CS195g Project 6: Panorama Stitching
Rudy Sandoval (rudy)
April 19, 2010
This project aims to automatically stitch overlapping photos into a panorama. Given two images capable of being a panorama, we first use the Harris feature detector to pick up important points in both images. Then, using a form of adaptive non-maximal suppression, we filter out less important features to end up with a very even distribution across the image space. Descriptors for each of these features are extracted from the image in the form of a 7x7 sampled patch of standardized intensity values surrounding each feature. Matches were found between the two images, and RANSAC was used to determine the best homography for warping one image to the other. Attempting to stitch more than 2 images resulted in no matching features found. Due to not keeping the camera at a relatively constant position, the warp from one image into the other is so strong, any features the second image may have shared with a third are lost. One 3-picture panorama was produced by creating two two-image panoramas with a shared middle picture, then stitching. The results are shown below.
- Globe panoramas were automatically created given a panorama shot. I took my image space to be -Pi to Pi on the x-axis, and 0 to the preferred radius in the y. By taking a second grid, whose origin was centered, and converting it into polar coordinates, I could interpolate between these values and my original image. Some problems with this algorithm include a seam along theta = Pi/-Pi. Poisson blending was tried, but could not effectively hide the edge. There is also a singularity at r = 0, where all points along y = 0 of my image converge to. After searching around, it seems most people solve this problem by taking my minimum r to be greater than 0, leaving a whole in my globe, which is then filled using some form of texture synthesis. One image in particular worked very well, where the ground was uniform enough not to notice the singularities, and having a full 360 degree panorama hides the seam. The results are shown below.
Results Images
Globe Panoramas
