Srikumar Ramalingam Department of Computer Science University of California, Santa Cruz srikumar @cse.ucsc.edu 3D Reconstruction from a Pair of Images
Dec 20, 2015
Srikumar Ramalingam Department of Computer Science University of California, Santa Cruz [email protected]
3D Reconstruction from a Pair of Images
- Problem Definition
- Previous Work
- Solution
- Experiments and Results
- Conclusion and Future work
Overview
Problem Definition
Image-1
Image-2
3D Texture-Mapped Model
Perspective View of the Model
Previous Work- Zhao, Aggarwal, Mandal and Vemuri, “3D Shape
Reconstruction from Multiple Views ”, Handbook of Image and Video Processing, pages 243-257, Al Bovik, 2000.
- Gang Xu and Zhengyou Zhang, “Epipolar Geometry in Stereo, Motion and Object Recognition”, Kluwer Academic Publishers, 1996.
- Zhang and Faugeras, “3D Dyanamic Scene Analysis-A Stereo Based Approach”, Springer-Verlag, 1992.
- Zhang, Deriche, Faugeras and Luong, “A Robust Technique for Matching Two Uncalibrated Images through the Recovery of the Unknown Epipolar Geometry”, INRIA Research Report, 1994.
- Zhang, “A New Multistage Approach to Motion and Structure Estimation: From Essential Parameters to Euclidean Motion Via Fundamental Matrix”, INRIA Research Report, 1996.
Previous Work Zhang, “Determining the Epipolar Geometry and its
Uncertainity: A Review”, INRIA Research Report, July 1996. Zhang, “A Flexible New Technique for Camera Calibration”,
Technical Report, Microsoft Research, 1998. Deriche and Giraudon, “A Computational Approach for Corner
and Vertex Detection”, INRIA Research Report, 1992.
Solution
-Feature Detection
-Getting Initial Set of Matches
-Medium Robust Correspondence
-Strong Robust Correspondence
-Camera Calibration
-3D Reconstruction
Feature Detection : Harris Corner Detection
Corner Threshold y)R(x,
)C(k.trace - )C(det y)R(x,
I II
II I C
^2
^
^2y
^
yx
^
yx
^ 2x^
Establishing Initial Set of Matches
Ambiguities in the Matches
Robust 1-1 Correspondence
-Medium Robust Matches
-Relaxation Techniques
-Strong Robust Matches
-Epipolar Geometry
Relaxation Techniques
- Winner-take-all
- Loser-take-nothing
- Some-winners-take-all - ( 1 – Max_Strength / Sec_Max_St)
Relaxation Strategies
End Result : No Ambiguities but False Matches
-Epipolar Geometry and Constraint
-Least Median of Squares
Strong Robust Estimation using Epipolar Geometry
Epipolar Geometry
0: 12Constraint mFmT
Fundamental Matrix (F) –3x3 matrix, which relates the corresponding points
Point corresponding to m lies on its epipolar line lm on the other image
Least Median of Squares – Removal of Outliers
),(),(
..3,2,1 )}({min
212
122
2
iT
iiii
ij
mFmdFmmdresidualr
mjrmed
- 8 Matches required for estimating F matrix
- Different combinations (m) of 8 matches selected
- Least median of squares algorithm is applied
If ri < Threshold, the match is discarded.
3D Reconstruction Problem is solved for Conventional Baseline Stereo System
X = b (xl+xr) / (2d)
Y = b (yl+yr) / (2d)
Z = bf / d
Intrinsic Parameters (5) Extrinsic Parameters (6)
f – focal length 3 rotational parameters, 3 translational parameters
u0, v0 – Center Intrinsic Matrix(A)
ku - unit length along x direction
kv – unit length along y direction
Angle between x and y direction
mnew(u,v) = A mold(x,y)
Need to conduct an experiment to calibrate the camera
Intrinsic and Extrinsic Parameters
0 1 0 0
0 v/sinθfk 0
0 u cotθfkfk
0v
0uu
3D Reconstruction- Triangulation
Robust Correspondence + Intrinsic Parameters Extrinsic Parameters
Robust Correspondence + Camera Parameters - 3D Points
TR
A
RTAP
IAP
tZYXPms
tZYXPms
TT
TT
,
]['
]0[
],,,[''
],,,[
2
1
Camera MatrixExtrinsic Parameters
Reconstructed 3D Model
Implementaton Pipeline
Matlab Implementations-Harris Corner Detection Algorithm (Deriche1992, Zhang1994)
-Initial Set of Matches Establishment (Zhang1994, Xu1996)
-Medium Set of Matches using Relaxation Techniques (Zhang1994, Xu1996)
-Strong Set of Matches using Epipolar Geometry (Zhang1994, Xu1996)
-Camera Calibration Experiment (Zhang1998)
-3D Points Reconstruction from Robust Matches and Camera Parameters (Zhang1994, Zhang1996, Xu1996)
-3D Polygonal Model Reconstruction (Delaunay Triangulation)
- Texture Mapping (OpenGL/C)
Standard Data Sets- Corner marked
Robust 1-1 Correspondence shown
Color Coding for Z Coordinates after 3D Reconstruction
3D Delaunay Triangulation
3D Texture Mapped Model – On Rotation
Real Data Sets and Results
Baskin Engineering Parking Scene – Two Images
Feature Points using Corner Detection process
Robust Set of Matches
Color Coding for Z Coordinates after 3D Reconstruction
Red-Max, Green – Intermediate, Blue – Min depths
3D Delaunay Triangulation
Texture Mapped 3D Model of the Scene
Perspective View of the Texture Mapped 3D Model
Camera Calibration Experiment
-Checker pattern
-3 images taken in different orientations
-Corners are marked
-Computation of camera parameters
Conclusion and Future Work
-Increasing the number of feature points
- Multiple Images
- Alternate Algorithms
- 3D Reconstruction of Urban Scenes (Faugeras 1995)
- Registration within GIS Data
Questions?