CNS/EE 148 - Spring 2000
Syllabus
|
Week 1
|
Introduction to recognition |
|
Topics
|
Recognizing objects and object classes in images. Invariance with respect to deformations, occlusion, viewpoint and lighting conditions. Detection and recognition using matched filters. Evaluation of detector performance (Receiver Operating Characteristic). Achieving contrast, orientation and scale invariance. Limitations of matched filters. |
|
Homework
|
Gather image database. Experiment on a simple recognition problem using matched filtering. Discover where matched filter breaks down. |
|
Week 2
|
Statistical Pattern Recognition |
|
Topics
|
Bayesian decision theory. Mahalanobis distance. Estimation of unknown probability densities. Discrimination versus Representation. Class separability and feature selection. Fischer's Linear Discriminant. |
|
Homework
|
Different normalization schemes. Experiment with matched filters on deformable patterns. Use Fisher linear discriminant theory to build better feature detectors. |
|
Week 3
|
Constellation models |
|
Topics
|
Modeling deformable objects with a probabilistic model which captures global geometry and local photometry. Elements of decision theory. Detecting objects with missing features. |
|
Homework
|
Implementation of a very simple translation invariant constellation recognizer. |
|
Week 4
|
Feature Detection |
|
Topics
|
Generic feature detectors: Foerstner, Lucas-Tomasi-Kanade, edge detection. Specific feature detectors: Linear Filtering, PCA, neural networks. Scale and orientation invariance for features. |
|
Homework
|
Extension of the constellation model to use generic and specific feature detection. |
|
Week 5
|
Scale, orientation and shape invariance |
|
Topics
|
Representing constellations in shape space. Dryden-Mardia theorem. Introduction to EM for learning unknown pdfs. Estimating Dryden-Mardia densities. |
|
Homework
|
Extension of the constellation model to perform scale and orientation invariant recognition using a learned shape model. Implementation and use of EM in a toy example. |
|
Week 6
|
Efficient search |
|
Topics
|
Combinatorial explosion of hypothesis space. Efficient single-object search strategies. Searching large model databases using geometric hashing. |
|
Homework
|
|
Week 7
|
Unsupervised model learning |
|
Topics
|
Supervised vs unsupervised training of models for recognition. Unsupervised learning of features. Unsupervised learning of shape and model structure. |
|
Homework
|
Final version of recognition system which uses specific feature detectors which are learned along with geometry from a set of training data. The final system will be characterized with a second set of test data. |
|
Week 8
|
Recognizing in 3D |
|
Topics
|
View-based vs 3D models. Rotation invariance of view-based models. The canonical view of an object. Theories of human object recognition and experimental evidence. |
|
Homework
|
|
Week 9
|
Miscellanea |
|
Topics
|
Review of course material. Project presentations. Overview of the literature. |
|
Homework
|