Higher-Order Active Shape Models: A general model-based multiple-instance object modeling and recognition vision framework
Model-based approaches have been established as a robust way to tackle computer vision and pattern recognition problems. Their practical performances are highly dependent on how efficient the representation (abstraction) of real-world objects on a computer is. Indeed, useful object representations need be general, discriminative, transformation invariant, numerically stable and easily implementable on a standard computer.
Throughout my talk, I will discuss the use of the higher-order active shape (HOAS) methodology for the representation and recognition of objects in vision problems. HOASes are indeed a particular class of Markovian interaction models and their basic claim is that objects along with their transformations can be represented intrinsically and modeled as stable objects (ie. minima) using weighted interactions between all possible n-uples of their points (n>=2). Namely, I will focus on the geometry of objects in an object detection and recognition perspective.
Therefore, after briefly presenting the general HOAS framework and providing its main stability (optimality) result, I will describe in detail three of its classes:
- The traditional SOAC (2nd order) class and a proof of its limitation.
- The FOAC (4th order) class as a general Euclidean invariant shape modeling framework of homogenous resp. heterogeneous shapes.
- The extended SOAC (2nd class) as a general transformation invariant resp. heterogeneous shape modeling framework.
For each HOAS class, after presenting its theory and interpretation, I will describe the efficient implementation of both its learning (modeling) and operational (on-line) steps. I will also describe the strength and the weakness of each class. The talk will be supported by practical results on 2D IR and stereo images.
If time allows it, I will devote a third part of my talk to describing the convex relaxation of HOAS models and how the latter compares with other numerical resolution approaches (such as level-sets and graph-cuts).