Philipp Eichmann (peichmann)

Finding interest points

In order to find suitable interest points for matching, I implemented a Harris Corner Detection. My implementation works as follows:

• For every input image
  1. Apply an initial gaussian blur
  2. Compute the X and Y derivatives Ix, Iy using a Sobel filter
  3. Compute the corners by using the following formula:

In a next step, I supress local non-maxima by simply nulling all non-maximas within regions of the size of 3x3. This approach, however, causes most of the interest points to be found in high contrast regions leads to an unevenly distributed set of interest points. This is why I use Adaptive Non-Maximum Surpression (ANMS).

Interest Points found by using Harris corner detection on two different images.

Describing features

Each interest point is then passed on the a function that tries to describe it by subdiving its surrounding area into 4x4 subareas and computing a histogram of all of their gradients.

According to the angle of the gradient in each subregion, the magnitude of the gradient is added to one of 8 bins. These bins can later be used to identify identical features.

Feature Matching

In a last step, the descriptor of both images are compared by using a nearest neighbor approach. More precisely, the ratio of the two nearest neighbors is observed in order to determine whether two features match using this forumla:

For the images I used, I found this approach to work best when features are accepted as match only if the ratio of the closest distance to the second closest is less than 0.7 - 0.9.

28 good matches and only 4 bad matches.

Results

The results vary on the chosen parameters and on the set of images being used. Namely the following parameters can have great effects on the outcome:

• alpha value when computing the Harris Corners
• Size and sigma of Gaussian Filter
• Threshold for interest points
• Maximum ratio for nearest neighbor distance

Also, since this implemenation is not scale and rotation invariant, images at different scales or taken from very different position won't lead to many matching features. As a consequence when comparing non-similar images, interest points should be analyzed at different scales and the gradient of the of a feature should be computed relatively to the features domninant orientation.