Quiz

In many practical applications, a target shape we want to analyse is given as an image. With the following questions you can check if you understand how this is done and why model fitting is so useful for image segmentation.

Question 1

Why is fitting shapes in images more difficult than fitting shapes that are represented as contours or surfaces? Multiple answers are possible.

Images are much more high-dimensional and hence the problem is computationally difficult.

Images are usually noisy and thus we might not find the right contour.

The contour points are not explicitly given but only represented as image intensities.

Question 2

A solution to the model fitting problem directly implies a solution to the segmentation problem. Why is this the case?

Because the fitted image intensities can be used to partition the images into object foreground and background.

Because the deformation field we find for a successful fit defines the contour of the target shape the image represents.

Because we have not only correspondence of the shape but also of the intensities.

Question 3

Why do we need a probabilistic model of the intensities and cannot use a fixed intensity pattern for describing the intensities around a point?

Because the intensities at two different points are not the same.

To become more robust against outliers.

Because the same point in two different images can have different intensities, depending on acquisition conditions, tissue properties, etc.

Question 4

Why is image segmentation such a difficult problem? Multiple answers are possible.

Depending on the acquisition device, the contour or surface of a shape is not always clearly visible.

Images are often noisy.

Different objects in an image can have the same intensities.

Question 5

In the Active Shape Model (ASM) algorithm, the intensities are often not modelled directly, but the gradient (i.e. the difference between intensities) is modelled instead. Why?

To make the model more robust against intensity variations.

Since we usually have a strong gradient at a contour, this will make it easier to detect the contour.

It reduces the effect of high-frequency noise in the image.