CSE576 Spring 2005

Project 2

T. Scott Saponas

Introduction

For project 2 I implemented a panormaic image stitcher. This image stitcher takes images in order with knowledge of the focal length and radial distortian coeffecients and in several steps stiches them into a panorama (360 degrees). I modified my project 1 feature detector and changed my descriptor to a simple grayscale window descriptor normalized for illumination.

The Test Squence

+ + + + + + + + + + + + + + + + +
=
[ZOOM]  Panorama

Taken from a Kaidan Head With Cannon PowerShot A10

+ + + + + + + + + + + + + + + + +
=
[ZOOM]  Panorama

Taken from a Kaidan Head With Cannon PowerShot A10

+ + + + + + + + + + + + + + + + +
=
[ZOOM]  Panorama

Handheld With Cannon PowerShot A10

+ + + + + + + + + + + + + + + + + + + +
=
[ZOOM]  Panorama

First Hand-Held Sequence With Nikon D100 & Tamron 28-75/F2.8 @ 75/F11

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +
=
[ZOOM]  Panorama

What Worked

I was particularly happy with how my vertical panorama came out (the first one listed after the test set). I thought it would be interesting to see a panorama using a diffrent degree of freedom that horizontal pan. Although there is a very noticible exposure diffrence between when the camera had mostly ground in the image and mostly sky in the image, the feathering did a darn good job of blending them together. If you look at the trunk of the big tree at the top of the image you'll see it going from light to dark in the bottom half of it. Other than that it's not too noticible, you can just see dark things (the path) and light things (the sky and clouds) in the same image. Even the feathering on the tree trunk doesn't look too bad - I think it looks cool.

In general my project 1 feature aligner seemed to do well. I'm sure my images would have aligned a little bit better had I used the provided SIFT binary - but I'm happy with the results.

The blending seemed to work well in most images.

What Didn't Work

Both my handheld sequences seemed to have much more vertical drift than my tripod sequences. However, this is not too surprising since I wasn't using a tripod.

In all the sequences shot with Cannon powershot there are exposure diffrences between the images. The feathering seemed to blend them together well although it's rather noticible in the test sequence.

I found that when the overlapping portion of two pictures was mostly water, horizon, grass, or tree branches there were some allignment problems. In the case of just horizon (sky, hills, and water) there just aren't that many good features from the standpoint of my project 1 corner detector to try to align. In some places in the hills it did actally work pretty well when you could make out the rectangular nature of a couple buildings. Where there was mostly close water, grass or tree branches the wind played havoc with my alignment. Not surprisingly, when those things were moving they weren't the same in adjacent images and were harder to align.

Extra Credit

In my handheld sequence with my camera the subject appears in multiple frames doing "YMCA". I think this looks kind of cool, however there is some ghosting and alignment issues from the subject moving - and not entirely following the photographer's directions :).

Easter Eggs and Screw Ups

What happens when you warp things wrong...
[ZOOM]