Multiview Geometry and Bundle Adjustment. CSE P576 David M. Rosen

Transcription

1 Multiview Geometry and Bundle Adjustment CSE P576 David M. Rosen 1

2 Recap Previously: Image formation Feature extraction + matching Two-view (epipolar geometry) Today: Add some geometry, statistics, optimization Turn it up to 11 N! 2

3 Motivating example: Photogrammetry The science of measurement using cameras 3

4 Application: Remote Sensing Mars Reconnaissance Orbiter Launched 12 Aug 2005 Entered orbit 10 Mar 2006 ~ 112 minute orbital period ~ 12.8 orbits / (Earth) day Sensors: High Resolution Imaging Science Experiment (HiRISE) Context Camera Mars Color Imager Image credit: NASA/JPL 4

5 Image credit: NASA/JPL-Caltech/MSSS 5

6 Image credit: NASA/JPL-Caltech/MSSS 6

7 Image credit: NASA/JPL/University of Arizona/USGS 7

8 Image credit: NASA / JPL-Caltech / UA / Kevin M. Gill 8

9 Image credit: NASA/JPL/University of Arizona/USGS 9

10 Image credit: NASA / JPL-Caltech / UA / Kevin M. Gill 10

11 Video credit: NASA / JPL / AU / Doug Ellison 11

12 Application: 3D Reconstruction Goal: Build a 3D model of a scene from a collection of images [S. Agarwal et al., Building Rome in a Day, Communications of the ACM, 2011] 12

13 13

14 14

15 15

16 Application: Robotics Video credit: MIT DRC team 16

17 Video credit: MIT DRC team 17

18 Video credit: MIT DRC team 18

19 19

20 In this lecture Photogrammetry: The problem of measurement using imagery Maximum-likelihood estimation and bundle adjustment: Solving the photogrammetry problem Practicalities Problem scale Robust estimation Representation of rotations 20

21 Photogrammetry: The Problem Given: A collection of images Estimate: 3D positions of imaged points Poses of the imaging cameras Intrinsic parameters of the imaging cameras 21

22 Photogrammetry: Generative model Q: How are variables of estimation: 3D point positions p j Camera poses x i = t i, R i Camera intrinsics K i related to images? p 1 u 11 u 21 u 31 u ij = f(x i, K i, p j ) x 1,K 1 x 3,K 3 where f is the camera projection function (from Lecture 1) 22 x 2,K 2

23 Photogrammetry: Estimation procedure Main idea: Given a set of images 1. Extract features 2. Match features (identify the set of 3D points) 3. Estimate parameters so that: u ij = f(x i, K i, p j ) for all point projections u ij x 1,K 1 x 2,K 2 x 3,K 3 23

24 The problem of measurement noise We want to find x i, K i, p j so that u ij = f(x i, K i, p j ) (i.e. our model matches the data) for all u ij. But: All real-world measurements have errors What we actually measure is: u ij = f x i, K i, p j + ε ij We cannot find parameters x i, K i, p j that fit the measured projections u ij exactly 24

25 Example: Linear regression 25

26 Maximum likelihood estimation MLE is a method for fitting parameters to noisy data y = y 1,, y n, given a sampling model y~p( θ). Basic idea: Choose the that best fits the data y. But: How can we measure goodness of fit? Likelihood function: L θ p y θ. Measures how likely the data y is for each choice of. MLE principle: Best fit Maximum likelihood Pick the that maximizes the likelihood of data y: መθ = max θ L θ 26

27 Example: Regression under Gaussian noise Consider fitting a function f ; θ to data D = y i = f x i ; θ + ε i, (x i, y i ) N i=1 ε i ~N 0, Σ i under the model: For each choice of, for each (x i, y i ): The pdf for ε i is: ε i = y i f x i ; θ p ε i = det(2πσ i ) 1 2exp 1 2 ε i T Σ i 1 ε i The likelihood of the ith data point (x i, y i ) is: p x i, y i θ = det(2πσ i ) 1 2exp 1 2 (y i f x i ; θ ) T Σ 1 i (y i f(x i ; θ)) The likelihood for the entire dataset D is: N L(D θ) exp 1 2 y i f(x i ; θ) T Σ 1 i y i f(x i ; θ) i=1 ε i 27

28 Example: Regression under Gaussian noise Consider fitting a function f ; θ to data D = y i = f x i ; θ + ε i, (x i, y i ) N i=1 ε i ~N 0, Σ i under the model: The likelihood for the entire dataset D is: Taking the logarithm: N L(D θ) exp 1 2 y i f(x i ; θ) T Σ 1 i y i f(x i ; θ) i=1 N log L(D θ) = c 1 2 i=1 y i f x i ; θ MLE under additive Gaussian noise is a nonlinear least-squares problem: 2 Σi θ = min θ N i=1 2 y i f(x i ; θ) Σi ε i 28

29 Exercise: Linear regression Consider fitting a linear function to the data: x y under the model: y i = ax i + b + ε i, ε i ~N(0,.35 2 ) 29

30 Exercise: Linear regression x y Model: y i = ax i + b Estimated: a = 1.73 b = 1.06 True: a = 1.70 b =

31 Bundle adjustment Recall: Given a set of point projections u ij, we want to estimate: 3D point positions p j camera poses x i camera intrinsics K i p j Assuming the measurement model: u ij = f x i, K i, p j + ε ij, ε i ~N 0, Σ i Maximum-likelihood estimation is then: x i, K i, pƹ j = min x i,k i,p j i,j u ij f(x i, K i, p j ) Σij 2 x 1,K 1 x 2,K 2 x 3,K 3 Minimize (weighted) reprojection error 31

