Arthur YIdi (ayidi)

Overview of the first project includes the results and is then followed by an overview of the process at the end.

First test with a different image pair provided goods results. Images shown here have only minor scale shifts and rotation, since both the interest point and feature descriptors implemented are too simple to cope with major transformation shifts.

Again good results when images are in the same viewpoint.

Now slight rotation and light intensity changes are throwing aways the descriptors.

Final results for Notre Dame base test case.

As the harris interest point yielded more and better corners the perfomance increased rapidly. When given about 3,000 interest points the matching performance increased to 81% (image above).

First detect interest points using a based interpertation of the Harris corner detector.

Here I decided to explore the gradient values of the images, to see how this would effect the descriptors.

Finally, here's an overview of the process of implementing the SIFT descriptor.