DSPose: Decoupled Real-Time Pose Estimation from Semantic-RGB Radiance Rays

Area of Science:

• Computer Vision
• Robotics
• Augmented Reality

Background:

• Six-degree-of-freedom (6-DoF) pose estimation is critical for embodied agents and augmented reality.
• Current ray selection methods struggle with RGB feature dependence and coupled translation-rotation optimization, leading to degeneracy.
• Existing prior-free methods face performance barriers in accuracy and speed.

Purpose of the Study:

• To introduce DSPose, a robust pose estimation approach addressing selection ambiguity and optimization coupling.
• To improve the quality of candidate rays and stabilize the selection process.
• To achieve state-of-the-art, real-time, and prior-free 6-DoF pose estimation.

Main Methods:

• Developed Geometric-Resampling Semantic Representation (GRSR) for geometrically accurate scene structure and enhanced ray discriminability.
• Introduced Euclidean Lie Decoupled (ELD) mechanism for stable, decoupled optimization of translation and rotation.
• Generated Semantic-RGB radiance rays for improved quality and reduced selection ambiguity.

Main Results:

• DSPose establishes a new state-of-the-art (SOTA) in 6-DoF pose estimation.
• Achieved an unprecedented combination of accuracy and real-time performance.
• Significantly reduced both angular and translation errors in real-time, breaking prior-free method performance barriers.

Conclusions:

• DSPose offers a robust and efficient solution for prior-free 6-DoF pose estimation.
• The method makes accurate and real-time pose estimation practical.
• This work represents a significant advancement in overcoming long-standing performance limitations.