Transparency-Aware Segmentation of Glass Objects to Train RGB-Based Pose Estimators.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Robotic manipulation necessitates precise object pose estimation for tasks like household chores.
• Estimating 6D poses for transparent objects (e.g., glasses) is challenging due to disturbed depth data and occlusions in RGB images.

Purpose of the Study:

• To improve RGB-based pose estimation for transparent objects by redefining ground-truth training data.
• To enhance the consistency of object scale and size representation in training data.

Main Methods:

• Proposed a transparency-aware multisegmentation approach where pixels can belong to multiple objects.
• Introduced transparency-aware bounding boxes that encompass entire objects, even when occluded.
• Trained an existing pose estimator with new ground-truth types on the ClearPose dataset.

Main Results:

• Achieved a 4.3% increase in ADD-S AUC by utilizing transparency-aware segmentation for training.
• Demonstrated performance improvement without modifying the core pose estimator architecture.
• Validated the effectiveness of the proposed ground-truth definition on a challenging dataset.

Conclusions:

• Transparency-aware ground-truth definitions significantly enhance RGB-based 6D pose estimation for transparent objects.
• The proposed multisegmentation and bounding box methods offer a generalizable approach applicable to datasets with 3D models and ground-truth poses.
• This research paves the way for more robust robotic manipulation in complex, real-world environments involving transparent items.