Real-Time Single Image Depth Perception in the Wild with Handheld Devices.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Depth perception is crucial for applications like autonomous driving and consumer tech.
• Monocular depth estimation offers versatility but faces challenges in reliability and real-time performance on mobile devices.

Purpose of the Study:

• To investigate and overcome limitations in monocular depth estimation for real-world mobile applications.
• To develop efficient network designs and training strategies for reliable, real-time depth estimation.

Main Methods:

• Investigated network design and training strategies for monocular depth estimation.
• Developed methods for mapping networks onto low-power embedded systems for real-time performance.
• Conducted thorough evaluations on generalization capabilities.

Main Results:

• Demonstrated that appropriate network design and training address reliability and resource constraints.
• Achieved real-time performance on handheld devices.
• Showcased the networks' ability to generalize well to diverse, real-world environments.

Conclusions:

• Optimized monocular depth estimation networks are feasible for real-world mobile applications.
• The proposed methods enable reliable, real-time depth-aware augmented reality and image blurring on smartphones.
• This research facilitates practical deployment of advanced computer vision on edge devices.