IMPORTANT

What is different compared to the solution binary?

We have one key difference for our Panoramic procedure.
We use our images clockwise relative to perspective point.
We had to renumber our images in reverse order compared to the order taken by the solution binary.
Because of this, we placed a set of the test panoramic images but in reverse order.
We also spoke to Avanish about this slight modification and he said it would be fine as long as we provide the images in reverse order.

Does our program work well?

We have tested our programs with different image sets including the given image set and everything works out well.

Small observation about handheld images

On some of our image sets, we get 0.00, 0.00 as the x and y feature alignment coordinate. When this happens
increasing the RANSAC threshold value fixes it. The fact that the same thing doesn't happen so often with image sets taken with a tripod
makes sense because the displacement between images is bigger when we take them without a tripod.
And to tolerate this we need to increase the RANSAC threshold value.

PANORAMA

Test sequence

Test image for Project 2, CSE 455 Winter 2012

Note that the image is a little different than what the solution binary generates.
The reason is because we proccess the images in a clockwise manner while the solution goes the other way around.

Sequence with Kaiden panorama head

SSD ratio matching, RANSAC threshold = 2.

"Green Lake" by Michael Hotan (photographer) and Antonius Harijanto, CSE 455 Winter 2012 (full size, 360° viewer)

Sequence taken by hand

SSD ratio matching, RANSAC threshold = 4.

Observe the black cutoff on the right part of the panorama.
Mike intentionally move the camera position when he took the part with black cutoff.
He wanted to see the resulting panorama with such pictures.
It turned out to be pretty great!

"Drumheller Fountain" by Michael Hotan (photographer) and Antonius Harijanto, CSE 455 Winter 2012 (full size, 360° viewer)

ROC comparison of our code vs. SIFT

The area under the ROC curve is the probablity of a randomly picked match to be a true match.
In other words, among all the matches, the probability that a match is a ground truth is the area
under the curve.

ROC comparison for Yosemite - MOPS+SSD (red), MOPS+ratio (green), SIFT+SSD (blue), SIFT+ratio(purple)

ROC comparison for graf (1 to 2) - MOPS+SSD (red), MOPS+ratio (green), SIFT+SSD (blue), SIFT+ratio(purple)

ROC comparison for graf (1 to 4) - MOPS+SSD (red), MOPS+ratio (green), SIFT+SSD (blue), SIFT+ratio(purple)

As expected, the state-of-the-art feature detection algorithm, SIFT, peforms better compared
to the relatively simpler harris corner detection that we implemented for this project.
Ratio test also performs better compared than regular SSD because it reduces the number of ambiguous matches.