Feature Detection

EE 576: Computer Vision

Project 1 Artifact

Jounsup Park

Feature Descriptor

1.      Normal 21x21 window feature descriptors : 21x21 window size is enough to get large AUCs. I got the window size by experiments.

2.      Dominant Orientation methods: Dominant orientation method is useful to detect the matching features in rotated image because it compensates the rotation. First, I have computed the dominant orientations by larger eigenvalue and adopted to the feature descriptors.

Harris Image

 HarrisImage (Yosemite1.ppm, Yosemite2.ppm)

 

Harris Image, graf (img1.ppm, img2.ppm, img3.ppm)

ROC curves

ROC curves and AUC of the feature matching process for graf and leuven.

 

ROC curves

My Feature Descriptor

I couldn’t finish my own feature descriptor using dominant orientation method. I have to resize and map the new axis after rotating the images. However, on the source, you could find the source to get the angle by compute the eigenvalues. It will make feature descriptor much stronger than without using the dominant orientations.