Secrets of the Sparse Point Cloud in Gaussian Splatting process

[@robertglas]

Fantastic as always! Minor note on your remarks of dense point clouds: You state the limit is around 1 gig. Sounds about right but one can be more precise: the limit is the same as the upper limit of the max splat count in Postshot. With my configuration i can train models up to 9 million splats and as long i make sure the dense cloud I start of with is no bigger then 9 M points, it works. You probably knew this, but it might help others.

[@pedroantonio5031]

Your content is so good and well produced!

[@DavidHeath-u8m]

Olli, love watching your videos, well produced and on focus concerning many of the questions folks are asking about 3DGS. XGRIDS is onto something with the hybrid combination of point cloud and 3DGS data. Exciting times we are living in!

[@erickgeisler]

Great video Olli. I just got the xgrids scanner and the output is pretty great. The capture time is really fast. I think your research is spot-on. Sparce points are the chicken-wire everything is built-on. My guess is the more even the distribution of points along the surface, the better the result. Time for some testing...

[@bottledwaterprod]

Camera tracking for average ppl is possible now though! It's still new so I forget the company making it, but through an app and connection interface, you can use an iPhone's cameras, sensors, gps, and lidar to become a position tracker for any other camera and lens you may own. It was demoed at NAB with the goal of making virtual production in Unreal 5 more accessible to enthusiasts and indie filmmakers. I bet that's the missing piece of this puzzle. Ever since I got into photogrammetry, I wondered why I couldn't track my camera's position when it seems all the tech we need to measure and make use of such data is right there.

[@cem_kaya]

Do you remember the name ?

[@GoceMilanoski]

​@@cem_kaya i believe the app is Jetset Cine by Lightcraft Pro

[@Grigoriy360]

Thank you, Olli. Like always a lot of new information to think about.

[@randomthing999]

Cool video, very instructive, thank you! Though, I dont really see that as a feature but more as an imprecision. Looks like I would just have to put one spare point and it will compute all the rest. I hoped that inputing a perfect sparse point would help him be less noisy but with its way of doing things, it doesnt change anything. Do you know if there is a way to decrease the splatting in the splats related to the sparse points? To tighten things up?

[@AndriiShramko]

Thank you.

[@francescoluciani3397]

Fantastic tutorial!!!!

[@edwardverbree9448]

Great video, again. Thanks. Would be nice to check the created .ply file as 'just' a point cloud and count the number of points. Are there points created to represent the skull? Are the point on the arm the same as the synthetic SfM points?

[@kasali2739]

Thank you! This is what i was waiting for. What is If you have a 3d synthetic Data AS a House With separate rooms and you produce pcloud by scattering Points ON geometry, how can a Training process know what picture to Match to what Points, since room structures With corridor AS whole are Not that continues space AS one single room is? I watched old Tutorial ON doing that With Kinect device and it used some Kind of GPS i belive

[@shiccup]

good to see

[@josealberto-rj1si]

Hi, great video, one question, can you transfer a scan from Jawset to Blender? With an addon, I don't care about the texture and it doesn't look that good either. If possible, can you make a tutorial? Thanks, and sorry for my English, it's the translator.

[@OlliHuttunen78]

No. Not from Postshot. But if you want to convert 3dgs to mesh you should check out the new features in Kiri Engine. They have that option in their service. Check this video: kzbin.info/www/bejne/f5m1o5J7n9OUjbMsi=DJU4ApbOplSpvqUE

[@josealberto-rj1si]

@@OlliHuttunen78 Thanks, it worked for me

[@tiagogiraovideos]

Superb!

[@epkostaring]

Love your content 🔥

[@sander-wit]

Hi Olli, I haven't yet watched the full video yet or tried your setup in Blender for creating an array of images from around an object, but seeing your setup at 2:50 I noticed you're using an uv-sphere for the positions of the cameras, but I think it would be more efficient to use an icosphere in stead because the cameras are than evenly spaced around the object where you have a lot of camers close to eachother at the poles of a uv-sphere.

[@OlliHuttunen78]

Watch the video.

[@AndriiShramko]

I've noticed the following issue and can't figure out how to resolve it. For example, when I create a point cloud and align cameras in RealityCapture, I always get worse GS results than if I align the cameras and create the point cloud in Postshot Colmap. I've tried all possible RC settings and created various point clouds, but I have never achieved a result better than Colmap. The result is always worse. What could be the reason for this? If point clouds aren't that important, then why is the GS result always worse when we import the camera alignment and point cloud from RC? I've done hundreds of different tests and still haven't found a way to create a GS of the same quality as with Colmap. Does RC always produce worse camera alignment than Colmap?

[@r.m8146]

It'd be awesome if you could review XGRIDS.

[@oebleh]

Testing the XGrid K1 right now , results are amazing. In rendering there is no time advantage because of the Lidar pointcloud by the way. ( or our models are too complex :-) )

[@davorbokun]

This is great analysis, thank you! How good are Gaussian Splats at representing volumetric data, like layered fog, or a volumetric hologram? Something without well defined surface. Also, a bit related question, does the algorithm get confused if there are many points in places where there isn't anything in the scene? For example if you were to include the skull in the point cloud but omit it in the renders?

[@davekite5690]

That was fascinating - I wonder how it might work, if recreating virtual exteriors with lots of 3d geometry (eg 3d poly trees )?

[@mousatat7392]

The reason that the skull appeared while it does not exist in the point cloud is the images has it, so the algorithm shifted some points from the hand to the position of the skull to reconstruct it, this is not so efficient it will reduce the quality of the skull and the hand a little bit, but still works enough.

[@jorcher]

nice!

[@taavetmalkov3295]

isnt the pointcloud a frame to which the splats land on? hmm

[@mousatat7392]

The problem in sfm is not the sparsity of the point cloud, but the disability of localising the images with low texture and details. Nowadays VGGsfm solves this issue but still needs 80 GB gpu to run it.😢

[@edh615]

what about ace0?

[@StevenClark-r9r]

Thanks for sharing such valuable information! I have a quick question: I have a SafePal wallet with USDT, and I have the seed phrase. (alarm fetch churn bridge exercise tape speak race clerk couch crater letter). How can I transfer them to Binance?

[@realhamza2001]

I know this is unrelated but give the Quran a read, also have a good day :)