Vision 3D articielle Disparity maps, correlation

Transcription

1 Vision 3D articielle Disparity maps, correlation Pascal Monasse IMAGINE, École des Ponts ParisTech

2 Contents Triangulation Epipolar rectication Disparity map

3 Contents Triangulation Epipolar rectication Disparity map

4 Triangulation Let us write again the binocular formulae: λx = K(RX + T ) λ x = K X Write Y T = ( X T 1 λ λ ) : ( ) KR KT x 0 3 K x Y = 0 6 (6 equations 5 unknowns+1 epipolar constraint) We can then recover X. Special case: R = Id, T = Be 1 We get: If also K = K, d is the disparity z(x KK 1 x ) = ( Bf 0 0 ) T z = fb/[(x x ) e 1 ] = fb/d

5 Triangulation Fundamental principle of stereo vision z h, B/H z = d H. f f focal length. H distance optical center-ground. B distance between optical centers (baseline). Goal Given two rectied images, point correspondences and computation of their apparent shift (disparity) gives information about relative depth of the scene.

6 Recovery of R and T Suppose we know K, K, and F or E. Recover R and T? From E = [T ] R, E T E = R T (TT T T 2 I )R = (R T T )(R T T ) T + R T T 2 I If x = R T T, E T E x = 0 and if y x = 0, E T E y = T 2 y. Therefore σ 1 = σ 2 = T and σ 3 = 0. Inversely, from E = Udiag(σ, σ, 0)V T, we can write: E = σu T U U T V = σ[t ] R { T = ±σu[e 3 ] U T Actually, there are up to 4 solutions: R = UR z (± π 2 )V T B A B B A A B A

7 What is possible without calibration? We can recover F, but not E. Actually, from x = PX x = P X we see that we have also: x = (PH 1 )(HX) x = (P H 1 )(HX) Interpretation: applying a space homography and transforming the projection matrices (this changes K, K, R and T ), we get exactly the same projections. Consequence: in the uncalibrated case, reconstruction can only be done modulo a 3D space homography.

8 Contents Triangulation Epipolar rectication Disparity map

9 Epipolar rectication It is convenient to get to a situation where epipolar lines are parallel and at same ordinate in both images. As a consequence, epipoles are at horizontal innity: 1 e = e = 0 0 It is always possible to get to that situation by virtual rotation of cameras (application of homography) Image planes coincide and are parallel to baseline.

10 Epipolar rectication Image 1

11 Epipolar rectication Image 2

12 Epipolar recti cation Image 1 Recti ed image 1

13 Epipolar recti cation Image 2 Recti ed image 2

14 Epipolar rectication Fundamental matrix can be written: 1 F = = thus T x Fx = 0 y y = 0 0 Writing matrices P = K K = f x s c x ( I 0 ) and P = K ( I Be 1 ) : f x s c x 0 f y c y K = 0 f y c y F = BK T [e 1 ] K 1 = B f y f y f y 0 f y c y f y c y f y We must have f y = f y and c y = c y, that is identical second rows of K and K

15 Epipolar rectication We are looking for homographies H and H to apply to images such that F = H T [e 1 ] H That is 9 equations and 16 variables, 7 degrees of freedom remain: the rst rows of K and K and the rotation angle around baseline α Invariance through rotation around baseline: T cos α sin α cos α sin α = [e 1 ] 0 sin α cos α sin α cos α Several methods exist, they try to distort as little as possible the image Rectif. of Gluckman-Nayar (2001)

16 Epipolar rectication of Fusiello-Irsara (2008) We are looking for H and H as rotations, supposing K = K known: H = K n RK 1 and H = K n R K 1 with K n and K of identical second row, n R and R rotation matrices parameterized by Euler angles and f 0 w/2 K = 0 f h/ Writing R = R x (θ x )R y (θ y )R z (θ z ) we must have: F = (K n RK 1 ) T [e 1 ] (K n R K 1 ) = K T R T z R T y [e 1 ] R K 1 We minimize the sum of squares of points to their epipolar line according to the 6 parameters (θ y, θ z, θ x, θ y, θ z, f )

