Structure from motion

Transcription

1 Structure from motion

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

3 Structure from motion Given: m images of n fixed 3D points λ 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 X j x 1j P 1 x 2j x 3j P 2 P 3

4 Outline Reconstruction ambiguities Affine structure from motion Factorization Projective structure from motion Bundle adjustment Modern structure from motion pipeline

5 Is SfM always uniquely solvable? Necker cube Source: N. Snavely

6 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 æ ç è 1 k ö P ( k ø X) It is impossible to recover the absolute scale of the scene!

7 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)

8 Projective ambiguity With no constraints on the camera calibration matrix or on the scene, we can reconstruct up to a projective ambiguity Q p é A ê T ëv tù v ú û x PX ( PQ -1 )( Q X) P P

9 Projective ambiguity

10 Affine ambiguity If we impose parallelism constraints, we can get a reconstruction up to an affine ambiguity Affine Q A é A ê T ë0 tù 1 ú û x PX ( PQ -1 )( Q X) A A

11 Affine ambiguity

12 Similarity ambiguity A reconstruction that obeys orthogonality constraints on camera parameters and/or scene Q s ésr ê T ë0 tù 1 ú û x PX ( PQ -1)( Q X) S S

13 Similarity ambiguity

14 Affine structure from motion Let s start with affine or weak perspective cameras (the math is easier) center at infinity

15 Recall: Orthographic Projection Image World Projection along the z direction

16 Affine cameras Orthographic Projection Parallel Projection

17 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 non-homogeneous coordinates ú û ù ê ë é ú ú ú û ù ê ê ê ë é ú ú ú û ù ê ê ê ë é affine] [ affine] 3 [ b A P b a a a b a a a x X a 1 a 2 b AX x + ø ö ç ç è æ + ø ö ç ç ç è æ ú û ù ê ë é ø ö ç è æ b b Z Y X a a a a a a y x Projection of world origin

18 Affine structure from motion Given: m images of n fixed 3D points: xij A i X j + b 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): éa -1 ê ë 0 bù 1 ú û éa ê ë 0 bù 1 úq û, æ Xö Qç è 1 ø 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 æ ç è X 1 ö ø

19 Affine structure from motion Centering: subtract the centroid of the image points in each view xˆ ij x ij - 1 n n å k 1 x ik A X i j + b i - 1 n n å k 1 ( A X + b ) i k i A i æ ç è X j - 1 n n å k 1 X k ö ø A Xˆ i j For simplicity, set the origin of the world coordinate system to the centroid of the 3D points After centering, each normalized 2D point is related to the 3D point X j by x ˆ A ij i X j

20 Affine structure from motion D Let s create a 2m n data (measurement) matrix: éxˆ ê ê xˆ ê ê ëxˆ m1 xˆ xˆ xˆ m2!! "! xˆ xˆ xˆ 1n 2n mn ù ú ú ú ú û cameras (2m) points (n) C. Tomasi and T. Kanade. Shape and motion from image streams under orthography: A factorization method. IJCV, 9(2): , November 1992.

21 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 ú ú ú ú û ù ê ê ê ê ë é ú ú ú ú û ù ê ê ê ê ë é cameras (2m 3) points (3 n) The measurement matrix D MS must have rank 3! C. Tomasi and T. Kanade. Shape and motion from image streams under orthography: A factorization method. IJCV, 9(2): , November 1992.

22 Factorizing the measurement matrix Source: M. Hebert

23 Factorizing the measurement matrix Singular value decomposition of D: Source: M. Hebert

24 Factorizing the measurement matrix Singular value decomposition of D: Source: M. Hebert

25 Factorizing the measurement matrix Obtaining a factorization from SVD: Source: M. Hebert

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

27 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) Source: M. Hebert

28 Eliminating the affine ambiguity Transform each projection matrix A to another matrix AC to get orthographic projection Image axes are perpendicular and scale is 1 a 2 a 1 x a 1 a 2 0 a 1 2 a X This translates into 3m equations: (A i C)(A i C) T A i (CC T )A i Id, i 1,, m Solve for L CC T Recover C from L by Cholesky decomposition: L CC T Update M and S: M MC, S C -1 S Source: M. Hebert

29 Reconstruction results C. Tomasi and T. Kanade, Shape and motion from image streams under orthography: A factorization method, IJCV 1992

30 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: cameras points Possible solution: decompose matrix into dense subblocks, factorize each sub-block, and fuse the results Finding dense maximal sub-blocks of the matrix is NPcomplete (equivalent to finding maximal cliques in a graph)

31 Dealing with missing data 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 (triangulation) (3) Solve for a new camera that sees at least three known 3D points (calibration) F. Rothganger, S. Lazebnik, C. Schmid, and J. Ponce. Segmenting, Modeling, and Matching Video Clips Containing Multiple Moving Objects. PAMI 2007.

32 Projective structure from motion Given: m images of n fixed 3D points λijxij Pi Xj, i 1,, m, j 1,, n Problem: estimate m projection matrices P i and n 3D points Xj from the mn correspondences xij X j x 1j P 1 x 2j x 3j P 2 P 3

33 Projective structure from motion Given: m images of n fixed 3D points λijxij Pi Xj, i 1,, m, j 1,, n Problem: estimate m projection matrices P i and n 3D points Xj from the mn correspondences xij 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

34 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

35 Incremental 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

36 Incremental 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 Refine and extend structure: compute new 3D points, re-optimize existing points that are also seen by this camera triangulation

37 Incremental 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 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

38 Bundle adjustment Non-linear method for refining structure and motion Minimize reprojection error m n w ij i1 j1 x ij 1 λ ij P i X j 2 X j visibility flag: is point j visible in view i? P 1 X j P 1 x 1j P 2 X j x 2j P 3 X j x 3j P 2 P 3

39 Representative SFM pipeline N. Snavely, S. Seitz, and R. Szeliski, Photo tourism: Exploring photo collections in 3D, SIGGRAPH 2006.

40 Feature detection Detect SIFT features Source: N. Snavely

41 Feature detection Detect SIFT features Source: N. Snavely

42 Feature matching Match features between each pair of images Source: N. Snavely

43 Feature matching Use RANSAC to estimate fundamental matrix between each pair Source: N. Snavely

44 Feature matching Use RANSAC to estimate fundamental matrix between each pair Image source

45 Feature matching Use RANSAC to estimate fundamental matrix between each pair Source: N. Snavely

46 Image connectivity graph (graph layout produced using the Graphviz toolkit: Source: N. Snavely

47 Incremental SFM Pick a pair of images with lots of inliers (and preferably, good EXIF data) Initialize intrinsic parameters (focal length, principal point) from EXIF Estimate extrinsic parameters (R and t) using five-point algorithm Use triangulation to initialize model points While remaining images exist Find an image with many feature matches with images in the model Run RANSAC on feature matches to register new image to model Triangulate new points Perform bundle adjustment to re-optimize everything

48 The devil is in the details Handling degenerate configurations (e.g., homographies) Eliminating outliers Dealing with repetitions and symmetries Handling multiple connected components Closing loops Making the whole thing efficient! See, e.g., Towards Linear-Time Incremental Structure from Motion

49 SFM software Bundler OpenSfM OpenMVG VisualSFM See also Wikipedia s list of toolboxes

50 Review: Structure from motion Ambiguity Affine structure from motion Factorization Dealing with missing data Incremental structure from motion Projective structure from motion Bundle adjustment Modern structure from motion pipeline