Johan Edstedt @Parskatt Twitter profile

Last Seen Profiles

@shai_wininger

@KsiazekPawel

@YomiDenzel96

@OmerFreixa

@UChicagoMed

@ensemble43530

@symphony_221B

@dahafazlasicin

@loveyoustepan

@sepiascribe

@bluesbuzzblog

@shercrushh

@AnnePip16

@warabe_show

@jordanrcrook

@RRG_Vlahos

@stoplovemeplase

@aintgeraldine

@urbabydrew

@PolskaSuzuki

@MEATBRICKMOLLY

@mcivorclarke

@wotzitt

@cosit2024

@Optional43

@PeachSooda

@elianadivakarr

@tigermaskv2

@iKtOVT6Ejum9ZaY

@sara_gerardina

@lukeitn

@blinkcotech

@londonbluenews

@momos55

@HurtzDonut

@PaircUiCha0imh

Johan Edstedt

@Parskatt

20 days

Actually this is the future

5

21

216

Johan Edstedt

@Parskatt

5 months

4

15

193

Johan Edstedt

@Parskatt

9 months

DeDoDe is now on arxiv! 🥳

Johan Edstedt

@Parskatt

10 months

Say hi to DeDoDe 🎶! DeDoDe is a keypoint detector trained to detect 3D tracks. The reverse, DeDoDe, is a descriptor that matches the tracks. DeDoDe and DeDoDe are simple to train, and show great performance 📈 Code:

7

12

52

3

25

97

Johan Edstedt

@Parskatt

2 months

👀

4

7

95

Johan Edstedt

@Parskatt

3 months

@frncswllms

2

16

75

Johan Edstedt

@Parskatt

1 month

Finally managed to cite myself 100 times

2

0

54

Johan Edstedt

@Parskatt

10 months

Say hi to DeDoDe 🎶! DeDoDe is a keypoint detector trained to detect 3D tracks. The reverse, DeDoDe, is a descriptor that matches the tracks. DeDoDe and DeDoDe are simple to train, and show great performance 📈 Code:

7

12

52

Johan Edstedt

@Parskatt

3 months

Objaverse is the most important 3D vision paper in last 5 years, if google made street view data accessible that would be the most important in 10 years.

4

3

44

Johan Edstedt

@Parskatt

2 months

DL3DV looks good :) @LuLing26466911 thanks for making the colmap caches available!

1

5

43

Johan Edstedt

@Parskatt

7 months

Yushan's work GMSF: Global Matching Scene Flow is accepted to NeurIPS 2023!🥳 We propose a simple but powerful approach to scene flow estimation through global matching that achieves state-of-the-art performance. Paper: Code:

0

10

42

Johan Edstedt

@Parskatt

1 month

DeDoDe v2 coming soon with some improvements to the detector! Colab with @BokmanGeorg and @zhenjun_zhao

3

7

41

Johan Edstedt

@Parskatt

14 days

Trying out, let's see.

2

1

36

Johan Edstedt

@Parskatt

3 months

@cHHillee @alicemazzy At least in Sweden it can also mean that the place is too local, only regulars go there and rate the place highly. Typical example: local pizzeria.

0

35

Johan Edstedt

@Parskatt

2 months

DeDoDe now in kornia 😀

Kornia

@kornia_foss

2 months

0.7.2 is out! - Added DeDoDe features (thanks @Parskatt ) - LightGlue models, available nowhere else - DeDoDe (B/G), KeyNet-HardNet - KMeans implementation - New augmentations: RandomGaussianIllumination, RandomLinearIllumination, RandomLinearCorner 1/2

1

28

130

1

5

35

Johan Edstedt

@Parskatt

4 months

RoMa can do MVS well :)

3

2

31

Johan Edstedt

@Parskatt

3 months

Checkout my "50% done" PhD seminar, where I talk about my recent works in 3D Reconstruction using Neural Networks! There are new things in there! Watch to the end! If you have questions, ask them in this thread :) Link:

2

29

Johan Edstedt

@Parskatt

4 months

Choose your imports carefully lol. Maybe I should change name to RoMatch... Or they should change to RotMan

PyTorch

@PyTorch

4 months

RoMa: an easy-to-to-use, stable and efficient library to deal with rotations and spatial transformations in PyTorch. Read all about this PyTorch Ecosystem Tool in our latest Medium post ⚡

18

118

648

3

1

26

Johan Edstedt

@Parskatt

15 days

The real problem with COLMAP is that it's impossible to ctrl+F stuff. There's always 3 layers of abstraction for everything. Makes it so hard when you don't have 5 years of experience with it.

