Presentation is loading. Please wait.

Presentation is loading. Please wait.

Active Appearance Models Computer examples A. Torralba T. F. Cootes, C.J. Taylor, G. J. Edwards M. B. Stegmann.

Similar presentations


Presentation on theme: "Active Appearance Models Computer examples A. Torralba T. F. Cootes, C.J. Taylor, G. J. Edwards M. B. Stegmann."— Presentation transcript:

1 Active Appearance Models Computer examples A. Torralba T. F. Cootes, C.J. Taylor, G. J. Edwards M. B. Stegmann

2 AAM = Analysis by synthesis Ingredients: 1) A database of annotated objects. 2) Synthesis method for generation of photo-realistic images from model parameters. 3) Analysis: extraction of model parameters from images.

3 Labeling the training data set is one of the main difficulties of the approach. RoboFaces 1) Toy training database

4 It is a function that applies a deformation to an image given a set of corresponding points: 2) Image warping y1 x1 The main building block of AAM is the image warping procedure. Synthesis method for generation of photo-realistic images from model parameters

5 The Matlab implementation is limited to convex objects but this is good enough for faces. = ImageWarp (,,, ) This function is used during the iterations of the AAM. background Background Original image 2) Image warping

6 We warp a “real” face into the roboFaces in order to have more realistic images. We have same modes of variation. 2) Upgrading the toy training database

7 Appearance Model Each image is represented as a collection of correspondence points (shape) and a texture image normalized in shape. Shape information (texture free) Texture information (shape free) Original image I x1 x2... xi = ImageWarp (,,, ) Original image Mean shape Shape free texture shape zeros Shape normalization is obtained by warping the image into the mean shape of the training database.

8 Shape model PCA of shape information for the training database: PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 + s 1 + s 2 + s 3 +... = Shape Mean shape Each shape can be decomposed as:

9 Texture model PCA of texture information for the training database: PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 Each texture (shape free) can be decomposed as: The PCA is done on the shape free images + t 1 + t 2 + t 3 = Shape free texture Mean texture

10 Original image + s 1 + s 2 + s 3 + = + t 1 + t 2 + t 3 = shape texture = ImageWarp (,,, ) Original image Mean shape Shape free texture shape zeros t s Appearance Model AAM uses an additional PCA, to reduce redundancy between texture and shape.

11 3) Active Appearance Model Search Given a “face” the model has to build an appearance model (shape + texture) that reproduces the original image. Shape = ? Texture = ? This is done in an iterative procedure that tries to minimize the reconstruction error.

12 Two elements of the iterative procedure: 1) given a set of shape parameters, warp input image into its shape free approximation: = ImageWarp (,,, ) estimated shape mean shape zeros Input image s i s i +1  s + s i t i +1  t + t i 3) ss tt = F - = The residual is function of errors in both shape and texture parameters t i Normalized input 2) measure the residual image and correct the appearance model.

13 ss tt = F - = Learning to correct model parameters t i Linear approximation: ss tt = A Column vector Matrix A is learned by adding perturbations to the parameters of the training set. The residual corresponds to the difference between the image obtained with the real parameters and the one perturbed.

14 ss = A vector s Learning to correct model parameters Each row of A s describes how the residual contributes to each shape mode: 1st row of A s 2nd row 3rd row 4th row 5th row 6th row Shape parameters:

15 Learning to correct model parameters tt A vector t = Texture parameters: Each row of A t describes how the residual contributes to each texture mode: 1st row of A t 2nd row 3rd row

16 Results Input image 5 10 1 Iter = Model Shape Residual 100 Convergence after 50 iterations

17 Results Even when the images have real parameters that deviate from the distribution of the training set, the algorithm seems to converge: Input image Model Shape

18 Adding priors to possible appearance parameters may prevent this. iter error


Download ppt "Active Appearance Models Computer examples A. Torralba T. F. Cootes, C.J. Taylor, G. J. Edwards M. B. Stegmann."

Similar presentations


Ads by Google