Jump to content

360 spherical panoramas, resolution and render time


Jon Berntsen
 Share

Recommended Posts

I also posted this to Virtual Reality Europe on facebook.

 

I want to initiate a discussion about resolutions and rendered 360 panoramas for HMD's. I have been doing some rendered tests, and fact is, it seems like I need to go to 12000x6000 px. That is insanely high for a render. But another thing as well: When viewing as your own eyes, you'd expect no noise at all. When looking at photos, we are used to some grain, but in the HMD we need to have no noise to have as good realism as possible.

 

So rendering 12000x6000px in a 3d software/render like max/vray with as little noise as possible, means noise treshold down to at least 0,0001 (yes), and then we are talking 48 hours of render with 70 buckets going, with noise treshold lowered to 0,001 and the new Vray denoiser on to compensate a little. So 48 hours for a still image! So imagine going for hi res 360 video, 25 x 48 hours per second.

 

I quickly realise this is not even possible if we bought another 10 or even 20 render nodes, so what do people here think about this issue?

Link to comment
Share on other sites

I guess that's why real-time VR would be the way to go over video! 12000px sounds like a lot, but considering your FOV is only around a quarter of that at any time then it makes sense. Optimisation of materials, reflections etc. becomes even more important. Geometry rather than maps, proxies where necessary. Worth keeping an eye on memory usage to make sure there isn't a bottleneck.

Link to comment
Share on other sites

its very render demanding

we recently did some basic 360 panoramas and it was 2 x 18k renders for stereo.....

denoising is pretty helpful, either in render or in post

 

for 360 video on one is rendering that res or its done on realtime engines. per frame 3D video is insanely demanding.

Link to comment
Share on other sites

You have to read the VR guide from Vray. The reason you need such high render sizes is you need to ensure a consistent image size when it is shown on screen. This is why the cube maps are so much better than spherical panos. In a spherical, you waste an easy 40% of your render time rendering the distortion at the top and bottom of the image at the poles. The cube maps, there is a lot less distortion and therefore is a much more effective use of pixels on your screen.

 

.001 seems like an insane noise threshold for that big of a render size. We do ours at .01 and if Bob Q Layman viewing our image can see (or care about) slight noise, then I'll give him a piece of candy.

 

Overall, no one ever said that the 6x1 or 12x1 cube maps at 18,000 some pixels wide was ever going to be a fast render. We'll put about 15-20 machines on each 12x1 cube map and our render times are on average 45 minutes to 2 hours depending on what the scene consists of.

Link to comment
Share on other sites

6000x12000 px is insane and is an overkill imho. Tried switching from hemispherical pano to a cube map? As rendering six images in about 1500x1500px should be way faster. Also use denoiser for noise reduction.

 

It's not insane. Just the render time :- ) { I did 12k equi pano test in corona, and it was about 40 hours with strong dual-xeon, with noise,yes, http://bit.ly/td_pano )

 

12k pano, has more information visually than 6x1.5k cubemaps, even though the pixel ratio makes for some loss there. Can be compensated with going with higher res for cube and still shaving some time though.

Never found the information 'inefficiency' at poles due to distortion to really matter, you almost never look straight up or down, but need the information on horizon the most. Equi pano does that.

 

One big reason equi pano isn't inferior to cubemaps, is the ease of color correction, and artifacts from effects like denoising, AA, glow/glare. Depend on situation, that might be minor but there are reasons when going pano equilateral is good choice.

 

But 12k pano needs to be converted to cube maps afterwards anyway since lot of gpus have limitation of 8196 output (tablets, mobile, low-cost laptops).

Link to comment
Share on other sites

40 hours rendertimes are insane, yep. It's not viable, especially not when you're gonna render out 20 of those in two weeks, along with a couple of other projects going. I never tested equirectangular images, and I can't find that option in vray, and google doesn't give me much info. But it do make sense that the extra resolution for the corners are not needed. I am not a fan of wasted time, and I am eager to get this optimized as much as possible. What you say about cube maps makes sense, Juraj, I just tested that, and I'm going to see if there's a possible equi solution for vray. Let me know if you are aware of such. Thanks.

Link to comment
Share on other sites

Jon, http://www.chaosgroup.com/public_images/VRguide_v1.pdf You might find some interesting thoughts in this document

 

Thank you. It would be wrong to say that I've read it all before, but I've scourged through it. I should perhaps read it in-depth.

However, as far as I can see, there's no info about equilateral setups there, and none in vray camera settings. I should go to the chaos group forum to check this out further, if no-one here have had success with equi camrea setup in vray.

 

Btw I am using Kolor Panotour Pro for tour software, and so far I am satisfied.

Link to comment
Share on other sites

I was able to get away with doing 5100x2250 for a recent project, render times were under an hour and the quality was good. The images weren't stereoscopic and they were being viewed in a web browser so this may not be an apples to apples comparison.

