Jerome Revaud @JeromeRevaud Twitter profile

Last Seen Profiles

@gemmanofficial

@MontemayorAnton

@BrendanBoylan

@BitesOfTime

@alkachka1

@TheRealYLT

@he11680

@AdekunleRoland5

@bswoffs

@FrankMozzicato

@EduardKukan

@KeskinBayindir

@weiz44

@insideloc

@kmtv

@freakingkiley

@Akadem_Campus

@jyunyank

@Kayne_Seys

@alper_tas

@the_original_b_

@klejdijd

@tanakabays

@RBesong

@SergioMicco

@sked_trojka

@abdusalamsalah

@macemasters

@SGolshaeian

@EscapistMag

@abhigyan

@skates_jaden

@CourtlynnSteele

@iam_josephsage

@anudeepd7

@inviolableblanc

Jerome Revaud

@JeromeRevaud

3 months

Another convenient usage of DUSt3R (kind of anecdotal): While looking for a rental apartment on airbnb for my vacations, i noticed I had difficulty grasping the layout and space of the apartment based on photos alone. Solution: put all photos in DUSt3R and voila :)

15

82

496

Jerome Revaud

@JeromeRevaud

3 months

An example of how DUSt3R can do "impossible matching": given two images without any shared visual content (my office, obviously never seen at training), it can output an accurate reconstruction (no intrinsics, no poses!) in seconds

18

109

459

Jerome Revaud

@JeromeRevaud

3 months

GitHub - naver/dust3r: DUSt3R: Geometric 3D Vision Made Easy

DUSt3R: Geometric 3D Vision Made Easy. Contribute to naver/dust3r development by creating an account on GitHub.

github.com

4

45

286

Jerome Revaud

@JeromeRevaud

5 months

We believe we did a breakthrough in Geometric 3D Vision: meet DUSt3R, an all-in-one 3D Reconstruction method

13

55

267

Jerome Revaud

@JeromeRevaud

1 month

our #DUSt3R is not even a spotlight, it's a poster. Even though we had excellent reviews... I don't get it

Kosta Derpanis

@CSProfKGD

1 month

#CVPR2024 presentation decisions are now posted on OpenReview.

1

4

44

30

3

130

Jerome Revaud

@JeromeRevaud

2 months

Some more impressive 3D reconstruction examples from #DUSt3R kindly prepared by our boss ;-) @martinhu @naverlabseurope

Vincent Leroy

@Vinc3nt_Leroy

2 months

Truly amazed by the response of the community to DUSt3R, thanks a lot everyone! The repo is on fire 🔥🔥 We are busy RN with the deadline but rest assured we will handle the opened issues ASAP!

0

7

68

2

20

119

Jerome Revaud

@JeromeRevaud

1 month

UniDepth is a super interesting approach for jointly predicting depth and camera intrinsics. Conceptually, it's very similar to #DUSt3R , except for the second view and the representation of space coordinates. We tried to use their (azimuth,elevation,log_z) for #DUSt3R

Ryohei Sasaki

@rsasaki0109

1 month

UniDepth: Universal Monocular Metric Depth Estimation Zero-Shot Visualization YouTube (The Office - Parkour)

1

33

261

2

15

111

Jerome Revaud

@JeromeRevaud

24 days

Coming back from London, after being kindly invited for a talk about DUSt3R + MASt3R in the Dyson lab of @AjdDavison at Imperial college and in the 3DV team of @LourdesAgapito at UCL. Had insightful discussions with all lab’s members, this was truly a great time spent there ❤️

5

2

84

Jerome Revaud

@JeromeRevaud

23 days

@WayneINR @AjdDavison @LourdesAgapito Here are the slides of my recent invited talk about CroCo + DUSt3R + MASt3R Note: It's a pdf, i know the videos won't work but there's really nothing hitherto unseen. I can point to where to find each individual video upon request.

