Thursday, October 2, 2003

A new stitching system: Matthew Brown, a PhD student at UBC in Vancouver, Canada, recently wrote the first system to recognise and stitch panoramas fully automatically from an unordered set of images. It uses a new class of generic image features and object recognition techniques to automatically detect matching images in an image set. It then automatically registers and stitches each panorama in the image set that it finds. The results of this are online. Also available at the same site is a paper entitled "Recognising Panoramas" by Matthew Brown and D. G. Lowe, which is being published at the International Conference on Computer Vision this year. From the webpage: "This allows photographers to input to the system the images on an entire flash card or film, and the system will automatically recognise and stitch panoramas without user input." The software can extract panoramas of differing orientation from a batch; i.e. some of the panoramas in the batch can be portrait, others landscape. The software is called AutoStitch, and can handle full 360x180 degrees panoramas, as well as cylindrical panoramas. It runs on Linux and is available on license from the University of British Columbia, and the "SIFT" feature matching is patented. The AutoStitch software sounds quite fast at what it does, yet it is a research implementation. It currently takes about 2 minutes to automatically match, stitch and blend panoramas from 50 input images.  My opinion: this is a fantastic advance in panoramic photography. The time savings alone make it an exciting development. Imagine no longer having to tediously copy each panorama's images into unique directories, tell the stitcher what kind of lens you used (if you can remember), then drag each image one at a time into some sort of stitcher "waiting room" panel, and then drag those images into position in the stitcher itself, then tell the stitcher to stitch each pair or set together. Instead, this software can look through a batch of images downloaded from a CF card and figure out which images belong in panoramas, infer the lens, and stitch them together. From looking at the sample images, it appears to do a very good job of stitching and blending. The amount of time saved by such a process is significant--I would guess it could save me 4-8 hours of mind-numbing computer tedium per day trip's worth of panoramas. This is the type of advance I've been calling for in my essay. I would guess that all stitching applications would want to have the auto-matching feature. I'll post more information as I get it. 7:00:00 AM

© Copyright 2006 erik goetze.