CV4‐문제 - waegari/waegari.github.io GitHub Wiki
컴퓨터 비전: Line Patterns 및 Ellipse Detection
1부: 문제
객관식 문제 (20문제)
1. Line fitting 과정에서 발생하는 어려움으로 맞는 것을 모두 고르시오.
- a) Clutter 문제로 인한 불필요한 edge point 존재
- b) Multiple models로 인한 점 할당 어려움
- c) 선의 일부만 감지되는 불완전한 데이터
- d) Edge detection 알고리즘의 계산 복잡성
2. RANSAC 알고리즘에 대한 설명 중 틀린 것을 모두 고르시오.
- a) Outlier의 영향을 줄이기 위해 inlier만 사용하는 방식이다
- b) 모든 포인트를 사용하여 초기 모델을 계산한다
- c) 무작위로 선택된 sample로 model을 계산한다
- d) Inlier 수가 충분하면 모든 inlier로 least-squares 추정을 재계산한다
3. RANSAC에서 필요한 iteration 수(k)를 결정하는 데 관련된 요소로 맞는 것을 모두 고르시오.
- a) Inlier 비율(w)
- b) 모델 정의에 필요한 point 수(n)
- c) 모델 파라미터 수
- d) Image 해상도
4. Hough Transform에 대한 설명으로 맞는 것을 모두 고르시오.
- a) 높은 비율의 outlier를 처리할 수 있는 voting 전략이다
- b) Line detection을 위해 parameter space를 이산화(discrete)한다
- c) 직선 표현을 위해 y = mx + b 방정식만 사용한다
- d) Accumulator array를 사용하여 vote를 계산한다
5. Line detection을 위한 Hough Transform에서 사용하는 polar representation은?
- a) y = mx + b
- b) x·cos θ + y·sin θ = ρ
- c) (x-h)²/a² + (y-k)²/b² = 1
- d) ax² + bxy + cy² + dx + ey + f = 0
6. Ellipse detection algorithm에 대한 설명으로 틀린 것을 모두 고르시오.
- a) Image의 edge point를 사용하여 ellipse의 center를 찾는다
- b) 모든 edge point를 동일한 weight로 처리한다
- c) Ellipse property: 평행한 chord의 mid-point들과 ellipse center는 collinear하다
- d) Hough Transform은 ellipse detection에 절대 사용할 수 없다
7. Edge detection threshold가 ellipse detection에 미치는 영향으로 맞는 것을 모두 고르시오.
- a) High threshold: noise 감소, inlier edge 감소, 낮은 recall
- b) Low threshold: 많은 noise, 많은 inlier와 outlier edge, 낮은 precision
- c) Threshold는 detection 성능에 영향을 미치지 않는다
- d) Mid threshold: 항상 최적의 detection 성능을 보장한다
8. Arc detection 과정에서 제거되는 noisy edge로 맞는 것을 모두 고르시오.
- a) Straight edge (Δy=0, Δx=0, Δy/Δx = constant)
- b) Small edge (적은 수의 pixel로 구성)
- c) Sharp turn이 있는 edge
- d) 이미지 경계에 위치한 모든 edge
9. Arc quadrant classification에서 arc gradient D(arc^m)와 arc convexity C(arc^m)에 대한 설명으로 맞는 것을 모두 고르시오.
- a) D(arc^m) = sign(Δy)·sign(Δx)
- b) C(arc^m)는 arc 위아래 영역의 면적 비교로 결정된다
- c) Area(U) = Area(O)일 경우 arc는 ellipse의 일부로 간주된다
- d) Gradient와 convexity는 arc의 quadrant 결정에 사용된다
10. Arc의 quadrant label 지정 방법으로 맞는 것은?
- a) Quadrant I: D(+), C(+)
- b) Quadrant II: D(-), C(-)
- c) Quadrant III: D(+), C(-)
- d) Quadrant IV: D(+), C(+)
11. Ellipse detection을 위한 arc selection rule로 틀린 것을 모두 고르시오.
- a) 최소 2개의 arc 필요
- b) 같은 quadrant level의 arc만 선택
- c) 두 method로 계산된 center가 threshold distance 내에 있어야 함
- d) Prior information을 활용할 수 있음
12. Ellipse parameter estimation 방법으로 맞는 것을 모두 고르시오.
- a) Ellipse property 활용: 평행 chord의 mid-point로 center 추정
- b) Ellipse fitting equation 사용: ax² + bxy + cy² + dx + ey + f = 0
- d)/(2(ae-b²/4))
