SENSE: Self-Evolving Learning for Self-Supervised Monocular Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Self-supervised depth estimation methods utilize unlabeled monocular videos but struggle with joint depth and pose uncertainty.
• Supervised methods offer superior performance but are constrained by the availability of labeled data.

Purpose of the Study:

• To introduce SENSE, a novel learning paradigm for self-supervised monocular depth estimation.
• To progressively enhance depth and pose estimation using supervised learning without requiring labeled data.

Main Methods:

• Leveraging pseudo-labels generated by self-supervised methods as a novel training signal.
• Employing a fully supervised depth estimation network trained on pseudo-labels.
• Developing a comprehensive training pipeline that iteratively refines both depth and pose estimation branches.

Main Results:

• Pseudo-labels from self-supervised methods can yield superior depth estimation results compared to ground truth.
• The SENSE approach effectively mitigates the challenges of multi-task training in self-supervised depth estimation.
• Achieved state-of-the-art performance on the KITTI dataset.

Conclusions:

• SENSE offers a robust framework for self-supervised monocular depth estimation by effectively utilizing pseudo-labels.
• The proposed method demonstrates significant improvements in depth and pose estimation accuracy.
• This approach advances the field by enabling high-performance depth estimation without reliance on extensive labeled datasets.