Occlusion-Aware Self-Supervised Monocular 6D Object Pose Estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• 6D object pose estimation is crucial but challenging.
• Convolutional Neural Networks (CNNs) show promise but require extensive annotated data.
• Manual annotation is labor-intensive and time-consuming.

Purpose of the Study:

• Develop a novel monocular 6D pose estimation approach using self-supervised learning.
• Eliminate the need for real-world annotations.
• Improve robustness and accuracy in challenging scenarios.

Main Methods:

• Initial supervised training on synthetic RGB data.
• Leveraging noisy student training and differentiable rendering for self-supervision on real RGB(-D) data.
• Utilizing both visible and amodal mask information for robustness.

Main Results:

• The proposed self-supervised method outperforms existing approaches relying solely on synthetic data or domain adaptation.
• The self-supervised approach shows consistent improvement over its synthetically trained baseline.
• The method nearly closes the performance gap with fully supervised counterparts.

Conclusions:

• Self-supervised learning offers a viable solution to the data annotation bottleneck in 6D pose estimation.
• The proposed method achieves robust and accurate pose estimation, even with occlusions.
• This approach significantly advances monocular 6D object pose estimation capabilities.