6D-ViT: Category-Level 6D Object Pose Estimation via Transformer-Based Instance Representation Learning.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Category-level object pose estimation is crucial for robotic manipulation and scene understanding.
• Existing methods often struggle with accuracy and generalization across diverse object categories and environments.

Purpose of the Study:

• To develop a novel transformer-based network for highly accurate category-level object pose estimation using RGB-D images.
• To explore effective instance representation learning by integrating appearance, geometry, and shape priors.

Main Methods:

• Introduced 6D Vision Transformer (6D-ViT), a two-stream encoder-decoder framework with Pixelformer and Pointformer branches.
• Utilized a multisource aggregation network to combine RGB appearance, point cloud geometry, and shape priors for dense instance representations (NOCS).
• Computed 6D pose by solving similarity transformation between observed point clouds and reconstructed NOCS representations.

Main Results:

• Achieved state-of-the-art performance on both synthetic and real-world datasets for category-level object pose estimation.
• Demonstrated the effectiveness of the proposed framework in extracting powerful instance representations.
• Validated the robustness and accuracy of 6D-ViT across various scenarios.

Conclusions:

• The proposed 6D-ViT framework significantly advances the state-of-the-art in category-level object pose estimation.
• The integration of multi-modal data through a transformer architecture proves highly effective for instance representation learning.
• The method offers a promising solution for real-world applications requiring precise 6D object pose estimation.