Bernhard Kerbl @Snosixtytwo Twitter profile | Pikagi

Pikagi

Bernhard Kerbl

@Snosixtytwo

1,366

Followers

60

Following

10

Media

107

Statuses

Visiting Researcher at Carnegie Mellon University

Pittsburgh, Pennsylvania

https://t.co/7EchdtAtMc

Joined September 2009

Don't wanna be here? Send us removal request.

Pinned Tweet

@Snosixtytwo

Bernhard Kerbl

7 days

📢 Train high-quality 3DGS models - but 5x faster ✨ with "Taming 3DGS"! The paper will be presented at SIGGRAPH Asia 2024, looking forward to a great time in Tokyo. Huge congrats to both Saswat @saswat_mallick and Rahul @__lankylad__ for this great academic collab.

Tweet media one

1

22

153

Last Seen Profiles

@0xCoha

@JackermanDev

@powerdy_CS

@morenofuerte12

@IndustryKE

@ogboss530

@NileetQte04

@SaisonFound

@prideofgyp67

@ArchCristao

@TshantWrdt92006

@bisho27

@RicoVegaXxx

@britneyambertv

@saudiunicorns

@SleepSinai

@mohmmaed2S

@dakuNSFW

@bilecikvaliligi

@ronenfield

@framele1082

@you73563

@Megsenmumdr

@asame_yuki

@LiloHafs

@erenbrave

@Therexvlr

@ramah_aqar

@yamao8o

@DrReneeMitchell

@WeblateOrg

@Lev_Ponomarev

@FormulaD

@rollingloudfest

@dwvSofie

@sanahmedd

@Snosixtytwo

Bernhard Kerbl

5 months

Happy to announce the results of our latest research, which takes 3D Gaussian Splatting to the next level: "A Hierarchical 3D Gaussian Representation for Real-Time Rendering of Very Large Datasets," which has been accepted at #SIGGRAPH2024 !🎉 Find it here:

64

494

3K

@Snosixtytwo

Bernhard Kerbl

1 year

I am delighted to say that today, we can present what has taken shape over the past 6 months at GraphDeco @inria_sophia . Point clouds, software rasterization, differentiable rendering, all working together to give you crisp, real-time rendering within minutes of training.

@GKopanas

George Kopanas

1 year

We will be in @siggraph 2023 with "3D Gaussian Splatting for Real-Time Radiance Field Rendering", have you ever seen radiance fields with 100+ FPS and MipNeRF360 quality? Check out our website here:

14

147

821

6

15

130

@Snosixtytwo

Bernhard Kerbl

5 months

🥳 Positively stoked about our 2 papers getting accepted at SIGGRAPH2024! Details coming soon. But this is probably a good time to announce that I am actively looking for faculty openings starting this fall 🥸 Always happy to learn about opportunities where I might be a good fit!

3

7

76

@Snosixtytwo

Bernhard Kerbl

4 months

Congratulations to the whole team for their awesome work on fixing temporal artifacts in 3DGS! This paves the way for light-weight, yet smooth 3DGS visualization. VR, anyone? Super happy to have been a part of this. If you want GPU done right, consider asking these guys😁

@rafourdl

Lukas Sebastian Radl

4 months

🚀 Excited to announce our latest research: “StopThePop: Sorted Gaussian Splatting for View-Consistent Real-Time Rendering”, which will appear in #SIGGRAPH2024 ! 🎉 Finally, fast and view-consistent rendering of 3D Gaussians, without popping artifacts!

6

26

180

0

10

69

@Snosixtytwo

Bernhard Kerbl

4 months

Congratulations to @JohannesUgb for receiving the best paper award at this year's EGPGV! "Fast Rendering of Parametric Objects on Modern GPUs" (link below), that's a great way to conclude your PhD🙂

Tweet media one

1

8

56

@Snosixtytwo

Bernhard Kerbl

5 months

We approach scalable 3DGS from the perspective of real-time graphics methodology. After training, a model is fitted with a hierarchical level-of-detail structure. We present answers to the 3 key challenges a good LOD solution should address: generation, selection, and switching.

