DPCN++: Differentiable Phase Correlation Network for Versatile Pose Registration.

Area of Science:

• Computer Vision
• Robotics
• 3D Reconstruction

Background:

• Pose registration is crucial for tasks in vision and robotics.
• Existing methods often require initialization or predefined correspondences.
• Handling heterogeneous measurements presents a significant challenge.

Purpose of the Study:

• To develop an initialization-free pose registration method for up to 7 Degrees of Freedom (7DoF).
• To address the challenge of registering homogeneous and heterogeneous measurements.
• To improve upon existing learning-based and classical pose registration techniques.

Main Methods:

• Propose DPCN++, a differentiable solver combined with feature extraction networks.
• Utilize phase correlation in the spectral domain for correspondence-free and initialization-free registration.
• Employ Fourier transform and spherical radial aggregation for translation and scale invariant spectrum representation.
• Independently estimate rotation, scale, and translation in the spectrum.
• Train the entire pipeline end-to-end.

Main Results:

• DPCN++ demonstrates strong generalization on unseen objects.
• The method successfully performs registration for both homogeneous and heterogeneous inputs.
• Outperforms classical and learning-based baselines across various tasks and modalities.
• Achieves superior performance on partially observed and heterogeneous measurements.

Conclusions:

• DPCN++ offers a robust and efficient solution for initialization-free pose registration.
• The spectral domain approach effectively decouples rotation, scale, and translation.
• The method shows significant potential for diverse applications including 2D images, 3D data, and medical imaging.