The technology is already available via photosynth: http://photosynth.net/preview/view/df869f96-2765-4939-8eb3-2...

It appears that what they're doing here is simply extracting keyframes from the video, using them to compose a photosynth, then converting the autoplay of the synth to a video. If you load a photosynth and press "c", you can even see a the same point clouds and scene reconstruction seen on the research page.

Source: I worked on photosynth.

To me it seems like they are just taking frames subject to three constraints: average must be one every 10 frames, maximum gap must be say 80 frames, and finally the aggregate distance is minimized. In other words minimizing that metric subject to those two constraints. That's all. It's a nonlinear minimization problem.

EDIT: After reading their description, I agree they are going the photosynth route. Why not, they have the technology that you worked on. And they say that the naive subsampling I described above doesn't work...

It's striking but it's far more believable when you realize that they need to play at a much faster speed than the source, so they have tons of extra samples from which to extract information. They basically use all that data in the extra frames (that would otherwise simply get tossed away in a regular time-lapse) to construct a 3D scene. This wouldn't look nearly as good if they had to play it at normal speed.

On the flip side, it sounds like shooting at higher fps and using this to end up with a 24 or 60 fps result might end up with some cool results.

Camera movement between each frame would be minimal though so there'd be a lot of overlap between frames so minimal extra information. I'd guess the key to improving this result would be multiple cameras at different angles; I imagine it only works as well as it does because the GoPro uses a fisheye lens.

Take a look at an older paper 'Content-Preserving Warps for 3D Video Stabilization'. That seems to do an ok job http://createdigitalmotion.com/2009/06/magical-3d-warping-te...

EDIT: After clicking through to their website, I found this updated technique which shares one principal author that seems to work even better: http://people.cs.nctu.edu.tw/~yushuen/VideoStabilization/

Wow. Actually, if they that add that technique to the mix it might solve the deformed "pop"-effect you see in some videos, like the deformed building you see around 16 seconds into this videos:

http://research.microsoft.com/en-us/um/redmond/projects/hype...

Cool, I'd not seen the updated work. I wonder how much can be done in realtime. I have no idea what the compute split is between between the different processes.

With sensors (gyros etc) the camera path would be trivial, instead of recovering that from the video. Rendering the results would be possible on a mobile GPU. Just leaving the frame conversion to a point cloud in terms of compute and memory.

Maybe some scheme where you down sample the input frames to create the deformation mesh, then apply that to the full size frame would be the way to go

> With sensors (gyros etc) the camera path would be trivial, instead of recovering that from the video.

Well... not quite trivial. They're calibrated differently per model, and it's actually quite tricky to reconstruct the path based on accelerometers and gyroscope alone. There's also the likely issue of synchronising the data from these sensors with the video input. If you solve that second issue however, it could in theory at least help with recovering the path from the video, creating better predictions where the point cloud has moved to for example.

I was being a little cheeky with that. I know position and orientation data can be quite noisy and getting clean good quality data is not simple

"This wouldn't look nearly as good if they had to play it at normal speed."

For normal speed you wouldn't need this :)

This technique actually works great for normal speed and can mimic the smoothness of a cinematographer's expensive dolly: https://www.youtube.com/watch?v=3TlCGh5Pc90

It's certainly way better than the source video, but it's nothing close to what would come from a steadycam or dolly. You couldn't use this finished product in any kind of real production.

That depends heavily on your definition of "real production," and probably quickly devolves to no true Scotsman. I absolutely think this could be used in productions I would consider real, particularly documentary/travel/reality programs and sports.

Yep, it has a definite 'look' to it, and it appears to work better for some types of material than others (the bicycling footage was far more watchable to my eyes than the climbing stuff), but the effect is engaging and not unpleasant to view at all.

Personally I thought the climbing video was breathtaking. With a decent film crew it could be an absolutely breathtaking scene in a climbing film.

There was a mildly annoying effect somewhat reminiscent to pop-in seen when terrain geometries go from lower to higher detail in video games. It was particularly evident here:

https://www.youtube.com/watch?v=SOpwHaQnRSY#t=170

It would, I think, be even more distracting if the video was higher resolution.

Yeah, I noticed that too. I wonder how much of that is really an artifact of the lighting and (relatively) low resolution of the camera. Something shot with a better camera and lighting that reveals more terrain might give the algorithms something better to latch onto so the terrain models more cleanly.

GoPros can shoot in 4K, they're not a low resolution camera by any means.

I wonder if you could apply it to higher-quality (i.e. already pretty stable) input to approach a steadycam though?

I think a higher frame rate would help a lot. It's a higher sampling rate basically.

Stereo input from two cameras might help with constructing the 3D scene.

For hobbyists like me, it is amazing.

Yeah, it looks amazing. If the video is taken at faster speed (like 10X), then they can get a smooth realtime result when slowing it down at post processing.

Because they drop frames, they aren't stabilizing they are throwing away frames that move too much.

This is good stuff, I like it, but it isn't as wow as the structure from motion work.

And for the folks saying just up the framerate, that won't really help because the head motion needs to back in the same position as a previous frame. It is a function of how much and at what frequency the motion you want to remove is.

This was on my todo list, item removed.

You think a week of post end processing? I didn't read the paper or anything so I could be way off base, but I would assume the algo simply has to choose which frames to keep and which to toss. I doubt this would take an enormous amount of time even with HD videos. The algo is most likely just really clever in how it chooses a good frame vs a bad one.

On the other hand, if it is actually generating a lot of "best guess" images to put between gaps that are too large to bridge (too many bad frames in a row) with the current frames I could see that taking a bit longer, but not a week.

It does 3D scene and camera path reconstructions then re-renders the scene from different perspectives. It's not just "picking the best frames". The technical explanation video goes into the details: https://www.youtube.com/watch?v=sA4Za3Hv6ng

Maybe you should watch the technical video or read the paper before commenting.

The paper says it took 23,700 minutes to do just the frame selection for the 13 minute bike input video. That's over two weeks.

Actually it says that it took 13 minutes to process 23,700 frames! :)

No, see table 1: "input duration (minutes and seconds), input frame count". It says the source file is 13 minutes 11 seconds and that it has 27,000 frames. In table 2 it says that source selection took 1 minute per frame. That's where I got 27,000 minutes from.

Ah, good call, thanks!

I now think we both got it wrong (but me more so than you): Table 2 specifies "1 min/frame", but the source frame selection happens for output frames, not input frames. Table 1 lists a total of 2189 output frames for the 23700 input frames of the "BIKE 3" sequence, so I guess we're looking at 2189 minutes?

Correct me if I misread anything. Again.