掌上视觉基础模型：优化移动增强现实的实时场景理解

Key Points

• The aim is to enhance real-time visual prediction for mobile AR applications using VFMs while addressing computational constraints.
• Developed ARIA, the first system for on-device VFM inference acceleration.
• Utilized a parallel and selective inference scheme leveraging mobile processor heterogeneity.
• Full-frame predictions are offloaded to GPUs, while dynamic region updates are handled by NPUs.
• Achieved significant improvements in prediction accuracy in real-world mobile AR scenarios.
• Increased the deadline success rate for real-time applications.

Abstract