EECS 442 Computer vision. Multiple view geometry Affine structure from Motion

Similar documents
Affine Structure from Motion

Two or more points can be used to describe a rigid body. This will eliminate the need to define rotational coordinates for the body!

Chimica Inorganica 3

STRAIGHT LINES & PLANES

Machine Learning for Data Science (CS 4786)

LECTURE 8: ORTHOGONALITY (CHAPTER 5 IN THE BOOK)

Notes The Incremental Motion Model:

Chapter 1. Complex Numbers. Dr. Pulak Sahoo

Matrix Algebra from a Statistician s Perspective BIOS 524/ Scalar multiple: ka

18.01 Calculus Jason Starr Fall 2005

Section 14. Simple linear regression.

R is a scalar defined as follows:

Math 21C Brian Osserman Practice Exam 2

Bivariate Sample Statistics Geog 210C Introduction to Spatial Data Analysis. Chris Funk. Lecture 7

Apply change-of-basis formula to rewrite x as a linear combination of eigenvectors v j.

Frequency Domain Filtering

Generalized Principal Component Analysis (GPCA)

AH Checklist (Unit 3) AH Checklist (Unit 3) Matrices

Singular value decomposition. Mathématiques appliquées (MATH0504-1) B. Dewals, Ch. Geuzaine

A widely used display of protein shapes is based on the coordinates of the alpha carbons - - C α

8. Applications To Linear Differential Equations

Lecture 8: October 20, Applications of SVD: least squares approximation

U8L1: Sec Equations of Lines in R 2

5.1 Review of Singular Value Decomposition (SVD)

Properties and Tests of Zeros of Polynomial Functions

U8L1: Sec Equations of Lines in R 2

24 MATH 101B: ALGEBRA II, PART D: REPRESENTATIONS OF GROUPS

Ma 530 Introduction to Power Series

Cov(aX, cy ) Var(X) Var(Y ) It is completely invariant to affine transformations: for any a, b, c, d R, ρ(ax + b, cy + d) = a.s. X i. as n.

Complex Analysis Spring 2001 Homework I Solution

Algebra of Least Squares

Centers of a Simplex. Contents

JEE(Advanced) 2018 TEST PAPER WITH SOLUTION (HELD ON SUNDAY 20 th MAY, 2018)

The Basic Space Model

Ma 530 Infinite Series I

Matrix Algebra 2.3 CHARACTERIZATIONS OF INVERTIBLE MATRICES Pearson Education, Inc.

[412] A TEST FOR HOMOGENEITY OF THE MARGINAL DISTRIBUTIONS IN A TWO-WAY CLASSIFICATION

CS 112 Transformations II. Slide 1

The Method of Least Squares. To understand least squares fitting of data.

The Pendulum. Purpose

Chapter 2. Periodic points of toral. automorphisms. 2.1 General introduction

CALCULUS BASIC SUMMER REVIEW

Linear regression. Daniel Hsu (COMS 4771) (y i x T i β)2 2πσ. 2 2σ 2. 1 n. (x T i β y i ) 2. 1 ˆβ arg min. β R n d

Geometry of LS. LECTURE 3 GEOMETRY OF LS, PROPERTIES OF σ 2, PARTITIONED REGRESSION, GOODNESS OF FIT

Machine Learning for Data Science (CS4786) Lecture 4

α x = α x for all scalars α

Mathematical Foundations -1- Sets and Sequences. Sets and Sequences

SAFE HANDS & IIT-ian's PACE EDT-10 (JEE) SOLUTIONS

Ellipsoid Method for Linear Programming made simple

Inverse Matrix. A meaning that matrix B is an inverse of matrix A.

Homework 1 Solutions. The exercises are from Foundations of Mathematical Analysis by Richard Johnsonbaugh and W.E. Pfaffenberger.

15.083J/6.859J Integer Optimization. Lecture 3: Methods to enhance formulations

Lecture #20. n ( x p i )1/p = max

Regn. No. North Delhi : 33-35, Mall Road, G.T.B. Nagar (Opp. Metro Gate No. 3), Delhi-09, Ph: ,

IIT JAM Mathematical Statistics (MS) 2006 SECTION A

Chapter 6 Principles of Data Reduction

Signal Processing in Mechatronics

(VII.A) Review of Orthogonality

2) 3 π. EAMCET Maths Practice Questions Examples with hints and short cuts from few important chapters

Machine Learning for Data Science (CS 4786)

Zeros of Polynomials

Math 451: Euclidean and Non-Euclidean Geometry MWF 3pm, Gasson 204 Homework 3 Solutions

NBHM QUESTION 2007 Section 1 : Algebra Q1. Let G be a group of order n. Which of the following conditions imply that G is abelian?

f(w) w z =R z a 0 a n a nz n Liouville s theorem, we see that Q is constant, which implies that P is constant, which is a contradiction.

3.2 Properties of Division 3.3 Zeros of Polynomials 3.4 Complex and Rational Zeros of Polynomials


MATH 205 HOMEWORK #2 OFFICIAL SOLUTION. (f + g)(x) = f(x) + g(x) = f( x) g( x) = (f + g)( x)

How to Maximize a Function without Really Trying

Chapter 4. Fourier Series

Definitions and Theorems. where x are the decision variables. c, b, and a are constant coefficients.

