EE 576 Project 1

EE 576
Project 1

Kyle Merry: kjmerry@uw.edu

April 18, 2013

1 Feature Descriptor

I modeled my feature descriptor after MOPS, although I haven’t implemented all of the features. After each Harris feature is detected, I take a 16x16 or 32x32 window of points around each feature and create a Gaussian pyramid, downsampling to 8x8. After normalizing these 64 points to zero mean and unit variance, they become the feature descriptor. This is the MOPS descriptor without orientation correction.

2 Design Choices

The first change I made was adding intensity correction to the basic 25x25 window. This marginally improved the ROC curves for both test sets (Yosemite and graffiti), but not enough! I added pyramids, and found that performance improved. 16x16 downsampled once performs better on the Graffiti dataset, 32x32 downsampled twice performs better on the Yosemite dataset.

3 Performance

3.1 ROC Curves

Figure 1: ROC curve for Yosemite dataset

Figure 2: ROC curve for graffiti dataset

3.2 AUC/Pixel Error

Table 1: Bikes

	SSD	Ratio

Window	0.54/399	0.71/399
MOPS	0.84/417	0.75/417

Table 2: Graffiti

	SSD	Ratio

Window	0.49/291	0.51/291
MOPS	0.64/320	0.555/318

Table 3: Leuven

	SSD	Ratio

Window	–/393	–/393
MOPS	0.79/226	0.88/226

Table 4: Wall

	SSD	Ratio

Window	0.39/367	0.58/367
MOPS	–	–

3.3 Harris Operator

Figure 3: Harris maxima for Yosemite 1

Figure 4: Harris maxima for graffiti 1

4 Strengths and Weaknesses

My algorithm handles translations and intensity/exposure well, but can’t handle rotations.

5 Extra Credit

I made my feature contrast invariant by rescaling all features for zero mean and unit variance. I also implemented image periods for slight performance increase.

6 Extra Images

Figure 5: Harris features w/o orientation

Figure 6: Harris features w/o orientation

Figure 7: Harris features w/o orientation