1

2

50

@Snosixtytwo

Bernhard Kerbl

1 year

"3D Gaussian splatting is noice and all, but no web viewer and the file sizes are just too damn high" @jakub_c5y : "Hold my beer"

@m_schuetz

Markus Schütz

1 year

Awesome browser-based gaussian splat implementation by @jakub_c5y :

Tweet media one

Tweet media two

Tweet media three

7

50

281

4

4

39

@Snosixtytwo

Bernhard Kerbl

5 months

With LOD enabled on them, your GPU stores and renders only what it needs! This way, we can have real-time flythroughs and paths through radiance fields that just go on and on and on and on...

1

2

36

@Snosixtytwo

Bernhard Kerbl

5 months

We note that there has been some exciting concurrent work! Praise goes to the authors for successfully making their mark in such a fast-moving field: (and others that we might have missed?)

1

2

33

@Snosixtytwo

Bernhard Kerbl

5 months

We have developed this method using a collection of truly large-scale datasets, which we will release with our code. The key to training them is a divide-and-conquer scheme, enabling independent optimization and consolidation of chunks, each much bigger than standard 3DGS scenes.

1

0

32

@Snosixtytwo

Bernhard Kerbl

1 year

Preliminary deadline for PhD applications is in one week! :) Think point-based graphics are neat and ready to join us in Vienna? There's tons of stuff we still haven't explored! Check out … and apply for a doctorate with us (and @m_schuetz ) at TU Wien.

@GKopanas

George Kopanas

1 year

We will be in @siggraph 2023 with "3D Gaussian Splatting for Real-Time Radiance Field Rendering", have you ever seen radiance fields with 100+ FPS and MipNeRF360 quality? Check out our website here:

14

147

821

0

6

29

@Snosixtytwo

Bernhard Kerbl

1 year

Yesterday we released the code of our prototype on Github, we hope you like it! If you pulled immediately after release, please pull again (and don't forget submodules!!!), or re-download the real-time viewers for a performance bump.

Tweet card media

GitHub - graphdeco-inria/gaussian-splatting: Original reference implementation of "3D Gaussian...

Original reference implementation of "3D Gaussian Splatting for Real-Time Radiance Field Rendering" - graphdeco-inria/gaussian-splatting

@Snosixtytwo

Bernhard Kerbl

1 year

I am delighted to say that today, we can present what has taken shape over the past 6 months at GraphDeco @inria_sophia . Point clouds, software rasterization, differentiable rendering, all working together to give you crisp, real-time rendering within minutes of training.

6

15

130

2

1

28

@Snosixtytwo

Bernhard Kerbl

6 months

Gradient descent with @JonathonLuiten at @3DVconf

2

0

27

@Snosixtytwo

Bernhard Kerbl

5 months

Each member of our team was absolutely essential in making this happen. My congratulations and thanks go out to shared first author @AndreasMeuleman , @GKopanas , @alanvinx , Michael Wimmer, and @GDrettakis .

1

0

24

@Snosixtytwo

Bernhard Kerbl

5 months

@pereirarb1 We use a couple of processing steps to get rid of moving cars and humans. It's not 100% bulletproof, but pretty solid! We had some challenges with camera calibration indoors, but we will announce if we add more datasets.

0

0

4

@Snosixtytwo

Bernhard Kerbl

5 months

This research was funded by the ERC Advanced grant FUNGRAPH No 788065 () for @inria_sophia , and WWTF project ICT22-055: Instant Visualization and Interaction for Large Point Clouds at @TUWienCG .

0

2

22

@Snosixtytwo

Bernhard Kerbl

5 months

@xik_le Way too large for a regular GPU (numbers in paper). That's where the LOD aspect comes in.

0

0

1

@Snosixtytwo

Bernhard Kerbl

10 months

What an amazing push for real-time rendering! Sublime quality with a smart divide-and-conquer for seamless, high-quality novel views on your mobile phone!

@PeterHedman3

Peter Hedman

10 months

Check our work! Explore big scenes on your phone with zip-NeRF quality. Please enjoy the real-time demos and marvel at SMERF! Huge congratulations to @duck @chrisjreiser @PeterZhizhin @jfthibert @mariolucic_ @RSzeliski @jon_barron

1

3

37

1

0

16

@Snosixtytwo

Bernhard Kerbl

5 months

@nnnghh Very large! We have detailed per-scene stats in the paper if you're interested.

