Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[deathevor]

How difficult would be to make a video recorder to output second stream with Z-depth data? (same question for screenshots).
And maybe possibility to output other channels like velocity pass?


Outerra can be a great engine for making shots/backgrounds for films. Especially the LOD system (with some tweaks) makes it perfect for long camera flights.

[cameni]

Videos might be problematic because of increased bandwidth and disk speed, but if you have a SSD drive it could be ok. But another question is the format.

For screens it should be considerably easier, once the format is determined.

[deathevor]

Using Revodrive for caching. But not very cheap solution.

Will the side-by-side (double width) writes will be any faster or easier to sync?
So 3840x1080 - Left side - complete pass. Right Side - Z-Depth.

Or another idea is to have an option to select different drives for different streams.


It's just if the Frame generation is not so difficult for you, would be nice not lo leave this option behind.

[deathevor]

About formats:

Cinema DNG is open format. It's 12bit "raw" File Sequence.
http://en.wikipedia.org/wiki/CinemaDNG

BlackMagic camera records 2.5k DNG file stream on standard SSD.
It's only single stream, so not sure how fast it would do 2.

[deathevor]

Outerra screenshots with Z-Depth:

[cameni]

Internally the videos capture frame in YUV422 format, but we'll switch to YUV420. One pixel comes with one luminance and one chroma sample. To pack it into the same frame, we could put depth into one of these and have the other for something else. However, the resolution is 8 bits there with compression, what resolution you need for the depth and what distribution? Linear would not be suitable.

[deathevor]

Bit-Depth:

The more bit depth is better as it gives better precision for depth.
For PNG's I've used 16bit Grayscale.
However some Kinect scans I had had 8-bit Grayscale - still had quite good result in 3D reconstruction.

Resolution:

The resolution would be nice if it would be the same as RGB pass.

Distribution:

Didn't understand that. But if you talking about Z Min/Max. Distance for black and white. Would be nice to have it customizable (somewhere in recording setup).

[cameni]

The question was whether the depth should be written linearly between min/max, or if it should be something that uses the values better, like a logarithmic function.
Anyway, please put a feature request into the tracker so that it's not easily forgotten.

[deathevor]

Right got it.
I think most of plugins use linear. In one I use has a curve option.
However standard  3d max Z-Depth uses linear depth calculation. So I stick with it.

But thanks for pointing at it, will pay more attention as currently working with 5k and some kind of exponential distribution could bring better precision.


So why linear doesn't suit you?

[deathevor]

Cameni, sorry for silly question - where's feature request tracker?
I thought "Ideas & Suggestions & Questions" is the place for feature requests.

[cameni]

The forum is for discussion, once something gets worthy of future implementation it should go to the issue tracker, see here. It's because topics in the forum tend to scroll away into oblivion soon, and we easily forget what was where - there's a lot of noise on the forums.

Linear distribution isn't very suitable for depth mapping because it doesn't reflect the changing detail of what you see with distance. If you have a resolution of 10m, it's 10m in front of you, and it's 10m also 1km away. Clearly you'd need the closer objects to have a better resolution, while the distant ones can have less values from the available range. Basically, you want the precision to be tied to the screen size of the objects, since you are doing post-process effects on the screen.

Since the projected screen size of objects is proportional to the reciprocal of geometry depth (1/depth), a function that maps the available range best is the one that has its derivative proportional to 1/depth. That function is the logarithm, and that's what we are using.

[deathevor]

Request Submitted

[boonitz]

Why not consider openEXR for stills? It's pretty much an industry standard, it allows you to embed unlimited AOVs with different channel definitions, bit depth and compression, plus metadata etc...
It'd allow you to encode additionally pretty much all you want in there (depth, world coordinates, normals, coverage, layers sorted by material, by entity, camera parameters, time of the day, you name it...)
Just my 50 cents