Regression-Based Camera Pose Estimation through Multi-Level Local Features and Global Features.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Accurate camera pose estimation is critical for augmented reality and autonomous driving.
• Existing methods struggle with illumination and viewpoint changes, and keypoint inaccuracies.

Purpose of the Study:

• To propose a novel relative camera pose regression framework.
• To enhance robustness and accuracy in challenging environmental conditions.

Main Methods:

• Utilizing a multi-level deformable network for rotation-variant local feature detection and description.
• Employing pixel correspondences to refine feature processing.
• Introducing a novel loss function combining relative and absolute regression with geometric constraints.

Main Results:

• Achieved an average mean translation error of 0.18 m and rotation error of 7.44° on the 7Scenes dataset.
• Demonstrated effectiveness in pose estimation and image matching tasks through ablation studies on 7Scenes and HPatches datasets.

Conclusions:

• The proposed framework offers a significant improvement in camera pose estimation accuracy and robustness.
• The combination of global and local features with a novel loss function effectively addresses limitations of prior methods.