0

0

0

@Snosixtytwo

Bernhard Kerbl

2 years

Looking for graphics PhD students! The Rendering and Modeling group at Vienna University of Technology @tu_wien is now accepting applications for a full-time PhD to explore rendering, interaction, and learning solutions for massive point-based data sets:

1

1

14

@Snosixtytwo

Bernhard Kerbl

5 months

@WayneINR Definitely, as soon as we cleaned it up properly (June/July)

1

0

12

@Snosixtytwo

Bernhard Kerbl

1 year

@docmilanfar @siggraph @GKopanas @docmilanfar Extremely humbled by such a nice take on our paper! But there is definitely work left to do: NeRFs are way more compact, and degrade very gracefully. There is vast potential to bring those properties to Gaussian splatting too, and we hope to see this in future work!

0

0

12

@Snosixtytwo

Bernhard Kerbl

6 months

Congratulations to @PapantonakisP for the first awesome paper toward his PhD, with more to follow for sure! Pleasure to work with him on our shot at making 3DGS more compact and portable, to be presented at I3D this year😀

@PapantonakisP

PanagiotisP

6 months

With our work "Reducing the Memory Footprint of 3D Gaussian Splatting," a method that reduces the size of 3DGS from several hundreds to just a few tens of MBs, you now have more space available for additional scenes! For more, check out our project page

Tweet media one

2

42

214

0

0

13

@Snosixtytwo

Bernhard Kerbl

1 year

@montyburgess @nobbis @GKopanas My personal take on the points above: - meh... - yes - YES - yes, but that's not strictly a pro - yes - Theoretically yes - oh no. No no NO. Not yet :P

1

1

11

@Snosixtytwo

Bernhard Kerbl

3 months

😮

@taiyasaki

Andrea Tagliasacchi 🇨🇦

3 months

📢📢📢 Introducing "𝐒𝐩𝐨𝐭𝐋𝐞𝐬𝐬𝐒𝐩𝐥𝐚𝐭𝐬: Ignoring Distractors in 3D Gaussian Splatting" lead by @sabour_sara and @lily_goli TL;DR: exploit pre-trained features to "recognize" what should be ignored. Source code will be released in a few days.

4

16

139

1

0

12

@Snosixtytwo

Bernhard Kerbl

1 year

@docmilanfar Funny enough: PSY literally was a keynote speaker at 2015 CHI conference.

0

0

8

@Snosixtytwo

Bernhard Kerbl

1 year

Looking for a PhD project, willing to travel and think this project is cool? In this first version, there are tons of things we haven't even tried yet! Check out (server unresponsive rn, we'll fix it) and apply for a doctorate with us at TU Wien (Vienna)!

1

1

6

@Snosixtytwo

Bernhard Kerbl

2 years

I am very happy to share this work, authored by the extremely smart and ridiculously diligent @clementjbn during his @inria_sophia internship. His vision is for users to explore and extend it, so don't hesitate to try the interactive prototype!

@clementjbn

Clément Jambon

2 years

📢 I am happy to share our paper "NeRFshop: Interactive Editing of Neural Radiance Fields" 🎨, to be presented at I3D 2023! 🌐: 🎥: 1/6

8

96

397

1

1

6

@Snosixtytwo

Bernhard Kerbl

5 months

@THERE_IS_N0D4T4 Not yet! But that would be a great next step 🙂

0

0

4

@Snosixtytwo

Bernhard Kerbl

5 months

@rooevans Hi, we have some additional details on the project page/paper. We used helmets with 5/6 cameras, including a backward facing. We also show a few scenes where we turn around / walk around. Quality there will depend on how carefully it was captured.

1

0

4

@Snosixtytwo

Bernhard Kerbl

3 months

@ana_dodik > In the distance, you hear the faint laugh of Don Knuth, echoing from atop his 300 foot pile of error messages that are equally creative and unhelpful

1

0

5

@Snosixtytwo

Bernhard Kerbl

1 year

We invite you to check out the project's website with the full paper, videos and code (soon) here:

0

0

4

@Snosixtytwo

Bernhard Kerbl

7 months

@PMel3D 😄While this would have been fun, it would also be very wrong, because it was an extremely collaborative effort, and the main contribution is at least as much @GKopanas ' as it is mine.