5

11

83

Jerome Revaud

@JeromeRevaud

2 months

Sth I find fascinating with DUSt3R is its uncanny ability to perform auto-calibration... even with single images. Is it precise? Quite so, as you can see below. This challenges the prevailing notion that single-image calibration is ill-posed (ambiguous & under-constrained)!

Jack Naylor

@nackjaylor

3 months

Sometimes you see a paper that just knocks your socks off in terms of coolness - this is one of those papers.

0

5

21

2

7

66

Jerome Revaud

@JeromeRevaud

3 months

done, DUSt3R now first of the leaderboard!

Eric Brachmann

@eric_brachmann

3 months

@JeromeRevaud @ducha_aiki @Parskatt Nice! Even if it's not in the paper, it cannot hurt to submit your best entry to the leaderboard to mark your territory 😉 You did the bulk of work anyway already.

1

0

5

3

4

58

Jerome Revaud

@JeromeRevaud

2 months

Nice! Note that DUSt3R is a very data-driven approach, and it has never seen any human bodies during training. More generally, it kind-of struggles with situations it has never scenes. Still, no so bad :)

cocktail peanut

@cocktailpeanut

3 months

Instant Video-to-3D with DUST3R Dust3r generates a whole 3D scene from just a couple of images. What if it could: 1. Accept a VIDEO 2. Extract the video frames 3. Turn them into 3D? So added a Gradio Video Component. Here's me generating 3D from a video cc: @naverlabseurope

6

50

279

2

7

49

Jerome Revaud

@JeromeRevaud

2 months

So InstantSplat obliterates all concurrent pose-free methods in terms of recovered pose poses and rendering quality. And 50~200x faster at that. Wow! Congrats @WayneINR @yuewang314 👏

MrNeRF

@janusch_patas

2 months

Tl:dr: some new black magic! "InstantSplat: Unbounded Sparse-view Pose-free Gaussian Splatting in 40 Seconds" Paper: not yet Project: Cc @JeromeRevaud they use DUSt3R Method 1 I 2 ⬇️

5

31

245

2

1

48

Jerome Revaud

@JeromeRevaud

3 months

For reference, here is how my office truly looks like

3

2

27

Jerome Revaud

@JeromeRevaud

1 month

@arankomatsuzaki Interesting how they excluded CroCo (NeurIPS’22), a visual foundation model for 3D vision which is *specifically designed* to remedy the lack of multi-view awareness. CroCo is powering #DUSt3R btw :)

Weinzaepfel Philippe

@WeinzaepfelP

11 months

Happy to present CroCo 🐊, our NeurIPS'22 paper on self-supervised learning tailored to geometric tasks by Cross-view Completion, during the 3DMV workshop at #CVPR2023 . Cross-View Completion extends Masked Image Modeling (MIM), eg MAE, with a second view of the same scene. 👇 1/8

1

20

3

0

26

Jerome Revaud

@JeromeRevaud

5 months

Existing image-based 3D reconstruction methods (e.g. COLMAP) rely on a complex pipelines with many handcrafted steps. In contrast, DUSt3R's main component is a deep network that directly outputs a 3D model from 2 input images (no pose, no calibration required)

2

1

26

Jerome Revaud

@JeromeRevaud

5 months

@ducha_aiki More examples, videos and demonstrations at

1

6

25

Jerome Revaud

@JeromeRevaud

5 months

The key insight is to output "pointmaps" = dense 2D-3D onoe-to-one mappings, that live in the same 3D coordinate space. From this output, we can extract all scene and camera parameters effortlessly. When there's more than 2 images, we align all pointmaps in a common coord frame

1

21

Jerome Revaud

@JeromeRevaud

3 months

@ddetone The answer is simple: witchcraft🪄🧙‍♀️

Chris Offner

@chrisoffner3d

3 months

@riverakid1 @JeromeRevaud @bchidlovskii @Vinc3nt_Leroy Witchcraft. 😄 I just ran it on my own images (after changing all mentions of CUDA to [device = "mps" if () else "cpu"]) and wow, this is super cool! 🤩

1

6

32

1

18

Jerome Revaud

@JeromeRevaud

5 months

DUSt3R is not only simple and fast (inference takes 40ms per image pair on H100 GPU, global alignment for >2 images takes a few seconds), it also outperforms the state of the art in monocular AND multiview depth estimation, and relative pose estimation.

