Yes. I wrote the code for the http://www.3d-4u.com 3D camera. You definitely can not use simple still image stitching algorithms. Per pixel depth estimation (via dense stereo matching) and subsequent blending must be used. I would say more, but I had to sign a long NDA before I worked on this.

i was expecting the cube to get solved once the page finished loading.. I'm dissapointer

This looks like 180, not 360.

Yep -- same seam issues though.

> get stereo from mono data

Why do you call it mono data? Given how close are each lense from another, I'm guessing, that anything far enough (0.5m?) from the cameras is beign recorded by at least two cameras at all times.

Because simply stitching it will create one image/video stream from one perspective. If you have 2 eyes, you have stereo vision so can perceive depth. This technique allows for the synthesis of stereo video from a rig that doesn't directly capture stereo video. Pretty cool stuff

The point the GP was making was that you can treat each pair of cameras as a pair of 'eyes' for the stereoscopic analysis. And you get every combination of adjacent cameras as a pair so a ring of 8 cameras represents 8 distinct pairs (call them camera 'a' and camera 'a' + 1 clockwise)

By processing those streams you should be able to derive high quality depth information from the combined video streams, and given the overlap in views re-compute a 'view' from any direction.

A lot of early work was done using the streetview cameras as well to do street facing feature extraction using similar techniques but as those cameras were essentially single shot (discrete pictures over time rather than video) you end up with the "slewing" effect of moving from one spot to the next.

Presumably Google will go out and do street view with this sort of setup and get a continuous terrain option so that you can really walk around as if you were there (caveat occlusion effects)

I don't know specs of the HERO4, but the GoPro HERO3 has a 170 degree view angle - so that 16 camera rig has a _significant_ overlap. I suspect you could generate both "adjacent camera" and "every second camera" stereoscopic pairs without too much difficulty, and you can probably even extract useable data out of the "every third camera" pairs and their ~30 degree view overlap. My guess is everything seen by that rig is viewed by 3 distinct pairs (a&a+1, a&a+3, and a&a+3), and that the entire scene from a 16 camera 170fov rig could be considered to have 16 a&a+1 pairs, 16 a&a+2 pairs and 16 a&a+3 pairs giving you three stereo views with different baselines for every point in the scene. I'm guessing there's a _lot_ of spatial information in there.

Actually, I just realised all those HERO4's are mounted in portrait orientation, so they won't be getting 170 degrees of horizontal view, and you probably aren't going to get useful overlap between the a&a+3 camera pairs. (but it does mean there's only two 20 degree wide conical polar "blindspots")

I had thought we would never get 360 + stereo. I wonder how much touch up work this will require and what kinds of scenes will provide problems with incorrect depth heuristics.