MYFix: Automated Fixation Annotation of Eye-Tracking Videos.

Area of Science:

• Computer Vision
• Human-Computer Interaction
• Mobile Eye-Tracking

Background:

• Automatic fixation point annotation is crucial for mobile eye-tracking research, particularly in dynamic urban environments.
• Existing methods face challenges due to observer and environmental movement, complicating data analysis.
• Accurate fixation point identification is essential for understanding visual attention in real-world scenarios.

Purpose of the Study:

• To develop a novel, automated pipeline for annotating fixation points in mobile eye-tracking data.
• To leverage foundation models (YOLOv8, Mask2Former) for robust annotation without additional training.
• To enhance the accuracy and efficiency of fixation point annotation in complex outdoor scenes.

Main Methods:

• Integration of YOLOv8 for object detection and Mask2Former for semantic segmentation into a unified pipeline.
• Utilizing pre-trained models on MS COCO and Cityscapes datasets, eliminating the need for fine-tuning.
• Testing the pipeline in both controlled data collection and real-world outdoor wayfinding experiments.

Main Results:

• Achieved 89.05% accuracy in a controlled environment and 81.50% accuracy in a real-world outdoor wayfinding scenario.
• Demonstrated consistent and reliable annotations even in scenes with multiple objects at varying depths.
• Reported an average processing time of 1.61 ± 0.35 seconds per frame.

Conclusions:

• The proposed YOLOv8 and Mask2Former pipeline offers a robust and efficient solution for automatic fixation point annotation in mobile eye-tracking.
• This approach significantly streamlines the annotation process, making it suitable for complex, dynamic environments.
• The method provides a reliable tool for researchers studying visual attention in real-world settings.