Recovering structure from a single view Pinhole perspective projection

Transcription

1 EPIPOLAR GEOMETRY The slides are from several sources through James Hays (Brown); Silvio Savarese (U. of Michigan); Svetlana Lazebnik (U. Illinois); Bill Freeman and Antonio Torralba (MIT), including their own slides.

2 Recovering structure from a single view Pinhole perspective projection P p O w Calibration rig Scene C Camera K Why is it so difficult? Intrinsic ambiguity of the mapping from 3D to image in 2D. see example...

3 Is this an illusion of 3D to 2D? Courtesy slide S. Lazebnik

4 Why multiple views? Structure and depth are inherently ambiguous from single views. P1 P2 P1 =P2 Optical center

5 Epipolar constraint Geometry of two views constrains where the corresponding pixel for some image point in the first view must occur in the second view. It must be carved out by a plane connecting the world point and the optical centers.

6 Epipolar geometry Epipolar Line Epipolar Line Epipolar Plane Epipole Baseline Epipole

7 Epipolar geometry: terms Baseline: line joining the camera centers. Epipole: point of intersection of baseline with image plane. Epipolar plane: plane containing baseline and world point. Epipolar line: intersection of epipolar plane with the image plane. All epipolar lines in an image intersect at the epipole... or, an epipolar plane intersects the left and right image planes in epipolar lines. Why is the epipolar constraint useful?

8 Epipolar constraint Reduces the correspondence problem to a 1D search in the second image along an epipolar line. Image from Andrew Zisserman

9 Two examples: Slide credit: Kristen Grauman

10 Converging cameras have finite epipoles. Figure from Hartley & Zisserman

11 Parallel cameras have epipoles at infinity. X at infinity e 1 x 1 x 2 at infinity e 2 O 1 O 2 Baseline intersects the image plane at infinity. Epipoles are at infinity. Epipolar lines are parallel to x axis.

12 In parallel cameras search is only along x coord. Figure from Hartley & Zisserman Slide credit: Kristen Grauman

13 Motion perpendicular to image plane

14 Motion perpendicular to image plane forward

15 Forward translation first e 2 second O 2 e 1 O 1 The epipoles have same position in both images. Epipole here is called FOE (focus of expansion).

16 A 3x3 matrix connects the two 2D images. This matrix is called the Essential Matrix, E when image intrinsic parameters are known the Fundamental Matrix, F more the general uncalibrated case

17 Essential matrix: E x T E x = l 2 - Two views of the same object - Suppose we know the camera positions and camera matrices ==> E matrix - Given a point on left image, how can we find the corresponding point on right image?

18 1 v u M P P 0 M K I O O p p P R, T Epipolar Constraint - E matrix T K R ' M 1 v u P M P homogeous coordinates '

19 P p p O R, T O M K I 0 ' K and K are known (calibrated cameras) M' ' K R T M I 0 M ' R T

20 In the epipolar plane we have P p p O R, T O T ( R p ) Perpendicular to epipolar plane first camera coordinates p T T (R p ) 0

21 Cross products can be written as matrix multiplication....verify it The matrix derived from a is skew-symmetric. The matrix is rank 2. The null vector is along the vector a.

22 T E is from p' to p T T p' E p = 0 Essential matrix P some denote E as from image 2 to 1 second camera coordinate p p then R and t are from image 1 to 2 O R, T O p T T (R p ) 0 p T T R p 0 X E is a rank 2 matrix! E = essential matrix (Longuet-Higgins, 1981)

23 Essential matrix properties X T (x_1) E x_2 = 0 calibrated x 1 l 1 l 2 e 1 e 2 x 2 O 1 O 2 E x 2 is the epipolar line associated with x 2 (l 1 = E x 2 ) E T x 1 is the epipolar line associated with x 1 (l 2 = E T x 1 ) E is singular (rank two) -- two equal singular values are one. E e 2 = 0 and E T e 1 = 0 l e = (E x ) e = 0 valid for any x E is 3x3 matrix with 5 DOF: 3(R) + 3(t) -1(scale) T 1 T

24 Fundamental matrix: F x l = F T x - Uncalibrated cameras. - No additional information about the scene and camera is given ==> F matrix - Given a point on left image, how can I find the corresponding point on right image?

