Compression performance achievable with low-motion scene?

[Volatris]

Hello, I have a research project I am trying to size the long-term storage for.

My question is what sort of compression performance is achievable, be it in bits per pixel or data rate for a given resolution. A good example scene is a surveillance camera, like on a rooftop, where small portions of the scene have motion of relatively stable objects moving on relatively stable trajectories, and over extremely long periods of time major portions of the scene only vary by lighting condition. The camera would be relatively large frame, medium rate (3 to 15 Hz).

I can find several papers and sites describing compression performance for high-motion scenes or standard DVD content, but very little on these sorts of low-motion scenes. If I had a 1080p, 10 Hz (20 megapixel/s) video of this sort of scene, how much below 20 Mbps can I get with my long-term average data rate? Is 2 Mbps or even 200 kbps possible with suffiently low motion? I don't want to give up image quality to achieve compression; I need to remain in the "visually lossless" territory. But it seems the low-motion, very long time frame nature of the content should lend itself to tremendous compression.

Would anything be better than H.264 for this case? Would multiple passes of compression help (assume I have no constraints on processing power)? Are there particular settings that would be especially useful? Is there an advantage to compressing a long period of time at once (vs. breaking it up into shorter slices of time)? Would really infrequent keyframes be smart here?

Precision answers are not really required, I would like to get a feel for the ballpark though.

Thank you so much for your time and thoughts.

[Nelson37]

The amount of compression is only limited by the quality desired. Asking this question is like asking "how much ketchup do you put on your french fries?" It is a matter of personal choice.

With a low-motion static scene, there is no advantage to breaking up the clip inot smaller units, this will result in a larger total file.

Long keyframe intervals will increase compression, with the disadvantage of making seeking in the file less accurate.

H.264 is currently the best compression algorithm easily available. Needs more CPU power to create and to display.

The choice of BW versus color cameras and clean lenses will have a large effect on PQ versus file size.

[Volatris]

OK, thanks for the reply, here is some elaboration:

I'm aware of the personal choice, I tried to state I was looking for the "visually lossless" range (distinguished being numerically lossless). So, very high spatial quality, but seeking all of my gains in compression through exploiting the highly compressible nature of the temporal content. This is something I can find no metrics or even ballpark figures on.

Good follow-up questions: so let me ask for the case of BW (not color), and very high quality lenses.

I tried to say that CPU power was effectively unlimited here, and many many compression passes on a very long clip/movie should be encouraged. And long keyframe intervals impacting seek times and accuracy is also not a concern due to the more archival nature of my need.

I'm trying to size the storage for a research project, and would love to have some basic performance metrics or ballpark figures to do some estimating before getting too far along.

[Nelson37]

"Visually lossless" requires human eyeballs and a human brain, and is therefore a judgement call. If you make 10 vids, each slightly more compressed and study them in sequence, quality loss can be almost invisible. However, compare the first and the last and the loss is usually obvious. To determine at what point "some loss" is "too much" is not, IMO, mathematically quantifiable.

ALL compression involves loss, except maybe minimal compression such as Lagarith or Huffy. If there was some simple way to break the clip into ZERO-motion segments, these could be represented with a single still with a time indicator,then take the SLIGHT-motion clips, somehow decide if that motion was important and eliminate it, then find the Important-motion clips and ccompress those, higher compression could be achieved. BUT, the cost of the human-judgement time involved would exceed the cost of the extra storage required by not doing this.

The commercial methods for doing this for surveillance involve identifying zones where any motion is important, such as doors and windows, combined with a very low framerate except where motion is identified in important zones. Compression limits are usually defined by the possibility of facial recognition or reading a license plate withing the field of view. Motion-capture accuracy is sacrificed for resolution. The most important part is done at the source.

It's a decision of what you need to keep vs what you can throw away. There are folks who put a 3-hour vid on a CD and are happy, there are those who need a DL disk for 30 minutes.

For multi-pass, most agree that anything over three passes is overkill, a few swear by five passes or more. For an interesting alternate view, look up Spectate-Swamp's posts on videography.

Take a few hours of representative source video, and start compressing it. Go to extremes on either end, then look at it. A key factor is also have other people look at it.

To mathematically answer "How close do you need to be to the Mona Lisa to get a good look at it, and how long will it take?" IMO just can't be done. An educated guess at "How much can you compress properly set-up, hi-quality surveillance cameras using motion-capture software for recording" would be 50-80%.