Real-Time RGB-D Camera Pose Estimation in Novel Scenes Using a Relocalisation Cascade.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Camera pose estimation is crucial for applications like simultaneous localisation and mapping (SLAM), virtual/augmented reality, and navigation.
• Traditional methods often require offline training on specific scenes, limiting their use in new environments.
• Regression forests offer accurate correspondence matching but have faced limitations in online adaptation.

Purpose of the Study:

• To significantly improve real-time camera pose estimation performance.
• To enable robust relocalisation in new environments without extensive offline training.
• To enhance the efficiency and accuracy of existing on-the-fly forest adaptation techniques.

Main Methods:

• Developed an extension to on-the-fly regression forest adaptation for camera pose estimation.
• Implemented a geometric scoring method to refine RANSAC pose hypotheses.
• Introduced a cascaded relocaliser approach, chaining multiple instances with varying parameters.
• Tuned cascade and individual relocaliser parameters for optimal performance.

Main Results:

• Achieved significantly better relocalisation performance compared to previous state-of-the-art methods.
• Maintained fully real-time performance, suitable for online applications.
• Demonstrated improved accuracy on the 7-Scenes and Stanford 4 Scenes benchmarks.
• Presented a novel visualization technique for understanding forest behavior.

Conclusions:

• The proposed cascaded approach with geometric scoring substantially enhances camera pose estimation accuracy and speed.
• The method effectively circumvents the need for pre-training on generic scenes, enabling rapid deployment in novel environments.
• This work advances the state-of-the-art in real-time, adaptable camera pose estimation for computer vision systems.