By: Alex Kopp (ark236) and Rudhir Gupta (rg495)

Design Choices

We did not use ImageLib transformation functions and implemented the rotation, translation and subsampling function ourselves.

Gaussian Kernel Calculation

We utilized a function to generate a gaussian kernel of arbitrary size (we used 5x5). While it doesn't really have much benefit on this project, it may be useful in future projects.

Edge Pixels

Discuss our reasoning behind our handling of edge pixels

Downsampling

Our process of downsampling consists of:

Blurring the image (Gaussian)
Scaling image by half (remove every other row and column)
Repeat until we have a 10x10 sample
Remove top and bottom rows, and left and right columns, leaving us with an 8x8 sample

Performance

Graffiti

Harris Operator

ROC Curve and Threshold

Area Under Curve

Descriptor	Matching Algorithm	Area Under Curve (AUC)
Simple	SSD	0.679276
Simple	Ratio	0.618478
MOPS	SSD	0.768960
MOPS	Ratio	0.885618

Yosemite

Harris Operator

ROC Curve and Threshold

Area Under Curve

Descriptor	Matching Algorithm	Area Under Curve (AUC)
Simple	SSD	0.933348
Simple	Ratio	0.883671
MOPS	SSD	0.813112
MOPS	Ratio	0.887718

Benchmarks

Leuven

Descriptor	Matching Algorithm	Average Error	Average AUC
Simple	SSD	403.275842 pixels	0.375975
Simple	Ratio	403.275842 pixels	0.538899
MOPS	SSD	306.945615 pixels	0.545819
MOPS	Ratio	306.945615 pixels	0.599089

Wall

Descriptor	Matching Algorithm	Average Error	Average AUC
Simple	SSD	344.541513 pixels	0.408929
Simple	Ratio	344.541513 pixels	0.542835
MOPS	SSD	330.393817 pixels	0.584562
MOPS	Ratio	330.393817 pixels	0.627402

Bikes

Descriptor	Matching Algorithm	Average Error	Average AUC
Simple	SSD	375.973564 pixels	0.439654
Simple	Ratio	375.973564 pixels	0.538266
MOPS	SSD	450.706956 pixels	0.584909
MOPS	Ratio	450.706956 pixels	0.555136

Strengths and Weaknesses

Stengths

Intensity and rotation invariant
Ratio test provides better match score than simple distance calculation.

Weaknesses

With Non-max suppression, feature points are evenly distributed and hence cover the full structure of image.
Scale variant
Inefficient matching algorithm (O(n^2)).

Random Images

Extra Credit

Non-Maximum Suppresion

We implemented the Non-max suppression algorithm as described in paper 'Multi-Image Matching using Multi-Scale Oriented Patches'. The purpose of the algorithm is to provide better spatial distribution of the features in the image which could improve image stitching operations.

Graf ROC

Descriptor	Matching Algorithm	Area Under Curve (AUC)
Simple	SSD	0.770376
Simple	Ratio	0.658044
MOPS	SSD	0.693452
MOPS	Ratio	0.762649

Yosemite ROC

Descriptor	Matching Algorithm	Area Under Curve (AUC)
Simple	SSD	0.933695
Simple	Ratio	0.854732
MOPS	SSD	0.806818
MOPS	Ratio	0.848011