cathedral.jpg Red shift: ( 8 , 3 ) Green shift: ( 3 , 1 ) |
|
icon.tif Red shift: ( 177 , 45 ) Green shift: ( 81 , 33 ) |
|
emir.tif Red shift: ( 67 , -782 ) Green shift: ( 49 , 24 ) This was the only image in the sample set that failed to align. It's fairly easy to see why, just by taking a look at the images: the blue channel and the red channel, over the entire expanse of the emir's robe, are basically inverts of one another, since the "real" color is a saturated blue. As such, the NCC for the correct alignment will be quite low, and the highest NCC value will have the least amount of overlap in that area. |
|
emir.tif (using Sobel horizontal edge kernel as feature) Red shift: ( 105 , 41 ) Green shift: ( 49 , 24 ) After experimenting with a few different convolutional matrices, I found that using Sobel edges as features instead of the original image data resulted in a correct alignment for the image of the Emir. Interestingly, generic edge detection with a ((-1,-1,-1),(-1,8,-1),(-1,-1,-1)) matrix failed to produce usable results for this image, and using vertical Sobel edges seemed to align the image offset in the X direction properly, but the images remained misaligned vertically. |
|
lady.tif Red shift: ( 119 , 12 ) Green shift: ( 56 , 8 ) |
|
melons.tif Red shift: ( 178 , 13 ) Green shift: ( 82 , 10 ) |
|
monastery.jpg Red shift: ( 3 , 2 ) Green shift: ( -3 , 2 ) |
|
nativity.jpg Red shift: ( 8 , 0 ) Green shift: ( 3 , 1 ) |
|
old_cross.tif Red shift: ( 47 , 2 ) Green shift: ( 20 , 7 ) |
|
three_generations.tif Red shift: ( 112 , 11 ) Green shift: ( 53 , 14 ) |
|
tobolsk.jpg Red shift: ( 6 , 3 ) Green shift: ( 3 , 3 ) |
|
train.tif Red shift: ( 87 , 32 ) Green shift: ( 42 , 6 ) |
|
village.tif Red shift: ( 138 , 22 ) Green shift: ( 64 , 12 ) |
|
workshop.tif Red shift: ( 105 , -12 ) Green shift: ( 53 , 0 ) |
|
monument.tif Red shift: ( 60 , 29 ) Green shift: ( 24 , 21 ) |
|
standard.tif Red shift: ( 120 , -50 ) Green shift: ( 50 , -25 ) |