100-Phones: A Large VI-SLAM Dataset for Augmented Reality Towards Mass Deployment on Mobile Phones

Area of Science:

• Computer Vision
• Robotics
• Mobile Computing

Background:

• Visual-inertial SLAM (VI-SLAM) is crucial for real-time 6-DoF motion recovery in Augmented Reality (AR).
• Smartphones are primary AR devices, but current VI-SLAM methods struggle with robustness on mid- and low-end models.
• Existing datasets lack diversity, failing to represent the wide range of phone hardware impacting VI-SLAM performance.

Purpose of the Study:

• To introduce the first VI-SLAM dataset encompassing a broad spectrum of mainstream mobile phones.
• To analyze the impact of diverse phone hardware on visual-inertial data quality.
• To identify and address robustness issues in VI-SLAM for mass mobile deployment.

Main Methods:

• Collected 350 sequences using 100 distinct smartphone models.
• Performed extensive analysis and experiments on the collected visual-inertial data.
• Evaluated the performance of current open-source VI-SLAM methods on the diverse dataset.

Main Results:

• Demonstrated significant variations in visual-inertial data quality across different phone models.
• Confirmed that existing open-source VI-SLAM methods exhibit critical robustness issues on a wide range of phones.
• Identified specific hardware-related factors contributing to VI-SLAM performance degradation.

Conclusions:

• The heterogeneity of mobile phone hardware poses a substantial challenge for robust VI-SLAM deployment.
• Further research and development are needed to improve VI-SLAM robustness for widespread AR adoption.
• The 100-Phones dataset provides a valuable resource for advancing VI-SLAM research and development.