KeyPointDiffuser

Unsupervised 3D Keypoint Learning via Latent Diffusion Models

KeyPointDiffuser learns meaningful 3D keypoints from point clouds without labels. These keypoints act as a compact structural representation that guides a diffusion model to reconstruct and generate 3D shapes.

What it does

The model extracts consistent keypoints from 3D objects, helping identify important geometric structure such as wings, chair legs, handles, and object boundaries.

Results

Across ShapeNet object categories, KeyPointDiffuser improves keypoint consistency and supports smooth shape generation from learned keypoint representations.

Key Features

• Unsupervised 3D keypoint discovery
• Diffusion-based point cloud reconstruction
• Compact and interpretable latent representation
• Shape generation without reference meshes
• Smooth interpolation in keypoint space

Example: Shape Interpolation

Interpolating between keypoints produces smooth transitions between generated 3D shapes, showing that the learned representation captures meaningful geometry.

Authors

Rhys Newbury, Juyan Zhang, Tin Tran, Hanna Kurniawati, Dana Kulić

Monash University & Australian National University