Project 1: Feature Detection and Matching
Cameron Lee
CSE 576
4/16/2008
The custom feature “
Procedure
The feature was designed based on the method described in class. This method was used because it seemed to be a simple way of describing the contours surrounding local extreme in an image. This method could easily be done without complicated math function such as transformations or iterative solving algorithms. I also liked the fact that the measure was immune to illumination offsets creating some robustness to illumination changes.
Figure 1. ROC curve for yosemite1.jpg and yosemite2.jpg
Table 1. Area Under the Curve (AUC) values for
|
|
Wind |
SIFT |
SSD
Metric |
0.7056 |
0.7975 |
0.9947 |
Ratio
Metric |
0.7783 |
0.8372 |
0.9955 |
Figure 2. ROC curves for graph1.jpg and graph2.jpg
Table 2. AUC of graph images
|
|
Wind |
SIFT |
SSD
Metric |
0.5455 |
0.5504 |
0.9678 |
Ratio
Metric |
0.4505 |
0.4933 |
0.9320 |
Figure 3. Simple discrimination test for setting feature matching threshold
Figure 4. yosemite1.jpg after the Harris operation
Figure 5. graph1.ppm after the Harris operation
Table 3. Average
AUC for benchmark images
Image Set |
AUC |
Graph |
0.6117 |
|
0.5887 |
Bikes |
0.6644 |
Wall |
0.5874 |
Obviously this descriptor could use some improvement. One of the major weaknesses of this metric is the in ability to decipher local maxima minima and saddle points as long as the respective slopes are similar. Another problem with this metric is that it is not rotation invariant, however, a simple matching metric change could easily be made to allow for rotation invariance. Another weakness it the feature being skewed by exposure effects; normalization of magnitude the data could help reduce error caused by exposure changes. This method on the other had lends it’s self easily to rotation invariance since rotation can be represented by barrel rolling the data. A circular correlation function cold be performed with multiple lags to supply two matching criteria: best match score and best match angle. This angle information could be used to adjust match scores after a data set is collected. Finally this feature smaller then the other 5x5 window proposed.
Here is an example of the feature detection running on an image I took of a crab. It is interesting to see how the feature detector is attracted to regions of high frequency such as the saw dust and twigs.