Overview
The goal of this project was to create image mosaics. I first took pictures of a scene, keeping the center of projection constant. I then selected a list of correspondences (points the images overlapped) and created a homography to warp one image into the geometry of the other. The images can then be blended together to form a seamless mosaic, emulating a wide angle lense.
The results came out fairly well, despite me not knowing how to use a camera more complicated than a point-and-shoot. Everything was set to auto on the hybrid DSLR I used, except flash, which was turned off. I also shot everything free-hand, without a tripod stand.
Rectified Images
To create these, I simply selected a square from the table top and carpet, and warped them into a square.
before |
after |
 |
 |
 |
 |
Warped Images
Here is an example of warping an image into another. Results of the blend are in the next section.
Image 1 |
Image 2 |
Image 2 warped |
 |
 |
 |
Blending Results:
I implemented linear, feathering, and multi-scale laplacian blending. Since I did not use a tripod, and took the pictures free-hand, some of the images don't correspond exactly for nearby things (A good example is the circle on the pavement in the 2nd mosaic), as well as high-frequency/detail objects. However, most of the image blends correctly.
The linear blending as done by just averaging the two images at the intersection point, leaving obvious artifacts. My feathering mask did very well, since I created it so that each image "owned" two edges that it was next to, and interpolated all other values in between. The reason I could do this was because I knew that only one image was warped, while the other would still be rectangular. I used the same feathering mask with laplacian blending. However, laplacian blending left some edges at the top and bottom of the intersection, because my input images were already warped to the final size (with a black background), causing gaussian blurring to darken the edge pixels, and the laplacian pyramid to detect the strong edges.
Image1 |
Image2 |
|
|
Linear |
Feathering |
Laplacian |
 |
 |
 |
Image1 |
Image2 |
 |
 |
Linear |
Feathering |
Laplacian |
 |
 |
 |
Image1 |
Image2 |
 |
 |
Linear |
Feathering |
Laplacian |
 |
 |
 |
Bells and Whistles: Times of Day
Image1 |
Image2 |
Combined |
 |
 |
 |
Bells and Whistles: Historic and Modern
Image1 |
Image2 |
Combined |
 |
 |
 |
Bells and Whistles: Two Level vs Multiscale Laplacian Blending
As can be seen from below, two levels of laplacian blending contains more blurs, since fewer of the high frequency components have been extracted. This is most obvious in the parts of the image that does not intersect. The images are shown full sized below, since resizing made the blur much less obvious.
This was the blend that showed the effect the best, because it had the sharpest input images.Two Level |
Multiscale |
 |
|