고정 헤더 영역
상세 컨텐츠
본문
TPAMI 2002
Drummond and Cipolla, Cambridge
개요
initialization: 역시 rough hand initialization. 거의 모든 페이퍼가 automatic initialization을 future work로 꼽고 있다는...
edge tracking의 장점:
1. edges are strong enough
2. takes advantage of the aperture problem - limited to 1D search. 라고 하는데, 그보다는 꼭 corner가 아니라도 사용할 수 있으니까 more features can be used 이런 얘기를 해야 하지 않을까?
projection: measurement vector → the subspace (안드로메다로... -.-)
The system projects this m-dimensional measurement vector onto the six-dimensional subspace corresponding to Euclidean transformations. (m = 400개로 edges를 샘플링)
우선, 변수 이해
ξ sample index (m = 400까지)
i group generator index
This subspace is given by the f_ξi which describe the magnitude of the edge normal motion that would be observed in the image at the ξth sample point for the ith group generator.
These can be considered as a set of six m-dimensional vectors which describe the motion in the image for each of the six modes (=basis?) of Euclidean transformation.
즉 f_ξi 는 (6)처럼 tx,ty,tz,rx,ry,rz에 의한 edge sample points의 변화량(?)이라고 볼 때,
결국 (14)를 minimize하는 coefficients의 solution인 β_i = α_i 를 구하는 것이 최종 목적이다.
(14) S = sum_ξ (d_i - β_i f_ξi )^2 → minimize
따라서 ∂S/∂β_i = 0 (15)-(17)
Robustness
substitute a robust M-estimator for the least-squares estimator by replacing the objective function with one that applies less weighting to outlying measurements.
This approach is known as iterative reweighted least squares (IRLS) since s depends on d, which changes with each iteration.
online camera calibation
online camera calibration이 가능. (Fig.10)
skew = 0. four parameters: f, a, u, v
This creates four new vector fields, L_i, (7 <= i <= 10).
Convergence is fast, taking only 5-10 iterations. (웬 iteration이지? 원래 iterative하게 푸는 건가!?)
ill-conditioned (degenerate) configurations: single plane parallel to the image plane
ill-condition 상황을 피하기 위해, well-conditioned 상황에서 먼저 calibration하고, 나머지는 parameter 고정.
기타
visual servoing system에 적용, results를 보임.
3장 complex configurations는 별로 가져다 쓸 것도 없고, 스킵.
Drummond and Cipolla, Cambridge
개요
initialization: 역시 rough hand initialization. 거의 모든 페이퍼가 automatic initialization을 future work로 꼽고 있다는...
edge tracking의 장점:
1. edges are strong enough
2. takes advantage of the aperture problem - limited to 1D search. 라고 하는데, 그보다는 꼭 corner가 아니라도 사용할 수 있으니까 more features can be used 이런 얘기를 해야 하지 않을까?
projection: measurement vector → the subspace (안드로메다로... -.-)
The system projects this m-dimensional measurement vector onto the six-dimensional subspace corresponding to Euclidean transformations. (m = 400개로 edges를 샘플링)
우선, 변수 이해
ξ sample index (m = 400까지)
i group generator index
This subspace is given by the f_ξi which describe the magnitude of the edge normal motion that would be observed in the image at the ξth sample point for the ith group generator.
These can be considered as a set of six m-dimensional vectors which describe the motion in the image for each of the six modes (=basis?) of Euclidean transformation.
즉 f_ξi 는 (6)처럼 tx,ty,tz,rx,ry,rz에 의한 edge sample points의 변화량(?)이라고 볼 때,
결국 (14)를 minimize하는 coefficients의 solution인 β_i = α_i 를 구하는 것이 최종 목적이다.
(14) S = sum_ξ (d_i - β_i f_ξi )^2 → minimize
따라서 ∂S/∂β_i = 0 (15)-(17)
Robustness
substitute a robust M-estimator for the least-squares estimator by replacing the objective function with one that applies less weighting to outlying measurements.
This approach is known as iterative reweighted least squares (IRLS) since s depends on d, which changes with each iteration.
online camera calibation
online camera calibration이 가능. (Fig.10)
skew = 0. four parameters: f, a, u, v
This creates four new vector fields, L_i, (7 <= i <= 10).
Convergence is fast, taking only 5-10 iterations. (웬 iteration이지? 원래 iterative하게 푸는 건가!?)
ill-conditioned (degenerate) configurations: single plane parallel to the image plane
ill-condition 상황을 피하기 위해, well-conditioned 상황에서 먼저 calibration하고, 나머지는 parameter 고정.
기타
visual servoing system에 적용, results를 보임.
3장 complex configurations는 별로 가져다 쓸 것도 없고, 스킵.