2

0

26

Johan Edstedt

@Parskatt

4 years

@get2rao @tg_bomze @Rob__Milliken @h_bash This is probably because it converts images to be square, in the preprocessing step. Try it with a cropped image instead and see if it improves.

0

20

Johan Edstedt

@Parskatt

26 days

2

1

20

Johan Edstedt

@Parskatt

17 days

@jxmnop No

1

0

17

Johan Edstedt

@Parskatt

1 month

Can someone pls make better MVS than patchmatch, I'm begging you. Otherwise I'm making it this autumn.

4

0

17

Johan Edstedt

@Parskatt

2 months

brb 15 min solving 3D vision

1

0

16

Johan Edstedt

@Parskatt

9 months

It do be like that

0

1

14

Johan Edstedt

@Parskatt

2 months

@docmilanfar Dust3r has some theories.

0

1

14

Johan Edstedt

@Parskatt

3 months

While you were partying I studied the Fundamental Matrix

0

1

13

Johan Edstedt

@Parskatt

1 month

Spring in Linköping

1

0

13

Johan Edstedt

@Parskatt

3 months

Wake up babe, new matching dataset just dropped

Lu Ling

@LuLing26466911

3 months

Our DL3DV-10K dataset paper has been accpeted by #CVPR2024 🎉! It provides scene-level videos at 4K resolution, RGB-images, camera pose, and point coulds. The DL3DV-3K is currently available and more versions come soon. Feel free to check our project page:

6

23

162

2

0

13

Johan Edstedt

@Parskatt

16 days

@ducha_aiki @JeromeRevaud @olliecrosen @AjdDavison @LourdesAgapito Ive been recommended this before

3

1

12

Johan Edstedt

@Parskatt

3 months

@docmilanfar The method of combining rocks into a henge is simple, and while a simple method is not grounds for rejection, the method is not general. If the authors put some henges in Sudan and Thailand I might reconsider.

0

12

Johan Edstedt

@Parskatt

10 days

Quaternions should never be exposed outside of library, happy that pycolmap banned it. Coordinate systems are difficult enough as is.

4

0

12

Johan Edstedt

@Parskatt

5 months

@pesarlin I agree for senior PhDs who already has several papers published. I don't agree for new students. Your first paper wont be your best resesrch (usually), and you need practice through quantity.

2

0

12

Johan Edstedt

@Parskatt

3 months

Every "we beat COLMAP" paper

Chinmaya Kausik

@ChinmayaKausik

3 months

For a small fee, I will go to your enemy's talk and ask "did you tune hyperparameters for all algorithms or just yours?"

2

6

52

1

11

Johan Edstedt

@Parskatt

7 months

Rotation equivariance/invariant learned descriptors. What exists?

7

1

11

Johan Edstedt

@Parskatt

9 months

0

5

11

Johan Edstedt

@Parskatt

2 months

This is the year of 3D, feelsgoodman

1

0

11

Johan Edstedt

@Parskatt

3 months

torchvision perspective is numerically unstable (see ) noted by @BokmanGeorg . Same in kornia for float32 (solve here should always be float64 imo) @ducha_aiki This particularly applies to large images with small crops.

_get_perspective_coeffs is numerically unstable · Issue #8248 · pytorch/vision

🐛 Describe the bug from torchvision.transforms.functional import _get_perspective_coeffs A = torch.tensor(((1,1),(0,1)), dtype=torch.float32) offsets = torch.tensor([[1, 1], [-1,-1], [-1, 1], [1,...

github.com

4

0

11

Johan Edstedt

@Parskatt

14 days

@AlbyHojel I assume this is real-time?

1

0

11

Johan Edstedt

@Parskatt

9 months

This means RoMa is now Apache 2.0! 🥳

GitHub - Parskatt/RoMa: [CVPR 2024] RoMa: Robust Dense Feature Matching; RoMa is the robust dense...

[CVPR 2024] RoMa: Robust Dense Feature Matching; RoMa is the robust dense feature matcher capable of estimating pixel-dense warps and reliable certainties for almost any image pair. - Parskatt/RoMa

github.com

Yann LeCun

@ylecun

9 months

DINOv2, the cutting-edge computer vision model trained through self-supervised learning to produce universal features, is now available under the Apache 2.0 license. Onward with open source AI.

