BIL Computer Vision Apr 16, 2014

Transcription

1 BIL Computer Vision Apr 16, 2014 Binocular Stereo (cont d.), Structure from Motion Aykut Erdem Dept. of Computer Engineering Hacettepe University

2 Slide credit: S. Lazebnik Basic stereo matching algorithm For each pixel in the first image Find corresponding epipolar line in the right image Examine all pixels on the epipolar line and pick the best match Triangulate the matches to get depth information Simplest case: epipolar lines are scanlines When does this happen?

3 Slide credit: S. Lazebnik Simplest Case: Parallel images Image planes of cameras are parallel to each other and to the baseline Camera centers are at same height Focal lengths are the same

4 Slide credit: S. Lazebnik Simplest Case: Parallel images Image planes of cameras are parallel to each other and to the baseline Camera centers are at same height Focal lengths are the same Then, epipolar lines fall along the horizontal scan lines of the images

5 Essential matrix for parallel images R t E E x x T ] [, 0 = = " " # $ $ $ % & = = ] [ T T R t E R = I t = (T, 0, 0) Epipolar constraint: " # $ $ $ % & = ] [ x y x z y z a a a a a a a t x x Slide credit: S. Lazebnik

6 Essential matrix for parallel images R t E E x x T ] [, 0 = = " " # $ $ $ % & = = ] [ T T R t E ( ) ( ) v T Tv Tv T v u v u T T v u = = " " " # $ % % % & ' = " " " # $ % % % & ' ) ) ) * +,,, R = I t = (T, 0, 0) The y-coordinates of corresponding points are the same t x x Slide credit: S. Lazebnik Epipolar constraint:

7 Stereo image rectification Slide credit: S. Lazebnik. S. Seitz

8 Slide credit: S. Lazebnik. S. Seitz Stereo image rectification reproject image planes onto a common plane parallel to the line between optical centers pixel motion is horizontal after this transformation two homographies (3x3 transform), one for each input image reprojection C. Loop and Z. Zhang. Computing Rectifying Homographies for Stereo Vision. IEEE Conf. Computer Vision and Pattern Recognition, 1999.

9 Rectification example Slide credit: S. Lazebnik. S. Seitz

10 Why rectification is useful? Makes the correspondence problem easier Makes triangulation easy Slide credit: S. Savarese

11 Slide credit: S. Lazebnik. S. Seitz Depth from disparity X x x z f f O Baseline O B disparity = x x" B = z f Disparity is inversely proportional to depth

12 Disparity map / depth map Disparity map with occlusions Slide credit: S. Savarese Disparity maps Stereo pair

13 Slide credit: S. Lazebnik Correspondence search Left Right scanline Matching cost disparity Slide a window along the right scanline and compare contents of that window with the reference window in the left image Matching cost: SSD or normalized correlation

14 Correspondence search Left Right scanline SSD Slide credit: S. Lazebnik

15 Correlation Method Pick up a window around p(u,v) Build vector W Slide the window along v line in image 2 and compute w Keep sliding until w w is maximized. Slide credit: S. Savarese

16 Slide credit: S. Savarese Correlation Method Normalized Correlation; minimize:

17 Correspondence search Left Right scanline Norm. corr Slide credit: S. Lazebnik

18 Slide credit: S. Lazebnik, S. Seitz Effect of window size Smaller window + More detail More noise Larger window + Smoother disparity maps Less detail W = 3 W = 20

19 Failures of correspondence search Textureless surfaces Occlusions, repetition Non-Lambertian surfaces, specularities Slide credit: S. Lazebnik. S. Seitz

20 Slide credit: S. Seitz Stereo results Data from University of Tsukuba Similar results on other images without ground truth Scene Ground truth

21 Slide credit: S. Lazebnik. S. Seitz Results with window search Data Window-based matching (best window size) Ground truth

22 Better methods exist... Graph cuts Ground truth Y. Boykov, O. Veksler, and R. Zabih, Fast Approximate Energy Minimization via Graph Cuts, PAMI 2001 For the latest and greatest: Slide credit: S. Lazebnik. S. Seitz

