CSCE 643 Computer Vision: Structure from Motion Jinxiang Chai
Stereo reconstruction Given two or more images of the same scene or object, compute a representation of its shape known camera viewpoints
Stereo reconstruction Given two or more images of the same scene or object, compute a representation of its shape known camera viewpoints How to estimate camera parameters? - where is the camera? - where is it pointing? - what are internal parameters, e.g. focal length?
Calibration from 2D motion Structure from motion (SFM) - track points over a sequence of images - estimate for 3D positions and camera positions - calibrate intrinsic camera parameters before hand Self-calibration: - solve for both intrinsic and extrinsic camera parameters
SFM = Holy Grail of 3D Reconstruction Take movie of object Reconstruct 3D model Would be commercially highly viable
How to Get Feature Correspondences Feature-based approach - good for images - feature detection (corners or sift features) - feature matching using RANSAC (epipolar line) Pixel-based approach - good for video sequences - patch based registration with lucas-kanade algorithm - register features across the entire sequence
A Brief Introduction on Feature-based Matching Find a few important features (aka Interest Points) Match them across two images Compute image transformation function h
Feature Detection Two images taken at the same place with different angles Projective transformation H3X3
Feature Matching ? Two images taken at the same place with different angles Projective transformation H3X3
How do we match features across images? Any criterion? Feature Matching ? Two images taken at the same place with different angles Projective transformation H3X3 How do we match features across images? Any criterion?
How do we match features across images? Any criterion? Feature Matching ? Two images taken at the same place with different angles Projective transformation H3X3 How do we match features across images? Any criterion?
Feature Matching Intensity/Color similarity The intensity of pixels around the corresponding features should have similar intensity
Feature Matching Feature similarity (Intensity or SIFT signature) The intensity of pixels around the corresponding features should have similar intensity Cross-correlation, SSD
Feature Matching Feature similarity (Intensity or SIFT signature) The intensity of pixels around the corresponding features should have similar intensity Cross-correlation, SSD Distance constraint The displacement of features should be smaller than a given threshold
Feature Matching Feature similarity (Intensity or SIFT signature) The intensity of pixels around the corresponding features should have similar intensity Cross-correlation, SSD Distance constraint The displacement of features should be smaller than a given threshold Epipolar line constraint The corresponding pixels satisfy epipolar line constraints.
Feature Matching Feature similarity (Intensity or SIFT signature) The intensity of pixels around the corresponding features should have similar intensity Cross-correlation, SSD Distance constraint The displacement of features should be smaller than a given threshold Epipolar line constraint The corresponding pixels satisfy epipolar line constraints. Fundamental matrix H
Feature-space Outlier Rejection bad Good
Feature-space Outlier Rejection Can we now compute H3X3 from the blue points?
Feature-space Outlier Rejection Can we now compute H3X3 from the blue points?
Feature-space Outlier Rejection Can we now compute H3X3 from the blue points? No! Still too many outliers…
Feature-space Outlier Rejection Can we now compute H3X3 from the blue points? No! Still too many outliers… What can we do?
Feature-space Outlier Rejection Can we now compute H3X3 from the blue points? No! Still too many outliers… What can we do? Robust estimation!
Robust Estimation: A Toy Example How to fit a line based on a set of 2D points?
RANSAC for Estimating Projective Transformation RANSAC loop: Select four feature pairs (at random) Compute the transformation matrix H (exact) Compute inliers where SSD(pi’, H pi) < ε Keep largest set of inliers Re-compute least-squares H estimate on all of the inliers For more detail, check - http://research.microsoft.com/en-us/um/people/zhang/INRIA/software-FMatrix.html - Philip H. S. Torr (1997). "The Development and Comparison of Robust Methods for Estimating the Fundamental Matrix". International Journal of Computer Vision 24 (3): 271–300
Structure from Motion Two Principal Solutions Bundle adjustment (nonlinear optimization) Factorization (SVD, through orthographic approximation, affine geometry)
Projection Matrix Perspective projection: 2D coordinates are just a nonlinear function of its 3D coordinates and camera parameters: K R T P 26
Nonlinear Approach for SFM What’s the difference between camera calibration and SFM?
