Maxwell Render | Official public forum

We all know and love conventional 360° panorama images but Lat-Long Stereo panorama images are a very different beast, primarily because the eye separation has to be faked somehow. The way that I understand it, is that there needs to be maximum parallax when looking at the equator and then a transition to ZERO parallax when looking at either of the two poles. I don't know if I'm missing something obvious but it looks to me like Maxwell's solution has this exactly backwards and creates stereo panoramas that don't seem to work.

To show you what I mean, you can download my sample MXS files from here:
https://drive.google.com/file/d/0B2THCA ... sp=sharing

To be able to see the expected parallax I created 2 concentric spherical cages and they are both centered on my non-stereo camera. The setup looks like this:


The non-stereo centered 360° panorama view looks like the following image and the main thing to notice is that the red cage is exactly behind the green cage and is not visible:


Now when it comes time to render the Right eye Lat-Long Stereo image, I would expect to see the red cage poking out on the right side of the green cage. Most obviously at the equatorial regions and petering out to nothing at the two poles. Unfortunately Maxwell's rendered image is almost the exact opposite of what I would have expected:


Here's a rough mock-up of what I was expecting to see for the Right eye Lat-Long Stereo image. Notice the parallax at the equatorial regions:


Another thing to notice in Maxwell's Right eye Lat-Long Stereo image is the warping of the image at the two poles. This creates a spiral effect at the poles which is clearly incorrect when viewed with panorama viewer software. Here is a screen capture of Maxwell's output:


And here is what we should be seeing:


I tried all the settings in the Lat-Long Stereo Lens options but nothing gave me what I wanted.
* I'm not sure what "Parallax Distance" does. It only goes up to 360 which makes me think that it is not a distance but rather an angle.
* I thought "Separation Map" might be my answer and I tried a few different grey-scale gradients but nothing made a difference.

So after all that, am I missing something obvious or is this thing really not working as expected?


Thanks,
Andrew.

Hi Andrew,

Thanks for taking the time of testing this and sharing the scenes. We are looking into it.

Regards,

Fernando

Hello Andrew,

I know exactly of those things, You are talking about. I asked some questions about stereo lenses and output on last webinar. There wasn`t too much informations about it, just that there is some kind of standard implementation of stereo pictures output. I asked, because I see definitely more complex situation to solve. As You illustrated on fourth picture how expected image should look from right eye position. There is even more - as You will rotate head with virtual helmet (not just on 3D Tv screen), there will be moving of camera positions of both left and right eye. Center of the rotation will be in your neck. So the effect of real head rotation will appear as objects are moving in the picture with speed based on z-depth (try to rotate your head in the room with looking on the same place with your eyes - room where are objects near and far). Closer object are "moving" faster against farther objects. More complex situation is if there are transparent objects as glass (put glass with water on the table and look at it and rotate your head)... ... ... So the solution could not be just two stereo pictures. You should render "twins" for each angle small enough to not see flicker, or another solution will be specialized viewer and specialized output. Renderer should render each voxel visible in each head position, then in the viewer each pixel will be moving based on its z-distance. As I said before, it is little problem with transparent objects. Without this in mind, virtual reality output will look somehow fake and unreal.

Patrik

PA3K wrote:I see definitely more complex situation to solve.

I know what you mean Patrick but I think that's beyond the scope of these simple stereo panoramas. What you're talking about sounds like 'Light Field' technology. In case you haven't seen it, there's a really exciting video here:
https://youtu.be/pyJUg-ja0cg

Andrew.

Thank You Andrew, I haven`t seen this video yet. That is exactly I was talking about.
You are right - these simple panoramas have long way to go to simulate reality. But it can be done by more ways as I said before. One way is specialized viewer with some kind of its own render engine with input prepared by main render engine. Other way (kind of "brute force") there must be a lot of "twins" and software must choose "twin" from database based on helmet or head position. This way it would be a lot of data in one panorama (360*360 "twins" * maybe 2MB = 259,2 GB) some motion blur algorithm to blur one degree transition... I can not imagine software and hardware to handle it ...
We will see ...

And I see on Your third picture (maxwell output for right eye) "classic" problem. Try to imagine that You are standing on south pole. Where is the north? Everywhere. If You look upside (90 degree up) with virtual helmet - where is "right" and where "left"? You are sitting on office chair looking upside and spinning. 180 degree rotation and You should see right where was left (absolutely can not be done with just two pictures) ... ... ...