Optimization Methods MIT 2.098/6.255/ Final exam

Image Spaces. What might an image space be

ANALYSIS OF EXPERIMENTAL ERRORS

M 340L CS Homew ork Set 6 Solutions

M 340L CS Homew ork Set 6 Solutions

CHAPTER 5. Theory and Solution Using Matrix Techniques

Matrix Algebra 2.2 THE INVERSE OF A MATRIX Pearson Education, Inc.

Ω ). Then the following inequality takes place:

Notes 8 Singularities

A) is empty. B) is a finite set. C) can be a countably infinite set. D) can be an uncountable set.

9.3 Power Series: Taylor & Maclaurin Series

ALGEBRAIC GEOMETRY COURSE NOTES, LECTURE 5: SINGULARITIES.

Orthogonal transformations

The Jordan Normal Form: A General Approach to Solving Homogeneous Linear Systems. Mike Raugh. March 20, 2005

( ) ( ) ( ) notation: [ ]

The Boolean Ring of Intervals

Math 203A, Solution Set 8.

Improvement of Generic Attacks on the Rank Syndrome Decoding Problem

Feedback in Iterative Algorithms

Chapter 2 The Solution of Numerical Algebraic and Transcendental Equations

1 Duality revisited. AM 221: Advanced Optimization Spring 2016

Solutions to Homework 1

Symmetric Matrices and Quadratic Forms

Math Solutions to homework 6

Early Work by D Arcy Thompson

Distributional Similarity Models (cont.)

APPLICATION OF YOUNG S INEQUALITY TO VOLUMES OF CONVEX SETS

The inverse eigenvalue problem for symmetric doubly stochastic matrices

MASSACHUSETTS INSTITUTE OF TECHNOLOGY 6.265/15.070J Fall 2013 Lecture 21 11/27/2013

If a subset E of R contains no open interval, is it of zero measure? For instance, is the set of irrationals in [0, 1] is of measure zero?

Transcription:

EECS 442 Computer visio Multiple view geometry Affie structure from Motio - Affie structure from motio problem - Algebraic methods - Factorizatio methods Readig: [HZ] Chapters: 6,4,8 [FP] Chapter: 2 Some slides of this lectures are courtesy of prof. J. Poce, prof FF Li, prof S. Lazebik & prof. M. Hebert

Structure from motio problem X j j M m M mj 2j M 2 Give m images of fied 3D poits ij M i X j, i,, m, j,,

Structure from motio problem X j j M m M mj 2j M 2 From the m correspodeces ij, estimate: m projectio matrices M i 3D poits X j motio structure

Applicatios Courtesy of Oford Visual Geometry Group

Affie structure from motio (simpler problem) Image World Image From the m correspodeces ij, estimate: m projectio matrices M i (affie cameras) 3D poits X j

Fiite cameras p q r R Q P O [ ]X T K R T R K M 3 3 M Caoical perspective projectio matri Affie homography (i 3D) Affie Homography (i 2D) y s K o y o α α

Trasformatio i 2D Affiities: y H y t A y' ' a -Preserve: - Parallel lies - Ratio of areas - Ratio of legths o colliear lies - others - 6 DOF

Weak perspective projectio Whe the relative scee depth is small compared to its distace from the camera P ~ ' m y' my where m f z ' is the magificatio.

Weak perspective projectio Whe the relative scee depth is small compared to its distace from the camera ' m y' my Scalig fuctio of the distace (magificatio)

Orthographic (affie) projectio Whe the camera is at a (roughly costat) distace from the scee ' y' y Distace from ceter of projectio to image plae is ifiite

Orthographic (affie) projectio Whe the camera is at a (roughly costat) distace from the scee

Affie cameras [ ]X T K R T R K M y s K o y o α α Projective case Affie case

Trasformatio i 2D Projective: y H y b v t A y' ' p Affiities: y H y t A y' ' a

Affie cameras [ ]X T K R y α s α o y o K T R K M T R K M y s K o y o α α Projective case Affie case Parallel projectio matri (poits at ifiity are mapped as poits at ifiity) Magificatio (scalig term)

Weak perspective projectio Qigmig Festival by the Riverside Zhag Zedua ~9 AD

Affie cameras [ ]X T K R y K α α T R K M b A 4affie] [4 3affie] 3 [ 2 23 22 2 3 2 b a a a b a a a M + + 2 23 22 2 3 2 X b AX Euc M b b Z Y X a a a a a a y [Homogeeous] [o-homogeeous image coordiates] [ ] b A M M Euc ; P M Euc

Affie cameras p P p M camera matri To recap: from ow o we defie M as the camera matri for the affie case p u v AP + b M P ; M [ A b]

The Affie Structure-from-Motio Problem Give m images of fied poits P j (X i ) we ca write N of cameras N of poits Problem: estimate the m 2 4 matrices M i ad the positios P j from the m correspodeces p ij. How may equatios ad how may ukow? 2m equatios i 8m+3 ukows Two approaches: - Algebraic approach (affie epipolar geometry; estimate F; cameras; poits) - Factorizatio method

Algebraic aalysis (2-view case) - Derive the fudametal matri F A for the affie case - Compute F A - Use F A to estimate projectio matrices - Use projectio matrices to estimate 3D poits

. Derivig the fudametal matri F A p P v p u Homogeeous system Dim? 44

Derivig the fudametal matri F A where The Affie Fudametal Matri!

