🔥 Feature Matching

One common problem in computer vision is to “match” two images by finding corresponding points between them. To do this, we first extract feature points from the image, usually with ⌛️ SIFT, and then match them to find correspondences.

SIFT provides us with key points where each point contains a feature—a grid of histograms that describe the local area around the point. We use these descriptors to match points across pictures; ideally, the histograms should be similar for correct correspondences.

Now, given points and descriptors across two images, we want to find matches between them. We go through each point in one image and compute the following:

1. Look through all points in the other image, rank them by feature similarity. Let the distances be where is the most similar (smallest distance).
2. Check the uniqueness of the most similar feature by checking : if this is true, then is much better than and will likely be correct.

We then repeat this computation in the reverse direction. Our final matches are those that were found in both directions, satisfying bi-directional consistency.

One direct application of these correspondences is calculating the 🖼️ Homography between these two pictures, allowing us to blend them together.