· Feature descriptor

The feature descriptor that I use is mainly based on the MOPS (multi-scale oriented patches) descriptor. With the feature point, first find out the dominant orientation of the image window, rotate to horizontal, and then save the 5*5 square window centered at feature point as the descriptor. At last, normalize the window by subtracting the mean, dividing by the standard deviation in the window.

It is important to find the dominant orientation of the image window around the feature point, which is given by , the eigenvector of H corresponding to the largest eigenvalue .

Assuming , in order to get the eigenvalue, which satisfies Hx=, let )=0, which means det

We can get the result that , considering b=c in H.

Assume that , so that orientation

· Why choices this design

This descriptor is invariant to rotation (by rotate the descriptor according to the domain orientation), translation and illumination (by normalizing).

· Performance (the ROC curve and AUC) on the provided benchmark image sets.

o Two sets of 6 ROC curves of Yosemite and graf, respectively.

Figure 1 ROC of Yosemite (left) and graf (right)

There is a large translation between Yosemite1 and Yosemite2, which may account for the poor performance of simple and MOPS descriptors. Furthermore, the performance of ratio test is not satisfying with lower AUC than SSD match method. SIFT always performs the best in both cases.

o For one image each in both the Yosemite and graf test pairs, an image of the Harris operator is showed below.

Figure 3 Harris feature image of Yosemite (left) and graf (right)

o Average AUC for two feature detecting (both simple 5x5 window descriptor and MOPS descriptor) and matching code (both SSD and ratio tests) on four benchmark sets -- graf, leuven, bikes and wall.

Descriptor	5x5 window descriptor		MOPS descriptor
Matching	SSD	ratio tests	SSD	ratio tests
graf	0.507238	0.500029	0.473587	0.484041
leuven	0.318475	0.420399	0.521641	0.461822
bikes	0.426898	0.484124	0.624961	0.484685
wall	0.258767	0.487066	0.573809	0.452281

· Strengths and weaknesses of my algorithms

From the above AUC table, we can see that AUC of different image database differ significantly from each other. The main difference of the database of graf is rotation, leuven is illumination, bikes is resolution, and wall is scale. MOPS descriptor is better than the simple 5*5 window descriptor when there is illumination, resolution and/or scale variance. Mostly, although ratio test performs better with 5*5 window descriptor as expected, ratio test dose not perform better with MOPS descriptor. However, with the rotation images (graf), ratio test do perform better with both descriptor.

· My own images and performance