1

0

16

Jerome Revaud

@JeromeRevaud

3 months

@janusch_patas @ssh4net I can understand if you look at it purely from an application-driven viewpoint. Clearly not useable yet for realworld use-cases. For me, it's a revolution because that's the first time 3D reconstruction is done so straightforwardly, and for the first time, without calibration 😎

2

0

16

Jerome Revaud

@JeromeRevaud

1 month

Fixed! RoMa is now under BSD license

GitHub - naver/roma: RoMa: A lightweight library to deal with 3D rotations in PyTorch.

RoMa: A lightweight library to deal with 3D rotations in PyTorch. - naver/roma

github.com

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 months

@romain_bregier @JeromeRevaud @Parskatt Oh, I see the catch now: non-commercial. Any reason why? I mean, I can see a reason for the new methods, like R2D2, etc, but for the basic 3d geometry code? Anyway, that's better for kornia then ;)

1

0

2

3

15

Jerome Revaud

@JeromeRevaud

5 months

Surprisingly, it can also establish "impossible" pixel correspondences between opposite viewpoints which share practically no visual content! (examples not cherry-picked) @ducha_aiki

2

1

15

Jerome Revaud

@JeromeRevaud

3 months

It's pronounced "duster"! lol

BSLIVE Pinokio Dust3R to turn 2D into 3D Mesh

#Dust3R #pointcloud In this episode, I am showing you a way to convert 2D photo or multiple 2D photo references into 3D mesh.This runs really fast on local m...

www.youtube.com

0

3

14

Jerome Revaud

@JeromeRevaud

2 months

@janusch_patas @WayneINR @PMel3D on our side, we are planning to make dust3r better than COLMAP in every respect

1

0

13

Jerome Revaud

@JeromeRevaud

1 month

@eric_brachmann Sorry the opportunity was too good to miss, so here we go with DUSt3R to the rescue with impossible pairs😉

1

0

13

Jerome Revaud

@JeromeRevaud

17 days

@jianyuan_wang @janusch_patas 😜

GitHub - nerlfield/wild-gaussian-splatting: DUSt3R + Gaussian Splatting

DUSt3R + Gaussian Splatting. Contribute to nerlfield/wild-gaussian-splatting development by creating an account on GitHub.

github.com

3

13

Jerome Revaud

@JeromeRevaud

3 months

@janusch_patas @ssh4net (SfM like COLMAP works without calibration, but it's a quite complex pipeline). Hopefully one day we'll beat COLMAP in term of precision too :)

1

0

12

Jerome Revaud

@JeromeRevaud

24 days

Many thanks again to @AjdDavison , @LourdesAgapito and Andrea Vedaldi @ VGG oxford, also thanks to @makezur for taking care of the organisation, and to all students who took the time to make a demo! @HideMatsu82 It was very impressive, really feeling a vibe of convergence in #3dv

1

0

12

Jerome Revaud

@JeromeRevaud

23 days

@vincesitzmann oh after ACE-Zero another deep-SfM method! That's exciting to be part of this race to beat COLMAP :)

1

0

11

Jerome Revaud

@JeromeRevaud

5 months

@ducha_aiki Work done with Shuzhe Wang, Vincent Leroy, Yohann Cabon, @bchidlovskii at @naverlabseurope cc @WeinzaepfelP @skamalas

1

11

Jerome Revaud

@JeromeRevaud

5 months

@ducha_aiki @chriswolfvision @naverlabseurope didn't have time yet... We plan on submitting more experimental results soon (e.g. visual localisation).

