Photogrammetry / NeRF / Gaussian Splatting comparison
Vložit
- čas přidán 30. 09. 2023
- Workflow and resources:
Photogrammetry model on sketchfab: skfb.ly/oLOQw
Church Rock dataset ZIP: drive.google.com/file/d/1ttkl...
Agisoft Metashape: www.agisoft.com/
NerfStudio: docs.nerf.studio/
3D Gaussian Splatting for Real-Time Radiance Field Rendering: github.com/graphdeco-inria/ga...
Aras-p’s Unity Project: github.com/aras-p/UnityGaussi...
I love how you can hear the fan going with the gaussian splatter
😂 my poor comp was breathing heavy
You can hear it from the jump when he switches to NeRF too.
@@MatthewBrennanThought it was raining there for a moment, then realized I’ve head that same sound in my office. Def taxing the GPU, but great video.
NeRF and gaussian splatting are so realistic that they even captured the wind.
Not exaggerating, when Gaussian Splatting became to sudden prominence, I was waiting for a video exactly like this.
For anyone interested, here's my TLDR of the video: Biggest advantage of photogrammetry method (polygonal): model is metrically valid, can be used for measurements, model is fairly lightweight as it's a model, polygonal, and also less computationally taxing to post-process for the same reason, can use in many software aand share easily. Advantage of NeRFs (radiance field): includes more scene information (distance, sky), but this info is estimated, generated (via color info aka a radiance field) in realtime so perhaps it's not the most scientifically accurate. But because a full scene is being computed, creating new "footage" from novel viewpoints is possible. Advantage of Guassian Splat (static point cloud): includes scene information like NeRFs, but is not being computed in realtime aka it's a static point cloud (or splat cloud). Because the visualization is static, Guassian Splats can be used in many visualization software (game engines like Unity, Unreal, as well as 3D software like Blender, Cinema 4D). It's "the best of both worlds". Also, of course, it's the most fun to say :)
well put!
Hi, great video - I think it clarifies to a lot of people the difference between the three different technologies.
Just one thing that would have been important to mention, is the biggest advantages of NERF and GS is the ability to capture reflections, and even transparency, which is just about impossible with Photogrammetrie. Granted that your example is, of course, not the right one to demonstrate these features but maybe you have a different set of images where you can demonstrate that difference in more detail.
You're absolutely right - in fact I just went out and took some video of my car and some reflective surfaces in the rain - which would be very hard for photogrammetry to reconstruct well - I'll do a side-by-side!
photogrammetry*
GS can also rasterize thin structures much better than mesh recon or nerf. Give it a few months my bet is GS can surpass meshing in rendering quality and file size.
Good point
I have never used either of these techniques, however ever since I've seen these technologies become popular, i had a really poor understanding and had no idea how they compared, but still remained curious. This video filled the gap in my understanding perfectly
This was exactly what i was looking for! I tried gaussian in unity in vr. I have to try with your conclusion. Thanks!!! Subscribed! :)
Thanks for making this video! I've been playing with photogrammetry for about 6 - 7 years, and have only been a curious bystander with respect to neRF and GS. It's interesting but I envisioned something like neRF and GS in conjunction with 3D meshes several years ago, before learning that they exist. Your video does a wonderful job of explaining the distinction between the three, especially between neRF and GS. I watched til the end! Thank you for creating it. I'll connect with you on your social accounts and hopefully we'll run into each other in the unfolding "metaverse".
The splatter is really cool looking. Looks like a hallucination or a dream
Very cool work thanks!
Could be cool to use an example with more variance in material properties, to showcase how different techniques deal with things like reflectivity and transparency.
See my video here :) Reflective Object: Gaussian Splatting radiance field vs. Photogrammetry mesh
czcams.com/video/gheD8vrOJNI/video.html
Thank you!! I’ve been looking for a video like this since I learnt about Gaussian splatting. Also the music you played in the showcases was chilling :))
Amazing comparison! And very cool flying scenes for NeRF and Gaussian Splats.
thanks this was super helpful
Thanks a lot! That was exactly the quick introduction to the subject that I needed!
