Equivariant Diffusion Model With A5-Group Neurons for Joint Pose Estimation and Shape Reconstruction.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Geometry

Background:

• Object pose estimation and shape reconstruction are often studied separately.
• Existing joint methods struggle with partial observations and shape ambiguities.
• A unified approach is needed for mutual task benefit and improved robustness.

Purpose of the Study:

• To develop a diffusion model for joint category-level object pose estimation and 3D shape reconstruction.
• To leverage diffusion models' iterative nature for optimizing both tasks simultaneously.
• To address ambiguity by enabling multiple plausible outputs from partial observations.

Main Methods:

• Proposed an equivariant diffusion model integrating feature extraction and a ShapePose diffusion model.
• Utilized A5-group neurons for SO(3)-equivariance, enabling rotation-aware processing.
• Implemented SO(3)-equivariant 3D point convolution and concatenation for network-wide equivariance.
• Introduced a geometry-based plausibility measure to select the best pose-shape combination.

Main Results:

• Achieved state-of-the-art performance on shape reconstruction and pose estimation across multiple datasets.
• Demonstrated the ability to generate multiple plausible outputs for ambiguous inputs.
• Showcased improved robustness in handling partial observations and shape ambiguities.

Conclusions:

• Diffusion models offer a powerful framework for jointly tackling pose estimation and shape reconstruction.
• The proposed SO(3)-equivariant architecture effectively handles 3D geometric transformations.
• The method provides a robust and versatile solution for complex object understanding tasks.