23 Slide credit: S. Lazebnik. S. Seitz How can we improve window-based matching? The similarity constraint is local (each reference window is matched independently) Need to enforce non-local correspondence constraints

24 Slide credit: S. Lazebnik. S. Seitz Non-local constraints Uniqueness For any point in one image, there should be at most one matching point in the other image

25 Slide credit: S. Lazebnik. S. Seitz Non-local constraints Uniqueness For any point in one image, there should be at most one matching point in the other image Ordering Corresponding points should be in the same order in both views

26 Ordering constraint doesn t hold Slide credit: S. Lazebnik. S. Seitz Non-local constraints Uniqueness For any point in one image, there should be at most one matching point in the other image Ordering Corresponding points should be in the same order in both views

27 Slide credit: S. Lazebnik. S. Seitz Non-local constraints Uniqueness For any point in one image, there should be at most one matching point in the other image Ordering Corresponding points should be in the same order in both views Smoothness We expect disparity values to change slowly (for the most part)

28 Slide credit: S. Lazebnik. S. Seitz Scanline stereo Try to coherently match pixels on the entire scanline Different scanlines are still optimized independently Left image Right image

29 Shortest paths for scan-line stereo Left image I Right image I S left Right occlusion q t Left occlusion correspondence C occl s p S right C occl C corr Can be implemented with dynamic programming Ohta & Kanade 85, Cox et al. 96 Slide credit: Y. Boykov Slide credit: S. Lazebnik. S. Seitz

30 Stereo as energy minimization What defines a good stereo correspondence? 1. Match quality Want each pixel to find a good match in the other image 2. Smoothness If two pixels are adjacent, they should (usually) move about the same amount Slide credit: S. Seitz

31 Smoothness Slide credit: S. Savarese

32 Stereo matching as energy minimization I 1 I 2 D W 1 (i ) W 2 (i+d(i )) D(i ) 2 ( W ) 1 ( i) W2 ( i + D( i)) + λ ( D( i) D( j) ) E( D) = ρ i neighbors i, j data term smoothness term Energy functions of this form can be minimized using graph cuts Y. Boykov, O. Veksler, and R. Zabih, Fast Approximate Energy Minimization via Graph Cuts, PAMI 2001 Slide credit: S. Lazebnik

33 Stereo matching as energy minimization

34 L. Zhang, B. Curless, and S. M. Seitz. Rapid Shape Acquisition Using Color Structured Light and Multi-pass Dynamic Programming. 3DPVT 2002 Slide credit: S. Lazebnik Active stereo with structured light Project structured light patterns onto the object Simplifies the correspondence problem Allows us to use only one camera camera projector

35 Active stereo with structured light L. Zhang, B. Curless, and S. M. Seitz. Rapid Shape Acquisition Using Color Structured Light and Multi-pass Dynamic Programming. 3DPVT 2002 Slide credit: S. Lazebnik

36 Active stereo with structured light Slide credit: S. Lazebnik

37 Kinect: Structured infrared light Slide credit: S. Lazebnik. S. Seitz

38 Slide credit: S. Lazebnik. S. Seitz Laser scanning Digital Michelangelo Project Levoy et al. Optical triangulation Project a single stripe of laser light Scan it across the surface of the object This is a very precise version of structured light scanning

39 Slide credit: S. Seitz Laser scanned models The Digital Michelangelo Project, Levoy et al.

40 Laser scanned models The Digital Michelangelo Project, Levoy et al. Slide credit: S. Seitz

41 Slide credit: S. Seitz Laser scanned models The Digital Michelangelo Project, Levoy et al.

42 Beyond two-view stereo Slide credit: S. Lazebnik

43 Slide credit: B. Freeman and A. Torralba Structure from motion Multiple views of a static scene from different cameras One camera, a moving object

44 Slide credit: N. Snavely Volumetric stereo Scene Volume V Input Images (Calibrated) Goal: Determine occupancy, color of points in V

45 Discrete formulation: Voxel Coloring Discretized Scene Volume Input Images (Calibrated) Goal: Assign RGBA values to voxels in V photo-consistent with images Slide credit: N. Snavely

