SIFT SURF Mixed Reality Laboratory Hanyang University http

SIFT & SURF Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

특징점의 종류 • Corner • Blob • Region Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Blob detector Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Blob detector: Lo. G • Laplacian of Gaussian (Lo. G) Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Blob detector: Hessian response • Hessian matrix (H) – 2 nd derivative – Hessian (determinant of H) – Laplacian (trace of H) • Blob detection – Do. H: detect maxima Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

대표적인 특징기술자 • SIFT (Scale-Invariant Feature Transform) – D. Lowe, “Distinctive image features from scaleinvariant keypoints, ” IJCV, vol. 60, no. 2, pp. 91110, 2004. • SURF (Speeded Up Robust Features) – H. Bay, T. Tuytelaars, and L. V. Gool, “SURF: Speeded Up Robust Features, ” European Conference on Computer Vision, 2006. Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

SIFT (SCALE-INVARIANT FEATURE TRANSFORM) Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

SIFT • 4단계 알고리즘 – Scale-space extreme detection – Keypoint localization and filtering – Orientation assignment – Descriptor construction Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Scale-space extreme detection • Gaussian scale-space – 입력 영상에 σ를 증가시켜가며 가우시안 필터 를 적용한 영상을 만든다. – σ가 두 배가 될 때마다 영상을 ½로 다운샘플링 하고 위의 과정을 반복한다. • Do. G (Difference of Gaussian) Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Blob detector: Do. G • Difference of Gaussian – Approximated Lo. G Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Do. G Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Local extreme detection of Do. G • Do. G 영상에서 x, y, s축으로 인접한 26지점보 다 Do. G값의 절대값이 큰(극값을 갖는) 지점 을 찾아 특징점 후보로 선택한다. Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Keypoint localization and filtering • 낮은 극값을 갖는(low contrast) 특징점 후보 제거 • 에지 응답을 갖는 특징점 후보 제거 – Principal curvature가 큰 후보 • 샘플링된 극점이 아닌 subpixel 극점을 찾는다 Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

검출 예 • (a) original image • (b) extreme points (832) • (c) low extreme threshold (729) • (d) large ratio of principal curvature threshold (536) Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Orientation assignment Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

SURF (SPEEDED UP ROBUST FEATURES) Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

SURF • SURF의 특징 – 특징점의 고속 검출 – Haar-wavelet 응답을 이용한 특징기술자 – Laplacian의 부호를 이용하여 매칭속도 향상 Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Fast hessian detector • 박스 필터를 이용하여 hessian 응답을 근사 Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Scale-space • 영상의 크기를 줄이는 대신 필터의 크기를 키움으로써 고속화 Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr

Local maxima detection • Thresholding • Non-maxima suppression • Interpolation (subpixel accuracy) Mixed Reality Laboratory | Hanyang University http: //mr. hanyang. ac. kr