1 score (F1), or an algorithm ranking.

The results document differences in event counts, temporal composition, duration coverage, and short-event structure across the four approaches. I-VT and I-BMM show many short-duration and single-sample events, while I-BDT is characterized by a dominance of internal smooth-pursuit states that are mapped to the shared non-fixation movement label. Reduced sampling leads to lower event counts across the methods and changes fixation-related temporal metrics in algorithm-specific ways. The findings show that the selected detector and sampling-rate representation are part of the event-level feature definition used in later analyses.