How can it tell the difference between true video movement and just camera shaking/twitching?
It seems like every stabilized video is perfectly stabilized, with no errors or mistaken movement.
In: 6
Did you ask the creator?
I believe it takes the center of the video and allows the frames to move around it. That is why the stabilized video has lots of black space and you can see the frames flying around.
If all the pixel data move together, then the frame is moving.
If some pixels are moving differently, then those things are moving.
Sussing out the details accurately obviously takes some math, and modern algorithms are good at it, but that’s the basic idea.
It looks at a handful of frames and tries to pick out the “objects” in the frame by looking for edges and discernible shapes. It does further analysis to try and figure out depth and distance. This sounds amazing but it’s more or less what your brain’s doing with your eyes right now. (I mean, that’s still amazing, it’s just not a new trick. What made it take so long was getting enough computational power to do it on the fly, not the math of how to do it in general.)
Then it looks at how those objects seem to have moved between several frames. If it looks like the frames are tilting one way, it can rotate the frame to try and compensate and keep things moving along a straighter path. This takes some fiddly math, as it has to try to figure out the general motion of the objects then start to work out how they’re deviating from a “straight” path. It’s guessing. But it gets to guess a lot of times every fraction of a second and given enough guesses it’s pretty good at figuring it out.
Latest Answers