CS 6670 Project 1

Greg Sadowski gjs33

Note: I did not do any extra credit (as far as I know).

My Feature Descriptor

I wanted to design a feature descriptor that was different from anyting we had done in class, and was also invariant to translation, rotation, and changes in intensity. Here's how it works:

• Get sets of points that are a certain distance from the descriptor point. For example, all points of distance 5, i.e. all points in the circle of radius 5 centered at the descriptor. Note: I used distances of 6 through 13
• For each of these circles of points and for each color channel, compute the minimum, maximum and mean values over all points in the circle.
• Intensity normalize by subtracting the mean and dividing by the standard deviation of all points in the circle
• Add the normalized min, max, and mean to the feature data container

This feature detector is invariant to translations because the circles of points around a point remain the same after a translation. It is invariant to rotation because when the circle of points is rotated around the feature point, the same points remain in the circle. Thus, the min, max, and mean values in the circle remain the same. Finally, it is invariant to intensity changes because of the normalization step.

The detector is not very good compared to the other detectors in this assignment. I think it could definitely be made better with parameter tuning. The idea itself could probably also be improved. The detector in its current state does better than the simple detector on some images if ratio test is used (usually worse with just SSD)

Design Decisions

Here are some of my major design decisions:

• Precomputed image derivatives for easier computation
• Used the image gradient at each point to compute orientation (to avoid eigenvector ambiguity). To compute this, I just took atan2(I_y, I_x) from my precomputed images.
• I did some trial and error with my Harris score threshold, and I used values between 0.025 and 0.005.
• Rather than sampling a 40x40 patch, I sampled a 46x46 patch so that I could accurately apply a 7x7 gaussian filter to each point in the 40x40 pixels in the center.
• I used bilinear interpolation when taking points from the rotated 46x46 patch to the actual descriptor.
• For my ratio test threshold, I used a value of 0.97. This was via trial and error, and it seemed to work better than the other values that I tried.
• For everything else, I followed the project description.

Performance

Here are my ROC plots as required:

Images: graf - img1.ppm and img2.ppm

Images: yosemite - Yosemite1.jpg and Yosemite2.jpg

Here are some of my Harris operator images as required:

Image: yosemite - Yosemite1.jpg

Image: graf - img1.ppm

Here are my AUC statistics (all values are average AUC):

bikes, SSD, simple: 0.266

bikes, ratio, simple: 0.463

bikes, SSD, MOPS: 0.602

bikes, ratio, MOPS: 0.653

bikes, SSD, mine: 0.706

bikes, ratio, mine: 0.553

graf, SSD, simple: 0.577

graf, ratio, simple: 0.513

graf, SSD, MOPS: 0.557

graf, ratio, MOPS: 0.545

graf, SSD, mine: 0.546

graf, ratio, mine: 0.571

leuven, SSD, simple: 0.174

leuven, ratio, simple: 0.515

leuven, SSD, MOPS: 0.574

leuven, ratio, MOPS: 0.700

leuven, SSD, mine: 0.525

leuven, ratio, mine: 0.526

wall, SSD, simple: 0.270

wall, ratio, simple: 0.533

wall, SSD, MOPS: 0.662

wall, ratio, MOPS: 0.702

wall, SSD, mine: 0.553

wall, ratio, mine: 0.569

Strengths and Weaknesses

It was interesting to note that on some of the image sets (for example, graf), the ratio test performed worse than the SSD test. This may have been because the transformations in some image sets were more extreme than others. Image sets with larger transformations seemed to perform worse on with the ratio test.

I didn't have great overall performance (as far as I can tell). It would have taken hours of fine tuning parameters (e.g. local maxima threshold, ratio test threshold) to improve performance.

My feature descriptor worked a lot better when paired with the ratio test (rather than just SSD). I looked into the files and noticed that a lot of the features are very similar (but not the same). This makes sense because there are many patches with similar colors in the images. If we ignore these patches via the ratio test and just match the distinct patches, we get a lot better performance.

Other Test Images

Here is a screenshot of another test that I ran. I was able to match two images that I took of my desk.