Pose-Perceptive Convolution: Learning Geometry-Adaptive Receptive Fields for Robust 6D Pose Estimation.

Area of Science:

• Computer Vision
• Robotics
• Machine Learning

Background:

• 6D object pose estimation is vital for robotics and AR but challenged by object aspect ratios and appearance variations.
• Existing methods often overlook geometric mismatches between fixed convolutional receptive fields and object morphologies.
• This mismatch limits the performance of current 6D pose estimation techniques.

Purpose of the Study:

• To propose a novel Pose-Perceptive Convolution (PPC) to resolve geometric mismatches in feature extraction.
• To introduce a new Pose-Perceptive Fusion Network (PPF-Net) for robust 6D object pose estimation.
• To demonstrate an efficient frontend feature extraction strategy for improved pose estimation accuracy.

Main Methods:

• Developed Pose-Perceptive Convolution (PPC) that dynamically adapts receptive field shape and sampling density.
• Constructed a Pose-Perceptive Fusion Network (PPF-Net) integrating PPC for feature extraction.
• Evaluated PPF-Net on four benchmark datasets, including MP6D and YCB-Video.

Main Results:

• PPF-Net achieved a 19.4% VSD score improvement over FFB6D on MP6D.
• Attained 96.7% ADD-S accuracy on YCB-Video, nearing state-of-the-art.
• Demonstrated significant accuracy gains with minimal computational overhead.

Conclusions:

• Pose-Perceptive Convolution effectively resolves geometric mismatches in 6D object pose estimation.
• PPF-Net offers a robust and computationally efficient solution for accurate 6D pose estimation.
• Frontend feature extraction is an efficient strategy for improving 6D pose estimation robustness.