holy schnoz. this kicks can.

If the accuracy of the orientation detection was better than that of an individual pixel in the image, you could make a higher quality image than the camera sensor could make on its own. That’s something that I wonder about how well it could be implemented.

must. build. immediately.
he seems to just use the angular information from his box — if this would only be used as a hint for a real stitcher, the output would be even better. or, use an optical flow algorithm in combination with that (can be done easily on the GPU in real time).

Magnetic Sensor is used for detecting where the North Pole is. It acts as a compass. It makes it to the orientation data is more accurate then with the accelerometers alone.

Quite a cool effect!
I wonder how this looks like with some more processing and lining up :-)

This is cool, I like it.

With this you could almost do 3d maps of rooms or houses like in movies like the italian job and stuff. Its pretty close anyway.

the picture was somewhat distorted, if the sample rate was increased would the picture quality improve?

Andrew,
Let me be the first to congratulate and thank you for letting us see this fantastic project, its not often you see videos online that gives you a wide smile, and makes you keep saying ‘that is fantastic’ every so often!
Everything about what you’ve done seems so obvious, but of course, it really isn’t obvious at all. Apart from general panoramic photography, I haven’t seen anything like this before, its so ‘immediate’ and rewarding! Especially where it can auto-correct camera orientation AND generate full 360degree views!
Its just inspired!
I for one though would love it if you would make it much more accurate, less jittery, and maybe faster, as I’m kind of thinking its got a sort of ‘unfinished’ look to it.
But I do know where you’re coming from with the preference of that abstract visual effect.
Thanks again! You’ve given me the same big smile I had when I first saw Jeff Hans Multi-touch reel, and Johnny Chung Lees WiiMote PC based projects.
Thanks.

Questioning the need for the magnetometer and assuming this could be done with only an accelerometer comes up a lot, so I’d like to correct that. Linear algebra says that you need two mutually orthogonal vectors to fully determine an orientation in 3D space. If I only had the accelerometers, I would only be able to determine pitch and roll, but never yaw. (Without looking, you can tell which position your body is in, but not which direction you’re facing, right?) Similarly with the magnetometers I could only determine yaw and pitch, but never roll. (Sort of.. Magnetic north actually points about 45 degrees into the ground around here.) So this wouldn’t really work with a wiimote, as it only has the accelerometers.
I’m honored to be compared with Johnny Chung Lee- I remember being blown away by the simplicity and brilliance in what he did. In fact, his video inspired me to make this one.
I think if I had a *MUCH* faster sample rate from the sensors, I could develop an algorithm to smooth out the jitters and improve accuracy. With such a slow sample rate, though, it doesn’t seem like there’s a lot I could do to improve the data from the current hardware without getting into some hardcore image analysis.

This is madness! Awesome!

Did you thought about combining it
with an ultrasonic range detector?

amazing. I had just googled for this kind of thing… and lo it brings me here before my bi-weekly check of hackaday.

Perhaps this could be hacked/used to make a non-iPhone acquire some tilt features?

I’m curious if the inaccuracy is really due to the sensors, or the simple fact that he is holding the camera. while it can sense up and down and its angle, it cant determine it’s coordinate location, so movement left or right would change the alignment of the pixels. perhaps a tripod mount is all it would take to fix this problem.

definitely an interesting project that i may be interested in building one day, when i have money.

You don’t need to go through the effort if you just want to make a quick panorama like that. Just shoot in video mode from one pivot point scanning the area, load the video frames as layers in Photoshop and have it auto-align them as a panorama. I’ve done this before – it was hi-res and the accuracy was near perfect.