I have absolutely no affiliation with anything 3D-related though somehow I find this Gaussian Splatting thing so intriguing though I had no clear understanding what it was haha. Just that you could make high quality looking 3D renders of things with just a video. Cool to actually see a more technical and comparison video about it. Thank you for making this video!
Thanks for the good comparison video! You did a great job highlighting many of the topics of interest! 👏👏👏
Great comparision of different methods! Looking at the drone video vs the output, it seems clear that all these technologies will get better when we get more processing power. The current output is nowhere close the detail level of input video but there's no reason to think it couldn't be given enough computing resources.
Thanks for the video! The best way to learn the difference about NeRF and gaussian splatting 🙏🏻
thank you very much for this video :D superb for future developments in archaeology. I am already checking NERF with Kiri Engine, but I loved to see how to do it with the tools you mentioned. Coool!
The last shot is really good!
Very valuable information for us 3D creators. Thank you very much!
This is absolutely amazing! I think Gaussian Splatting might replace rasterised/polygonal rendering in the near future!
GS is great, but it will be a real challenge to invent an optimized way of applying real time lighting
@@dmitriytuchashvili8594 agree, interactive lightning is the main issue
Curious to see how GS handles reflections, normals, displacements, transluscency and so on.
There's also a problem with physics, interactivity and the rendering of anything besides real world data (procedural generation would be hell), which altogether makes it likely to only have niche use cases, though I do admit it's cool
@@constantinosschinas4503 see this video: czcams.com/video/gheD8vrOJNI/video.html
I think photogrametry and some kind of Gaussian splatting will be fused one day
Wonderful work, really appreciate it
Thank you, it was very educational and enjoyable video!
That's an amazingly beautiful rock formation, not that it's the topic of the video but just felt like mentioning it
It has an interesting backstory too! The legend goes that a utopian community wanted to hollow out the rock to use as a church - and even went so far as to begin chiseling a doorway (you can see the opening in the video/model).
Very intriguing comparison. Had not heard of Gaussian Splatting previously. With Photogrammetry it would be possible to model different sun angle and lighting effects. Could be interesting to explore how the various could take advantage of video captures taken at different times of day.
For example, could allow for animating time and motion. Just a thought that might be interesting to explore.
While it's true with photogrammetry you could relight the scene, the shadows are still "baked" into the texture. So in this case, Church Rock would still be casting that shadow even if you put virtual lights into the scene to reimagine it.
superb explanation...thank you for sharing your knowledge.
Ooh thanks for this overview. Very helpful.
Nobody gonna talk about how the rock looks like a space ship that turned into stone?
🚀
honestly I love the gaussian splatting look. Especially when you move "too" close :D
Thank you for explaining brother! 💪
Good comparison and review. Good notes, links in the video info grey area.
Great comparison! Could you have cropped the splats that were interrupting the sky in your Unity example for a cleaner look - or is that limited?
In theory it should be possible, because the splats are directly related to the sparse cloud from COLMAP - I'm planning to investigate this.
I love the examples, and the discussion on the technical differences. Though could definitely use better quality screen capture.
You're right - I captured at 1080 (my screen's max res) and upscaled it, but didn't realize it because I had my premiere clip previews set to 1/8 res, so didn't notice the blurriness! Sorry!
@@MatthewBrennan Some great points in there. Thanks for the vid.
Looks more like 360p mate, maybe you set premiere to render using preview cache @@MatthewBrennan
@@lemovision could be- I think I fixed it for subsequent exports, at least 🙃
CZcams compression likely not helping either.
you can also compare nvidia's neuralangelo. that would be great
Great work! Keep doing the stuff! Like and subcribed!
interesting , good informative vid , thankyou, as a side note, lets consider Bob ross once painted 50 sheds and 1 tree.
Great video, I think if you reuploaded this with better quality it could gather even more attention.
Unfortunately I don’t think you can replace previous uploads on CZcams
Great job, subbed. 🎉
Well described ❤ Appreciate ❤
intersting fact is that nerf technically uses ml to search for the camera positions. the process stops when the render of the estimated volume matches the photo itself. ( within a threshold )
Oh hey I've been there. Recognized it before you even said "Church Rock". Nice stretch of road just outside Canyonlands National jawdropping park.
