Cross attention-based prior deformation for category-level 6D pose estimation.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Category-level 6D pose estimation relies on deforming shape priors to match object instances.
• Current methods using global features for deformation lose critical local details, impacting accuracy.
• This limitation hinders precise alignment and robust pose estimation.

Purpose of the Study:

• To propose a novel method, Cross Attention-based Prior Deformation (CAPD), to enhance 6D pose estimation.
• To address the loss of local information in shape prior deformation.
• To improve the accuracy and robustness of category-level 6D pose estimation.

Main Methods:

• Developed the Cross Attention-based Prior Deformation (CAPD) method.
• Utilized a cross-attention mechanism to model point-to-point correlations between shape priors and instance point clouds.
• Enabled propagation of global and local information for more expressive deformation.

Main Results:

• CAPD effectively preserves local details during shape prior deformation.
• The deformed shape prior exhibits improved geometric fitting to the instance.
• Significantly enhanced robustness and accuracy in category-level 6D pose estimation.

Conclusions:

• CAPD outperforms existing shape-prior-based methods for category-level 6D pose estimation.
• The cross-attention mechanism is key to mitigating local information loss.
• The proposed method offers a more precise and robust approach to 6D pose estimation.