Histogram Processing Histogram hix x0 255 i1 N

256 N 밝기값 h 1(x) h 2(x) h. N(x) 0 1 2 400 106

256차원의 히스토그램을 n차원으로 바꿈 en Depth from the center of mass Let f(x) be

Visual Correspondence Search • It is implicitly assumed that all pixels in a support

Visual Correspondence Search • Dissimilarity computation and disparity selection Right image i’ Left image

Slides: 9

Download presentation

Histogram Processing • Histogram hi(x), x=0, …, 255, i=1, …, N • Hi(x)=hi(x)-mean of hi(x), x=0, …, 255 • Covariance matrix C = (HTH)/(N-1), H=N 256 matrix • PCA (Principal Components Analysis) • C의 256개의 eigenvalue를 구하고 이 가운데 상위 n개를 선정 • 상위 n개의 eigenvalue에 대응되는 eigenvector로 eigenspace 구축 (모든 아이겐벡터들은 서로 직교함): Feature vector matrix (F) = [e 1 e 2 … en] (256 n matrix) • Final data = FT HT

256 N 밝기값 h 1(x) h 2(x) h. N(x) 0 1 2 400 106 261 3 36 4 5 6 409 1255 1867 252 253 254 255 202 25 54 20 333 506 767 803 1212 2467 4334 33 122 50 각각의 밝기값에서 돗수의 평균을 구하고 돗수로부터 그 평균을 빼줌. 각 밝기값 돗수의 평균이 0이 됨

256차원의 히스토그램을 n차원으로 바꿈 en Depth from the center of mass Let f(x) be a histogram of a new image e 1 e 2

Visual Correspondence Search • It is implicitly assumed that all pixels in a support window are from similar depth in a scene and therefore, they have similar disparities • Support weight between pixel i and j within a window pixel i, the pixel under consideration is determined empirically. : color difference : position difference

Visual Correspondence Search • Dissimilarity computation and disparity selection Right image i’ Left image i j d j’