Object Pose Estimation Using Edge Images Synthesized from Shape Information.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Estimating object pose from monocular images is crucial for robotics and computer vision.
• Deep learning methods require extensive datasets, which are costly to acquire.
• Simulation-based methods using computer graphics (CG) offer a cost-effective alternative for dataset generation.

Purpose of the Study:

• To develop a more accurate method for six Degrees of Freedom (6DoF) pose estimation of texture-less objects from monocular images.
• To improve upon existing simulation-based deep learning approaches by incorporating edge information.
• To reduce the dataset collection cost by leveraging CG data.

Main Methods:

• A deep learning model trained on CG-rendered images.
• Utilizing edge information extracted from object ridgelines as input features.
• Comparing the proposed method against silhouette-based methods like Pose Interpreter Networks.

Main Results:

• Significant accuracy improvements in pose estimation compared to previous methods.
• Error rate reduction of 22.9% for translation and 43.4% for rotation on simulation data.
• Further refinement through fine-tuning on physical data, achieving 20.1% translation and 57.7% rotation error reduction.

Conclusions:

• Edge information from ridgelines enhances the robustness and accuracy of 6DoF pose estimation.
• The proposed method offers a practical solution for pose estimation using limited real-world data.
• Fine-tuning is effective in bridging the gap between simulation and physical data for improved real-world performance.