EBFA-6D: End-to-End Transparent Object 6D Pose Estimation Based on a Boundary Feature Augmented Mechanism

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Transparent objects pose significant challenges for 6D pose estimation due to their unique optical properties, lacking texture and color, and interfering with depth sensors.
• Traditional vision methods struggle with transparent objects, hindering accurate geometric detail capture and pose determination.

Purpose of the Study:

• To propose EBFA-6D, a novel end-to-end framework for direct 6D pose estimation of transparent objects from single RGB images.
• To address the inherent difficulties in estimating the pose of transparent objects by leveraging their boundary features.

Main Methods:

• Developed EBFA-6D, an end-to-end framework for 6D pose estimation from single RGB images.
• Implemented a boundary feature augmented mechanism to exploit high contrast at object boundaries.
• Utilized bottom-up feature fusion to enhance localization accuracy.

Main Results:

• EBFA-6D demonstrated superior accuracy compared to existing methods on the ClearPose dataset.
• The framework achieved near real-time inference speeds.
• Successfully provided efficient and effective 6D pose estimation for transparent objects.

Conclusions:

• EBFA-6D offers a robust solution for the challenging problem of 6D pose estimation of transparent objects.
• The proposed boundary feature augmentation and feature fusion techniques significantly improve performance.
• The method presents a practical advancement for applications requiring accurate transparent object pose estimation.