When I first learnt about Microsoft Live Labs Photosynth I was impressed. The idea of connecting together photos of a scene based on their points of view and placing them appropriately in 3D seemed to have potential and the demo video of St. Mark's Square was impressive. I eagerly anticipated being able to use the technology with my own photos.
Recently I found out that Photosynth is finally available for use by anyone via the photosynth.net web site. It's all done with a Windows application which integrates with the website as a browser plugin. Windows XP or Vista is required, and it seems the application does all the work, so the website integration doesn't seem necessary. However, I'm not really complaining because it's by Microsoft and the overall experience is quite smooth.
I first tried it out with photos of the Nanaimo waterfront downtown. That was an almost total failure. It was too much to ask it to really understand a 3D scene, and I bet the water, sky and and changing lighting also confused it. After this I decided to try other examples where there is more overlap. My second attempt was with the downtown Detroit viewed from across the river in Windsor. There I had plenty of overlap and the angles weren't too different. This didn't work too well either. I was surprised and disappointed by how shots from almost identical angles in different lighting conditions didn't connect.
Finally I had some success with photos from the Georgian Bay shore of the Bruce Peninsula. There, Photosynth seemed to work quite well, dealing with seasonal differences and obscuration. I guess what made it work was consistent lighting and certain obvious features. This showed that photosynth had some potential.
I then made some more synths. It wasn't easy to find candidate images. In most cases I didn't have enough photos of the same thing to attempt to make a synth because I didn't keep many similar viewpoints. Then I remembered that I saved some photos to disc before selecting which ones I wanted to keep and I found candidate images there. I only considered groups involving views from multiple points; I feel photos taken from a single point should be joined into a panorama instead. I only got a 100% synthy result once, but with the other results I re-did the synth with only the images that would link up because I felt it's pointless to keep unlinked images. You can see all of my results at my profile page at photosynth.net.
This was fun, but it wasn't too impressive. My main complaint is that when I view synths they don't feel very 3D. A lot of viewpoint changes feel like transforming 2D images and switching between them rather than navigation through a 3D space, and the point clouds don't seem impressive. This even holds true with many examples by others which involve far more photos. I prefer panoramas for a single viewpoint and Panoramio's Look Around feature for multiple viewpoints. The Look Around feature seems to be more effective at linking up images and it links up images taken by different people.