Probing Effective and Efficient Category-Level Articulated Object Pose Perception.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Category-level articulated object pose perception is crucial for embodied AI.
• Existing methods struggle with kinematic constraints, self-occlusions, and optimization.
• Limitations hinder robust interaction in complex environments.

Purpose of the Study:

• Introduce CAPER++, a unified framework to overcome limitations in articulated object pose perception.
• Improve geometric consistency, rotation representation, and tracking stability.
• Advance real-time, accurate pose estimation and tracking for AI applications.

Main Methods:

• Joint-centric hierarchical model for explicit kinematic constraint embedding.
• SE(3) manifold formulation with Lie algebra for stable, singularity-free optimization.
• Proxy canonicalization and dynamic keyframes for robust pose tracking and drift suppression.

Main Results:

• State-of-the-art accuracy and robustness demonstrated on diverse benchmarks (synthetic, semi-synthetic, real-world).
• Achieved real-time inference at 50 FPS without post-processing.
• Significantly improved articulated object pose estimation and tracking capabilities.

Conclusions:

• CAPER++ offers a robust and efficient solution for category-level articulated object perception.
• The framework's innovations enable geometrically consistent and stable pose recovery.
• Enables advanced embodied AI interactions in complex, real-world scenarios.