RAFusion: Integrating Residual Attention for Global Context-Aware 6D Pose Estimation

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• 6D object pose estimation from RGB-D data is crucial for robotic perception.
• Existing methods like DenseFusion struggle with occlusion and cluttered environments.

Purpose of the Study:

• To enhance the DenseFusion architecture for more robust 6D object pose estimation.
• To improve feature representation and global context modeling in RGB-D based pose estimation.

Main Methods:

• Proposed RAFusion network integrating Squeeze-and-Excitation (SE) blocks for RGB feature enhancement.
• Incorporated a RealFormer-based block for residual self-attention on fused RGB-D features to model global context.

Main Results:

• RAFusion demonstrated improved performance over the DenseFusion baseline.
• Achieved competitive results compared to other state-of-the-art 6D pose estimation methods on LINEMOD and Occlusion LINEMOD datasets.

Conclusions:

• The proposed SE blocks and RealFormer-based module effectively enhance feature representation and global context modeling.
• RAFusion offers a more robust solution for 6D object pose estimation, particularly in challenging conditions.