The custom descriptor generates 10px*10px samples from a 41px*41px patch samples. It uses a 7x7 Gaussian kernel to blur the gray image before sampling.

ROC curves

Yosemite

$Description: C:\Documents and Settings\yl844\Desktop\gnuplot\binary\plot.roc.png$

graf

$Description: C:\Documents and Settings\yl844\Desktop\gnuplot\binary\plot.roc.png$

Harris operator

Yosemite

$Description: C:\Documents and Settings\yl844\Desktop\Dropbox\CV HW1\Skeleton Code\Features\Debug\Yosemite.png$

graf

$Description: C:\Documents and Settings\yl844\Desktop\Dropbox\CV HW1\Skeleton Code\Features\Debug\graf.png$

Average AUC

Graf

Simple 5x5 window descriptor, SSD

average error: 294.271096 pixels

average AUC: 0.527415

Simple 5x5 window descriptor, ratio test

average error: 294.271096 pixels

average AUC: 0.536091

MOPS, SSD

average error: 295.068524 pixels

average AUC: 0.505722

MOPS, ratio test

average error: 295.068524 pixels

average AUC: 0.530100

Custom, SSD

average error: 290.061380 pixels

average AUC: 0.499081

Custom, ratio test

average error: 290.061380 pixels

average AUC: 0.523495

Average AUC: 0.520317333

Leuven

Average AUC: 0.687962

Bikes

Average AUC: 0.782125

Wall

Average AUC: 0.713481

Describe strengths and weaknesses

Strengths: The program can identify features with different orientation and exposure.

Weaknesses: when the image is scales, the program may not be able to identify the feature. Also it could mismatch the features sometimes.

Personal Testing Pictures

Original pictures:

$Description: C:\Documents and Settings\yl844\Desktop\Dropbox\CV HW1\A1.jpg$ $Description: C:\Documents and Settings\yl844\Desktop\Dropbox\CV HW1\A2.jpg$

Features:

Matches: