7.5
While manual selection is a common practice in object recognition systems, this chapter has presented a novel theory for automated, optimal parameter estimation in optimization-based object recognition. The theory is based on learning from examples. Mathematical principles of correctness and instability are established and defined for the evaluation of parameter estimates. A learning algorithm is presented for computing the optimal, i.e. minimal-instability, estimate. An application to the MRF-based recognition is given. Experiments conducted show very promising results. Optimal estimates automatically learned from examples can well be generalized for recognizing other scenes and objects.
The exemplary instances are given to reflect the designer's judgment of desirable solutions. However, a recognizer with a given functional form can not be trained by arbitrary exemplar. The exemplar should be selected properly to reflect the correct semantics, in other words, they should be consistent with the constraints with which the functional form is derived. Assuming the form of the objective function is right and the exemplar contains useful information, then the more exemplary instances are used to train, the more generalizable is the learned parameter estimate.
The learning procedure also provides a means of checking the validity of the energy function derived from mathematical models. An improper mathematical model leads to an improper functional form. If no correct parameter estimates can be learned, it is a diagnostic symptom that the assumptions used in the model are not suitable for modeling the reality of the scene. The procedure also provides useful information for feature selection. Components of the optimal parameter estimate will be zero or near zero for instable features.