0

0

4

@Snosixtytwo

Bernhard Kerbl

4 months

@janusch_patas We tried it on the scenes that can finish with < 80Gb VRAM. With the --eval flag properly enabled, we see much lower quality metrics. Test renderings look very close to the quality in the video. Ofc maybe we missed something!! But it looks like it might be test set pollution.

1

0

3

@Snosixtytwo

Bernhard Kerbl

1 year

This was an intense, formative, and fun project to work on, with a world-class team consisting of @GKopanas , Thomas Leimkühler and George Drettakis.

1

1

4

@Snosixtytwo

Bernhard Kerbl

1 year

@simesgreen @janusch_patas @8Infinite8 @GKopanas That being said, resolving this randomness is unlikely to give you desired results. Even if the same frame would yield identical Gaussians, it doesn't mean that a slightly different frame will have a similar "structure" of Gaussians. @JonathonLuiten has ideas!

1

0

4

@Snosixtytwo

Bernhard Kerbl

1 year

@seidtgeist @GKopanas Hi, we're doing our best to make the code readable and easy to build. It will be out either June or July, unfortunately, we can't be more precise right now. With the released paper code, an RTX 3090 should pretty much yield the numbers we report.

1

0

4

@Snosixtytwo

Bernhard Kerbl

5 months

@AndreWeinhold Excellent point. Our capture setup has some backward-facing cameras, so it is possible to turn around / walk around (project page videos contain examples). It depends a lot on how carefully it's captured, but for sure it looks best along the path.

0

0

3

@Snosixtytwo

Bernhard Kerbl

3 months

This is super-impressive stuff! 😮

@taiyasaki

Andrea Tagliasacchi 🇨🇦

3 months

Talk about to start! (Just after novotni) We are VERY excited to launch this paper ☺️☺️☺️ Few seats left (not many)

1

0

18

0

1

3

@Snosixtytwo

Bernhard Kerbl

1 year

@charshenton Super cool results, thanks a lot for the thread! For performance: usually with triangles, most are rendered opaque. Here, everything has a smooth falloff and some amount of transparency. I believe trying to do this with the conventional graphics pipeline wouldn't work that well.

1

0

3

@Snosixtytwo

Bernhard Kerbl

1 year

@alexcarliera Very cool! Quick tip: there's a viewing option other than Trackball called "Interpolate". It might just do exactly what you want for this.

1

0

3

@Snosixtytwo

Bernhard Kerbl

11 months

@SophieScarlette Still slaps

0

0

1

@Snosixtytwo

Bernhard Kerbl

2 years

Additional thanks go out to a great team of co-authors @GKopanas @StavrosDiol and Thomas Leimkühler, as well as George Drettakis, leader of the GraphDeco team, for seamless supervision. The research was enabled by the ERC Fungraph grant.

0

0

2

@Snosixtytwo

Bernhard Kerbl

1 year

@MattNiessner That's one way to curb global warming

0

0

2

@Snosixtytwo

Bernhard Kerbl

11 months

@buitengebieden Short kong

0

0

1

@Snosixtytwo

Bernhard Kerbl

1 year

@nobbis @montyburgess @GKopanas Oh, we are watching that closely😃 I think with some pushed bug fixes our PSNR is better now, but their savings in Gaussians are great for the truck scene. AND they beat us to our own code release, which is insane😆

1

0

2

@Snosixtytwo

Bernhard Kerbl

6 months

@IDEAS_NCBR @tu_wien Thank you for hosting! I just wanted to point out that I ran out of time in this talk and didn't get to the 3DGS funding slide, nor was I able to talk about the great work on "NeRFshop" with @clementjbn

1

0

2

@Snosixtytwo

Bernhard Kerbl

2 years

@SIGGRAPHAsia @wenzeljakob @SA2023Sydney PaperS. Plural. This is Wenzel you are talking about.