Affie Epipolar Geometry Note: the epipolar lies are parallel.

Estimatig F A From at least 4 correspodeces, we obtai a liear system o the ukow alpha, beta, etc Measuremets: u, u, v, v v u v u v u v u f M M M M M Computed by least square ad by eforcig f SVD

Estimatig projectio matrices from epipolar costraits p P p

Affie ambiguity Affie PX ( PQ - )( Q X) A A

Estimatig projectio matrices from epipolar costraits p P p

Estimatig projectio matrices from epipolar costraits Choose Q such that A ~ ~ b [ ] T A ~ a ~ b b [ d] T c Caoical affie cameras Fuctio of the parameters of F

Estimatig projectio matrices from epipolar costraits Choose Q such that By re-eforcig the epipolar costrait, we ca compute a, b, c, d directly from the measuremets

Remider: epipolar costrait p P v p u Homogeeous system

Estimatig projectio matrices from epipolar costraits Choose Q such that ~ A ~ b Re-eforce the Epipolar costrait

Estimatig projectio matrices from epipolar costraits Choose Q such that A b

Estimatig projectio matrices from epipolar costraits Liear relatioship betwee measuremets ad ukow Ukow: a, b, c, d Measuremets: u, u, v, v From at least 4 correspodeces, we ca solve this liear system ad compute a, b, c, d (via least square) The cameras ca be computed How about the structure?

4. Estimatig the structure from epipolar costraits A b Ca be solved by least square agai

A factorizatio method Tomasi & Kaade algorithm C. Tomasi ad T. Kaade. Shape ad motio from image streams uder orthography: A factorizatio method. IJCV, 9(2):37-54, November 992. Ceterig the data Factorizatio

Ceterig: subtract the cetroid of the image poits ( ) j i k k j i k i k i i j i k ik ij ij ˆ ˆ A X X X A b A X b A X + + A factorizatio method - Ceterig the data X k ik i ^

Ceterig: subtract the cetroid of the image poits ( ) j i k k j i k i k i i j i k ik ij ij ˆ ˆ A X X X A b A X b A X + + A factorizatio method - Ceterig the data

Ceterig: subtract the cetroid of the image poits ( ) + + k k j i k i k i i j i k ik ij ij ˆ X X A b A X b A X j i ij X A ˆ A factorizatio method - Ceterig the data Assume that the origi of the world coordiate system is at the cetroid of the 3D poits After ceterig, each ormalized poit ij is related to the 3D poit X i by

A factorizatio method - Ceterig the data X ˆ A ij i X j

Let s create a 2m data (measuremet) matri: m m m D ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ 2 2 22 2 2 L O L L cameras (2m ) poits ( ) A factorizatio method - factorizatio

Let s create a 2m data (measuremet) matri: [ ] m m m m X X X A A A D L M L O L L 2 2 2 2 22 2 2 ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ cameras (2m 3) poits (3 ) The measuremet matri D M S has rak 3 (it s a product of a 2m3 matri ad 3 matri) A factorizatio method - factorizatio (2m ) M S

Factorizig the measuremet matri Source: M. Hebert

Factorizig the measuremet matri Sigular value decompositio of D: Source: M. Hebert

Factorizig the measuremet matri Sigular value decompositio of D: Sice rak (D)3, there are oly 3 o-zero sigular values Source: M. Hebert

Factorizig the measuremet matri Obtaiig a factorizatio from SVD: Motio (cameras) structure What is the issue here? D has rak>3 because of - measuremet oise - affie approimatio

Factorizig the measuremet matri Obtaiig a factorizatio from SVD: structure D D

Affie ambiguity The decompositio is ot uique. We get the same D by usig ay 3 3 matri C ad applyig the trasformatios M MC, S C - S We ca eforce some Euclidea costraits to resolve this ambiguity (more o et lecture!)

Algorithm summary. Give: m images ad features ij 2. For each image i, ceter the feature coordiates 3. Costruct a 2m measuremet matri D: Colum j cotais the projectio of poit j i all views Row i cotais oe coordiate of the projectios of all the poits i image i 4. Factorize D: Compute SVD: D U W V T Create U 3 by takig the first 3 colums of U Create V 3 by takig the first 3 colums of V Create W 3 by takig the upper left 3 3 block of W 5. Create the motio ad shape matrices: M M U 3 ad S W 3 V 3 T (or U 3 W 3½ ad S W 3½ V 3T ) 6. Elimiate affie ambiguity

Recostructio results C. Tomasi ad T. Kaade. Shape ad motio from image streams uder orthography: A factorizatio method. IJCV, 9(2):37-54, November 992.

Net lecture Multiple view geometry Perspective structure from Motio