My last thesis update was way back in November '08. I gave the most recent description of my research, and talked about a small bit of theory that I needed to confirm before moving on. Back then (and up until now), my thesis was supposed to be a simple experimentation of a particular class of interest point detectors to show how well they worked for matching photos with cube and cylinder panoramic images.
Unfortunately, I was never quite able to get the photos and panoramas to match automatically in any sort of reasonable way. I tried a few different experiments to show this, but it was difficult to convince ourselves that the strategy used just wasn't working until I really pared it down to the bare essentials. Basically, what I ended up showing was the inadequacy of the interest point descriptors in the context of comparing them in the panoramic and photographic images. You can see some of the results here.
Now I must seek another reasonable method for matching these particular images, given the following:
- The panoramic images are not terribly sharp in their quality. This is related to the fact that some areas of the image are sampled from several sources. Fine features are thus unlikely to be reliably detectable.
- In most comparisons, buildings will be the only structure that will appear in a scene consistently. The ambiguity from repeated structures in these buildings (such as windows) will pose a challenge.
- It may be reasonable to assume that we are always matching images of the same scene thanks to a previously run recognition algorithm and/or GPS location attempt.
- The matching process must result in being able to find the geometry between the cameras of the two scenes.