PA3K wrote:There is even more - as You will rotate head with virtual helmet (not just on 3D Tv screen), there will be moving of camera positions of both left and right eye. Center of the rotation will be in your neck. So the effect of real head rotation will appear as objects are moving in the picture with speed based on z-depth (try to rotate your head in the room with looking on the same place with your eyes - room where are objects near and far). Closer object are "moving" faster against farther objects.

Otherwise known as parallax?

This is really why I haven't understood too much the hype over stereo 360 renders - is the sense of depth really increased so much if there is no parallax, compared to just a regular 360 view?

Yes, You are right Mihai . Parallax or however called will add much more 3D resolution to images. Without it 3D images look like there are flat cardboards with various distances. Parallax add "plasticity". Even with one eye opened, i can tell if the object is in the same distance as other one. I tried to put my finger next to my laptop display edge. With just one eye opened I rotated my head. If the distance between my finger and edge remained the same I can tell they are in the same distance from me. Without parallax or without head rotation and one eye opened You can not tell even if objects are in the same distance. I feel little bit crazy when I am doing these tests in reality

I'm glad I saw this post before downloading and spending time to figure out the same things. Thank you for posting!

I have a question. Andrew Hazelden's mentalray shader for doing stereo lat-longs and fisheyes has the option of using maps to define areas of no-parallax. This is useful for removing or "flattening" areas where disparity doesn't make sense... like at north and south poles. Does Maxwell have something like this?

Thanks!

Thanks for all your feedback. It is the first time we enter in the stereo lens world and it is normal that some things are still missing/failing in this beta. We are looking on it.

BradT wrote:I'm glad I saw this post before downloading and spending time to figure out the same things. Thank you for posting!

I have a question. Andrew Hazelden's mentalray shader for doing stereo lat-longs and fisheyes has the option of using maps to define areas of no-parallax. This is useful for removing or "flattening" areas where disparity doesn't make sense... like at north and south poles. Does Maxwell have something like this?

Andrew is helping us a lot (hi Andrew, if you are reading this), our lat-long implementation is based on his work and all the parameters should be identical. So yes, you will be able to use maps to define no-parallax areas.

Excellent news! Thank you.

Ive just tried Maxwell v3.1.99.10 on the sample file that I provided above and even though there now is some parallax at the equator, the two poles are even more warped than they were before, when they really should have NO parallax at all. I'm using the "PSViewer 2" app on my Android phone and the poles have an extreme spiral in them and are so far apart for the left and right eyes that it's impossible to make the two images converge.

When I look at stereo panoramas done with Octane Render and now Vray, the stereo effect works really well and the poles don't cause any convergence problems at all. There is something still very wrong with Maxwell's implementation of the Stereo Panorama lens

I tried different values for the "Parallax Distance" but nothing seemed to converging on a usable solution.

From what I've read about Andrew Hazelden's work elsewhere, it feel like the "Separation Map" is the magic ingredient to progressively reduce the amount of parallax as you approach the poles but no matter what kind of grey scale gradient I try to plug into that slot, it seems to make no difference to the output at all.

What kind of image are we supposed to use for the "Separation Map" and how are you supposed to make it work. The online help doesn't seem to offer any clues at all.

Am I wrong in thinking that a correct solution should look something like this mockup which doesn't have any warping at the poles?


Cheers,
Andrew.

I feel like I'm being a pest about this but the latest version of Maxwell (v3.1.99.11) hasn't done anything to improve the Stereo Panorama Lat-Long render situation. The geometry at the poles is very warped and no matter what I do I can't seem to make the "Separation Map" work at all.

Is the "Separation Map" fully implemented at this stage in the beta cycle? Can we have some instruction on how to make it work? Should I even be turning to the "Separation Map" as the solution to the problem with the warping at the poles? Can we have some feedback on this? I feel like I'm groping around in the dark here.


Andrew.

Well here we are again and Maxwell v3.2 RC1 hasn't improved the Lat-Long Stereo panorama bug. Can someone from Next limit please comment on this issue of the warped zenith & nadir in the Left/Right stereo panorama images?

When I look at stereo panoramas created with V-Ray or OctaneRender, they handle the need for NO parallax at the poles perfectly well but Maxwell introduces counter rotating warped spirals at the zenith & nadir which makes it impossible for your eyes to merge the two Left/Right eye images as your gaze moves towards either of the two poles.

If stereo panoramas are going to be listed as a new feature in v3.2 then this issue needs to be addressed because as it stands now, I don't believe that it is working properly. If others are able to make it work and I'm just doing something stupidly wrong then please let me know, I would dearly love to be able to use this new feature.


Thanks,
Andrew.

Contact them directly via the website so you can confirm if this is a bug or need special setup