MOSAIC

CS 766 Computer Vision Assignment 2 - Spring 2014

Project done by : Pradeep Rajendran, Gautam Umesh Sargur

Aim

This project aims to construct a Mosaic (Panorama) from a set of images acquired about the optical center of a camera of known focal length using a kaidan head by estimating the homographies between pairs of images from features extracted using SIFT algorithm.

Algorithms implemented

Mapping images to cylindrical co-ordinates by inverse warping of the input images.

RANSAC algorithm to compute homographies from matches acquired by using SIFT features and David Lowe's matching algorithm.

Aligning all cylindrically mapped images from the estimated homographies.

Blending images using laplacian blending (bonus).

Image Set used to get the Mosaic image

We took one set of photographs. It gives a full 360 degree view of the street. There are 15 images. These images are taken in the landscape orientation with the camera having the following parameters:
Camera Model: Canon A640
Identification Tag: 4726208879
Resolution: 480x640
Focal Length: 660.8799
k1: -0.18533
k2: 0.21517

Approach

Mapping images to cylindrical co-ordinates: We used inverse warping function to map each point in the input image to its corresponding cylindrical co-ordinates. This is implemented in the MATLAB file cylinderTransform.m. It takes as input a directory in which input images are placed and a focal length and returns a set of mapped images.

RANSAC algorithm: We first extract features from the input set of images which have been warped by using the VLFEAT library's vl_sift function which returns the position and the descriptor of each interest point. The function getSiftFeatures.m does this for all input images. Once SIFT features are computed, we use the vl_ubdmatch function which gives initial estimates of matches for RANSAC to work. The file computeHomographies.m implements the RANSAC algorithm. We used 11 point correspondences instead of the required 4 to get better and more accurate results and run RANSAC for 1000 iterations.

Aligning Images: We then transform the images according to the homographies that are obtained from the above step in a pairwise fashion.

(Bonus) We implemented the laplacian blending algorithm to blen the images that result from the above operation in a smooth and non-glitchy fashion. Pyramidal blending involves collecting multiple resolutions of the same image and blending them at these different resolutions. The blending is based on a "alpha" mask with a feathered edge at the line of blending. Each level in the pyramid is produced by first blurring the image and down-sampling it by a factor of 2. And, then up-sampling the image via interpolation. This yields a resultant image which has the same dimensions as the original image. This process is also done for the "alpha" mask. Each level is multiplied by its mask to produce a masked-level. The left and right masked-levels are then added to produce the final image at a particular level. The final image is a sum over all such levels. The code for this blending is present in alignblend.m, getPyr.m, pyr_expand.m, pyr_reduce.m and pyrReconsturct.m files.