Proj4: automatic mosaic stitching

The goal of this project is to automatically detect matching feature points between mosaic candidates.
With matching feature points, we may use the method of proj3 to create a mosaic.
The work is based on Brown et al, "Multi-Image Matching using Multi-Scale Oriented Patches".

Feature points from Harris corner detection

We would like to decrease the number of feature points in a nice way

ANMS prune

We create a feature descriptor for each feature point.
This is done by sampling an area around the pixel at a higher level of the gaussian pyramid
Then we compare the feature descriptors and do further pruning based on the nearest neighbors.

1-NN/2-NN based prune

notice the lonely outlier on the left side of the bottom left picture.

Now we run RANSAC to get rid of any remaining outliers

After RANSAC

Finally, we can use the feature points that are remaining and do mosaicing!

Here some others, source images taken from classmates.

Heegun Lee

Last modified: Wed Nov 16 13:43:06 EST 2005