Unsupervised Learning of Depth and Camera Pose with Feature Map Warping

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Estimating image depth and egomotion is crucial for autonomous systems to navigate and avoid collisions.
• Current unsupervised methods often rely on photometric error, which is sensitive to real-world variations like lighting changes and occlusions.

Purpose of the Study:

• To develop a more robust unsupervised method for estimating image depth and egomotion.
• To improve the accuracy of depth and pose estimation in the presence of challenging conditions such as brightness changes and occlusions.

Main Methods:

• Proposed a feature pyramid matching loss (FPML) to capture trainable feature errors, offering greater robustness than traditional photometric error.
• Introduced an occlusion-aware mask (OAM) network to identify and mitigate the impact of occluded regions on estimation accuracy.

Main Results:

• The proposed unsupervised approach demonstrates high competitiveness against state-of-the-art methods.
• Achieved significant quantitative improvements, reducing absolute relative error (Abs Rel) by 0.017-0.088.

Conclusions:

• The novel method effectively addresses limitations of existing unsupervised depth and egomotion estimation techniques.
• The combination of FPML and OAM leads to more accurate and reliable scene understanding for autonomous agents.