That's why I'm here, I have driven past that rock hundreds if not thousands of times in my life. I have no knowledge about the actual subject matter of this video but I'll watch the whole thing just out of gratitude for the nostalgia.
Cool. Thanks for sharing.
It seems that the next step is a neural/AI enhanced version of guassian splatting to improve quality/render performance/data size.
Thanks - I'd like to see a comparison of how you would actually use these in production - especially NERF and splatting.
Right now I get the feeling it's very much a "solution" in search of a problem. But here's another video with some thoughts on the potential practicality: czcams.com/video/Ksi_RfY77SI/video.html
Hey matthew, thank you very much for making this video comparison. I have been wondering about Gaussian Splating all week long. But still I fail to see how is this a 3D rendering revolution like many in the community are saying, I don't understand the hype. Which could be the applications for this? Is it possible to make meshes from this point clouds? Can we make it collide with other objects in a scene? Can it be re lighten? To me it feels like baked lighting textures without a body. Please take me out of my miserable ignorance, what are the possible/theoretical applications?
Right now, I think it's a 'solution in search of a problem'. As I mentioned in the video - virtual production could benefit greatly from this: you can plan camera shots well in advance using limited input data to get a realistic representation of scene. Likewise, if you capture a scene well, you can do all of your camera moves virtually, and composite with actors, etc... and not have to worry about "getting the perfect shot" on location.
It is possible to mesh NeRFs. But the volumetric clouds (NeRFs and GS radiance fields) cannot be re-lit (by their nature). In a Unity or Unreal scene, they don't have collisions, but you could add some invisible geometry to act as colliders. I'll be making another video demonstrating some of that this week!
@@MatthewBrennan What an excellent answer. Thank you so much!
its essentially faster and cheaper photogramery which is in itself already quite useful. its better to ignore self compounding twitter hype but the "revolutionary" part is in that its a entirely new real time rendering technique that is entirely based on "splatted" point clouds not on polygons or anything so its a whole new area of research its possible to come up with an entirely different real time rendering pipeline based on "splatted" points with physics and everything and it might even be good or better! but will that actually happen? will it be actually good? will it fall to same problems as all other point/voxel based rendering systems did? i dunno
@@camerbot that adds a lot of insight. Thank you very much
Thanks for the good video and for providing the nice dataset. What is the focal length of your drone's camera (full frame equivalent)?
35mm equivalent is ~28mm. It's a 1" CMOS 20mpx sensor (for still images). However this dataset uses 4k video.
@@MatthewBrennan thanks for your answer. I'm trying to run your dataset using micmac photogrammetry. I'll let you know if I get good results!
Hi Matthew. Great video explanation. Which drone did you use for this test? Do you have by any chance any more drone footage? Have you use the DJI Mavick 3 pro Cine for photogrammetry? Thanks.
I used a Mavic 2 for this model. I have used a number of different drones for photogrammetry in the past, but haven't tried the Mavic 3 yet, although I don't think the Cine model adds anything particularly useful for traditional photogrammetry.
Thx for the vids - deep respect for your knowledge and sharing it. (needed to watch it in 1.5x tho - otherwise it would have been kind of as if told in slow motion xD)
😂
Seeing Matthew's analysis of these three technologies was quite insightful. It brought to mind a question an architect friend once posed: Could one feasibly craft an architectural blueprint of a home or apartment using video footage? Considering your experience with these three tech contenders, would you say we're on the brink of making this potential a reality?
You could definitely build a rough model from video footage (provided that the video entered every room). No digitization technology (yet) will output a plan useful to an architect without substantial work by hand- however the power of these techs is that you can achieve results based on very little information (I.e. a series of photos or video) that can then be interpreted by an architect or draftsman and turned into a polished representation, like a plan or section.
@@MatthewBrennan Thank you very much for your response; I have more clues so my friend can make things more efficient.
It's a matter of time before new papers are published. 🔜
very very interesting
Interesting how the NeRF has captured the changing cloud shadow over time as a positional aspect
3DGS does something similar, as everything is based on the interpolated “viewing angle”/position of the scene, which of course is tied to the conditions/time (shadow or sun) that each photo was captured under/during.
Big up my friend!
