SynPo-Net-Accurate and Fast CNN-Based 6DoF Object Pose Estimation Using Synthetic Training

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Estimating object poses (6 Degrees of Freedom - 6DoF) is crucial for robotics and augmented reality.
• Current deep learning methods often require real images, facing challenges in ground truth acquisition and dataset scalability.
• Synthetic data offers a scalable alternative but often struggles with domain gap issues.

Purpose of the Study:

• To develop a novel approach for 6DoF object pose estimation using exclusively synthetic data.
• To introduce SynPo-Net, a Convolutional Neural Network (CNN) architecture optimized for direct pose regression.
• To propose a domain adaptation technique to bridge the gap between synthetic and real image domains.

Main Methods:

• Training a Convolutional Neural Network (CNN) exclusively on single-channel synthetic images.
• Designing a specialized network architecture (SynPo-Net) for direct 6DoF pose regression.
• Implementing a domain adaptation scheme to align features from synthetic and real images.

Main Results:

• SynPo-Net significantly outperforms state-of-the-art methods trained on synthetic data.
• The proposed approach achieves higher accuracy and faster processing speeds.
• The system demonstrates effective 6DoF pose estimation from single frames.

Conclusions:

• Training deep neural networks solely on synthetic data is viable for 6DoF pose estimation.
• SynPo-Net offers a robust and efficient solution for robotic interaction and augmented reality applications.
• The domain adaptation scheme enhances the generalization capability of models trained on synthetic data.