17 Ruins E 0 = 3.21 pixels. E 6 = 0.12 pixels.

18 Ruins E 0 = 3.21 pixels. E 6 = 0.12 pixels.

19 Cake E 0 = 17.9 pixels. E 13 = 0.65 pixels.

20 Cake E 0 = 17.9 pixels. E 13 = 0.65 pixels.

21 Cluny E 0 = 4.87 pixels. E 14 = 0.26 pixels.

22 Cluny E 0 = 4.87 pixels. E 14 = 0.26 pixels.

23 Carcassonne E 0 = 15.6 pixels. E 4 = 0.24 pixels.

24 Carcassonne E 0 = 15.6 pixels. E 4 = 0.24 pixels.

25 Books E 0 = 3.22 pixels. E 14 = 0.27 pixels.

26 Books E 0 = 3.22 pixels. E 14 = 0.27 pixels.

27 Contents Triangulation Epipolar rectication Disparity map

28 Disparity map z = fb d Depth z is inversely proportional to disparity d (apparent motion, in pixels). Disparity map: At each pixel, its apparent motion between left and right images. We already know disparity at feature points, this gives an idea about min and max motion, which makes the search for matching points less ambiguous and faster.

29 Stereo Matching Principle: invariance of something between corresponding pixels in left and right images (I L, I R ) Example: color, x-derivative, census... Usage of a distance to capture this invariance, such as AD(p, q) = I L (p) I R (q) 1

30 Stereo Matching Principle: invariance of something between corresponding pixels in left and right images (I L, I R ) Example: color, x-derivative, census... Usage of a distance to capture this invariance, such as AD(p, q) = I L (p) I R (q) 1 Left image Ground truth Min AD

31 Stereo Matching Post-processing helps a lot! Example: left-right consistency check, followed by simple constant interpolation, and median weighted by original image bilateral weights Min CG Left-right test Post-processed

32 Stereo Matching Post-processing helps a lot! Example: left-right consistency check, followed by simple constant interpolation, and median weighted by original image bilateral weights Min CG Left-right test Post-processed Still, single pixel estimation not good enough Need to promote some regularity of the result

33 Global vs. local methods Global method: explicit smoothness term arg min E data (p, p + d(p); I L, I R ) d p + p p E reg (d(p), d(p ); p, p, I L, I R ) Examples: E reg = d(p) d(p ) 2 (Horn-Schunk), E reg = δ(d(p) = d(p )) (Potts), E reg = exp( (I L (p) I L (p )) 2 /σ 2 ) d(p) d(p )...

34 Global vs. local methods Global method: explicit smoothness term arg min E data (p, p + d(p); I L, I R ) d p + p p E reg (d(p), d(p ); p, p, I L, I R ) Examples: E reg = d(p) d(p ) 2 (Horn-Schunk), E reg = δ(d(p) = d(p )) (Potts), E reg = exp( (I L (p) I L (p )) 2 /σ 2 ) d(p) d(p )... Problem: NP-hard for almost all regularity terms except E reg = λ pp d(p) d(p ) (Ishikawa 2003) Aternative: sub-optimal solution for submodular regularity (graph-cuts: Boykov, Kolmogorov, Zabih), loopy-belief propagation (no guarantee at all), semi-global matching (Hirschmüller)

35 Global vs. local methods Local method: Take a patch around p, aggregate costs E data (Lucas-Kanade) No explicit regularity term Example: SAD(p, q) = r P I L(p + r) I R (q + r), SSD(p, q) = r P I L(p + r) I R (q + r) 2, SCG(p, q) = CG(p + r, q + r)... r P Can be interpreted as a cost-volume ltering. dmax dmin * 1 Px{0} Increasing patch size P promotes regularity.

36 Global vs. local methods Local method: Take a patch around p, aggregate costs E data (Lucas-Kanade) No explicit regularity term Example: SAD(p, q) = r P I L(p + r) I R (q + r), SSD(p, q) = r P I L(p + r) I R (q + r) 2, SCG(p, q) = CG(p + r, q + r)... r P Can be interpreted as a cost-volume ltering. Increasing patch size P promotes regularity. p p Proportion of common pixels between p + P and p + P: 1 1 n if P is n n

37 Local search At each pixel, we consider a context window and we look for the motion of this window. Distance between windows: d(q) = arg min (I (q + p) I (q + de 1 + p)) 2 d p F Variants to be more robust to illumination changes: 1. Translate intensities by the mean over the window. I (q + p) I (q + p) r F I (q + r)/#f 2. Normalize by mean and variance over window.

