1. Active Appearance Models based on the article: T.F.Cootes, G.J. Edwards and C.J.Taylor. "Active Appearance Models", 1998. presented by Denis Simakov

2. Mission Image interpretation by synthesis Model that yields photo-realistic objects Rapid, accurate and robust algorithm for interpretation Optimization using standard methods is too slow for realistic 50-100 parameter models Variety of applications Face, human body, medical images, test animals

3. Building an Appearance Model

4. Appearance Appearance = Shape + Texture Shape: tuple of characteristic locations in the image, up to allowed transformation Example: contours of the face up to 2D similarity transformation (translation, rotation, scaling) Texture: intensity (or color) patch of an image in the shape-normalized frame, up to scale and offset of values

5. Shape Configuration of landmarks Good landmarks – points, consistently located in every image. Add also intermediate points Represent by vector of the coordinates: e.g. x=(x 1,...,x n, y 1,...,y n ) T for n 2D landmarks Configurations x and x' are considered to have the same shape * if they can be merged by an appropriate transformation T (registration) Shape distance – the distance after registration: * Theoretical approach to shape analysis: Ian Dryden (University of Nottingham)

6. Shape-free texture An attempt to eliminate texture variation due to different shape (“orthogonalization”) Given shape x and a target “normal” shape x' (typically the average one) we warp our image so that points of x move into the corresponding points of x' warped tobecomes

7. Modeling Appearance training set (annotated grayscale images) shape (tuple of locations)texture (shape-free) model of shapemodel of texture model of appearance PCA

8. Training set Annotated images Done manually, it is the most human time consuming and error prone part of building the models (Semi-) automatic methods are being developed ** ** A number of references is given in: T.F.Cootes and C.J.Taylor. “Statistical Models of Appearance for Computer Vision”, Feb 28 2001; pp. 62-65 * Example from: Active Shape Model Toolkit (for MATLAB), Visual Automation Ltd. *

9. Training sets for shape and texture models From the initial training set (annotated images) we obtain {x 1,...,x n } – set of shapes, {g 1,...,g n } – set of shape-free textures. We allow the following transformations: S for the shape: translation (t x,t y ), rotation , scaling s. T for the texture: scaling , offset  (T g = ( g –  1 )/  ). Align both sets using these transformations, by minimizing distance between shapes (textures) and their mean Iterative procedure: align all x i (g i ) to the current ( ), recalculate ( ) with new x i (g i ), repeat until convergence.

10. Examples of training sets Shapes The mean shape Textures * From the work of Mikkel B. Stegmann, Section for Image Analysis, The Technical University of Denmark *

11. Model of Shape Training set {x 1,...,x n } of aligned shapes Apply PCA to the training set Model of shape: where (the mean shape) and P s (matrix of eigenvectors) define the model; b s is a vector of parameters of the model. Range of variation of parameters: determined by the eigenvalues, e.g. Example: 3 modes of a shape model

12. Model of Texture Training set {g 1,...,g n } of shape-free normalized image patches Apply PCA Model of texture: Range of variation of parameters Example: 1st mode of a texture model:

13. Combining two models Joint parameter vector where the diagonal matrix W s accounts for different units of shape and texture parameters. Training set For every pair (x i,g i ) we obtain: Apply PCA to the training set {b 1,...,b n } Model for parameters: b = P c c, P c = [P cs |P cg ] T Finally, the combined model: where Q s = P s W s -1 P cs, Q s = P g P cg.

14. Examples (combined model) Several modes Color model (by Gareth Edwards) Self-portrait of the inventor His shapeA mode of the modelTim Cootes

15. Exploiting an Appearance Model

16. Generating synthetic images: example By varying parameters c in the appearance model we obtain synthetic images:

17. Active Appearance Model (AAM) Difference vector:  I = I i – I m I i – input (new) image; I m – model-generated (synthetic) image for the current estimation of parameters. Search for the best match Minimize  = |  I| 2, varying parameters of the model Approach: Given: 1)an appearance model, 2)a new image, 3)a starting approximation Find: the best matching synthetic image

18. Predicting difference of parameters Knowing the matching error  I, we want to obtain information how to improve parameters c Approximate this relation by  c = A  I Precompute A: Include into  c extra parameters: translations, rotations and scaling of shape; scaling and offset of gray levels (texture) Take  I in the shape-normalized frame i.e.  I =  g where textures are warped into the same shape Generate pairs (  c,  g) and estimate A by linear regression.

