By Yan Huang (yh553) and Sean Ryan (ser99)

Design Decisions

We split out part of computeHarrisValues into a computeHarrisMatrices function. This function takes a grayscale image and produces three images filled with A, B, and C values for each pixel; thus it effectively computes a Harris matrix at each pixel. We did this so that we could compute Harris matrices in the computeHarrisValues function and in ComputeHarrisFeatures (for rotation) using the same code.

In ComputeHarrisValues, we compute the maximum eigenvalue and the corresponding eigenvector for each feature to get the feature's angle. We do this here because we consider the angle to be fundamental to the Harris feature. Note that this means we must force the angle to zero in ComputeSimpleDescriptors.

We wrote our MOPS descriptor before we knew about WarpGlobal and thus did the transformations and interpolation using our own code. Since it worked well enough we didn't go back and rewrite it all. Our MOPS code does the following

Make a copy of enough data to contain the rotated feature.
Applies a 7x7 Gaussian blur.
Uses inverse warping to get a 8x8 image, scaling and rotating the pixels; bilinear interpolation is used.
Subtracts the mean and divides by the standard deviation.

We've made the bugfix changes on Piazza, including the benchmark fixes from @53. Thus our benchmark figures can be trusted.

Performace

ROC Curves

Here are the two sets of six ROC curves. The first is between graph/img1.ppm and graph/img2.ppm; the second is between the Yosemite images. Thresholds of 60 and 3 were used for SSD and the ratio test respectively.

Harris Images

Here are the Harris operator images for graph/img1.ppm and Yosemite1.jpg

Benchmarks

Bikes

	AVG Error	AVG AUC
Simple SDD	403	0.379
Simple Ratio	403	0.571
MOPS SDD	390	0.831
MOPS Ratio	390	0.882

Leuven

	AVG Error	AVG AUC
Simple SDD	361	0.121
Simple Ratio	361	0.261
MOPS SDD	214	0.827
MOPS Ratio	214	0.925

Wall

	AVG Error	AVG AUC
Simple SDD	384	0.238
Simple Ratio	384	0.684
MOPS SDD	328	0.820
MOPS Ratio	328	0.838

Strengths & Weaknesses

Clearly the simple descriptor with SDD matching is quite bad. However, investigation with the GUI shows that the simple descriptor is decent at matching things in the Yosemite images because they are translation only.

Our MOPS descriptor has a good AUC and acts as expected in the GUI. It is very good at working under different lighting conditions and under translations. It is decent at matching under rotation. However, it is defeated by any transformation that significantly changes the plane of the image (graf/img4) and by images with high amounts of blur.