0

0

1

@Snosixtytwo

Bernhard Kerbl

6 months

@IDEAS_NCBR @tu_wien @clementjbn "3D Gaussian Splatting for Real-time Radiance Fields", authored by Georgios Kopanas & Bernhard Kerbl (INRIA Sophia-Antipolis), Thomas Leimkühler (MPII Saarbrücken) and George Drettakis (INRIA Sophia-Antipolis) Research was funded by the ERC Advanced Grant FUNGRAPH of GraphDeco

1

0

1

@Snosixtytwo

Bernhard Kerbl

1 year

@SebAaltonen @NOTimothyLottes Definitely overkill, mostly bc research focuses on prototypes. There's tons of opportunity there, we appreciate the suggestions :) But the SHs are converted to RGB once per point before binning, that stage rn is much cheaper than the actual blending that comes afterward.

1

0

1

@Snosixtytwo

Bernhard Kerbl

5 months

@ManbearpigAus That should be very feasible! I'll let @alanvinx answer with details if he finds the time!

1

0

1

@Snosixtytwo

Bernhard Kerbl

1 year

@nobbis @montyburgess @GKopanas Thanks so much, and yes, I totally agree with that. My notes on the benefits are for the method "as published". If there are things that can improve, we need to prove that first 🙂

0

0

1

@Snosixtytwo

Bernhard Kerbl

3 years

@McFunkypants "0 bytes free on drive C" sounds like a fantastic title for a song about the impending apocalypse.

0

0

1

@Snosixtytwo

Bernhard Kerbl

1 year

@SebAaltonen @NOTimothyLottes +cough+

2

0

1

@Snosixtytwo

Bernhard Kerbl

1 year

@SebAaltonen @NOTimothyLottes Thanks for this interesting discussion. We see that some pixel in a tile is usually affected even by the farthest splats, so afaict occlusion culling would not do much or need to be aggressive -> lower quality. Exploiting it without degradation probably implies different training

1

0

1

@Snosixtytwo

Bernhard Kerbl

1 year

@simesgreen @janusch_patas @8Infinite8 @GKopanas Yup, can confirm. Even if everything else (numpy, torch) have the randomization set to a known value, this will be a source of randomness. If the backward was designed differently, it could be avoided. However, I can't guarantee that e.g. Pytorch itself is 100% deterministic.

2

0

1

@Snosixtytwo

Bernhard Kerbl

2 years

@twominutepapers @twominutepapers for you in particular, "Science of Discworld" might be a good title, it is quite unique and outside the main storylines. But depending on your reading habits, the science aspects of it might be a bit simplistic for you.

1

0

1

@Snosixtytwo

Bernhard Kerbl

3 months

@GeneralMCNews Ppl: Saving the environment is boring, let science figure it out Science: Ppl: NO NOT LIKE THIS

0

0

1

@Snosixtytwo

Bernhard Kerbl

4 months

@RadioGenoa This looks like it's Kizomba, a dance of Angolan (South African) origin.

0

0

0

@Snosixtytwo

Bernhard Kerbl

5 months

@PMel3D @janusch_patas We tested it on smaller scenes (MipNerf360). The LOD does not improve detail but helps raise resource flexibility, e.g., if you need more FPS or VRAM.

0

0

1

@Snosixtytwo

Bernhard Kerbl

3 months

@EverythingKeanu Confirmation bias

0

0

0

@Snosixtytwo

Bernhard Kerbl

1 year

@makeshifted @GKopanas @siggraph It should definitely work with hardware rasterization once the depth order is taken care of. We had prototypes with a compute sorting pass based on depth, followed by hardware rasterization with blending enabled. Not sure about performance, wanted to try this again at some point.

0

0

1

@Snosixtytwo

Bernhard Kerbl

1 year

@TokyoWarfare @BrianKaris @inria_sophia Hi, we get some decent results from random points. Anything that is better than random makes the final result better. So sampled depth maps and other point scans can definitely be used to initialize it.

0

0

1

@Snosixtytwo

Bernhard Kerbl

1 year

@SebAaltonen @NOTimothyLottes Yes, we tried reducing it to 2 recently, made things noticeably worse.

1

0

1