Video stabilization

[Eupeptic]

Hello all!

Next door at the RL hardvore thread, there are a lot of videos that are shot by people who aren't professional videographers. While it's true that for RL stuff, you sort of have to take what you can get, I spent a few minutes Googling to see if there is any good way to post-process video to stabilize it. The feeling I get (again just from a couple of minutes Googling) is that this technology is not quite there yet. What I *want* is a program that I can give a YouTube URL (or maybe a video file downloaded from YouTube), push the button, have it grind for a while, and then spit out a stabilized video. What I can *get* are video editing programs that a) cost money and b) appear to require a bunch of manual intervention to do stabilization. Mostly it seems that you have to pick out an item that doesn't move in the frame very much, and then the program tries to keep that item in the place that you specify. (How this works if your subject is also moving is unclear to me.)

If done well, the stabilization can really improve a film. It was done on this well-known film of "Bigfoot" (no vore content) to produce this image, although I suspect that this stabilization may have been done by hand. It probably helps that in this particular film, there are a lot of trees and branches near the edges of the frame, which can be easily matched from one frame to the next. I have a feeling that starting with Flash video from YouTube is going to be sub-optimal due to low resolution, and of course it is preferable to start with less-shaky video in the first place. Also, stabilization won't help the problem of the camera being pointed at completely the wrong thing, as seen (and commented on) in a couple of recent videos. But, again, we have to take what we can get. :)

Has anyone out there ever experimented with video stabilization software (or maybe a "stabilize" plugin for a video editing suite?) Was it easy, hard? DId the results look good, bad, about the same, etc? Just curious.

Eupeptic

[Eka]

Video stabilization is a manually done thing. That bigfoot stabilization one, for example, were basically done frame by frame manually.

There are software out there that does what you are looking for but they won't give you nearly the perfect result comparable to your example.

Different stabilization technique are good for different things, so for different problem you might need a different software. Unfortunately, until you try a particular example on a particular software, you won't be able to tell how it will come out.

[oldman40k2003]

... The feeling I get (again just from a couple of minutes Googling) is that this technology is not quite there yet. What I *want* is a program that I can give a YouTube URL (or maybe a video file downloaded from YouTube), push the button, have it grind for a while, and then spit out a stabilized video. ...

Unfortunately I don't know of any programs that do what you want, but I suspect that some exist, though they probably cost big bucks. The reason I think they probably do exist is because the S.I.F.T. algorithm exists and has been used in programs to allow the construction of a 3d model from a set of photographs. This is at least as difficult as stabilizing a video, because in essence it's the same thing (IE: in both you are finding the common points in pictures and then using that information to place the picture at a particular place and rotation.). The reason I think that the programs are probably expensive is because S.I.F.T. is not a simple algorithm at all, so it will take considerable (and costly!) programmer hours to implement it.

A (partial)thought did occur to me though: automatic image stitching programs exist, and some of them produce not only a stitched image, but also a text file telling you where and at what angle each separate input file got stitched into the composite picture, and video files can be rendered as a series of images. With that information you can then reconstruct a virtual camera that is level and then project the images at their proper rotations, the end result being a stabilized camera view. Sounds like a project for an ambitious college CS student, so maybe it has been done already.

[Journeyman]

Adobe systems is working on a phenomenal bit of software for fully automatic video stabilization--the results have been absolutely stunning...but it is not available in the current release of Premiere or After Effects. One may speculate that this technology will become publicly available in the next release, although there have been no confirmed features thus far--only tech demos.

So....there's definitely nothing out there that just does this stuff, current spftware can reduce "camera shake" automatically, but not full stabilization, and better stuff is coming, but isn't here yet

[vxcv]

SIFT isn't really that out of reach, even to an intermediate programmer. Even though there are a lot of complex steps in the algorithm, free, open-source toolkits like OpenCV provide high level functions that do SIFT for you... i.e. you give it an image and it spits out the feature points.

The real challenge is correlating those feature points in some way -- and quickly, too, since videos have tens or even hundreds of thousands of frames -- without intervention from the user. Comparing every feature point to every other feature point is computationally burdensome, even on today's machines.

[PreyKill]

After having played with a cracked pirate copy of Sony Vegas 10 over the last year or so, I tried applying the program's video stabilization feature to a youtube video recently. It took a couple of tries before it gave me watchable results.

Here's the original: http://youtu.be/anB0nq7O0pM?t=12s

Here's the subclip I stabilized: http://youtu.be/48McBhGEnEo

[yummyyoungnorski]

It really should exist. Some simple edge finding algorithms running over your RGB matricies should do it, probably using the eigenvalues. Find the objects, by finding their edges, and then crop to some reasonable height and width region around the important elements whereby the "wobbling" of the region will always remain within the images in the frames. The only downside of this extremely simple approach is you will end up with an image that is smaller than what you started with, and will be smaller in approximate proportion to how much wobbling there really was, but as the old Russian Navy saying goes, "better is the enemy of good enough."

[Eupeptic]

After having played with a cracked pirate copy of Sony Vegas 10 over the last year or so, I tried applying the program's video stabilization feature to a youtube video recently.

That's actually reasonably watchable, given the original. It took me a second to realize that the "zoom" was due to it finding the part of the scene that is always in the shot. It gives some odd "blooming" effects when the frame size changes, but those don't happen too often.

Did you start with the 1080p original or something smaller? Do you have to hold Vegas' hand a lot while it works, or is more like "set everything up, push the button, wait, see what you get"? It also looks like you slowed down your copy a little vs. real time - was that a choice or is it inherent in the way the stabilization works?

Thanks!

Eupeptic

[PreyKill]

After having played with a cracked pirate copy of Sony Vegas 10 over the last year or so, I tried applying the program's video stabilization feature to a youtube video recently.

That's actually reasonably watchable, given the original. It took me a second to realize that the "zoom" was due to it finding the part of the scene that is always in the shot. It gives some odd "blooming" effects when the frame size changes, but those don't happen too often.

Did you start with the 1080p original or something smaller? Do you have to hold Vegas' hand a lot while it works, or is more like "set everything up, push the button, wait, see what you get"? It also looks like you slowed down your copy a little vs. real time - was that a choice or is it inherent in the way the stabilization works?

Thanks!

Eupeptic

I had started by doing everything manually with a pan/crop function - with unsatisfying results. Some algorithmic assist was needed. The stabilization control Vegas offered was adjustable, and turning everything on "full" wasn't as effective as throttling back some. There was no preview for this sort of thing, so it was a bit of trial and error, see what comes out, undo, adjust control inputs, run, see what comes out again.

That video was forgiving in that the handheld motion wasn't particularly dreadful to start with and stayed fairly widely framed throughout, so the instability wasn't exaggerated by tele/zoom lens magnification. Higher original resolution also helped. While Vegas did some crop-and-zoom as part of its function, I made manual adjustments to the stabilized footage to crop the scene further. The time-dilation was also my manual editing.

[PreyKill]

I have a few more attempts at stabilized clips to share:

Crocodiles eat zebra alive http://youtu.be/B4WoeLMKcLA
Lions eating zebra Serengeti Tanzania May 2014 http://youtu.be/-R2wOZYglo8
Three lionesses eat a young zebra alive Serengeti Tanzania http://youtu.be/lesxFy_p4vU

At this point I'm willing to take requests.