46 Complexity and computability Discretized Scene Volume 3 N voxels C colors True Scene Photo-Consistent Scenes All Scenes (C N3 ) Slide credit: N. Snavely

47 K.N. Kutulakos and S.M. Seitz, A Theory of Shape and Space Carving, ICCV 1999 Slide credit: S. Lazebnik Space Carving Image 1 Image N... Space Carving Algorithm Initialize to a volume V containing the true scene Choose a voxel on the outside of the volume Project to visible input images Carve if not photo-consistent Repeat until convergence

48 Space Carving Results: African Violet Input Input Image (1 (1 of 45) of 45) Reconstruction Reconstruction Reconstruction Slide credit: S. Seitz

49 Structure from motion Given a set of corresponding points in two or more images, compute the camera parameters and the 3D point coordinates? Camera 1? Camera 2 Camera 3 R 1,t 1?? R 2,t R 2 3,t 3 Slide credit: N. Snavely

50 Slide credit: S. Lazebnik Structure from motion Given: m images of n fixed 3D points x ij = P i X j, i = 1,, m, j = 1,, n Problem: estimate m projection matrices P i and n 3D points X j from the mn correspondences x ij X j x 1j x 2j P 1 P 3 x 3j

51 Slide credit: S. Lazebnik Structure from motion ambiguity If we scale the entire scene by some factor k and, at the same time, scale the camera matrices by the factor of 1/k, the projections of the scene points in the image remain exactly the same: x = PX = & $ % 1 k # P ( kx) " It is impossible to recover the absolute scale of the scene

52 Slide credit: S. Lazebnik Structure from motion ambiguity If we scale the entire scene by some factor k and, at the same time, scale the camera matrices by the factor of 1/ k, the projections of the scene points in the image remain exactly the same More generally: if we transform the scene using a transformation Q and apply the inverse transformation to the camera matrices, then the images do not change x = PX = ( PQ -1)( QX)

53 With no constraints on the camera calibration matrix or on the scene, we get a projective reconstruction Need additional information to upgrade the reconstruction to affine, similarity, or Euclidean Slide credit: S. Lazebnik Types of ambiguity Projective 15dof & A $ T % v t# v " Preserves intersection and tangency Affine 12dof & A $ T % 0 t# 1 " Preserves parallellism, volume ratios Similarity 7dof & s R $ T % 0 t# 1 " Preserves angles, ratios of length Euclidean 6dof & R $ T % 0 t# 1 " Preserves angles, lengths

54 Slide credit: S. Lazebnik Projective ambiguity Q p = & A $ T % v t# v " x = PX = ( PQ -1)( Q X) P P

55 Projective ambiguity Slide credit: S. Lazebnik

56 Slide credit: S. Lazebnik Affine ambiguity Affine Q A = & A $ T % 0 t# 1 " x = PX = ( PQ -1 )( Q X) A A

57 Affine ambiguity Slide credit: S. Lazebnik

58 Slide credit: S. Lazebnik Similarity ambiguity Q s = & sr $ T % 0 t# 1 " x = PX = ( PQ -1)( Q X) S S

59 Similarity ambiguity Slide credit: S. Lazebnik

60 Slide credit: S. Lazebnik Structure from motion Let s start with affine cameras (the math is easier) center at infinity

61 Recall: Orthographic Projection Special case of perspective projection Distance from center of projection to image plane is infinite Projection matrix: Image World Slide credit: S. Seitz

62 Slide credit: S. Lazebnik Affine cameras Orthographic Projection Parallel Projection

63 Affine cameras A general affine camera combines the effects of an affine transformation of the 3D space, orthographic projection, and an affine transformation of the image: Affine projection is a linear mapping + translation in inhomogeneous coordinates " # $ % & = " # $ $ $ % & = " # $ $ $ % & = 1 0 b A P affine] [ affine] 3 [ b a a a b a a a x X a 1 a 2 b AX x + = " # $ $ % & + " # $ $ $ % & ' ( ) * +, = " # $ % & = b b Z Y X a a a a a a y x Projection of world origin Slide credit: S. Lazebnik