- d) Parameter estimation은 항상 edge point 수에 비례하는 계산 복잡성을 가짐
13. Ellipse detection의 validation 단계에서 수행하는 작업으로 맞는 것을 모두 고르시오.
- a) 결과 ellipse가 선택된 arc에 잘 맞는지 확인
- b) Fitting error 계산
- c) Fitting error가 threshold보다 크면 ellipse 폐기
- d) 모든 ellipse를 무조건 검증 없이 결과로 사용
14. Ellipse detection의 clustering 단계에 대한 설명으로 맞는 것을 모두 고르시오.
- a) Multiple valid detection 처리
- b) Center 간 distance, axes, rotation으로 similarity 판단
- c) Cluster 내 ellipse averaging 또는 maximum fitting score 선택
- d) Clustering은 ellipse detection에서 선택적 단계이다
15. Ellipse detection이 직면하는 어려움으로 맞는 것을 모두 고르시오.
- a) Incomplete edges
- b) Occlusion
- c) Outlier edges
- d) Hough Transform의 적용 불가능성
16. RANSAC과 Hough Transform을 비교할 때 맞는 설명을 모두 고르시오.
- a) RANSAC은 무작위 sampling, Hough Transform은 voting 기반
- b) RANSAC은 outlier가 많을 때 비효율적, Hough Transform은 높은 outlier 비율에서도 유용
- c) RANSAC은 model parameter를 직접 추정, Hough Transform은 parameter space에서 최대값 찾기
- d) RANSAC은 line detection에만 적용 가능, Hough Transform은 다양한 shape detection에 적용 가능
17. Ellipse detection performance 평가 지표로 사용되는 것을 모두 고르시오.
- a) Precision
- b) Recall
- c) F-measure
- d) Signal-to-noise ratio
18. Line detection에서 voting 기법의 특성으로 맞는 것을 모두 고르시오.
- a) 특징점들이 호환 가능한 모든 model에 vote
- b) 많은 vote를 받은 model parameter 선택
- c) Noisy feature들의 vote는 영향을 미치지 않음
- d) 일부 feature가 missing되어도 model이 여러 fragment를 span 가능
19. Ellipse detection pipeline의 올바른 순서는?
- a) Edge detection → Arc detection → Arc quadrant classification → Arc selection → Parameter estimation → Validation → Clustering
- b) Edge detection → Arc detection → Arc selection → Arc quadrant classification → Parameter estimation → Validation → Clustering
- c) Edge detection → Arc detection → Arc quadrant classification → Parameter estimation → Arc selection → Validation → Clustering
- d) Edge detection → Parameter estimation → Arc detection → Arc quadrant classification → Arc selection → Validation → Clustering
20. Arc convexity classification에서 Area(U)와 Area(O)를 비교할 때, 다음 중 맞는 것을 모두 고르시오.
- a) Area(U) > Area(O) → Upper Convex(+)
- b) Area(O) > Area(U) → Lower Convex(-)
- c) Area(U) = Area(O) → Straight line으로 폐기
- d) Area(U) = Area(O) → 특별한 ellipse로 분류
주관식 단답형 문제 (5문제)
21. Line detection을 위해 unbounded parameter space [m,b]의 문제를 해결하기 위해 사용하는 polar representation 방정식을 쓰시오.
22. RANSAC 알고리즘에서 모든 sampling이 실패할 확률을 나타내는 수식을 쓰시오.
23. Ellipse property에서 평행한 chord의 mid-point들과 ellipse의 center가 같은 선 위에 있는 특성을 무엇이라고 하는가?
24. Arc의 quadrant classification에서 D(arc^m)가 positive이고 C(arc^m)가 negative일 때, 이 arc는 어떤 quadrant에 속하는가?
25. Ellipse detection performance 평가에서 precision과 recall을 결합한 단일 지표는 무엇인가?
2부: 정답
객관식 정답
- a, b, c
- b
- a, b
- a, b, d
- b
- b, d
- a, b
- a, b, c
- a, b, d
- a, c
- b
- a, b, c
- a, b, c
- a, b, c
- a, b, c
- a, b, c, d
- a, b, c
- a, b, d
- a
- a, b, c
주관식 정답
- x·cos θ + y·sin θ = ρ
- (1-w^n)^k
- collinear
- Quadrant III
- F-measure
3부: 해설
객관식 해설
1. a, b, c
- Line fitting의 주요 어려움은 clutter(불필요한 edge point), multiple models(여러 선 존재), 불완전한 데이터(missing line part)이다.
- d는 edge detection 알고리즘의 계산 복잡성이 아닌 noise in measured edge points가 어려움이다.
2. b
- RANSAC은 모든 포인트가 아닌 무작위로 선택된 최소한의 point로 초기 모델을 계산한다.
- 나머지 설명은 모두 RANSAC의 특성과 일치한다.
3. a, b
- RANSAC에서 필요한 iteration 수(k)는 inlier 비율(w)과 model 정의에 필요한 point 수(n)에 의해 결정된다.
- c와 d는 iteration 수 결정과 직접적인 관련이 없다.
