Domain-Translated 3D Object Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Synthetic 3D models are vital for object pose estimation, generating annotated data.
• Current methods enrich real-world training data with synthetic images but ignore domain differences.
• Real and synthetic image data distributions often differ, impacting pose estimation performance.

Purpose of the Study:

• To propose a novel framework for 3D object pose estimation that leverages synthetic data more effectively.
• To address the domain gap between real and synthetic images in object pose estimation tasks.
• To develop a method that improves 3D object pose estimation accuracy by translating real images to a synthetic domain.

Main Methods:

• A two-step framework is proposed: 1) pose-oriented image-to-image translation, 2) 3D object pose estimation.
• The image-to-image translation uses a novel objective function to preserve pose-relevant characteristics.
• The pose estimation is performed on the translated synthetic images, eliminating the need for real training data.

Main Results:

• The proposed framework significantly enhances 3D object pose estimation performance.
• Experimental results demonstrate superior performance compared to existing state-of-the-art methods.
• The pose-oriented translation effectively bridges the domain gap between real and synthetic images.

Conclusions:

• The novel framework offers a more effective approach to 3D object pose estimation by utilizing synthetic data.
• Translating real images to synthetic ones with a pose-oriented objective improves estimation accuracy.
• The method achieves high performance without requiring real data for training the pose estimation network.