0

10

Jerome Revaud

@JeromeRevaud

1 month

@WayneINR When there's that many details, it's clear that the resolution of #DUSt3R is not sufficient to capture them. As a result, the pointmap initialization is faulty on many small parts. Still, most of the scene is ok... I feel we're almost there, just need to train a better DUSt3R

2

0

9

Jerome Revaud

@JeromeRevaud

3 months

I guess the output is not so bad for a "no-code no-calibration" type-of-approach 😉

GitHub - naver/dust3r: DUSt3R: Geometric 3D Vision Made Easy

DUSt3R: Geometric 3D Vision Made Easy. Contribute to naver/dust3r development by creating an account on GitHub.

github.com

camenduru

@camenduru

3 months

@riverakid1 @Vinc3nt_Leroy @bchidlovskii2 @JeromeRevaud 3D modeling with 2D vs 3D reference 🤯

0

7

2

1

9

Jerome Revaud

@JeromeRevaud

1 month

oh wow

Zhiwen(Aaron) Fan

@WayneINR

1 month

Exciting update: Our latest results using THREE training views! Despite unknown poses and intrinsics, #instantsplat achieve this in under 20 seconds— even faster under sparser views.

10

12

137

0

1

9

Jerome Revaud

@JeromeRevaud

3 months

@ducha_aiki Agreed While DUSt3R is impressive with extreme viewpoint changes, it's still relatively sensitive to "long-term" changes (e.g. summer vs winter), which dominates the WxBS benchmark. A robust matcher like Roma based on Dinov2 makes more sense @Parskatt

1

0

8

Jerome Revaud

@JeromeRevaud

3 months

@janusch_patas Note: DUSt3R also has a problem with the sky. It predicts confidently some finite distance. Basically there was never any annotation to tell "this is infinite depth" We're working to fix it

1

0

8

Jerome Revaud

@JeromeRevaud

3 months

@eric_brachmann @ducha_aiki @Parskatt Yes we tried that... It's a pity we didn't have time to do it for the CVPR submission

3

0

8

Jerome Revaud

@JeromeRevaud

1 month

@eric_brachmann Yes that’s definitely a much more scalable approach than current pairwise matcher (including DUSt3R)

1

0

8

Jerome Revaud

@JeromeRevaud

2 months

@janusch_patas The results are in plain sight on a public leaderboard… now waiting for someone to notice 😛

2

0

7

Jerome Revaud

@JeromeRevaud

3 months

@jasonyzhang2 Nice paper, congratulations! Note that we @naverlabseurope pushed the idea a bit further with DUST3R

Jerome Revaud

@JeromeRevaud

5 months

We believe we did a breakthrough in Geometric 3D Vision: meet DUSt3R, an all-in-one 3D Reconstruction method

13

55

267

1

7

Jerome Revaud

@JeromeRevaud

22 days

I honestly find this concerning. This essentially means that in the hidden preprompt, the LLM is asked to deny at all cost that it has access to private info, even though it clearly has. Basically someone at meta decided « ok we’re gonna lie to the face of our users »

Jerrod Lew

@OnwardsProject

23 days

This is so creepy from Meta AI. It clearly can use my location but doesn’t admit to knowing this and instead completely lies about it.

536

2K

25K

0

7

Jerome Revaud

@JeromeRevaud

1 month

@habib__slim oh i see. Now i understand😆

0

6

Jerome Revaud

@JeromeRevaud

3 months

@Rafael_L_Spring something like 20ms on a A100 GPU?

0

6

Jerome Revaud

@JeromeRevaud

2 months

@ducha_aiki @LingjieLiu1 @KostasPenn Looks great. I really like this idea of optimizing a "canonical 3D space". It drops many constraints from the hard 4D (3D+t) reconstruction problem

0

4

Jerome Revaud

@JeromeRevaud

24 days

@eric_brachmann That’s pretty impressive. Awesome work, congrats! Also really happy you compared to the DUSr3R baseline :)

1

0

6

Jerome Revaud

@JeromeRevaud

24 days

@WayneINR @AjdDavison @LourdesAgapito Ill post the slides today

2

0

6

Jerome Revaud

@JeromeRevaud

4 months

@nburrus @luxonis Here is foretaste. DUSt3R outputs this quasi-instantaneously (<100ms). No camera parameters, no calibration file, just 2 pngs given as input! :)

1

0

6

Jerome Revaud

@JeromeRevaud

20 days

@olliecrosen @AjdDavison @LourdesAgapito There's two things: - first, CroCo is bad at semantic tasks, as mentioned above. - second, we also tried to use a DINOv2 encoder (instead of CroCo) in DUSt3R and MASt3R, either frozen or finetuned, and that was kind of a failure, but we only evaluated for geometric tasks.

