Reading #12: Constellation Models for Sketch Recognition

Comments:
Sam!

Summary:
A constellation model is a 'pictorial structure' model used to recognize strokes of particular classes. Each model is trained with labeled data in order to provide a higher probability of a successful match with testing sketches. This model works by looking for required and optional parts of each sketch with consideration to how each part is related to others. It is based on two key assumptions: a single instance of a mandatory part is in a sketch, and that similar parts will be drawn with similar strokes.

The constellation model was tested on facial recognition. Parts of a face (eyes, mouth, beard) were checked for both existence and spacial relation to other parts. A probability distribution for each object was calculated by training the recognizer with labeled data. A maximum likelihood search is then run to determine what an object 'is'. A sketch is checked multiple times as new objects are labeled so as to take advantage of the relational nature of the recognizer.

Discussion:
At first, I did not understand how only one of a required object could ever get the job done. Cyclops-only facial recognizer? But the authors state that each eye is treated as a different required object, thus bypassing this limitation. If the authors carry out their idea of having primitives be constructed from multiple strokes, then this model-based approach would afford a larger degree of freedom. Regardless, I like the idea.