You are on page 1of 1

Summary #9

Benedikt Willecke
IES19372

In this lecture two methods for feature matching/fitting where proposed.


RANSAC (Random Sample Consensus):
RANSAC is used in order to be robust against the many outliers that might exist when matching two
images.

1. Create a subset by randomly sampling data points need to fit the model.

2. Build a model to fit the randomly sampled points

3. Measure how well the model fits the overall population by using a threshold to check how many
inliers fit in the model compared to the total number of samples

4. Repeat step 1-3 until the best(the max amount of points within the threshold) is found

It is important to note that the more points in a sample we use to create the model the higher the
risk is to contain outliers. We should choose the number of trials so that we have a high probability
that at least one sample is free from outliers. This is a outlier robust model, however very costly in
terms of computational time.

Hough Transform:

This is another algorithm that is robust against outliers, but more efficient than RANSAC.
Given data points and a line to fit them we should transform them into Hough space. For this we
draw a perpendicular line from the origin to the given line. Then we can define two values. θ is the
angle between the given line and the line we just drew. ρ is the length of the line which we drew.
These two parameters are used as bases in the new coordinate system (Hough space). On a side note
we have to actually quantize this space because otherwise there would be too many possibilities.
Then we scan through each point and see how many parameter values fit with the (x,y) point. We
write and increment these into a table and look for maxima. This represents the maximum amount of
edge points that voted of an edge line. This can also be represented graphically. In that case we get
curves/lines in the Hough space for each edge point. The intersection of these lines is the vote given
for the edge line. To handle noisy data we can either adjust the grid size or use smoothing. However
we have to make a trade off here between noise tolerance, precision and speed.

You might also like