[weinzierl]

The article talks about image comparisons but does not say what the best way to extract an image is.

If I want the best possible quality image at a precisely specified time, what would I do?

Can I increase quality if I have some leeway regarding the time (to use the closest keyframe)?

Is there a way to "undo" motion blur and get a sharp picture?

[latexr]

I usually use a shortcut in mpv to extract the screenshot. If I want to do it via the command-line:

  ffmpeg -ss 00:00:12.435 -i '/Users/weinzieri/videofile.mp4' -vframes 1 '/Users/weinzieri/image.png'

The means “go to 00:00:12.435 on the file /Users/weinzieri/videofile.mp4 and extract one frame to the file /Users/weinzieri/image.png”.

[ErroneousBosh]

> Is there a way to "undo" motion blur and get a sharp picture?

Not really, no, any more than there is a way to unblur something that was shot out of focus.

You can play clever tricks with motion estimation and neural networks but really all you're getting is a prediction of what it might have been like if the data had really been present.

Once the information is gone, it's gone.

[weinzierl]

If the estimation is good it might be enough for some use cases. Is there any software out there that specializes in this? Similarly to maybe AI colorizing or upscaling, which both guess information that is not there anymore.

[memoriuaysj]

it's not gone, just more difficult to extract

video has certain temporal statistics which can allow you to fit the missing information

only true blurred white noise is impossible to recover

[ErroneousBosh]

It really is gone. You can predict what you think it might have been, but you can't know what it was.

[memoriuaysj]

it's gone in a single still frame

but across many consecutive frames, the information is spread out temporaly and can be recovered (partially)

the same principle of how you can get a high resolution image from a short video, by extracting the same patch from multiple frames

https://en.wikipedia.org/wiki/Video_super-resolution

[ErroneousBosh]

No, it's not "restoring detail". The information is gone.

It is predicting what the information might maybe have been like.

[memoriuaysj]

you are arguing with math proofs here, the information is not gone, if it was a real video (as opposed to adversarily generated video)

[weinzierl]

That looks interesting. Is there ready-made software that can do this? Doesn't have to be easy to use just useable with a time commitment of a few days.

[crazygringo]

> Not really, no, any more than there is a way to unblur something that was shot out of focus.

This is actually possible:

https://en.wikipedia.org/wiki/Deconvolution

If you have a high-quality image (before any compression) with a consistent blur, you can actually remove blur surprisingly well. Not completely perfectly, but often to a surprising degree that defies intuition.

And it's not a prediction -- it's recovering the actual data. Just because it's blurred doesn't mean it's gone -- it's just smeared across pixels, but clever math can be use to recover it. It's used widely in certain types of scientific imaging.

For photographers, it's most useful in removing motion blur from accidentally moving the camera while snapping a photo.

[perching_aix]

You'll need to settle on a decoder. I personally just use my video player for this, mpc-hc.

In mpc-hc, you can framestep using CTRL+LeftArrow (steps a frame backward) or CTRL+RightArrow (steps a frame forward). This lets you select the frame you want to capture. You do not need to be on a keyframe. These keybinds are configurable and may be different on the latest version.

Then in the File menu, there's an export image option. It directly exports the frame you're currently on, to disk. Make sure to use a lossless format for comparisons (e.g. PNG).

I'm aware this can be done in other players - like mpv - as well, although there I believe no keybinds are set up for this by default, and the default export format is JPEG.