Conditional Random Fields (for labeling)
- Doesn’t use Neural Networks but is a Supervised Learning method.
- Probabilistic graphical model
- Node in the graph == Face in the mesh
- Naively: Edges in the graph to and from neighbors of the face in the mesh
For:
: prediction
: face area for face 
: edge length connecting face to face 
: taking into account face features:
- area
- normal
- curvature
- spin images, ...
: face-face relationships
Optimization problem: - solve with: - Gradient Descent - Other methods
Limitations #
- Labeling cannot distinguish instances (e.g., legs are labeled legs without notion of individual legs) (This is true for all part segmentation methods)
- Need sufficient training data
- Sensitivity to topology:
