Here are my results for the panorama project for CSE490CV.

I didn't do a whole lot more than the basic requirements. The biggest problems I ended up having were exposure differences, as can easily be seen in the test sequence. And while I even explicitly set the expoosure settings on the camera to manual, I still got a wide variety of exposures in the images I took outdoors.

As a result, some of my best panoramas are the ones taken wither indoors where the lighting is pretty constant, or at night. Further descriptions are below...

As far as problems went, the biggest one I had was of course exposure differences, some of them due to the camera and some to my own photography mistakes. I also had issues with the supplied camera distortion parameters; it turns out that the ones in the sample script file require you to divide by the distortion parameters, whereas the supplied ones require you to multiply instead. So I ended up switching back and forth some, which was a pain. Same for the shear correction term (to correct for y drift); I found that depending on whether the pictures were in clockwise or counterclockwise order I had to change this. Also, the Lukas-Kanade algorithm here seems pretty sensitive to initial conditions, which I discovered while working on my handheld sequence.


Test sequence : [panoramic] [jpeg image]
This one has occasional fuzziness along the transitions thanks to exposure differences, and at one spot there is a flat-out error due to this. I thought about doing a color correction in Photoshop but then got too wrapped up in my own panoramas and forgot about it.

Outside Meany Hall (at night) : [panoramic] [jpeg image]
This one really turned out quite well. I'm turning it in as my handheld image, although I actually used the chair that is featured so prominently in the foreground to steady the image. I then used a timer to take a few shots of myself in various poses; you should be able to find 3 of me in the panorama. I ended up with about 3 times as many pictures as I actually needed to make the panorama and had to spend some time adjusting the initial parameters to Lukas-Kanade to get things to align. Also, when I stitched the shots together I deleted myself from the alpha channel so it wouldn't consider the image of me in the Lukas-Kanade algorithm; this was the only way I was able to get decent blends. I then had to modify the alpha channels of the images that didn't have me in them, so that it would not try to blend me with the background. The results turned out pretty well, I think. The image here is in black and white; you can look at a color version if you like.

Inside Mary Gates : [panoramic] [jpeg image]
This is one I used the Kaidan head for. When I first tried to stitch this one together, I got really horrible error. Removing the moving people from the alpha channel helped a bunch, but alignment is still a little bit imperfect in spots. It almost looks here like the camera warp parameters are off by a small amount or something because I wasn't able to hand-align the pictures either (which is strange because I took the Meany Hall sequence with the same camera and got far better results) In general, though, the effect is nice.

Building at Gasworks : [panoramic] [jpeg image]
This is a strange sort of indoor gym/playing facility that is located at Gasworks park. I thought it was somewhat architecturally interesting so I took a shot. Thanks to it being indoors, the result is pretty good.

Outside at Gasworks : [panoramic] [jpeg image]
Here is a shot for the water overlook at gasworks park. As you can see, the exposure differences are pretty awful, but oddly enough it aligned decently.