CPPF++: Uncertainty-Aware Sim2Real Object Pose Estimation by Vote Aggregation.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Robotics

Background:

• Object pose estimation is crucial for 3D vision tasks.
• Real-world pose annotation data is expensive to acquire.
• Existing methods often require extensive real-world training data.

Purpose of the Study:

• To develop a novel sim-to-real category-level pose estimation method.
• To overcome limitations of current methods using only 3D CAD models.
• To improve pose estimation accuracy without real-world annotations.

Main Methods:

• Introduced CPPF++ based on a probabilistic reformulation of the point-pair voting scheme.
• Modeled voting uncertainty by estimating probabilistic distributions of point pairs.
• Augmented contextual information using N-point tuples.
• Incorporated noisy pair filtering, online alignment optimization, and tuple feature ensemble.

Main Results:

• CPPF++ significantly outperforms previous sim-to-real approaches.
• The method achieves comparable or superior performance on novel datasets.
• Introduced the DiversePose 300 dataset for category-level pose estimation.

Conclusions:

• CPPF++ offers a robust and accurate solution for sim-to-real category-level pose estimation.
• The probabilistic approach effectively handles vote collision challenges.
• The method demonstrates strong generalization capabilities on new datasets.