Nonlinear Approach for SFM What’s the difference between camera calibration and SFM? - camera calibration: known 3D and 2D
Nonlinear Approach for SFM What’s the difference between camera calibration and SFM? - camera calibration: known 3D and 2D - SFM: unknown 3D and known 2D
Nonlinear Approach for SFM What’s the difference between camera calibration and SFM? - camera calibration: known 3D and 2D - SFM: unknown 3D and known 2D - what’s 3D-to-2D registration problem?
Nonlinear Approach for SFM What’s the difference between camera calibration and SFM? - camera calibration: known 3D and 2D - SFM: unknown 3D and known 2D - what’s 3D-to-2D registration problem?
SFM: Bundle Adjustment SFM = Nonlinear Least Squares problem Minimize through Gradient Descent Conjugate Gradient Gauss-Newton Levenberg Marquardt common method Prone to local minima
Count # Constraints vs #Unknowns M camera poses N points 2MN point constraints 6M+3N + 4 (unknowns) Suggests: need 2mn 6m + 3n+4 But: Can we really recover all parameters???
Count # Constraints vs #Unknowns M camera poses N points 2MN point constraints 6M+3N+4 unknowns (known intrinsic camera parameters) Suggests: need 2mn 6m + 3n+4 But: Can we really recover all parameters??? Can’t recover origin, orientation (6 params) Can’t recover scale (1 param) Thus, we need 2mn 6m + 3n+4 - 7
Are We Done? No, bundle adjustment has many local minima.
SFM Using Factorization Assume an orthographic camera Image World
SFM Using Factorization Assume orthographic camera Image World Subtract the mean
SFM Using Factorization Stack all the features from the same frame:
SFM Using Factorization Stack all the features from the same frame: Stack all the features from all the images: W
SFM Using Factorization Stack all the features from the same frame: Stack all the features from all the images: W
SFM Using Factorization Stack all the features from all the images: Factorize the matrix into two matrix using SVD: W
SFM Using Factorization Stack all the features from all the images: Factorize the matrix into two matrix using SVD:
SFM Using Factorization Stack all the features from all the images: Factorize the matrix into two matrix using SVD: How to compute the matrix ? W
SFM Using Factorization M is the stack of rotation matrix:
SFM Using Factorization M is the stack of rotation matrix: 1 Orthogonal constraints from rotation matrix 1 1 1
SFM Using Factorization M is the stack of rotation matrix: 1 Orthogonal constraints from rotation matrix 1 1 1
SFM Using Factorization Orthogonal constraints from rotation matrices: 1 1 1 1
SFM Using Factorization Orthogonal constraints from rotation matrices: 1 1 1 QQ: symmetric 3 by 3 matrix 1
SFM Using Factorization Orthogonal constraints from rotation matrices: 1 1 1 QQ: symmetric 3 by 3 matrix 1 How to compute QQT? least square solution - 4F linear constraints, 9 unknowns (6 independent due to symmetric matrix)
SFM Using Factorization Orthogonal constraints from rotation matrices: 1 1 1 QQ: symmetric 3 by 3 matrix 1 How to compute QQT? least square solution - 4F linear constraints, 9 unknowns (6 independent due to symmetric matrix) How to compute Q from QQT: SVD again:
SFM Using Factorization M is the stack of rotation matrix: 1 Orthogonal constraints from rotation matrix 1 1 1 QQT: symmetric 3 by 3 matrix Computing QQT is easy: 3F linear equations 6 independent unknowns
SFM Using Factorization Form the measurement matrix Decompose the matrix into two matrices and using SVD Compute the matrix Q with least square and SVD Compute the rotation matrix and shape matrix: and
Weak-perspective Projection Factorization also works for weak-perspective projection (scaled orthographic projection): z0 d
Factorization for Full-perspective Cameras [Han and Kanade]
SFM for Deformable Objects For detail, click here 55
SFM for Articulated Objects For video, click here
SFM Using Factorization Bundle adjustment (nonlinear optimization) - work with perspective camera model - work with incomplete data - prone to local minima Factorization: - closed-form solution for weak perspective camera - simple and efficient - usually need complete data - becomes complicated for full-perspective camera model Phil Torr’s structure from motion toolkit in matlab (click here) Voodoo camera tracker (click here) 57
All Together Video Click here - feature detection - feature matching (epipolar geometry) - structure from motion - stereo reconstruction - triangulation - texture mapping 58