25 uncalibrated camera Epipolar Constraint - F matrix P uncalibrated camera M' = K'[R t] M p R t p O O P M P u p M KI 0 v 1 homogeous coord. unknown (3x4)

26 F matrix derived from E matrix p K 1 p P p K 1 ' p p p O O [T]!! x p T T R p 0 (K 1 p) T 1 T R K p 0 p T K T 1 T R K p 0 p T F p 0 rank 2

27 T F is from p' to p T T p' F p = 0 second camera coordinate Fundamental matrix P some denote F as from image 2 to 1 p R t p then R and t are from image 1 to 2 O O p T F p (Faugeras and Luong, 1992) 0 The fundamental matrix has a projective ambiguity. Two pairs ~ ~ of camera matrices (P, P') and (P, P') give the same F if ~ ~ P = PH and P' = P'H where H is a 4x4 nonsingular matrix.

28 Fundamental matrix properties fundamental matrix is much more used X T (x_1) F x_2 = 0 uncalibrated x 1 x 2 e 1 e 2 O 1 O 2 F x 2 is the epipolar line associated with x 2 (l 1 = F x 2 ) F T x 1 is the epipolar line associated with x 1 (l 2 = F T x 1 ) F is singular (rank two) F e 2 = 0 and F T e 1 = 0 F is 3x3 matrix with 7 DOF: 9-1(rank 2) - 1(scale)

29 The eight-point algorithm of F (linear) (Hartley, 1995) x = (u, v, 1) T, x = (u, v, 1) T Minimize: N ( x T i F xi ) 2 i= 1 under the constraint F 33 = 1 can be an other F_ij constraint also

30 Estimating F W Homogeneous system Rank 8 If N>8 Wf 0 A non-zero solution exists (unique) Lsq. solution by SVD f 1 rank 3 solution Fˆ f

31 Taking into account the rank-2 constraint. p T Fˆ p 0 ^ The estimated F have full rank (det(f) 0) but F should have rank=2 instead. ^ Find F that minimizes F Fˆ 0 Frobenius norm (*) subject to det(f)=0 Taking the first two s.v. and the three equal zero. (*) Sqrt root of the sum of squares of all entries

32 Example of F recovery Data courtesy of R. Mohr and B. Boufama.

33 This are large errors... Mean errors: 10.0pixel and 9.1pixel

34 The problem with eight-point algorithm Poor numerical conditioning. Can be fixed by rescaling the data before estimation. Can be used for any DLT type algorithm. More sophisticated nonlinear methods after the 8-point algorithm exist, but we will not cover.

35 RESCALING BY NORMALIZATION You have i = 1,..., n points x i. points is x = 1 n x i n i=1 The mean of these which is translated to the origin (0, 0) by the vector x. The new coordinated of a point are x i = x i x. Compute the mean squared distance of the points from the center a 2 = 1 n x i n x i i=1 and move the square norm equal to 2 by multiplying the components of the original point 2/a. This is the scaling. The translation are the mean coordinates with opposite sign multiplied with 2/a. In the homogeneous 2D coordinates 1 0 x 2 1 T = 0 1 x 2 Tx i = a a a 2D similarity transformation. 2 a (x i1 x 1 ) 2 a (x i2 x 2 ) 1

36 The normalized eight-point algorithm (Hartley, 1995) Center the image data at the origin, and scale it so the mean squared distance between the origin and the data points is 2 pixels. Use the eight-point algorithm to compute F from the normalized points, n_1 and n_2. Enforce the rank-2 constraint. For example, take SVD of F and throw out the smallest singular value. Transform fundamental matrix back to original units: if T and T are the normalizing transformations in the two images, than the fundamental matrix in original coordinates is T T F T. Isotropic translation (mean to origin) and scale in each image separately x_1 = T n_1 x_2 = T' n_2 T -T -1 (n_1) T F T' n_2 = 0 => final F given

37 With transformation Without transformation Mean errors: 10.0pixel and 9.1pixel Mean errors: 1.0pixel and 0.9pixel

38 Comparison of estimation algorithms 8-point Normalized 8-point Nonlinear least squares Av. Dist pixels 0.92 pixel 0.86 pixel Av. Dist pixels 0.85 pixel 0.80 pixel

39 From epipolar geometry to camera calibration Estimating the fundamental matrix is known as weak calibration. If we know the calibration matrices of the two cameras, we can estimate the essential matrix: E=K T FK (see F from E slide) The essential matrix can give us the relative rotation and translation between the cameras, with 5 point pairs.

40 Example: Parallel calibrated images X e 1 x 1 x 2 t e 2 y z x O 1 O 2 K 1 =K 2 = known Hint : E=? x parallel to O 1 O 2 R = I t = (T, 0, 0)

41 see cross-product as matrix, before X e 1 x 1 x 2 e 2 y z x O 1 O K 1 =K 2 = known x parallel to O 1 O 2 E=? E [ t ] R 0 0 T 0 T 0

42 X e 1 x 1 x 2 e 2 y z x O 1 O 2 Epipolar constraint reduces to y = y In stereo vision that will be a big help.