Varghese Mathew and Faisal Khan, CS 766 Project 2

Source File	Explanation
load_images.m	Loads all images in a given directory.
cylinder_projection.m	Projects an image onto cylindrical coordinates.
model_homography.m	Computes the homography between two images using SIFT.
sift	The SIFT executable from the SIFT Program.
sift.m	Modified version of Matlab SIFT.
sift_matcher.m	Our code that does feature matching using SIFT features. Based on match.c in SIFT package.
stich.m	Given two images and the homography between the, this code stitches them together using the feathering blending method.
run.m	This is the core module. This code takes a set of images, and an array of pairwise homographies, and creates the panorama from them. It can create both partial and complete (full 360 degree) panoramas.
run_normal.m	This is a wrapper over run.m. This code take a directory as argument, loads all the images in the directory, rotates them as specified, computes pairwise homographies and then calls run to create the complete panorama.
run_hdr.m	This code is aimed towards creating high dynamic range panoramas. This takes a list of directories, and a key directory as argument. For creating HDR panorama, all constituent panoramas at different exposures should be stitched together using the same homographies. So we compute homography using the images in the key directory and then use that homography to create panoramas for images in each of the given directories in the list.

We captured 18 images each for 3 full view (360 degree) panoramas
We also captured 10 images for one vertical partial panorama of a building.
We tried to capture 11 (different exposures) * 18 images for an HDR panorama. However, fully charged batteries only last for 11 x 13 images on the given camera model. Therefore we could only gather 11 x 13 images for a partial HDR panorama.
We captured all images in Manual mode, fixing all tunable parameters to reduce vignetting.

We captured one of the panoramas with "Canon SX100 IS, tag 6418204024". However, the Kaidan head provided with it was causing a huge Y-drift. So we had to get that replaced. (Still in the extra time we were given, we coded up a complete shear algorithm which recovered the panorama from this image set.
We captured the rest of the images with "Canon SX100 IS, tag 6418203978".

For this, we take each point (x,y) on the cylindrical surface, and compute the inverse transform using the formulae discussed in the class to get (x', y') which is the coordinate on the source image. Now we copy pixel (x', y') of source image onto pixel (x,y) on target cylindrical image.
The advantage of doing it this way is that, we ensure that every pixel on the cylinder surface get assigned a value.
We also incorporate the lens distortion model into our cylindrical projection code, to remove all distortions.

We use SIFT to get us a list of key points.
Then we use the match algorithm to get potential matches. We model our match algorithm based on the match.c code provided with SIFT program.
After the match, we evaluate the best fitting homography using RANSAC algorithm iterating 1 to 100.
We also note that the shear in Y direction is given by the ratio of the cumulative homography translation along Y direction to that along the X direction.

Here we stitch one image at a time onto the cululative result.
We also do feathering to blend the images so that there's no intensity change bands visible.
We actually stitch half of the first image at the beginning and half at the end (for full view panoramas) to improve accuracy.

Often we notice that the images have a shear in the Y direction (due to bad kaidan heads). We incorporate a shearing algorithm to completely remove this effect.
This step is done between Stitching and cropping of the images.