Randomized trial demonstrates effective video anomaly detection in classroom examinations, suggesting a new benchmark for capturing subtle cheating behaviors.
Key Points
This research aims to develop a method for detecting cheating in classroom examinations by utilizing video analysis and weak supervision.
Introduced Cheatomaly, a curated video dataset for examining cheating behavior.
Utilized multiple instance learning with vision transformer features to rank videos based on anomaly.
Implemented a margin-based ranking objective focusing on segment-level representations.
Achieved strong video-level discrimination and effective frame-level localization in analyses.
Showed that temporal aggregation impacts the balance between ranking and localization without consistent gains.
Highlighted challenges in modeling context-dependent temporal behavior during anomaly detection.