Very well done. You mentioned using data resulting from aligning photos in Metashape in place of ColeMap for NeRF studio. Can this substitution of processing also be done for Gaussian Splatting? I'd love to use the aligned photo data & camera model data from many Metashape (standard edition) projects for GS without using ColeMap as Im already confident in the alignment. Any suggestions on how to accomplish that?
Yes - I use this script: github.com/agisoft-llc/metashape-scripts/blob/master/src/export_for_gaussian_splatting.py
@@MatthewBrennan Thanks! Would I be correct in assuming that this must be run in (or at least with) the professional edition (not the standard edition) of Metashape?
hmm... I'm actually not sure if the standard version supports python scripts. I am using the professional version to process this data.
Is this purpose-driven for a particular field? I've always been fond of the intersection between geospatial tech and the broader CS/gaming world.
Right now, I think the primary application is in virtual production. This is a relatively new method, so I'm sure as it evolves, new applications will develop. At the moment it is not a straight replacement for any existing digitization or visualization technology.
Hi Matt, are you testing with NV Omniverse? Great video!
No. I tried it about a year ago but didn't find it very compelling.
Cool stuff! I was thinking regarding the fact that extracting images from a movie gives poor photos: is there a way to extract higher quality images that you're aware of? I'm thinking it would be fairly straight forward to create an ai that takes five images in a row and creates a de-blurred version of the middle image using data from all five images. Or are there additional issues?
The best type of image is a high resolution digital still photo, with low iso (low sensor noise) and high sharpness (I.e. high f-number… usually f/8 or higher)
@@MatthewBrennan I see - but much could be gained if we were able to produce similar images from moving images - it would just be way faster to capture that way, not only for drones. Probably someone will have a go at it. I'm feel better images than a random frame from a movie could be generated, but never as good as the type of image you're describing.
BTW, someone mentioned that within a few years, a team shooting footage on location will map the site using photogrammetry as a matter of course. We're used to voice dubbing in post production, but it's hard to shoot a new scene on a site that's distant - or in different weather/season. I think there are interesting things ahead!
This is nuts for set design, remodels... Be interesting to see what could be done with scopes for plumbers, both doctors and the ones with the butt crackin.
Hi Matthew! Thanks for this video. I'm new and a total noob to this field. I'm a Marketer and my line of work I'm having to learn a lot of these things including reality capture, photogrammetry, NeRFs, 3D GS, Digital Twin. Do you have videos that are educational about these aspects which would help a beginner like me to understand the basics?
Ok. Why did the psychedelic sky of the gaussian splat give me a panic attack? I'm going to call it a Splat Attack.
Imaging google earth with gaussian splatting.
Pity about the low resolution. I assume it's low resolution as that's how it appears on my computer, even at 4K. Very interesting though!
Yea, I accidentally screen captured at 1080, but rendered everything at 4K!
I don't know what is the third method more than I did before starting your video
This wasn't meant to be a technical video, but if you want to know more about the technicals behind 3DGS, this is a good one: czcams.com/video/HVv_IQKlafQ/video.html
That's a rock
Working with photogrametry for 15 years or more to. And i tested Nerf as soon as they got out, and all nice and cool untill you give them uniform reflective surfaces. Photogrametry and Nerfs do not like uniform glossy surfaces. GS have no problm, they even get transparency,. Is an insane technology, far from perfect, have very limited use case, but i can see already GS being perfect solution for then next generation of realistic 3D google maps. Are very small and a lot easyer to do then NerFs and can be used with night shots to.
I wish the source videos were higher quality... Especially the Nerf part is painful to watch
Ah yeah, I see that now. Unfortunately I screencaptured at 1080 (my monitor's max) and then upscaled to match the NeRF/Gaussian videos (4k), and had my premiere preview set to 1/8 res. Whoops. I'll fix it for future ones - thanks for pointing it out!
I'd say, if your using reflective or translucent surfaces,i 100% recommend nerfs instead of photogrammatry, cuz it does it somuch better
I made another video explicitly comparing the two- the photogrammetry model actually turned out pretty well.
Where is the panorama sphere texture coming from that you used in the end?
It’s a panorama I took using a UAV directly above church rock.
It's me or from 8:30 and a half the quality of video seems 480?
