Category-Level 6-D Object Pose Estimation With Shape Deformation for Robotic Grasp Detection.

Area of Science:

• Robotics
• Computer Vision
• Machine Learning

Background:

• Category-level 6-D object pose estimation is vital for robotic grasp detection.
• Synthetic-to-real data domain gap limits model transferability, hindering real-world application.
• Large-scale real-world datasets are costly and time-consuming to acquire.

Purpose of the Study:

• Introduce CatDeform, a novel network for category-level object pose estimation.
• Enable effective model training on synthetic data for superior real-world performance.
• Reduce dependency on extensive real-world annotated datasets.

Main Methods:

• Transformer-based fusion module for multi-source information integration and enhanced prediction.
• Transformer-based attention module for deforming prior point clouds geometrically and feature-wise.
• Two-branch network for supervised learning and self-supervised training to bridge the synthetic-real data gap.

Main Results:

• CatDeform demonstrates strong performance on real datasets using primarily synthetic training data.
• Outperforms state-of-the-art methods in both self-supervised and supervised training paradigms on CAMERA25 and REAL275 datasets.
• Achieves high-precision pose estimation and improved grasp success rates in real-world robotic experiments.

Conclusions:

• CatDeform effectively addresses the synthetic-to-real domain gap in object pose estimation.
• The proposed transformer-based approach enables robust and accurate pose estimation with reduced reliance on real data.
• CatDeform shows significant potential for enhancing robotic grasping capabilities in practical applications.