Augmented Reality with Planar Homographies

CMU16-720, Computer Vision

October 2020

Abstract: 

In this project, my goal was to estimate the planar homography between two images. Using this homography, I could then warp images and implement an AR application using two video clips. 

Process: 

Before I was able to estimate homography, I had to implement a feature detector and descriptor to find interest points. For this application I implemented a FAST feature detector and a BRIEF feature descriptor. I began by comparing the points of a CV textbook head-on with the same textbook on a desk. From here, I was able to estimate the homography needed to transform a head-on image to this transformed frame. 

Result: 

Once I was able to successfully transform a single image, I applied the same homography estimation to a video segment. The result is a scene from Kung-Fu Panda overlaid on the same CV Textbook moving in its own video.