1
Autocalibration & 3D Reconstruction with Non-central Catadioptric Cameras Branislav Miˇ cuˇ ık Tom´ s Pajdla Center for Machine Perception, Czech Technical University in Prague, CONTRIBUTIONS Precise non-central models for real catadioptric cameras. Central approximations for ransac-based estimation of epipolar geometry. 3D metric reconstruction from uncalibrated (slightly) non-central catadioptric images by hierarchical approach: 1. Auto-calibration using central model to validate tentative correspondences and to obtain initial estimate of camera motion. 2. Bundle adjustment on inliers using the precise non-central model. REAL CATA-DIOPTRIC CAMERAS parablic mirror hyperbolic mirror spherical mirror + perspective camera PROBLEM FORMULATION 3D reconstruction from real para-catadioptric images using a) central model b)non-central model SKEWED CORRECT Input images: real para-catadioptric camera F p u π F u p π NON-CENTRAL & CENTRAL APPROXIMATION -800 -600 -400 -200 0 200 400 600 800 0 100 200 300 400 500 600 700 800 1mm ~ 36.6 pxl, f = 70 mm kuk[pxl] z[pxl] -600 -400 -200 0 200 400 600 -200 -100 0 100 200 300 400 500 1mm ~ 24.4 pxl, f lens = 75.0 mm kuk[pxl] z[pxl] -800 -600 -400 -200 0 200 400 600 800 0 100 200 300 400 500 600 700 800 1mm ~ 36.6 pxl, f = 70.0 mm kuk[pxl] z[pxl] 5 10 15 20 25 0.63 0.635 0.64 0.645 0.65 0.655 0.66 0.665 a k NON-CENTRAL MODEL X x y z u v u p F π γ γ x C W ν c , t c n m , t m Perspective projection + law of reflection x w = f (u, a, m , t m , c , t c , ) p w = f (u, a, m , t m , c , t c , ) x w = -1 m ( λ > c -1 u + t c ) + t m p w = -1 m > c -1 u - 2 ( > c -1 u ) · n knk · n knk λ, n depend on the shape of a mirror AUTO-CALIBRATION WITH CENTRAL MODEL x z p q X g u S 3 sensor plane g/h C h h(kuk)u s p = g(kuk,a) u h(kuk,a) for perspective proj. g(.),h(.)=1 p x p y p z a=a 0 +(a - a 0 ) ˙ p x ˙ p y ˙ p z a=a 0 = = p x - ˙ p x a 0 p y - ˙ p y a 0 p z - ˙ p z a 0 a=a 0 + a ˙ p x ˙ p y ˙ p z a=a 0 = = x + a s, Using epipolar constaint p > 2 p 1 =0 (x 2 + as 2 ) > (x 1 + as 1 )=0 leads to the Quadratic Eigenvalue Problem (QEP): ( + a + a 2 )f =0 where f =[ F 11 F 12 F 13 F 21 ... F 33 ] > and i are functions of coordinates of point correspon- dences (we used WBS [2]) Solving the QEP gives: mirror parameter a essential matrix inliers/outliers (ransac) 3D REC. WITH NON-CENTRAL MODEL x i w 1 p i w 1 x i w 2 p i w 2 X W C C Since cameras are calibrated, then R M = argmin a, c ,t c , m ,t m , N X i=1 | ( x i w 1 - x i w 2 ) · ( p i w 1 × p i w 2 ) | |p i w 1 × p i w 2 | ! 2 RESULTS REFERENCES [1] Fitzgibbon A., Simultaneous linear estimation of multiple view geometry and lens distortion. CVPR, 2001. [2] Matas J., Chum O., Urban M., and Pajdla T., Robust wide baseline stereo from maximally stable extremal regions. BMVC, 2002. [3] Micusik B. and Pajdla T., Para-catadioptric Camera Auto-calibration from Epipolar Geometry ACCV, 2004. [4] Swaminathan R. and Grossberg M. D. and Nayar S. K., Caustics of Catadioptric Cameras, ICCV, 2001. This research was supported by the following projects: MSM 212300013, M ˇ SMT Kontakt 22-2003-04, BENOGO–IST–2001–39184 Presented at the Conference on Computer Vision and Pattern Recognition (CVPR), 2004, Washington D.C., USA

Autocalibration & 3D Reconstruction with Non-central Catadioptric …cmp.felk.cvut.cz/ftp/articles/micusik/Micusik-CVPR2004-poster-a4.pdf · Autocalibration & 3D Reconstruction with

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Autocalibration & 3D Reconstruction with Non-central Catadioptric …cmp.felk.cvut.cz/ftp/articles/micusik/Micusik-CVPR2004-poster-a4.pdf · Autocalibration & 3D Reconstruction with

Autocalibration & 3D Reconstruction with Non-central

Catadioptric Cameras

Branislav Micusık Tomas Pajdla

Center for Machine Perception, Czech Technical University in Prague,

�� � �� � �� � ��� �� � � � � � � �

CONTRIBUTIONS

• Precise non-central models for real catadioptric cameras.

