Kidai Kwon (kk36)

1. get_interest_points.m

   I first blurred the image with a Gaussian filter and implemented the Harris corner detector. Then I applied colfilt on Har to do non max suppression and returned points whose Har value is greater than some threshold.

2. get_features.m

   Basic SIFT implementation. In addition, I raised each element of the final feature vector to some power that is less than one. Also after normalization, I clamped gradients that are greater than 0.2 and renormalized the vector.

3. match_features.m

   Basic ratio test implemented with KNN - Euclidean distance.

Final Results

87 total good matches, 16 total bad matches (good matches %: 0.844660) with NNDR = 0.79

39 total good matches, 2 total bad matches (good matches %: 0.951220) with NNDR = 0.7

With a lower ratio test threshold, the algorithm only returns highly confident matches.

Experiements

1. Basic Implementation

   71 total good matches, 33 total bad matches (good matches %: 0.682692)

2. Applied a Gaussian blur before applying the Harris corner detector.

   82 total good matches, 25 total bad matches (good matches %: 0.766355) -> (Improved)

3. In addition, I raised each element of the final feature vector to some power that is less than one. Also after normalization, I clamped gradients that are greater than 0.2 and renormalized the vector.

   87 total good matches, 16 total bad matches (good matches %: 0.844660) -> (Improved)