2

0

5

Jerome Revaud

@JeromeRevaud

2 months

@eric_brachmann @ducha_aiki Something tells me it might come sooner than you think.

1

0

5

Jerome Revaud

@JeromeRevaud

3 months

@ducha_aiki @Parskatt @jasonyzhang2 @amyxlase @RamananDeva @shubhtuls Working on it ... sometimes the results are relatively impressive, sometimes it totally fails (especially the RGB-IR pairs! wtf is this?) Anyway, DUSt3R code and model weights will be released this week.

2

0

5

Jerome Revaud

@JeromeRevaud

5 months

@ducha_aiki @bchidlovskii @naverlabseurope @WeinzaepfelP @skamalas cc @chriswolfvision

1

0

5

Jerome Revaud

@JeromeRevaud

3 months

@janusch_patas @Atehortuajf @_linus_franke Totally agree. Let's say that directly outputting 3D Gaussians with DUSt3R is on my todo list :)

1

0

5

Jerome Revaud

@JeromeRevaud

3 months

@janusch_patas @Atehortuajf We did not try yet to optimize 3D Gaussians from the output of DUSt3R. Most likely it will fail, because the camera and reconstruction are somewhat sloppy, and 3DGS need a really precise geometry at initialization. We're working on making it a lot more precise...

1

0

5

Jerome Revaud

@JeromeRevaud

2 months

@chriswolfvision he deserves some extra-reward

0

5

Jerome Revaud

@JeromeRevaud

2 months

@Parskatt oh that looks very interesting indeed

0

5

Jerome Revaud

@JeromeRevaud

11 days

@ducha_aiki Not for long. Wait for it

1

0

5

Jerome Revaud

@JeromeRevaud

4 months

@ducha_aiki @jonas_eschmann @chriswolfvision @danish_nazir1 @XXXLutzDE DUSt3R gives a focal of 1506.52

0

5

Jerome Revaud

@JeromeRevaud

2 months

@ptrck_brgr @WayneINR Also Colmap-free-3DGS requires the camera intrinsics. How do you get the intrinsics? ... COLMAP Honestly, I don't know of a single deep-learning-based 3D vision method that works without the intrinsics... except for #DUSt3R ofc :D

1

0

5

Jerome Revaud

@JeromeRevaud

2 months

(the castle pic (second image pair) shows our actual Lab! )

1

0

5

Jerome Revaud

@JeromeRevaud

2 months

@camenduru please try to put the resulting images in #DUSt3R and voila you should have a nice 3D model

Vincent Leroy

@Vinc3nt_Leroy

3 months

Been waiting for the code release of DUSt3R? Wait no more, there it is! Reminiscent of RayDiffusion (), we present a well behaved variant to Bundle Adjustment, this time with pointmaps.

3

86

471

0

5

Jerome Revaud

@JeromeRevaud

2 months

@ducha_aiki @Vinc3nt_Leroy By the way, we’re also working to release the full DUST3R training set

1

0

5

Jerome Revaud

@JeromeRevaud

2 months

@yuewang314 No need to thank, I’m just really happy and excited to see DUST3R contributing to the 3DV community

0

5

Jerome Revaud

@JeromeRevaud

1 month

@Michael_J_Black Well language is already expressed in a pure semantic manner (by definition) so that effectively removes a big part of the perception problem that we have with images. Also, massive text data are available. Much more than images, in terms of « semantic quantity » imo

