My Feature Descriptor

I did something similar to the MOPS descriptor, except that instead of storing the actual pixel intensities, I instead stored the I_x and I_y values. I chose this representation because the gradients in a descriptor window should be relatively robust to changes in illumination, as well as position and orientation. Unlike SIFT, my descriptor contains the exact location of gradients instead.

I initially wanted to simply store the gradient angle at each pixel location, but this would introduce a discontinuity in the gap from –pi to pi (or wherever the angles wrapped around). To get around this I chose to represent the gradient angle using Euler angles, which are sin and cosines of the gradient angle at a pixel. These turn out to be simply the I_x and I_y values, which can easily be found by convolving the original gray-scale image with a SobelX and SobelY kernel.

The rest of my descriptor was found the same way as the MOPS descriptor. I used the same oriented 41x41 pixel window around the feature, and down-sampled it to an 8x8 feature, first blurring the I_x and I_y images to avoid aliasing. I stored both the x and y components of the angle in the descriptor for a total of a 2*64 = 128 dimension feature vector.

After some testing I discovered this feature representation was slightly better than MOPS when using SSD matching, and actually worse than MOPS when using ration matching. Thus, I decided to supplement my feature with the MOPS descriptor data, and combined them into one descriptor. The results of this descriptor combining can be seen in Figures 1.

Figure 1: Comparison of MOPS descriptor, my initial descriptor, and a combination of the two.

$Description: Description: Description: C:\Documents and Settings\scm42\Desktop\Yosemite Mine SSD.png$ $Description: Description: Description: C:\Documents and Settings\scm42\Desktop\Yosemite Mine Ratio.png$

Thus, I changed my feature to include the descriptors found from both the descriptor described above, and the MOPS descriptor. I ran the following benchmarks on this combined feature, which I will refer to as my feature.

Performance

Figure 2: Yosemite ROC Curves

$Description: Description: Description: C:\Documents and Settings\scm42\Desktop\Yosemite ROC.png$

Figure 3: Graf ROC Curves, image 1 to 2

$Description: Description: Description: C:\Documents and Settings\scm42\Desktop\Graf 1to2 ROC.png$

Figure 4: Graf ROC Curves, image 1 to 4

$Description: Description: Description: C:\Documents and Settings\scm42\Desktop\Graf 1to4 ROC.png$

Figure 5: Threshold for Graf (left) and Yosemite (right)

$Description: C:\Documents and Settings\scm42\My Documents\My Dropbox\Spring2010\CS6670ComputerVision\Project1\Features\ImageSets\graf\plot.threshold.png$ $Description: C:\Documents and Settings\scm42\My Documents\My Dropbox\Spring2010\CS6670ComputerVision\Project1\Features\ImageSets\yosemite\plot.threshold.png$

Figure 6: Harris operator images for graf (top) and Yosemite (bottom)

$Description: Description: Description: C:\Documents and Settings\scm42\My Documents\My Dropbox\Spring2010\CS6670ComputerVision\Project1\Features\Benchmarks\graf\img2.jpg$ $Description: Description: Description: C:\Documents and Settings\scm42\Desktop\harris Graf.jpg$

$Description: Description: Description: C:\Documents and Settings\scm42\My Documents\My Dropbox\Spring2010\CS6670ComputerVision\Project1\Features\ImageSets\yosemite\Yosemite2.jpg$ $Description: Description: Description: C:\Documents and Settings\scm42\Desktop\harris Yosemite.jpg$

Benchmarks

		Average Error (px)				Average AUC
Descriptor	Match	graf	leuven	bikes	wall	graf	leuven	bikes	wall
Simple	SSD	259	397	483	351	48.0%	33.0%	42.8%	43.0%
MOPS	SSD	288	375	475	366	58.7%	30.8%	64.2%	57.6%
Mine	SSD	313	278	538	342	48.8%	48.0%	52.7%	63.9%

Simple	Ratio	259	397	483	351	51.4%	47.2%	51.6%	53.6%
MOPS	Ratio	288	375	475	366	54.9%	55.8%	58.5%	54.9%
Mine	Ratio	313	278	538	342	58.8%	64.3%	53.8%	62.4%

My descriptor outperformed MOPS on three out of four benchmarks when using ratio matching. It did significantly better in the leuven benchmark, which tests variances in lamination. My descriptor includes pixel gradient orientation information, which should remain constant under different lighting. This explains the almost 10% difference between my descriptor and MOPS in the leuven benchmark.

The instance where it did not do as well was the bikes benchmark. This benchmark tested the effects of blurring. Blurring an image removes the high frequency information, which can greatly change a pixel’s gradient orientation. My descriptor differs from MOPS by including this gradient orientation information, so when the orientations change due to blurring my feature detector will perform worse than MOPS.