GFA-Net: Geometry-Focused Attention Network for Six Degrees of Freedom Object Pose Estimation

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• Six degrees of freedom (6-DoF) object pose estimation is crucial for robotics and autonomous driving.
• Estimating pose from single RGB images is challenging due to limitations in deep neural network feature extraction.
• Existing methods often rely on supplementary data like depth, but robust RGB-only solutions are needed.

Purpose of the Study:

• To develop a novel framework, GFA-Net, for enhanced feature extraction in 6-DoF object pose estimation from RGB images.
• To improve the accuracy and robustness of pose estimation by focusing on geometric and textural object characteristics.
• To address the limitations of current deep neural networks in extracting salient features from object regions in RGB data.

Main Methods:

• Introduced the Geometry-Focused Attention Network (GFA-Net) for comprehensive feature extraction.
• Employed Point-wise Feature Attention (PFA) to capture subtle pose variations and localize object regions.
• Integrated a Geometry Feature Aggregation Module (GFAM) for multi-scale geometric feature distillation.
• Utilized a Perspective-n-Point (PnP) module for final 6-DoF pose computation based on 2D-3D correspondences.

Main Results:

• GFA-Net achieved competitive performance against state-of-the-art methods on LINEMOD and Occlusion LINEMOD datasets.
• Attained 96.54% accuracy on LINEMOD and 49.35% on Occlusion LINEMOD using the ADD-S metric (0.10d threshold).
• Demonstrated effective feature extraction by analyzing geometric and textural properties for pose estimation.

Conclusions:

• GFA-Net offers a promising approach for accurate 6-DoF object pose estimation using only RGB images.
• The proposed attention and feature aggregation mechanisms enhance the network's ability to discern pose-critical information.
• The method shows significant potential for real-world applications in robotics and autonomous driving.