64 Affine structure from motion Given: m images of n fixed 3D points: x ij = A i X + b j i, i = 1,, m, j = 1,, n Problem: use the mn correspondences x ij to estimate m projection matrices A i and translation vectors b i, and n points X j The reconstruction is defined up to an arbitrary affine transformation Q (12 degrees of freedom): - A 1 +, 0 b* 1 ( ) -A +, 0 b* ( Q 1), & X# Q$ % 1 " We have 2mn knowns and 8m + 3n unknowns (minus 12 dof for affine ambiguity) Thus, we must have 2mn >= 8m + 3n 12 For two views, we need four point correspondences & $ % X 1 # " Slide credit: S. Lazebnik

65 Affine structure from motion Centering: subtract the centroid of the image points For simplicity, assume that the origin of the world coordinate system is at the centroid of the 3D points After centering, each normalized point x ij is related to the 3D point X i by ( ) j i n k k j i n k i k i i j i n k ik ij ij n n n A X X X A b A X b A X x x x ˆ ˆ = " # $ % & = + + = = = = = j i ij X x = A ˆ Slide credit: S. Lazebnik

66 C. Tomasi and T. Kanade. Shape and motion from image streams under orthography: A factorization method. IJCV, 9(2): , November Slide credit: S. Lazebnik Affine structure from motion Let s create a 2m n data (measurement) matrix: D = & xˆ $ $ xˆ $ $ % xˆ m1 xˆ xˆ xˆ m2 " xˆ xˆ xˆ 1n 2n mn # " cameras (2 m) points (n)

67 Affine structure from motion Let s create a 2m n data (measurement) matrix: [ ] n m mn m m n n X X X A A A x x x x x x x x x D " # ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ ˆ " # $ $ $ $ % & = " # $ $ $ $ % & = cameras (2 m 3) points (3 n) The measurement matrix D = MS must have rank 3 (Because the motion matrix M and the structure matrix S are rank 3) C. Tomasi and T. Kanade. Shape and motion from image streams under orthography: A factorization method. IJCV, 9(2): , November Slide credit: S. Lazebnik

68 Factorizing the measurement matrix Slide credit: M. Hebert

69 Slide credit: M. Hebert Factorizing the measurement matrix Singular value decomposition of D:

70 Slide credit: M. Hebert Factorizing the measurement matrix Singular value decomposition of D:

71 Factorizing the measurement matrix Obtaining a factorization from SVD: Slide credit: M. Hebert

72 Factorizing the measurement matrix Obtaining a factorization from SVD: This decomposition minimizes D-MS 2 Slide credit: M. Hebert

73 Affine ambiguity The decomposition is not unique. We get the same D by using any 3 3 matrix C and applying the transformations M MC, S C -1 S That is because we have only an affine transformation and we have not enforced any Euclidean constraints (like forcing the image axes to be perpendicular, for example) Slide credit: M. Hebert

74 Eliminating the affine ambiguity Orthographic: image axes are perpendicular and scale is 1 a 1 a 2 = 0 x a 1 2 = a 2 2 = 1 a 2 a 1 This translates into 3m equations in L = CC T : A i L A i T = Id, i = 1,, m X Solve for L Recover C from L by Cholesky decomposition: L = CC T Update M and S: M = MC, S = C -1 S Slide credit: M. Hebert

75 Algorithm summary Given: m images and n features x ij For each image i, center the feature coordinates Construct a 2m n measurement matrix D: Column j contains the projection of point j in all views Row i contains one coordinate of the projections of all the n points in image i Factorize D: Compute SVD: D = U W V T Create U 3 by taking the first 3 columns of U Create V 3 by taking the first 3 columns of V Create W 3 by taking the upper left 3 3 block of W Create the motion and shape matrices: M = U 3 W 3 ½ and S = W 3 ½ V 3 T (or M = U 3 and S = W 3 V 3T ) Eliminate affine ambiguity Slide credit: M. Hebert

76 Reconstruction results C. Tomasi and T. Kanade. Shape and motion from image streams under orthography: A factorization method. IJCV, 9(2): , November Slide credit: S. Lazebnik

