Enhancing object pose estimation for RGB images in cluttered scenes.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Estimating an object's 6D pose is vital for robotic interaction.
• Cluttered scenes and occlusions pose significant challenges for existing 6D object pose estimation methods.
• Current two-stage methods using key-point localization are vulnerable to occlusion and struggle with multi-object scenarios.

Purpose of the Study:

• To propose an end-to-end framework for single and multi-object 6D pose estimation using RGB images.
• To address limitations of existing methods, particularly in cluttered and occluded environments.
• To develop a computationally efficient and scalable solution.

Main Methods:

• A novel framework combining Convolutional Neural Networks (CNNs) and self-attention mechanisms.
• Feature fusion for local feature extraction.
• Multi-head self-attention (MHSA) integrated with iterative refinement for enhanced pose estimation.
• Scalable architecture adaptable to computational resources.

Main Results:

• Achieved high performance on benchmark datasets (Linemod and Occlusion Linemod).
• Reached 97.45% accuracy on Linemod and 84.84% on Occlusion Linemod using the ADD(-S) metric.
• Demonstrated effectiveness in challenging conditions with occlusion and clutter.

Conclusions:

• The proposed CNN and self-attention framework offers an effective end-to-end solution for 6D object pose estimation.
• The method shows robustness against occlusion and clutter, outperforming traditional approaches.
• The framework's scalability and low computational cost make it suitable for real-world robotic applications.