1

0

5

Jerome Revaud

@JeromeRevaud

2 months

lol neat

Bernhard Egger

@VisionBernie

2 months

It’s time to revolutionize authorship order in publications across fields #ScienceTwitter ! Our work “AMOR: Ambiguous Authorship Order” resolves author ordering problems once and for all! We just implement animated author lists – that's it @SIGBOVIK : 1/n

8

12

98

0

5

Jerome Revaud

@JeromeRevaud

11 days

@ducha_aiki Exactly haha 👍

1

0

5

Jerome Revaud

@JeromeRevaud

24 days

@WayneINR @AjdDavison @LourdesAgapito Arxiv paper coming soon (we’re polishing some details)

2

0

5

Jerome Revaud

@JeromeRevaud

1 month

@ducha_aiki @Parskatt @SattlerTorsten We tried different flavors of 3DGS (like SuGaR) to convert small videos to dense 3D point-clouds. Too many floaters.

0

4

Jerome Revaud

@JeromeRevaud

3 months

@Parskatt @ducha_aiki @jasonyzhang2 @amyxlase @RamananDeva @shubhtuls In other words, the representation space is ill-posed. Therefore this kind of approach can only be successful in constrained settings, e.g. with object-centric datasets where the focal is small and roughly constant. Modelling rays rather than extrinsic is already better posed!

1

0

4

Jerome Revaud

@JeromeRevaud

23 days

@jianyuan_wang I like the very honest « limitations » section at the end of appendix :)

0

4

Jerome Revaud

@JeromeRevaud

2 months

@Parskatt True, in many cases, the problem is clearly ill-posed. One clue that we suspect helps DUSt3R are right angles. In most images you can find right angles, even for outdoor images in the wild: trees are verticals on a horizontal ground. That's enough to solve for the focal

1

0

4

Jerome Revaud

@JeromeRevaud

1 month

but contrary to the claims in UniDepth that decoupling rays/depth is important, we observe a decrease in performance. Here the regression loss on Habitat and MegaDepth. Not sure why... Note: log(z) diverges (the 2nd view can be behind the cam), so we added log(distance_to_origin)

0

4

Jerome Revaud

@JeromeRevaud

1 month

@jfbrly Thanks. We are!

0

4

Jerome Revaud

@JeromeRevaud

1 month

@Michael_J_Black @AjdDavison Thankfully that's what our plan always was... starting from #CroCo in 2022😀 Thanks for the tip! 🙏

0

4

Jerome Revaud

@JeromeRevaud

1 month

@eric_dexheimer @AjdDavison Great work, congrats (and thanks for releasing the code so fast!)

2

0

4

Jerome Revaud

@JeromeRevaud

5 months

@mayfer Yes, there were similar attempts in the past, like DeMoN and DeepV2D, but these approaches still require camera intrinsics. Performing 3D reconstruction without knowing the intrinsics was deemed infeasible (unless explicit BA), well we show it's not.

1

0

4

Jerome Revaud

@JeromeRevaud

1 month

@eric_brachmann @NianticLabs @axelbarrosotw @scriptide @viprad Congrats, that's a really cool paper further pushing the boundaries of image matching! Did you try optimizing the pose accuracy instead of VCRE?

1

0

4

Jerome Revaud

@JeromeRevaud

3 months

@ducha_aiki @Parskatt also, DUSt3R seems more robust with indoor data. I think there might be a training imbalance in that regard... we need more public outdoor datasets :)

1

0

3

Jerome Revaud

@JeromeRevaud

3 months

@FangQ Thanks 🙏 but note DUST3R has never seen human bodies during its entire training. Since Dust3r is a very data-driven approach by nature, reconstruction results on human bodies can be relatively sloppy and disappointing.

