Yuhang Tang (yt23)

CS 143 / Project 2 / Local Feature Matching

Algorithm

A Harris corner detector is used to find the points of interest in the picture. The image is first blurred with a Gaussian kernel of stdev 2 before being taken the second order derivative. The difference operator for the derivative is 4th order accurate. The cornerness value is normalized to [0,1] before applying the bwlabel function to search for the local maxima.

The featues are normalized image patches extracted from around the points of interest.

L2 norm is used to calculate the distance between features. The ratio of the first nearest neighbor to the second neighbor is used to thresholding the false postives. A threshold of 0.8 is used.

Result

For the Norte Dame scene, 32 total good matches and 14 total bad matches were obtained, resulting in an accuracy of 69.6%.