15-462 Project4: Automatic Panorama Stitching

Jing Jin

Disclaimer: Full automation ahead. Please keep your hands and feet inside of the vehicle at all times.

Additional Implementation Details:

I modified the thresholds for both the feature matching and the RANSAC elimination. For feature matching, I found that just using a 40x40 descriptor was only reliable for a small range of image sizes, so to accomodate the larger images, I decided upon the descriptor size according to the size of the image. By using a descriptor that's 0.4% of the size of the entire image gives reasonable matches. For the RANSAC elimination, it was also unreasonable to use a static number because the error is directly correlated with the size of the image. After a little trial and error, it seems that setting the threshold (x/9)^2 where x is the width of the image produces almost 100% correct decisions on which points to keep and which points to eliminate. I wonder if anyone's still reading this. Well, I'm done, you're free to gawk at the pictures now~

Panoramas:

Interest Poins:
The blue plus signs are the 500 weighted interest points. The yellow circles are the points matched, and the red points are the points confirmed using RANSAC:



Result:

P.S. You did find the people in it last time, riiiight?

Here are some more results:

And here's the prettiest one: