Tutorials on the Shape modelling toolbox

Back to AAMToolbox Documentation

The models shown in these tutorials illustrate features of the AAMToolbox software. They are not designed to understand the shape and appearance modelling which is better done from the published literature for example.
Viewing these pages. Some versions of Firefox and Explorer do not create satisfactory prints even though you can view the pages with no problems. Chrome does appear to produce good printouts.

Fives ways to use AAMToolbox

1) Analysing shapes. i.e. the arrangement of points around a shape

2) Viewing the data in shape space. i.e. approximating the data with two principle components

3) Comparing shapes from samples of different groups for example, comparing faces from different cartoon characters

4) Analysing shape and appearance. In addition to the points around a shape, analyse the appearance (grey scale or colour) within the shape.

5) Analysing 3D shapes

How to use these tutorials. First download and install the AAMToolbox. A zip file containing the project (PRJ_CartoonFaces) is available here. Download and unzip into a directory. Then, from Matlab, change directory into the project and launch the AAMToolbox

cd PRJ_CartoonFaces
AAMToolbox 

This project contains as set of faces that have been analysed using 2D shape models

• The set of faces
• 
• 
• 
• 
• 
• 
• 
• 

1 How to analyse 2D shapes using the Graphical User Interface

The process of analysing a set of images is:-

Create a new project. AAMToolbox project names are automatically prefaced with PRJ_. They have a particular directory structure and the images to be analysed need to be copied into the subdirectory called Cropped. It is best if they are all the same size. Tutorial on making a new project. Create a point model template. Points are placed around the object of interest, i.e. around a face or leaf. The set of points constitute the point model. Every image will be marked up in the same way. Tutorial on point model template. To digitise each image, move the points to the corresponding positions in each image in turn. The positions must correspond to the same material points in each image, i.e. the tip of the leaf, the corner of an eye, or halfway along a line between the two ends of the mouth. Tutorial on point model editor. Generate the shape model using principal component analysis (PCA) Tutorial on statistical model generator. View the result by varying each important component in turn. We call this walking the shape model. This movie shows a walk. Tutorial on viewing statistical model. Best fit point model using only the principle components. Tutorial on fitting a model to a particular image.

<wikiflv width="300" height="300" logo="false" loop="true" background="white">CartoonPC1.flv|CartoonPC1.png</wikiflv> Mean shape (points) joined by lines. The movie shows deviations from the mean by varying the principle component.

2 Viewing the results in Shape-Space. What is Shape-Space?

PCA allows us to view the data from a new angle: Shape Space rather than normal viewing space. This movie shows a trajectory through shape space that is projected back into normal viewing space. Tutorial on shape space.

<wikiflv width="300" height="300" logo="false" loop="true" background="white">CartoonPC2.flv|PC2_cartoon_2.png</wikiflv> Interpolating a walk from one point model shape to another in shape space.

---

Seen these elsewhere?

3 Comparing shapes in Shape-Space.

None of the above portraits exist. They were created by placing point models on samples of each painters work and computing the PCA of both the shape and the appearance. Each image is the mean shape coloured with the mean appearance. They are arranged in date order: Giotto (~1300), Leonardo da Vinci (~1600), Valazquez (~1630), Rembrandt (~1650), Modigliani (~1900), Soutine (~1930). Compare shapes of different populations in shape space. Groups of samples can be projected into a common shape space and compared. In this case portraits. By focussing on the main contributions to shape it might be possible to relate positions in shape space to, for example, genotype in leaves (Langlade et al 2005.,Bensmihen et al. 2010) and petals (Whibley et al 2006,Feng et al. 2010)

Portraits by different painters projected into a shape space created by PCA of all the results. Conclusions?