Absolute Camera Pose Regression Using an RGB-D Dual-Stream Network and Handcrafted Base Poses

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Absolute Pose Regression (APR) is a single-shot method for camera localization using end-to-end neural networks.
• Existing APR methods rely on learned base poses from training data, limiting generalization to unseen scenes.
• Overfitting to training data restricts the performance and applicability of current APR techniques.

Purpose of the Study:

• To overcome the generalization limitations of learned base poses in APR.
• To improve camera localization accuracy by preventing overfitting to training data.
• To enhance the robustness of APR methods on diverse and challenging datasets.

Main Methods:

• Implemented handcrafted base poses instead of learning-based ones to avoid overfitting.
• Utilized a dual-stream network architecture to process color and depth images independently.
• Evaluated the method on the 7 Scenes and Oxford RobotCar datasets for localization performance.

Main Results:

• Achieved top-tier performance in median rotation error on the 7 Scenes dataset.
• Outperformed previous APR methods in median translation error on the 7 Scenes dataset.
• Demonstrated significant improvements in both median translation and rotation errors on the Oxford RobotCar dataset compared to state-of-the-art methods.

Conclusions:

• Handcrafted base poses enhance the generalization capability of APR methods.
• The dual-stream network architecture contributes to more accurate camera localization.
• The proposed method offers a robust and improved solution for camera pose estimation in challenging environments.