pointfeaturematchingmethods1

pointfeaturematchingmethods1内容摘要：

tection 1 (scan NHBDs) CVPR04 Stockman (27 June 04) 17 Interesting neighborhood detection 2 (rate interest) v1 v2 v3 v4 CVPR04 Stockman (27 June 04) 18 Interest point detection • summary of matches between two images is shown • endpoints are from image 1 and image 2 • endpoints represent high interest values • endpoints also have similar neighborhood structure measured by cross correlation (see below) Change is so large that this would not be regarded as image registration. “background” unmatched CVPR04 Stockman (27 June 04) 19 Matching by cross correlation  best matching neighborhood is that which has smallest energy in the squared difference image ( N1 – N 2)  can normalize to benefit from the CauchySchwartz theorem  don’t do it unless both N1 and N2 have magnitude significantly above noise level CVPR04 Stockman (27 June 04) 20 Crosscorrelation matching  Correspond pixels of the 2 neighborhoods  sum the squares of the differences of pixel values  best match has least sum  (L2 norm) a b c d e f g h 2222 )()()()( hdgcfbeav Best match has the lowest v value. Can search Image 2 for neighborhood that best matches the one in Image 1. Usually, neighborhoods are much larger than 2 x 2. 16 x 16 in MPEG Image 1 Image 2 CVPR04 Stockman (27 June 04) 21 MPEG motion pression Video frames N and N+1 show slight movement: most pixels are same, just in different locations. 16 x 16 block A in image 2 has best match to 16 x 16 block B in image 1. A B CVPR04 Stockman (27 June 04) 22 Best matching blocks between video frames N+1 to N (“motion” vectors) 16x16 blocks are matched from I1 to I2. The bulk of the vectors show the true motion of the airplane taking the pictures. The long vectors are incorrect motion vectors, but they do work well for image pression! Best matches from 2nd to first image shown as vectors overlaid on the 2nd image. (Work by Dina Eldin.) CVPR04 Stockman (27 June 04) 23 Radial mass transform Useful for (a) det。