This lecture set introduces a probabilistic approach to object class recognition, based on a Bayesian classifier that uses the properties that were previously extracted from the binary images. The Bayesian approach has wide applicability, even though the shapes used here are quite simple. The set finished with a geometric recognition matching approach based on aligning model shape edges to image edges.
This lecture introduces the five classes of object recognition algorithms, and a top-level model of the algorithm used.
This lecture introduces the probabilistic approach to recognizing an object, and the Bayes Classifier as the standard algorithm.
This lecture introduces the multivariate Gaussian distribution. This is useful because: 1) it encodes the variation in the observed values and 2) it allows one to combine different numerical properties in the same framework. The lecture concludes with an example of how the multivariate Gaussian distribution can be used in a Bayes Classifier.
This lecture segment gives a concrete example of the training and recognition algorithms, using Matlab. Almost all details are presented. Discussion of why data is split into training, validation and test sets is included.
This lecture demonstrates the performance of the Bayes classifier, using the invariant shape features, when classifying the 3 simple shapes. The lecture also introduces the concept of using a 'confusion matrix' to report the performance in a compact way.
This video describes a template matching approach to shape matching, where the edges from a known shape are swept across the image to find a good match.