CS576 PA1
20113252 HyoSub Park
[Download] Source Code & ExecutableProject Description
The goal of this project is to find interest point using Harris corner detector and making a feature descriptor for those key points. In order to achieve this, functions such as computeHarrisFeatures, computeHarrisValues and computeLocalMaxima were modified from skeleton code. Also, descriptor function that defines the characteristic of interest point was also implemented.
Feature detection
For detection of interest points, Harris corner detector was used.
I used 5X5 gaussian grid for weight function of Harris matrix H. Threshold of 0.3 and 5x5 grid was used for detecting local maximum. Only the local maximum points were used for feature point.
I used 5X5 gaussian grid for weight function of Harris matrix H. Threshold of 0.3 and 5x5 grid was used for detecting local maximum. Only the local maximum points were used for feature point.
Feature description
I approached this problem as simple weighting problem. I transformed RGB color space into LUV and I gave 2:1:1 importance to L, U, V channel, respectively. I did this because the similarity of features are more described by luminance than chromacity. After changing the color space, I turned 5X5 nearby pixel into vector describing the pixel at the center. Using this metric, distance between 2 feature points are largely dominated by L factor, but chromacity still plays its roll.Results
1) ROC curves
| Yosemite image | Graf image | |
| SSD Distance |
|
|
| AUC |
My = 0.816089 Simple Window = 0.624433 |
My = 0.675134 Simple Window = 0.605403 |
| Ratio Test Distance |
|
|
| AUC |
My = 0.900098 Simple Window = 0.670327 |
My = 0.701104 Simple Window = 0.652893 |
2)Benchmark Results