 

This sounds viable. Is it possible to see it/test it? I'm curious on what other people achieves with 1 hour rendering time regarding noise on such image. Also, how many buckets worked on it?

Link to comment
Share on other sites

You have to read the VR guide from Vray. The reason you need such high render sizes is you need to ensure a consistent image size when it is shown on screen. ...

 

I don't get this, because the screen resolution is not higher than what you have on new phones?

This is also as a question for Juraj, what digital media did you show the 18000px image on to not go "overkill"?

Link to comment
Share on other sites

I am a little scared of the render times I am seeing here. Granted, we all have a very different threshold for quality control. Having come from Neoscape where quality standards are set really high and having moved on over to a traditional architecture company for whom I am their first rendering artist, having to find a balance producing images at the last minute with no render farm, has been an interesting juggling act.

 

With that in mind, this is a quick proof of concept project I turned around http://removed using Pano2vr Pro. Highly recommend it by the way. I tried KRPano and this is just so much more intuitive to use at the moment at least. The panoramas were 10x5k and took 1 - 1.5 hours to render. Having stored the HDRI contribution with irradiance, I was able to cut down on that time from 5 hours. Yes, the shadow quality is not the same, especially to a trained and discerning audience found here. However, proving that this is a viable form of communication that does not necessitate more resources and more time on the part of the firm, was the main objective.

 

Personally, I can die happy, and probably somewhat tormented, if we could have infinite/vector like zoom on our renderings. Until then, drawing a line on quality vs time vs optimization will have to take precedent.

Edited by nikanikitina
Link to comment
Share on other sites

I don't get this, because the screen resolution is not higher than what you have on new phones?

This is also as a question for Juraj, what digital media did you show the 18000px image on to not go "overkill"?

 

I think they call it the "optimal size" for on screen display. You also have to remember that you are not showing this image on the full display of the phone. Rather on those squares for each eye that the VR display's lenses merge into one image that appears to be the full size of the display.

 

We do all of our VR stuff on the Gear VR. So for us it is easier of a workflow to go 12x1 cube images from vray, to phone, to Gear VR. All of our adjustments for these are done in the frame buffer. We'll put about 20 machines on them, with one 32 core thrown in there for good measure.

Link to comment
Share on other sites

I don't get this, because the screen resolution is not higher than what you have on new phones?

This is also as a question for Juraj, what digital media did you show the 18000px image on to not go "overkill"?

 

The screen res for something like the Samsung Galaxy S6 is 2,560 x 1440; which is very close to the recommended resolution of 1536. The real problem comes when viewing VR on the iPhone6. At half the resolution of the Galaxy, when your eyeballs are 2 or 3 inches away from the screen AND the image is being magnified via bi-convex lenses, you see all the individual screen pixels. Many folks don't notice but it drives me crazy. haha.

Link to comment
Share on other sites

Thanks for replies.

 

Do anybody know how I best can shoot images at the actual site for use in a domelight? Say, the client takes photos that covers 360/180 in all spots. Is that a good way to go?

 

I'm not thinking about lighting the scene with it, so exposures is not essential. I am using another HDR for lighting as for now. The point is getting the correct view, also with the sky, for each of the 14 spots in my exterior tour. :)

 

AND I need to get the customer to do those shots because of the location.

 

It has to be first time for everything, yeh?

Link to comment
Share on other sites

I am a little scared of the render times I am seeing here. Granted, we all have a very different threshold for quality control. Having come from Neoscape where quality standards are set really high and having moved on over to a traditional architecture company for whom I am their first rendering artist, having to find a balance producing images at the last minute with no render farm, has been an interesting juggling act.

 

With that in mind, this is a quick proof of concept project I turned around http://bit.ly/2bCJqUW using Pano2vr Pro. Highly recommend it by the way. I tried KRPano and this is just so much more intuitive to use at the moment at least. The panoramas were 10x5k and took 1 - 1.5 hours to render. Having stored the HDRI contribution with irradiance, I was able to cut down on that time from 5 hours. Yes, the shadow quality is not the same, especially to a trained and discerning audience found here. However, proving that this is a viable form of communication that does not necessitate more resources and more time on the part of the firm, was the main objective.

 

Personally, I can die happy, and probably somewhat tormented, if we could have infinite/vector like zoom on our renderings. Until then, drawing a line on quality vs time vs optimization will have to take precedent.

 

Fantastic work. Wish it had been stereoscopic! lol. Can I ask where you sourced the large maple trees from out front? Those are fantastic.

Link to comment
Share on other sites

Fantastic work. Wish it had been stereoscopic! lol. Can I ask where you sourced the large maple trees from out front? Those are fantastic.

 

Thanks so much. I wish it was stereoscopic, and infinitely zoomable, more flushed out and and and .... the list never ends. The trees are Evermotion I think. I've had them proxied in another project from a while back so I can't be certain from which collection they are.

Link to comment
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now
 Share

×
×
  • Create New...