Design decisions

We decided to ignore all features whose descriptor window would go out of bounds. Since the pictures will be relatively large as compared with the windows, we will still have many features. This makes determining the descriptors easier and more robust. When calculating the harris values, we used the boarder values (clamping) for values out of the pictures. We choose this becuase having 0 might produce extraneous edges.

Performance

ROC

Curves for Graffiti:

AUC for comparison between 1 and 2:

Simple: 0.774559

Ratio: 0.818082

AUC for comparison between 1 and 4:

Simple: 0.691226

Ratio: 0.739354

$Description: Description: Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\grafplot1to2.png$ $Description: Description: Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\grafplot1to4.png$ $Description: Description: Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\grafthreshold1to2.png$ $Description: Description: Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\grafthreshold1to4.png$

Harris operator for img1.ppm:

$Description: Description: Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\harris.jpg$

Yosemite:

AUC for comparison between 1 and 2:

Simple: 0.905444

Ratio: 0.945654

$Description: Description: Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\Yosemiteplot1to2.png$

$Description: Description: Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\Yosemitetheshold1to2.png$

Harris image for Yosemite:

$Description: Description: Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\harrisy.jpg$

Average AUC

Leuven:

Simple:

SSD: 0.287985

Ratio: 0.605873

MOPS:

SSD: 0.6863996

Ratio: 0.870592

Bikes:

Simple:

SSD: 0.170991

Ratio: 0.509464

MOPS:

SSD: 0.657506

Ratio: 0.816662

Wall:

Simple:

SSD: 0.316945

Ratio: 0.650645

MOPS:

SSD: 0.692474

Ratio: 0.816313

Strengths/Weaknesses

MOPS seems invariant to translation, intensity changes, rotations, and limited amount of affine changes as shown below. However, it did badly with large perspective changes such as those in graffiti.

Our own pictures!

A Ferris wheel!

$Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\Ferris Wheel\Ferris Wheel 1.jpg$

$Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\Ferris Wheel\Ferris wheel 2.jpg$

The harris matrix:

$Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\Ferris Wheel\harris.jpg$

And the comparison:

$Description: Description: C:\Users\Nick\Documents\GitHub\ComputerVision\ROC\Ferris Wheel\Ferriscomparison.png$

The MOPS descriptor worked reasonably well on this image, but it’s invariance to rotation resulted in a lot of false matches with the various metal tubes forming the Ferris Wheel’s structure. The uniform grey background also made finding matches harder, though the threshold on local max did prevent the clouds from being marked as features outright.