DON6D: a decoupled one-stage network for 6D pose estimation.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Six-dimensional (6D) pose estimation is crucial for robotic manipulation and grasping.
• Existing two-stage methods suffer from slow inference speeds and require refinement for variations in lighting, noise, occlusion, and truncation.

Purpose of the Study:

• To propose a novel decoupled one-stage network (DON6D) for efficient and accurate 6D pose estimation.
• To improve inference speed while maintaining high accuracy in challenging robotic scenes.

Main Methods:

• A two-dimensional detection network is used for object localization in RGB-D images.
• A feature extraction and fusion module captures color and geometric information.
• Dual data augmentation enhances model generalization.
• An attention residual encoder-decoder refines pose estimation.

Main Results:

• The DON6D model demonstrates superior performance compared to state-of-the-art methods.
• Evaluated on LINEMOD and YCB-Video datasets, DON6D achieved better ADD(-S) and ADD(-S) AUC metrics.

Conclusions:

• The proposed DON6D model offers an effective solution for fast and accurate 6D pose estimation.
• DON6D successfully addresses the limitations of existing methods in complex environments.