77 Slide credit: S. Lazebnik Dealing with missing data So far, we have assumed that all points are visible in all views In reality, the measurement matrix typically looks something like this: points cameras

78 Dealing with missing data Possible solution: decompose matrix into dense sub-blocks, factorize each sub-block, and fuse the results Finding dense maximal sub-blocks of the matrix is NP-complete (equivalent to finding maximal cliques in a graph) Incremental bilinear refinement (1) Perform factorization on a dense sub-block (2) Solve for a new 3D point visible by at least two known cameras (linear least squares) (3) Solve for a new camera that sees at least three known 3D points (linear least squares) F. Rothganger, S. Lazebnik, C. Schmid, and J. Ponce. Segmenting, Modeling, and Matching Video Clips Containing Multiple Moving Objects. IEEE PAMI Slide credit: S. Lazebnik

79 Projective structure from motion Given: m images of n fixed 3D points z ij x ij = P i X j, i = 1,, m, j = 1,, n Problem: estimate m projection matrices P i and n 3D X j points X j from the mn correspondences x ij x 1j x 3j P 1 x 2j P 2 P 3 Slide credit: S. Lazebnik

80 Slide credit: S. Lazebnik Projective structure from motion Given: m images of n fixed 3D points z ij x ij = P i X j, i = 1,, m, j = 1,, n Problem: estimate m projection matrices P i and n 3D points X j from the mn correspondences x ij With no calibration info, cameras and points can only be recovered up to a 4x4 projective transformation Q: X QX, P PQ -1 We can solve for structure and motion when 2mn >= 11m +3n 15 For two cameras, at least 7 points are needed

81 Projective SFM: Two-camera case Compute fundamental matrix F between the two views First camera matrix: [I 0] Second camera matrix: [A b] Then b is the epipole (F T b = 0), A = [b ]F F&P sec Slide credit: S. Lazebnik

82 Sequential structure from motion Initialize motion from two images using fundamental matrix Initialize structure by triangulation points For each additional view: Determine projection matrix of new camera using all the known 3D points that are visible in its image calibration cameras Slide credit: S. Lazebnik

83 Sequential structure from motion Initialize motion from two images using fundamental matrix Initialize structure by triangulation points For each additional view: Determine projection matrix of new camera using all the known 3D points that are visible in its image calibration Refine and extend structure: compute new 3D points, re-optimize existing points that are also seen by this camera triangulation cameras Slide credit: S. Lazebnik

84 Slide credit: S. Lazebnik Sequential structure from motion Initialize motion from two images using fundamental matrix Initialize structure by triangulation For each additional view: Determine projection matrix of new camera using all the known 3D points that are visible in its image calibration Refine and extend structure: compute new 3D points, re-optimize existing points that are also seen by this camera triangulation Refine structure and motion: bundle adjustment cameras points

85 Bundle adjustment Non-linear method for refining structure and motion Minimizing reprojection error m E( P, X) = D n i= 1 j= 1 X j ( ) 2 x, P X ij i j P 1 X j x 1j x 3j P 1 P 2 X j x 2j P 3 X j P 2 P 3 Slide credit: S. Lazebnik

86 Slide credit: S. Lazebnik Self-calibration Self-calibration (auto-calibration) is the process of determining intrinsic camera parameters directly from uncalibrated images For example, when the images are acquired by a single moving camera, we can use the constraint that the intrinsic parameter matrix remains fixed for all the images Compute initial projective reconstruction and find 3D projective transformation matrix Q such that all camera matrices are in the form P i = K [R i t i ] Can use constraints on the form of the calibration matrix: zero skew Can use vanishing points

87 Summary: 3D geometric vision Single-view geometry The pinhole camera model Variation: orthographic projection The perspective projection matrix Intrinsic parameters Extrinsic parameters Calibration Multiple-view geometry Triangulation The epipolar constraint Essential matrix and fundamental matrix Stereo Binocular, multi-view Structure from motion Reconstruction ambiguity Affine SFM Projective SFM Slide credit: S. Lazebnik