3D Shape Segmentation

Overview #

Goal: Assign label to every Point / Face of the 3d shape.

PointNet can also be used for shape segmentation. There are also 3d Convolutional Neural Networks that do shape segmentation.

Traditional methods #

• K-Means
• Normalized Cuts
• Primitive Fitting
• Random Walks

Other methods #

• Conditional Random Fields (for labeling)

DL Methods #

Supervised Methods #

1. 3D CNN for Part Segmentation
2. PointNet
3. StructureNet
4. BAE-NET

Co-Segmentation #

• Given a collection of shapes, find a consistent way to segment all of them into their pieces

Unsupervised Methods #

Using shape descriptors #

• Manually defined
• Something like:
  • Geodesic distance to the bottom of the face
  • Angle of the normal of the face
  • …
• orientation inveriant

DL based #

• BAE-NET
• AdaCoSeg (weakly supervised)

Active Part Segmentation #

• Active Part Segmentation