1

0

4

Jerome Revaud

@JeromeRevaud

23 days

@AjdDavison Same goes in the other direction. I wasn’t expecting that many excellent, really top-notch papers and demos. Super interesting research overall! Also really liked the atmosphere in your lab, for sure one of the best lab i visited

1

0

4

Jerome Revaud

@JeromeRevaud

4 months

@chriswolfvision @danish_nazir1 @ducha_aiki Same story with DUSt3R, which challenges the prevailing notion that camera intrinsics are indispensable for 3D reconstruction and a whole bunch of 3DV tasks (unless using costly bundle adjustment). Yet, here we are😎

Jerome Revaud

@JeromeRevaud

5 months

We believe we did a breakthrough in Geometric 3D Vision: meet DUSt3R, an all-in-one 3D Reconstruction method

13

55

267

0

4

Jerome Revaud

@JeromeRevaud

1 month

@dreamingtulpa haha. It's trying

0

4

Jerome Revaud

@JeromeRevaud

1 month

@chriswolfvision « In the spirit of friendly academic shitposting » 🤣

1

0

4

Jerome Revaud

@JeromeRevaud

3 months

@georgiagkioxari Oh, ok! We actually borrowed this formulation from "Multi-task learning using uncertainty to weigh losses for scene geometry and semantics", CVPR'18 from Kendall et al. And thanks to you, I just notice we didn't properly cite their paper 😕

1

0

3

Jerome Revaud

@JeromeRevaud

2 months

@Parskatt @janusch_patas Who knows ^^

1

0

4

Jerome Revaud

@JeromeRevaud

1 month

@WayneINR I was waiting for this result... The first video that you posted was static, but this is even more impressive!

1

0

3

Jerome Revaud

@JeromeRevaud

3 months

@AndrewVoirol The only issue I see is that DUSt3R is not metric, so you'll have to find a way to properly scale the reconstructed 3D model

0

3

Jerome Revaud

@JeromeRevaud

4 months

@ducha_aiki HPatches is ... how to say politely... a bit outdated

0

2

Jerome Revaud

@JeromeRevaud

2 months

@JsonBasedman Same thing in Korea with Naver and Kakao (big IT competitors there) whereas they're actually close friends from the same school (they worked together at Samsung too)

0

3

Jerome Revaud

@JeromeRevaud

1 month

@hupobuboo @janusch_patas Great presentation, thank you for the coverage!

0

3

Jerome Revaud

@JeromeRevaud

3 months

@kysucix @chriswolfvision Thanks 😀 Kuddos to @riverakid1 @Vinc3nt_Leroy Yohann Cabon, and @bchidlovskii

0

3

Jerome Revaud

@JeromeRevaud

4 months

@ducha_aiki hahaha🤣

0

3

Jerome Revaud

@JeromeRevaud

1 month

@taiyasaki Thanks for letting us know 🙏

0

3

Jerome Revaud

@JeromeRevaud

2 months

@Parskatt In this example, for instance, easy-peasy: lots of right angles everywhere :)

Jerome Revaud

@JeromeRevaud

4 months

@ducha_aiki @jonas_eschmann @chriswolfvision @danish_nazir1 @XXXLutzDE DUSt3R gives a focal of 1506.52

0

5

0

3

Jerome Revaud

@JeromeRevaud

2 months

@Parskatt @chrisoffner3d Also, note that you can load and store .exr with opencv, which you most likely have installed already

1

0

2

Jerome Revaud

@JeromeRevaud

12 days

@erhartford @maximelabonne That’s very interesting if true

0

3

Jerome Revaud

@JeromeRevaud

2 months

@Poyonoz @WayneINR @yuewang314 yeah... but everything is too fast nowadays anyway

0

2

Jerome Revaud

@JeromeRevaud

28 days

@zhenjun_zhao @romain_bregier please check this and answer. It looks awfully similar to your 2021 paper

Deep Regression on Manifolds: A 3D Rotation Case Study

Many machine learning problems involve regressing variables on a non-Euclidean manifold -- e.g. a discrete probability distribution, or the 6D pose of an object. One way to tackle these problems...

arxiv.org

2

0

3

Jerome Revaud

@JeromeRevaud

23 days

@Parskatt @vincesitzmann @eric_brachmann haha literally yes

0

3

Jerome Revaud

@JeromeRevaud

2 months

@AjdDavison @HideMatsu82 @rmurai0610 that's really cool!

1

0

3