38 Distance between patches Several distances or similarity measures are popular: SAD: Sum of Absolute Dierences d(q) = arg min I (q + p) I (q + de 1 + p) d p F SSD: Sum of Squared Dierences d(q) = arg min (I (q + p) I (q + de 1 + p)) 2 d p F CSSD: Centered Sum of Squared Dierences d(q) = arg min (I (q + p) Ī F I (q + de 1 + p) + Ī F )2 d p F NCC: Normalized Cross-Correlation d(q) = arg max d p F (I (q + p) Ī F )(I (q + de 1 + p) Ī F ) (I (q + p) Ī F ) 2 (I (q + de 1 + p) Ī F )2

39 Another distance The following distance is more and more popular in recent articles: with ɛ(p, q) = (1 α) min ( I (p) I ) (q) 1, τ col + ( ) α min I I (p) (q), τ grad x x I (p) I (q) 1 = I r (p) I r (q) + I g (p) I g (q) + I b (p) I b (q) Usual parameters: α = 0.9 τ col = 30 (not very sensitive if larger) τ grad = 2 (not very sensitive if larger) Note that α = 0 is similar to SAD.

40 Varying patch size P = {(0, 0)}

41 Varying patch size P = [ 1, 1] 2

42 Varying patch size P = [ 7, 7] 2

43 Varying patch size P = [ 21, 21] 2

44 Varying patch size P = [ 35, 35] 2

45 Problems of local methods Implicit hypothesis: all points of window move with same motion, that is they are in a fronto-parallel plane. aperture problem: the context can be too small in certain regions, lack of information. adherence problem: intensity discontinuities inuence strongly the estimated disparity and if it corresponds with a depth discontinuity, we have a tendency to dilate the front object. C C patch rooftop O: aperture problem A: adherence problem ground

46 Example: seeds expansion We rely on best found distances and we put them in a priority queue (seeds) We pop the best seed G from the queue, we compute for neighbors the best disparity between d(g) 1, d(g), and d(g) + 1 and we push them in the queue. Right image

47 Example: seeds expansion We rely on best found distances and we put them in a priority queue (seeds) We pop the best seed G from the queue, we compute for neighbors the best disparity between d(g) 1, d(g), and d(g) + 1 and we push them in the queue. Left image

48 Example: seeds expansion We rely on best found distances and we put them in a priority queue (seeds) We pop the best seed G from the queue, we compute for neighbors the best disparity between d(g) 1, d(g), and d(g) + 1 and we push them in the queue. Seeds

49 Example: seeds expansion We rely on best found distances and we put them in a priority queue (seeds) We pop the best seed G from the queue, we compute for neighbors the best disparity between d(g) 1, d(g), and d(g) + 1 and we push them in the queue. Seeds expansion

50 Example: seeds expansion We rely on best found distances and we put them in a priority queue (seeds) We pop the best seed G from the queue, we compute for neighbors the best disparity between d(g) 1, d(g), and d(g) + 1 and we push them in the queue. Left image

51 Adaptive neighborhoods To reduce adherence (aka fattening eect), an image patch should be on the same object, or even better at constant depth Heuristic inspired by bilateral lter [Yoon&Kweon 2006]: ( ω I (p, p ) = exp p ) p 2 γ pos ( exp I (p) ) I (p ) 1 γ col Selected disparity: d(p) = arg min E(p, q) with d=q p r F E(p, q) = ω I (p, p + r) ω I (q, q + r) ɛ(p + r, q + r) ω r F I (p, p + r) ω I (q, q + r) We can take a large window F (e.g., 35 35)

52 Bilateral weights (a) (b) (c) (d)

53 Results Tsukuba Venus Teddy Cones Left image Ground truth Results

54 What is the limit of adaptive neighborhoods? The best patch is P p (r) = 1(d(p + r) = d(p)) Suppose we have an oracle giving P p Use ground-truth image to compute P p Since GT is subpixel, use P p (r) = 1( d(p + r) d(p) 1/2)

55 Test with oracle image ground truth oracle patches

56 Test with oracle image ground truth oracle patches

57 Conclusion We can get back to the canonical situation by epipolar rectication. Limit: when epipoles are in the image, standard methods are not adapted. For disparity map computation, there are many choices: 1. Size and shape of window? 2. Which distance? 3. Filtering of disparity map to reject uncertain disparities? You will see next session a global method for disparity computation Very active domain of research, >150 methods tested at

58 Practical session: Disparity map computation by propagation of seeds Objective: Compute the disparity map associated to a pair of images. We start from high condence points (seeds), then expand by supposing that the disparity map is regular. Get initial program from the website. Compute disparity map from image 1 to 2 of all points by highest NCC score. Keep only disparity where NCC is suciently high (0.95), put them as seeds in a std::priority_queue. While queue is not empty: 1. Pop P, the top of the queue. 2. For each 4-neighbor Q of P having no valid disparity, set d Q by highest NCC score among d P 1, d P, and d P Push Q in queue. Hint: the program may be too slow in Debug mode for the full images. Use cropped images to nd your bugs, then build in Release mode for original images.