Cycle-SfM: Joint self-supervised learning of depth and camera motion from monocular image sequences

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• 3D scene geometry understanding is crucial for computer vision tasks like depth prediction and visual odometry.
• Deep learning enables end-to-end solutions by framing 3D understanding as nonlinear optimization problems.
• Existing methods often require labeled data or complex pipelines.

Purpose of the Study:

• To develop a self-supervised framework for joint monocular depth and camera ego-motion estimation.
• To leverage unlabeled, unstructured monocular video sequences for 3D scene understanding.
• To improve the accuracy and efficiency of 3D geometry estimation.

Main Methods:

• A self-supervised framework utilizing forward-backward consistency constraint on view reconstruction.
• Exploiting bidirectional projection information across adjacent video frames.
• Introducing a lightweight and generalizable cost function improvement for enhanced accuracy.

Main Results:

• The proposed framework achieves comparable depth estimation results to existing methods.
• Demonstrates superior performance in pose estimation compared to current approaches on the KITTI dataset.
• The cost function improvement module is shown to be effective and seamlessly integrable.

Conclusions:

• The developed self-supervised method effectively estimates monocular depth and camera ego-motion from unlabelled video.
• The novel cost function enhancement offers a practical and efficient way to boost accuracy in self-supervised 3D vision tasks.
• The framework presents a significant advancement in efficient and accurate 3D scene geometry understanding.