15-463 - Computational Photography - Automatic Panorama Stitcher

Girish Jattani gsj«at»ece.cmu.edu


General methods and approach

I first chose to make sure that the feature descriptor vectors were being generated correctly, at the cost of not being able to select feature points that were not spaced out over the image. As it turned out, the top harris corner matches were not the same for both images because of intensity differences in the images (due to different ligthing angles). I eventually realized that to fully test the program I would have to implement adaptive non-maximal suppression, instead of selecting the strongest corners from both images for feature points. Once this was implemented, I used RANSAC to get rid of the outlying feature point matches.

Autostitching results:


Hiking in upstate New York (around Saugerties- near where Woodstock was held):

Correspondences:


Resulting auto-stitch


Safeco Field, Seattle, Washington (I think this is the left field entrance)

Correspondences:


Resulting auto-stitch


Morewood Parking Lot (I'm not going to explain where this is :P) [full parking lot stitch]

Correspondences:


Resulting auto-stitch


Lake Tahoe, Utah

Correspondences:


Resulting auto-stitch