• Central approximations for ransac-based estimation of epipolar geometry.

• 3D metric reconstruction from uncalibrated (slightly) non-central catadioptric images byhierarchical approach:

1. Auto-calibration using central model to validate tentative correspondences and to obtaininitial estimate of camera motion.

2. Bundle adjustment on inliers using the precise non-central model.

REAL CATA-DIOPTRIC CAMERAS

parablic mirror hyperbolic mirror spherical mirror+ perspective camera

PROBLEM FORMULATION

3D reconstruction from real para-catadioptric images using

a) central model b)non-central model

SKEWED CORRECT

Input images:

real para-catadioptriccamera

Fp

uP

Sfra

grep

lacem

ents

π

F

u

p

PSfra

grep

lacem

entsπ

π

NON-CENTRAL & CENTRAL APPROXIMATION

−800 −600 −400 −200 0 200 400 600 8000

100

200

300

400

500

600

700

800

1mm ~ 36.6 pxl, flens

= 70 mm

PSfrag replacements

‖u‖[pxl]

z[p

xl]

−600 −400 −200 0 200 400 600−200

−100

0

100

200

300

400

500

1mm ~ 24.4 pxl, flens

= 75.0 mm

PSfrag replacements

‖u‖[pxl]

z[p

xl]

−800 −600 −400 −200 0 200 400 600 8000

100

200

300

400

500

600

700

800

1mm ~ 36.6 pxl, flens

= 70.0 mm

PSfrag replacements

‖u‖[pxl]

z[p

xl]

5 10 15 20 250.63

0.635

0.64

0.645

0.65

0.655

0.66

0.665

PSfrag replacements

a

k

NON-CENTRAL MODEL

X

PSfrag replacements

xy

z

uv

u

p

F

π

γ γ

x

C

W ν

c, t

c

n

m, t

m

Perspective projection + law of reflection

xw = f (u, a,

m, tm,

c, tc,

)pw = f (u, a,

m, tm,

c, tc,

)

xw =

�−1m

(

λ

�>c

�−1u + tc

)

+ tm

pw =

�−1m

(

�>c

�−1u − 2

(

( �>c

�−1u)

·n

‖n‖

)

·n

‖n‖

)

λ, n depend on the shape of a mirror

AUTO-CALIBRATION WITH CENTRAL MODEL

PSfrag replacements x

z

pq X

g

u

S3

sensor plane

g/h

C

h

h(‖u‖)u

s

p =

(

g(‖u‖, a)uh(‖u‖, a)

)

for perspective proj.g(.), h(.) = 1

px

py

pz

a=a0

+ (a − a0)

px

py

pz

a=a0

=

=

px − pxa0

py − pya0

pz − pza0

a=a0

+ a

px

py

pz

a=a0

=

= x + a s,

Using epipolar constaint

p>2

p1 = 0

(x2 + as2)> �

(x1 + as1) = 0

leads to the Quadratic Eigenvalue Problem (QEP):

(

�� + a

�� + a2 �� )f = 0

where f = [ F11 F12 F13 F21 . . . F33 ]> and�

i are functions of coordinates of point correspon-dences (we used WBS [2])

Solving the QEP gives: • mirror parameter a• essential matrix

• inliers/outliers (ransac)

3D REC. WITH NON-CENTRAL MODEL

PSfrag replacements

xi

w1

pi

w1

xi

w2

pi

w2

XW

CC

Since cameras are calibrated, then

RM = argmina,

�c,tc,

�m,tm,

�N∑

i=1

(

|(

xiw1 − xi

w2

)

·(

piw1 × pi

w2

)

|

|piw1 × pi

w2|

)2

RESULTS

REFERENCES

[1] Fitzgibbon A., Simultaneous linear estimation of multiple view geometry and lens distortion. CVPR, 2001.

[2] Matas J., Chum O., Urban M., and Pajdla T., Robust wide baseline stereo from maximally stable extremalregions. BMVC, 2002.

[3] Micusik B. and Pajdla T., Para-catadioptric Camera Auto-calibration from Epipolar Geometry ACCV, 2004.

[4] Swaminathan R. and Grossberg M. D. and Nayar S. K., Caustics of Catadioptric Cameras, ICCV, 2001.

This research was supported by the following projects: MSM 212300013, MSMT Kontakt 22-2003-04, BENOGO–IST–2001–39184 Presented at the Conference on Computer Vision and Pattern Recognition (CVPR), 2004, Washington D.C., USA