VR.net: A Real-world Dataset for Virtual Reality Motion Sickness Research.

Area of Science:

• Virtual Reality (VR) and Human-Computer Interaction
• Machine Learning and Data Science
• Biomedical Engineering and Neuroscience

Background:

• Machine learning (ML) models for predicting virtual reality (VR) motion sickness require large, diverse, and accurately labeled datasets.
• Existing datasets are often limited in scope, hindering the development of generalizable ML models for VR experience.

Purpose of the Study:

• To introduce VR.net, a comprehensive dataset designed to facilitate the development of robust ML models for VR motion sickness detection.
• To enable accurate, frame-by-frame labeling of motion sickness indicators within VR gameplay.

Main Methods:

• Compiled 165 hours of gameplay videos from 100 diverse VR games across ten genres, evaluated by 500 participants.
• Developed an automated tool to extract ground truth data from 3D engines' rendering pipelines, enabling precise labeling of 24 motion sickness-related features (e.g., camera movement, depth of field).
• Ensured dataset diversity and scale to support generalizable ML model development without requiring access to VR game source code.

Main Results:

• VR.net provides 165 hours of labeled VR gameplay data with 24 detailed motion sickness-related labels per frame.
• Demonstrated the dataset's utility through applications in risk factor detection and sickness level prediction.
• The automated data extraction tool allows for precise ground truth data collection.

Conclusions:

• VR.net offers an unprecedented resource for VR motion sickness research due to its scale, accuracy, and diversity.
• The dataset and accompanying data collection tool can accelerate advancements in understanding and mitigating VR-induced discomfort.
• Encourages community contribution to further expand the VR.net dataset.