Depth Completion With Multiple Balanced Bases and Confidence for Dense Monocular SLAM.

Area of Science:

• Computer Vision
• Robotics
• Augmented and Virtual Reality

Background:

• Monocular dense Simultaneous Localization and Mapping (SLAM) is crucial for AR/VR on mobile devices.
• Existing sparse SLAM systems lack dense mapping capabilities, limiting AR/VR applications.
• Real-time dense mapping on mobile platforms remains a significant challenge.

Purpose of the Study:

• To develop a novel method for enabling dense mapping within sparse SLAM systems on mobile devices.
• To introduce a lightweight depth completion network, BBC-Net, optimized for sparse SLAM integration.
• To enhance the efficiency and mapping quality of monocular SLAM for AR/VR applications.

Main Methods:

• Integration of a lightweight depth completion network (BBC-Net) into a sparse SLAM system.
• Utilizing a multi-basis depth representation with predicted bases and confidence maps.
• Designing depth weight factors for seamless integration and optimization within SLAM bundle adjustment.

Main Results:

• BBC-Net successfully predicts multiple depth bases and confidence maps from monocular images.
• The proposed method achieves superior monocular dense mapping performance compared to state-of-the-art techniques.
• An online demo on a mobile phone validates the method's efficiency and mapping quality in real-world scenarios.

Conclusions:

• The proposed BBC-Net offers a versatile plug-in module for enhancing existing sparse SLAM systems.
• The method significantly improves global depth consistency and enables real-time dense mapping on mobile phones.
• This advancement holds substantial value for AR/VR applications requiring accurate 3D scene reconstruction.