3DPointCaps++: Learning 3D Representations with Capsule Networks.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Shape Analysis

Background:

• Conventional 3D generative models often require extensive supervision or annotations.
• Learning generalizable and robust 3D object representations remains a significant challenge.

Purpose of the Study:

• To develop a self-supervised method for learning disentangled 3D object representations.
• To enable robust part segmentation, interpolation, and correspondence estimation for 3D shapes.

Main Methods:

• Introduced 3DPointCaps++, a novel algorithm for unsupervised 3D representation learning.
• Employed a structured latent space with independent sub-spaces (capsules) for shape variations.
• Utilized a novel decoder with deconvolution operators and a cluster loss for self-supervised 3D point reconstruction.

Main Results:

• Demonstrated the ability to disentangle object parts within the latent space.
• Achieved robust performance in part segmentation, interpolation, and correspondence estimation.
• Showcased generalizability across rigid/non-rigid shapes and within/across categories.

Conclusions:

• 3DPointCaps++ learns effective and generalizable 3D object representations without heavy supervision.
• The proposed method offers a robust foundation for various downstream 3D computer vision tasks.