IMAGE WARPING and MOSAICING
Nick Vandal
Overview
The goal of this assignment is to get your hands dirty in different
aspects of image warping with a “cool” application -- image mosaicing.
Image Rectification
Input |
|
Output |
|
Input |
|
Output |
|
Part I: Blend the images into a mosaic
The images below correspond to Part I of the project, where hand labeled correspondences are defined using a MATLAB GUI. Two band blending was found to provide the best results, although the full multi-scale pyramid blending algorithm was also implemented.
Feature matching for Autostitching
The images below correspond to Part II of the project, interest points are detected and matched, and homographies are determined using RANSAC.
Step 0: Input Images
Step 1: Single Scale Harris Interest Points
Step 2: Non-maximal Adaptive Supression
Step 3: Matching of Features
Step 4: Warping using RANSAC-determined Homograpy
Step 5: Blending using 2-band method
Gallery of Mosaics
Part III: Tour into the Picture: Single View Modeling
The goal of this project is to create a simple, planar 3D scene from a single photograph. The project will follow the description in Tour into the Picture by Horry et al. in modeling the scene as a 3D axis-parallel box. First, we will let the user specify simple constraints on that box (the back wall plus thevanishing point).
All code, including the 3D rendering is performed with native MATLAB code.
Bells and Whistles
Better Blending
I implemented full multi-scale Laplacian pyramid blending. Comparisions between the two blending techniques are shown below. In general, it appears as if the two-band blending technique (where low frquencies are linearly blended and high frequencies are taken from the source image with the largest alpha weighting) performed better than the linear blend at all frequencies.
Two Band Blending | Mult Scale Laplacian Blending |
|
|
Movie Fly-through of Oxford image
Panorama Recogntion
Given a set of unordered images, some of which might form panoramas, my code automatically extracts features from each image, and performs pairwise matches between all images. Interest points are matched and RANSAC is performed to determine a homography between each pair of images. The determination of which pairs are valid correspondences is based on the number of inliers (probabilistic model from M. Brown and D.G. Lowe "Recognising Panoramas" paper). I then warp all other images to the perspective of the most connected image and find the shortest path from each of the other images to this most connected image using the pairwise homographies. If there is no path from a particular node, it is considered to not be a part of the panorama.