ViDDAR: Vision Language Model-based Task-Detrimental Content Detection for Augmented Reality

This IEEE conference talk presents ViDDAR (Vision Language Model-based Task-Detrimental Content Detection for Augmented Reality), the first system that leverages Vision Language Models to detect content that can negatively impact AR experiences. Learn how researchers from Duke University developed this innovative solution to identify both obstruction and information manipulation attacks in augmented reality environments. The system achieved impressive results with 92.15% accuracy in obstruction detection and 82.46% accuracy in identifying information manipulation. Discover how ViDDAR was tested on real-world datasets and implemented as a mobile application, offering practical protection against virtual content that might block critical real-world information or present misleading data to users. Part of the "Evaluation methods" session at IEEE VR 2025, this 11-minute presentation demonstrates significant advances in ensuring safer and more reliable augmented reality experiences.