computer visioncs-courses.mines.edu/csci508/schedule/17/statisticalmodels.pdf · a. szymczak, w....

27
Computer Vision Colorado School of Mines Colorado School of Mines Computer Vision Professor William Hoff Dept of Electrical Engineering &Computer Science http://inside.mines.edu/~whoff/ 1

Upload: others

Post on 18-Jul-2020

1 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Colorado School of Mines

Computer Vision

Professor William HoffDept of Electrical Engineering &Computer Science

http://inside.mines.edu/~whoff/ 1

Page 2: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Statistical Models for Shape and Appearance

2

Note – some material for these slides came from• Computer Vision Algorithms and Applications, by Szeliski, Springer 2011• Computer vision: models, learning and inference, by Prince, Cambridge University, 

2012

Page 3: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Statistical Models for Shape and Appearance

• We often have knowledge about the possible ways that shape and appearance can vary

• When matching to a new image, we can use this to constrain the shape and possibly give us better results

3

Page 4: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Principal Components

• One way to model the variation in your data, is to use principal components

• Principal components:– Assume we have a population of measurements vectors – The covariance matrix is – Find the eigenvectors  and eigenvalues  of – The  are called “principal components”

• We can express any of the original vectors  as a linear combination of PCs:   

where , , … , ,     , , … ,

4

Page 5: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Generative Model

• This is a “generative” model (as opposed to a “discriminative” model)

• Given coefficients h (which represent the parameters of the model in the world), we can generate examples of measurements x:  

where  is normally distributed noise with covariance 

5

Page 6: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Approaches

• We’ll look at 3 approaches that use PC‐based statistical models:1. Appearance‐based (use images)

• Also called  “eigenimages”

2. Shape‐based (use contour points or surface points)• Also called “active shape models” or “point distribution models”

3. Shape and appearance‐based (use both)• Also called “active appearance models”

6

Page 7: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Appearance‐Based (eigenimages)

• We have a collection of images, I1..IK– We subtract off the mean of the collection of images

• We transform each MxN image Ii into a column vectorxi = [Ii(1,1), Ii(1,2), … Ii(1,N), Ii(2,1), … Ii(M,N)]T

7

I1I2

I3

IK

• Find the covariance matrix Cx

• The eigenvectors (principal components) represent “basis” images (or “eigenimages”)

x1 x2 x3 xK

Page 8: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Application ‐ Face images

from: http://www.cs.princeton.edu/~cdecoro/eigenfaces/

Input images (Yale database)

8

Page 9: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Eigenfaces

• 10 PC’s capture 36% of the variance

• 25 PC’s capture 56% of the variance

9

Page 10: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Reconstruction with a small number of Principal Components

Each new image from left to right corresponds to using 1 additional PC for reconstruction

Each new image from left to right corresponds to using 8 additional PC for reconstruction

10

Page 11: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Recognition

• To recognize a new image x, project it onto the linear subspace spanned by the eigenimages (i.e., find the coefficients  ):

11

DFFS:  test if x is a face

DIFS:  measure how close x is to a training image face 

Page 12: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Shape‐based (Active Shape Models)

• Represent a shape with a set of points– For example, evenly spaced points along a contour

• Let x be the vector consisting of the set of points• We compute the covariance of a set of training vectors, and 

get the eigenvectors (principal components)

12

Page 13: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Examples of contours

• Resistors

13

Page 14: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

• Hand 

14

Page 15: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Recognition 

• We assume that the coefficients of the PCs (the shape parameters) are normally distributed

• To recognize a new shape x, we project it onto the linear subspace formed by the PCs, and compute the Mahalanobisdistances to each category center

15

Page 16: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Alignment of shapes

• The shapes must be registered and points sampled in the same places

16

Page 17: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Registration to image

• To register a point distribution model with an image, each control point searches in a direction normal to the contour to find the most likely corresponding image edge point

• Then search over (a) the shape parameters and (b) the parameters of a transform to align the shapes

17

Page 18: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Examples

18

Page 19: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines19

Page 20: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

3D Shape Models

• Example:  Creating bone models from 2D silhouettes using a statistical atlas 

20

Josh King (MS 2007), Colorado School of Mines.  Thesis title:  “Bi‐Planar Image Registration and Modeling of Bones”

A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using Conjugate Gradient Method,” Medical Imaging 2010: Image Processing. Proceedings of the SPIE, Volume 7623, pp. 762316‐762316‐8, 2010.

Page 21: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Shape and Appearance‐Based

• You get better results if you model both shape andappearance– This is called an “Active Appearance Model (AAM)”

• We use principal components for shape (encoded by a set of control points) and appearance (ie, the image texture)– The texture is normalized to a canonical shape before being analyzed

21

Page 22: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Active Appearance Model

• Both shape and texture are represented as deviations from a mean shape  and mean texture 

• Note – the same appearance parameters h simultaneously control both the shape and texture deformations – these are correlated!

22

Page 23: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Example – face modeling

23

Page 24: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines24

Page 25: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Warping images

25Computer vision: models, learning and inference.  ©2011 Simon J.D. Prince

To align the image texture according to shape displacements, use a piecewise affine transformation

Triangulate image points using Delaunay triangulation.

Image in each triangle is warped by an affine transformation. 

Page 26: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Registering a new image

26Computer vision: models, learning and inference.  ©2011 Simon J.D. Prince

To register the face model to a new image,  search over (a) the appearance parameters and (b) the warping parameters of a transform to align the shapes

Page 27: Computer Visioncs-courses.mines.edu/csci508/schedule/17/StatisticalModels.pdf · A. Szymczak, W. Hoff and M. Mahfouz, “3D Shape from Silhouette Points in Registered 2D Images Using

Computer VisionColorado School of Mines

Example

• AAMs have been used directly for recognition, but their main use is to align faces into a canonical pose so that other methods of face recognition can be used

27