Harris Corner Detection
We find all interest points in an image using harris corner detection. Here's a photo from last project with it's interest points plotted. As you can see, we have way too many points.
|
|
Adaptive Non-Maximal Suppresion
To resolve this issue of having too many points, we make sure the points are more evenly distributed across the image. We find the minimal distance that harris values are "suppressed" for each point, and take the interest points that have the highest of these distances. Here's what this looks like for the 500 'best' points.
|
|
|
Feature Descriptors
To identify corresponding points, we sample each interest point at its lower frequencies. This involves taking a 40x40 subsection of the image centered at our interest point, passing it through guassian low pass filter, and then taking an 8x8 sample of the subsection. Here are a couple of the features:
|
|
|
|
|
|
Welp.
Due to thanksgiving break, I didn't really have time to finish the rest of this project. Sorry, thats it :(