I can appreciate the quality correctly because seems a 480p video rescalated.
Aside of that the video is very interesting.
It would be nice if in a few years the Gaussian splatting has a way to erase things you don't want, or correct blurry parts of a scene. Gaussian to mesh would be the holy grail.
It's possible (albeit somewhat crudely) to edit the Gaussian cloud now, in Unity.
Great video, thank you. Could you advise me how to scan an object very sharply for gaussian splatting? As it happens, I can't get it sharp. The letters are not clear. I own an android (s22 samsung, nikon z50, insta 360 x3). Which one would be the most accurate? As I am using the texture for training purpose. Thanks in advance. Greetings, Sebas
The Nikon z50 would likely be the best (physical shutter + megapixels), although of course it depends on what lens you are using. In my experience, I get the best results using a high-resolution mirrorless camera (compared to an iphone or action camera). Of course - more data = longer processing times, so there is always a trade off or compromise.
Hello Matthew, thank you very much for your response. I assume I need to create an mp4 file if I want to scan a gaussian splatter. What are the settings? I am really a novice in this. Previously I was doing an fps 30. Is it convenient to set fps60? Thanks in advance.
In my experience, still images (photographs) work much better than video! Follow good photogrammetric practice for capture, and then process as a 3DGS.
which would work best for a house with trees all around? and is there a way to see past the trees in any way? or would i need to fly my drone down into and past the tree line?
I would use photogrammetry, then edit out/delete the trees
so go in front of the trees to get the house? drone whole area and then camera to move into more detailed spots? @@MatthewBrennan
Yes, you can also combine the drone data with photos taken by hand using a high quality mobile phone or still camera
@@MatthewBrennan awesome thank youf or the help.
hey :) question regarding the 3Dmodel in the sphere. did u shoot the panorama exactly on top and center of the poi (the rock)? and is there a tutorial how to do the hole 3D process in blender? thx, kind regards sebastian
Yes, directly over it. I don’t know if there’s a tutorial…I had the idea years ago as a way of “faking” a more extensive scene (see some of ‘digital Hadrians villa’ project videos). Basically you just create a sphere/geo-sphere, invert the normals, and apply a spherically mapped panorama. Hope that helps!
thx for your response and infos! have tried it right now on one of my models and it was a success :) @@MatthewBrennan
@@klangr4usch Awesome! Send me a link if you post it on sketchfab!
It should be no problem to import the photogrammetry model into blender, import a world texture (photosphere) and render a similar camera path.
What do you thing is currently the best technology to scan faces to create highly accurate 3d files? Which are also future proof in terms of working with them? The priority is accuracy and feasibility for a non-professional. If we are limiting to cost to maybe 5 to 10 grand? Photogrammetry, a 3D scanner like from Einscan, NeRF or anything else? Do you know of any software that could benefit from having more than 1 camera at a time for photogrammetry? When I look at professional studios, they have like a hundred cameras surrounding a person that gets scanned.
Photogrammetry would fit the bill, but you'd need a multi-camera rig. I've scanned a live subject (just the bust - shoulders + head) with a single camera, but it required quite a bit of cleanup in 3D sculpting software. A calibrated high-resolution, multi-camera solution would be the way to go.
great video. except that the definition is not good even in 2160, it appeared blurry ?
Matthew, may i know what is your computer specification when creating this?
The photogrammetry and NeRF were processed on a desktop computer with a Ryzen 9 3900X CPU + 4070ti GPU. The Gaussian Splatting was processed with an nVidia A100 GPU.
Hi, you said we can export camera positions from 1st method and import it into NeRF to reduce amount of work for neural network and speed up a process. Don't you know how to do this with MeshRoom and colmap (or instant ngp)? Recently I was trying to make a NeRF with those tools but found that even after 2 days of work it was still processing the data (which are around 1200 frames from video).
I use metashape instead of COLMAP, because COLMAP is very slow and gives subpar results (it’s open source though, which is nice).
@@MatthewBrennan I'll check this as well, thanks.
Hello, Brennnan, May I ask you how long did the Gaussian Splatting method reconstruction takes you to get that church rock? I used a RTX 3080 machine and it will take me about 24 hours to reconstruct it(I start reconstruction last night, and it is still running now finished 45%). I dont know if it is OK......