4. a, b, d
- Hough Transform은 voting 전략으로, discrete parameter space와 accumulator array를 사용한다.
- c는 틀린 설명으로, Hough Transform은 주로 polar representation(x·cos θ + y·sin θ = ρ)을 사용한다.
5. b
- Line detection을 위한 Hough Transform의 polar representation은 x·cos θ + y·sin θ = ρ이다.
6. b, d
- Ellipse detection에서는 edge point를 동일 weight가 아닌 다양한 특성(gradient, convexity 등)으로 분석한다.
- Hough Transform은 ellipse detection에 사용될 수 있다.
7. a, b
- High threshold는 noise를 줄이지만 inlier edge도 줄여 recall이 감소한다.
- Low threshold는 많은 noise와 edge를 포함하여 precision이 감소한다.
8. a, b, c
- Noisy edge로 제거되는 것은 straight edge, small edge, sharp turn이 있는 edge이다.
- 이미지 경계에 위치한 모든 edge를 제거하는 것은 아니다.
9. a, b, d
- Arc gradient는 D(arc^m) = sign(Δy)·sign(Δx)로 계산된다.
- Convexity는 arc 위아래 영역 면적 비교로 결정된다.
- c는 틀린 설명으로, Area(U) = Area(O)일 경우 straight line으로 간주되어 폐기된다.
10. a, c
- Quadrant I: D(+), C(+)
- Quadrant III: D(+), C(-)
- Quadrant II는 D(-), C(+)이고, Quadrant IV는 D(-), C(-)이다.
11. b
- Arc selection에서는 다른 quadrant level의 arc를 선택해야 한다(같은 quadrant가 아님).
12. a, b, c
- Ellipse center 추정 방법으로는 ellipse property 활용과 fitting equation 사용이 있다.
- Center 좌표 계산식은 x₀ = (cd-be)/(2(ae-b²/4)), y₀ = (ae-bd)/(2(ae-b²/4))이다.
- d는 틀린 설명으로, parameter estimation의 계산 복잡성은 다양한 요소에 의존한다.
13. a, b, c
- Validation 단계에서는 ellipse가 arc에 맞는지 확인, fitting error 계산, threshold 초과 시 폐기를 수행한다.
- 모든 ellipse를 무조건 사용하지는 않는다.
14. a, b, c
- Clustering은 multiple detection 처리, similarity 판단, 최종 ellipse 선택을 위한 필수 단계이다.
- 선택적 단계가 아니라 pipeline의 필수 부분이다.
15. a, b, c
- Ellipse detection이 직면하는 어려움은 incomplete edges, occlusion, outlier edges 등이다.
- Hough Transform은 적용 가능하다.
16. a, b, c, d
- RANSAC은 무작위 sampling, Hough Transform은 voting 기반이다.
- RANSAC은 outlier가 많을 때 비효율적이고, Hough Transform은 높은 outlier에서도 유용하다.
- RANSAC은 직접 parameter 추정, Hough Transform은 parameter space에서 최대값을 찾는다.
- RANSAC은 다양한 model에, Hough Transform도 다양한 shape에 적용 가능하다.
17. a, b, c
- Ellipse detection 성능 평가에는 precision, recall, f-measure가 사용된다.
- Signal-to-noise ratio는 일반적으로 사용되지 않는다.
18. a, b, d
- Voting 기법에서는 모든 feature가 호환 가능한 model에 vote하고, 많은 vote를 받은 parameter가 선택된다.
- c는 틀린 설명으로, noisy feature도 vote에 영향을 미친다.
- 일부 feature가 missing되어도 model이 여러 fragment를 span할 수 있다.
19. a
- Ellipse detection pipeline의 올바른 순서는: Edge detection → Arc detection → Arc quadrant classification → Arc selection → Parameter estimation → Validation → Clustering이다.
20. a, b, c
- Area(U) > Area(O)는 Upper Convex(+)
- Area(O) > Area(U)는 Lower Convex(-)
- Area(U) = Area(O)는 straight line으로 간주되어 폐기된다.
주관식 해설
21. x·cos θ + y·sin θ = ρ
- 이것은 line detection을 위한 Hough Transform의 polar representation이다.
22. (1-w^n)^k
- RANSAC에서 모든 k sampling이 실패할 확률을 나타내는 수식으로, w는 inlier 비율, n은 model 정의에 필요한 point 수이다.
23. collinear
- 평행한 chord의 mid-point들과 ellipse center가 한 직선상에 있는 특성을 collinear라고 한다.
24. Quadrant III
- D(arc^m)가 positive이고 C(arc^m)가 negative일 때 해당 arc는 Quadrant III에 속한다.
25. F-measure
- F-measure는 precision과 recall을 조화평균하여 단일 지표로 결합한 값이다.