Chapter Questions
Explain how human vision differs from computer vision. Why is the problem of image understanding so difficult when every small child "knows how to do it"?
Explain with examples the differences between lower and upper processing levels.
Explain the main ideas of the following image understanding control strategies; if possible, provide a block diagram. Specify their primary applicability within the image understanding process.
What is conflict resolution? When is it needed?
What is the difference between short-term and long-term memory?
Give a real-world example of an image understanding application (other than those in the text) in which each of bottom-up, top-doun and combined control strategies may be used.
Summarize the key location, feature extruction and matching phases of SIFT.
Take a linear 2D data set: add noise to selected data points until a least-squares fit fails to come close to 'correct' (many easily available packages will do this). What can you say about the perturbations you have made to achieve this?
What kind of information can be represented by point distribution models?
Explain the process of determining the modes of variation represented by point distribution models.
Considering information provided in Table 10.1, how many principal components must be used in the point distribution model to leave less than $5 \%$ variation unexplained?
What is the main difference between the PDM and AAM?
List and explain all steps needed for(a) AAM training and(b) AAM-based segmentation
Give the most obvious example of how statistical pattern recognition can be used in classification/labeling of multi-spectral satellite image data.
Give the essential rationale why a boosted cascade of classifiers is an efficient approach to object detection.
Considering AdaBoost, when and why do you stop boosting rounds?
Explain principles of the joint classification-regression process when using random forests for object detection and localization.
Explain the main strategy of discrete labeling Give a real-world example of image interpretation to which discrete labeling can be applied, providing details.
Explain why discrete relaxation is a special case of probabilistic relaxation.
Explain the principle of genetic image segmentation and interpretation. Why are primary and specific adjacency graphs used?
Name some shortcomings of conventional image segmentation based on region growing. and on semantic region growing. Does genetic image segmentation and interpretation solve these shortcomings? Which ones and how?
Define an order-k Markov model.
Define an order-k Hidden Markov model.
Define in the context of HMMs:$\bullet$ Evaluation$\bullet$ Decoding$\bullet$ Learning
Cite with illustration some problems in which equation (10.85) well describes a vision related problem.
Generate and plot some mixtures of 2,3 and 4 1D Gaussians. Demonstrate to yourself that recapturing the original parameters from observations of the mixture would be difficult.