It takes about 15 minutes using a 40GB a100… sounds like you could be running out of vram and instead of crashing it’s just hanging. Shouldn’t take that long…
@@MatthewBrennan Yes. I changed a 3090 machine and now it is less in one hour
Man is that a rocket fan?
GPU goes brrrrrzzzzz
I hereby dub this rock...The 890 JUMP 😉
How exactly do I make the sphere itself ? Can you make a step by step please. I do lots of photogrammetry models but I laways fail to add some background
I'll make a quick video outlining it, but the gist is: place your 3D model at the center of a sphere geometry created in blender, for example (you may have to invert the normals so that the polygons are "facing" inwards). Then use a spherical UV-map to apply your 360-image.
@@MatthewBrennan The "create sphere part" is not a problem I donßt know how to make a 360 photo or to project it. I use CAD software and Agisoft Metashape.
@@sierraecho884 You need to take a panoramic photograph at/on location... Or you can find a spherical photograph online (google street view). Equirectangular images are typically 2:1 aspect ratio. These will easily UV map using a spherical projection onto sphere geometry.
Good video, but I have some tips. 1: Get a capture card so your system usage doesn't effect quality, it's hard to tell which method is better when the resource
intensive ones are in 240p. And 2: either write a script or, if you want to maintain your natural manner of speaking, do a dry run first. Go though all your talking points once, then immediately start recording and do it again. Should help you move along a little faster and keep your pauses and "ums" to a minimum.
Why the video is blurred at 1440p?
does anybody know if nerf or gaussian spalatting work with 360 images?
It will, but you need to split them into overlapping frames (8-14 per spherical image).
what is the name of the music used at 14:30?
What dhrone and camera you used? :)
DJI Mavic 2
4k60fps video. But you recorded your Desktop in 720p or what?
yep. Oops. I've fixed it in subsequent videos.
you can make a camera path around photogrammetry just as easily though, no?
Yes- but you won’t have the sky/context. The photogrammetry data is essentially limited to just the object of interest. I think it’s illustrated slightly better in some of my other videos- 3DGS is great for capturing “unbounded scenes”, whereas photogrammetry is better suited to discrete objects (where you don’t need context).
@@MatthewBrennan the sky is captured very bad though
The capture of the SW is really blurry 😞
15 years ago, you say?
Were you a user of the original Photosynth desktop app?
Yes, and 123D Catch, original photoscan back in 2010, etc…
Could you release the generated models in this video? I’m interested to run my own model for comparisons
There's a google drive link to the Church Rock dataset in the description
@@MatthewBrennan Found it! cheers
I write papers related to NeRF and published them in CVPR. This field is always hungry for new datasets and if you're keen to release some scenes (say around 5-8 real scenes), do let me know! I'll be happy to collaborate with you to publish these datasets for the benefit of the research community!
So basically , photogrammetry is still king when it comes to realtime applications (games) as it holds the detail even up close and you can retopo it later, and things like gausiang splatting are currently good for video production/vfx, correct ?
Yep, essentially :)
@@MatthewBrennan In that case subscribing as I don't want to be left behind on the info XD (and difficult to keep track on everything) . I suspect that when a good and quality method for converting gaussian splatting to 3d mesh comes out, I will see the news here to keep up to date .
Cheers mate !
the main problem with GS its that i cant work with the model,,, i have no geometry to work with... at the end of the day y have a geometro model with a texture that i can modify and use in a 3d enviroment i create...
Yep - see this video for a discussion of Neural Surface Reconstruction: czcams.com/video/qFkCGvscsMQ/video.html
I don't think NeRFs or GS will replace meshes anytime soon, but I do like Gaussian Splat point clouds for video rendering: czcams.com/video/Mi27jpUC5nU/video.html
19:32 Music name please ?
This video is rendered in 4K, yet the original video quality is more like potato 480p
Yep. I rendered the NeRF/3DGS at 4k, without realizing my screen recordings were set to 1080. In premiere I had the video previews at 1/8 quality, so I didn't notice the blurriness. Lesson learned! I fixed the problem in subsequent uploads :(
am i the only one who thought the image at the beginning was a rock that was only like a meter long