19. Checking the quality of linear prediction We can check our linear prediction  c = A  g by perturbing the model Translation along one axis In the multi-resolution model (L0 – full resolution, L1 and L2 – succesive levels of the Gaussian pyramid)

20. AAM Search Algorithm For the current estimate of parameters c 0 Evaluate the error vector  g 0 Predict displacement of the parameters:  c = A  g 0 Try new value c 1 = c 0 – k  c for k=1 Compute a new error vector  g 1 If |  g 1 |<|  g 0 | then accept c 1 as a new estimate If c 1 was not accepted, try k=1.5; 0.5; 0.25, etc. Iterate the following: until |  g| is no more improved. Then we declare convergence.

21. AAM search: examples Model of faceModel of hand From the work of Mikkel B. Stegmann, Section for Image Analysis, The Technical University of Denmark

22. AAM: tracking experiments Extension of AAM By Jörgen Ahlberg, Linköping University AAM Done with AAM-API (Mikkel B. Stegmann)

23. AAM: measuring performance Model, trained on 88 hand labeled images (about 200 pixels wide), was tested on other 100 images. Convergence rateProportion of correct convergences

24. View-Based Active Appearance Models based on the article: T.F. Cootes, K.N. Walker and C.J.Taylor, "View-Based Active Appearance Models", 2000. presented by Denis Simakov

25. View-Based Active Appearance Models Basic idea: to account for global variation using several more local models. For example: to model 180  horizontal head rotation exploiting models, responsible for small angle ranges

26. -40  - +40  ±(40  - 60  ) ±(60  - 110  ) View-Based Active Appearance Models One AAM (Active Appearance Model) succeeds to describe variations, as long as there no occlusions It appears that to deal with 180  head rotation only 5 models suffice (2 for profile views, 2 for half-profiles, 1 for the front view) Assuming symmetry, we need to build only 3 distinct models

27. Estimation of head orientation Assumed relation: where  is the viewing angle, c – parameters of the AAM; c 0, c c and c s are determined from the training set For the shape parameters this relation is theoretically justified; for the appearance parameters its adequacy follows from experiments Determining pose of a model instance Given c we calculate the view angle  : where is the pseudo-inverse of the matrix [c c |c s ] T

28. Tracking through wide angles Match the first frame to one of the models (choose the best). Taking the model instance from the previous frame as the first approximation, run AAM search algorithm to refine this estimation. Estimate head orientation (the angle). Decide if it’s time to switch to another model. In case of a model change, estimate new parameters from the old one, and refine by the AAM search. To track a face in a video sequence: Given several models of appearance, covering together a wide angle range, we can track faces through these angles

29. Tracking through wide angles: an experiment 15 previously unseen sequences of known people (20-30 frames each). Algorithm could manage 3 frames per second (PIII 450MHz)

30. Predicting new views Given a single view of a face, we want to rotate it to any other position. Within the scope of one model, a simple approach succeeds: Find the best match c of the view to the model. Determine orientation . Calculate “angle-free” component: c res = c – (c 0 +c c cos(  )+c s sin(  )) To reconstruct view at a new angle , use parameter: c(  ) = c 0 +c c cos(  )+c s sin(  ) + c res

31. Predicting new views: wide angles To move from one model to another, we have to learn the relationship between their parameters Let c i,j be the “angle-free” component (c res ) of the i’th person in the j’th model (an average one). Applying PCA for every model, we obtain: c i,j = c j +P j b i,j where c j is the mean of c i,j over all people i. Estimate relationship between b i,j for different models j and k by linear regression: b i,j = r j,k +R j,k b i,k Now we can reconstruct a new view:...

32. Predicting new views: wide angles (cont.) Now, given a view in model 1, we reconstruct view in model 2 as follows: Remove orientation: c i,1 = c – (c 0 +c c cos(  )+c s sin(  )) Project into the space of principle components of the model 1: b i,1 = P 1 T (c i,1 – c 1 ). Move to the model 2: b i,2 = r 2,1 +R 2,1 b i,1. Find the “angle-free” component in the model 2: c i,2 = c 2 +P 2 b i,2. Add orientation: c(  ) = c 0 +c c cos(  )+c s sin(  ) + c i,2.

33. Predicting new views: example Training set includes images of 14 people. A profile view of unseen person was used to predict half-profile and front views.