2010. 7. 18. 17:05
Computer Vision
Accurate Camera Calibration from Multi-View Stereo and Bundle Adjustment
Yasutaka Furukawa & Jean Ponce, CVPR 2008
Yasutaka Furukawa
http://www.cs.washington.edu/homes/furukawa/
Jean Ponce
http://www.di.ens.fr/~ponce/
http://www-cvr.ai.uiuc.edu/ponce_grp/
DxO Labs (렌즈 왜곡 교정 소프트웨어)
http://www.dxo.com/
PMVS (Patch-based Multi-view Stereo Software) version 2
http://grail.cs.washington.edu/software/pmvs/
CMVS (Clustering Views for Multi-view Stereo)
http://grail.cs.washington.edu/software/cmvs/
sba : A Generic Sparse Bundle Adjustment C/C++ Package Based on the Levenberg-Marquardt Algorithm (Manolis Lourakis)
http://www.ics.forth.gr/~lourakis/sba/
image-based modeling
multi-view stereovision (MVS)
> two main approach camera calibration problem
1) chart-based calibration (CBC)
2) structure from motion (SFM) + auto-calibration + bundle adjustment (BA)
* selection of feature correspondences (SFC)
eg. RANSAC
DxO Optics Pro - Lens Distortion
Yasutaka Furukawa & Jean Ponce, CVPR 2008
yasutakaFurukawa_2008cvpr.pdfYasutaka Furukawa
http://www.cs.washington.edu/homes/furukawa/
Jean Ponce
http://www.di.ens.fr/~ponce/
http://www-cvr.ai.uiuc.edu/ponce_grp/
DxO Labs (렌즈 왜곡 교정 소프트웨어)
http://www.dxo.com/
PMVS (Patch-based Multi-view Stereo Software) version 2
http://grail.cs.washington.edu/software/pmvs/
CMVS (Clustering Views for Multi-view Stereo)
http://grail.cs.washington.edu/software/cmvs/
sba : A Generic Sparse Bundle Adjustment C/C++ Package Based on the Levenberg-Marquardt Algorithm (Manolis Lourakis)
http://www.ics.forth.gr/~lourakis/sba/
image-based modeling
multi-view stereovision (MVS)
> two main approach camera calibration problem
1) chart-based calibration (CBC)
2) structure from motion (SFM) + auto-calibration + bundle adjustment (BA)
* selection of feature correspondences (SFC)
eg. RANSAC
Standard BA algorithms optimize both the scene point and camera parameters by minimizing the sum of squared reprojection errors.
Unlike BA algorithms, multi-view stereo algorithms are aimed at recovering scene information alone given fixed camera parameters.
DxO Optics Pro - Lens Distortion
In practice, we have found PMVS to be robust to errors in camera parameters as long as the image resolution matches the corresponding reprojection errors - that is, when features to be matched are roughly within two pixels of the corresponding 3D points.
> Algorithm
1) initializing feature correspondences
1-1) building image pyramids
1-2) for the given camera parameters, obtaining a conservative estimate of the expected reprojection error
1-3) sub-sampling
2) refining feature correspondences
2-1) determining a patch and the local image texture inside
2-2) optimizing the reference camera with the conjugate gradient method
2-3) removing outliers and updating the corresponding visibility information
3) updating the camera parameters with the SBA bundle adjustment software
4) repeat 1)-3) 4 times with the updated expected reprojection error and the fixed pyramid level
1) initializing feature correspondences
1-1) building image pyramids
1-2) for the given camera parameters, obtaining a conservative estimate of the expected reprojection error
1-3) sub-sampling
2) refining feature correspondences
2-1) determining a patch and the local image texture inside
2-2) optimizing the reference camera with the conjugate gradient method
2-3) removing outliers and updating the corresponding visibility information
3) updating the camera parameters with the SBA bundle adjustment software
4) repeat 1)-3) 4 times with the updated expected reprojection error and the fixed pyramid level