Lifelong-MonoDepth: Lifelong Learning for Multidomain Monocular Metric Depth Estimation

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Lifelong learning (LL) models are crucial for autonomous driving and robot navigation, requiring metric depth estimation.
• Existing LL methods struggle with sensor-dependent depth map variations and large domain gaps, limiting them to relative depth estimation.
• Challenges include depth scale variation, significant domain gaps, and the need for automated domain-aware inference.

Purpose of the Study:

• To develop a lifelong learning model for accurate metric depth estimation.
• To address challenges of depth scale variation, domain gaps, and real-time inference in autonomous systems.
• To improve the efficiency, stability, and plasticity of lifelong depth learning models.

Main Methods:

• A lightweight multihead framework for scale-aware depth learning to address depth scale imbalance.
• An uncertainty-aware lifelong learning solution to handle significant domain gaps.
• An online domain-specific predictor selection method for real-time, automated domain-aware inference.

Main Results:

• The proposed method demonstrates efficiency, stability, and plasticity in lifelong depth learning.
• Achieved benchmark improvements of 8%-15% in metric depth estimation.
• Successfully addresses depth scale variation and large domain gaps.

Conclusions:

• The developed framework facilitates lifelong metric depth learning, overcoming limitations of existing approaches.
• The method offers a practical solution for real-time, domain-aware depth inference in autonomous applications.
• The approach enhances the performance and adaptability of depth estimation models in dynamic environments.