Unsupervised Monocular Depth and Camera Pose Estimation with Multiple Masks and Geometric Consistency Constraints.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Estimating scene depth and camera pose from video is crucial for 3D reconstruction, visual navigation, and augmented reality.
• Existing unsupervised learning methods struggle with dynamic objects and occlusions in challenging visual scenes.

Purpose of the Study:

• To develop a novel unsupervised learning framework for robust depth and camera pose estimation.
• To address limitations of current methods in handling dynamic objects and occluded regions.

Main Methods:

• Implemented multiple mask technologies to identify and exclude outliers from loss computation.
• Utilized identified outliers as supervised signals to train a mask estimation network.
• Introduced geometric consistency constraints to mitigate illumination variations and enhance pose estimation.

Main Results:

• The proposed framework effectively mitigates the negative impacts of challenging scenes on depth and pose estimation.
• Experimental results on the KITTI dataset show superior performance compared to other unsupervised methods.
• The mask estimation network and geometric constraints act as effective supervised signals.

Conclusions:

• The novel unsupervised framework significantly enhances the accuracy and robustness of depth and camera pose estimation.
• The integration of mask technologies and geometric consistency offers a promising direction for future research in visual SLAM and related fields.