32 Photogrammetry and Bundle Adjustment Given: A set of images 1. Extract features 2. Match features (identify 3D points) 3. Bundle adjust (minimize reprojection error): p j x i, K i, pƹ j = min x i,k i,p j i,j u ij f(x i, K i, p j ) Σij 2 x 1,K 1 x 2,K 2 x 3,K 3 32

33 Special Case: Perspective-n-Point (PnP) Given: Known point positions p j Known camera intrinsics K Estimate: Camera pose x = (R,t) x = min x N j=1 2 u j f(x, K, p j ) Σj Image credit: OpenCV 33

34 Special Case: Camera calibration Given: Known point positions p j (on calibration object) Estimate: Camera poses x i Intrinsic parameters K x i, K = min x i,k i,j u ij f(x i, K, p j ) Σij 2 Image credit: OpenCV 34

35 Given: A pair of cameras with: Known poses x 1, x 2 Known intrinsics K 1, K 2 Special Case: Stereovision Estimate: 3D point positions p j 2 pƹ j = min σ p j u 1j f(x 1, K 1, p j ) j Σ1j + u 2j f(x 2, K 2, p j ) Σ2j 2 Image credit: MIT Space Systems Laboratory pƹ j = min p j independently for all j u 1j f(x 1, K 1, p j ) Σ1j 2 + u 2j f(x 2, K 2, p j ) Σ2j 2 35

36 Practicality: Problem Scale Piazza San Marco reconstruction: ~14,000 images ~4.5 million points Assuming each point is observed 20x, what is the size of the BA problem? 36

37 Practicality: Problem Scale Camera variables (0-skew, equal pixel scaling): 14, = 126,000 Point variables: 4,500,000 3 = 13,500,000 Point observations: 4,500, = 180,000,000 Totals: 13,626,000-dimensional state vector 180,000,000-dimensional residual vector This is a huge optimization problem! 37

38 Practicality: Feature mismatches Recall: We construct the bundle adjustment problem: x i, K i, pƹ j = min x i,k i,p j i,j u ij f(x i, K i, p j ) Σij 2 Image credit: Tian Zhou using estimated feature matches. But: What happens if these are mis-estimated? 38

39 Example: Contaminated linear regression Consider fitting the linear model: y i = ax i + ε i, ε i ~N 0,.35 2 to a contaminated data set x y

40 Exercise: Contaminated linear regression x y Model: y i = ax i + b Estimated: a = 1.37 b = 1.58 True: a = 1.70 b =

41 The problem of outliers Recall: We assumed additive Gaussian image noise: u ij = f x i, K i, p j + ε ij, ε ij ~N(0, Σ ij ) and obtained a nonlinear least-squares problem: x i, K i, pƹ j = min x i,k i,p j i,j u ij f(x i, K i, p j ) Σij 2 ε i NB: This loss weights extreme errors very heavily. This estimator is not robust! ε i 41

42 Robust loss functions One solution: Replace quadratic loss with a function that attenuates gross errors Tradeoff: More robust to outliers (Slightly) less statistical power Quadratic Huber Cauchy Geman-McClure 42

43 Example: Contaminated linear regression Least-squares loss Huber loss 43

44 Photogrammetry and Bundle Adjustment: Summary Given: A set of images 1. Extract features 2. Match features (identify 3D points) 3. Bundle adjust using a robust loss function : p j x i, K i, pƹ j = min x i,k i,p j i,j ρ u ij f(x i, K i, p j ) Σij x 1,K 1 x 2,K 2 x 3,K 3 44

45 Practicality: Representing rotations So far, we ve represented rotations using rotation matrices: R = r 11 r 12 r 13 r 21 r 31 r 22 r 32 r 23, r 33 R T R = I 3, det R = +1 Pro: Trivial point operations Con: Over-parameterized Not super convenient for optimization (requires constraints) 45

46 Euler s Rotation Theorem Theorem: Every rotation of 3D space has a fixed axis e. We can describe a rotation using: An axis (unit vector e) (Right-handed) rotation angle This is the axis-angle parameterization of rotations. Can also combine these into a single axis-angle vector: θ = θe 46

47 Rodrigues Formula How does a rotation parameterized as θ = θe act on points? Rodrigues formula: Given a vector v, v rot = v cos θ + sin θ e v + 1 cos θ e v e NB: Any 3D vector θ determinates a valid rotation Rodrigues formula is differentiable in θ Axis-angle is much more convenient for use in optimization! 47

48 Rodrigues Formula How does a rotation parameterized as θ = θe act on points? Rodrigues formula: Given a vector v, v rot = v cos θ + sin θ e v + 1 cos θ e v e Matrix form: where R θ = I + (sin θ) E + 1 cos θ E 2, E = 0 e 3 e 2 e 3 0 e 1 e 2 e

49 Photogrammetry and Bundle Adjustment: Summary Given: A set of images 1. Extract features 2. Match features (identify 3D points) 3. Bundle adjust using a robust loss function : p j x i, K i, pƹ j = min x i,k i,p j i,j ρ u ij f(x i, K i, p j ) Σij x 1,K 1 x 2,K 2 x 3,K 3 49