39

229

2K

0

2

10

Johan Edstedt

@Parskatt

20 days

Faster colmap MVS available from docker pull parskatt/colmap:12.2.2-sm_80

4

1

10

Johan Edstedt

@Parskatt

3 months

When your paper gets accepted after 47 rounds of rejection and rebuttal

1

0

9

Johan Edstedt

@Parskatt

3 months

kornia things: transform_points(extrinsics, points) project_points(points, intrinsics) keeps you on your toes while coding :D

1

0

9

Johan Edstedt

@Parskatt

3 months

No one cares about statistical singificance in computer vision, and that's a good thing.

1

2

9

Johan Edstedt

@Parskatt

14 days

Could we get a better example of a good review? This one is just someone stating that they love the paper with (actually) no justification.

3

0

9

Johan Edstedt

@Parskatt

10 days

mamba install cuda restored some of my sanity

0

9

Johan Edstedt

@Parskatt

20 days

@vincesitzmann @eric_brachmann @JeromeRevaud Sky's the limit with conventional sfm 😉

1

0

8

Johan Edstedt

@Parskatt

28 days

Enjoying learning rust, really nice language and packaging!

0

8

Johan Edstedt

@Parskatt

3 months

Me frfr

1

0

8

Johan Edstedt

@Parskatt

18 days

@ducha_aiki @CSProfKGD @ftm_guney @dimadamen @chriswolfvision @taiyasaki Please accept my cockroach with 6 toes.

1

0

8

Johan Edstedt

@Parskatt

4 months

Is there any modern MVS (not patchmatch) that works out-of-the-box for colmap SfM outputs?

2

3

8

Johan Edstedt

@Parskatt

10 months

Any recent good keypoint detectors except SiLK? @ducha_aiki @zhenjun_zhao

2

4

8

Johan Edstedt

@Parskatt

10 months

@zhenjun_zhao DINOV2 features are the future :)

0

8

Johan Edstedt

@Parskatt

10 months

@ducha_aiki I made a meme for this response ;)

1

8

Johan Edstedt

@Parskatt

1 year

@zhenjun_zhao Code now available (beware of 🐛) at

GitHub - Parskatt/RoMa: [CVPR 2024] RoMa: Robust Dense Feature Matching; RoMa is the robust dense...

[CVPR 2024] RoMa: Robust Dense Feature Matching; RoMa is the robust dense feature matcher capable of estimating pixel-dense warps and reliable certainties for almost any image pair. - Parskatt/RoMa

github.com

0

7

Johan Edstedt

@Parskatt

10 months

Showerthought: We can improve matching by learning from neighbours (superglue, loftr, dkm), or by better descriptions of the points. I think the latter has more potential, and maybe we should als rethink what we mean by "point".

3

1

7

Johan Edstedt

@Parskatt

7 months

I know RELF exists

GitHub - bluedream1121/RELF: [CVPR 2023] Official PyTorch implementation of "Learning Rotation-Eq...

[CVPR 2023] Official PyTorch implementation of "Learning Rotation-Equivariant Features for Visual Correspondence" - bluedream1121/RELF

github.com

1

0

7

Johan Edstedt

@Parskatt

30 days

@JeromeRevaud @arankomatsuzaki Unfortunately people are not aware of croco/croco-v2. I think it's a shame, since they perform similarly to DINOv2 in my tests. Perhaps we/you can try out their benchmark with croco?

0

7

Johan Edstedt

@Parskatt

2 months

Code available now

Jerome Revaud

@JeromeRevaud

2 months

4

45

286

0

7

Johan Edstedt

@Parskatt

10 days

I think roboticists have gaslighted themselves that quats are nice.

1

0

7

Johan Edstedt

@Parskatt

3 months

Vscode just broke remote for all machines with glibc <= 2.28...

vscode server won't start: "Waiting for server log..." · Issue #203967 · microsoft/vscode

Does this issue occur when all extensions are disabled?: Yes/No VS Code Version: 1.86.0 OS Version: Windows 10.0.19044 Steps to Reproduce: Start vscode Observe log of remote SSH I just restarted vs...

github.com

2

1

7

Johan Edstedt

@Parskatt

7 months

@ducha_aiki Cool! I think the remaining issue is large rotations, working on releasing a model more robust to that.

1

0

6

Johan Edstedt

@Parskatt

10 days

@chrisoffner3d I'm team 3x3 matrix as what you show externally, and converting it to whatever you want internally. As long as you know the convention I think you're fine, but I'm seriously going crazy from convertions.

2

0

6

Johan Edstedt

@Parskatt

2 months

Causal next word prediction like pixelcnn doesn't seem popular in vision, reason for this?

1

0

6

Johan Edstedt

@Parskatt

7 months

@giffmana sOTa, if you're going to do it wrong, go all out.

0

5

Johan Edstedt

@Parskatt

4 months

This (slightly older work) seems to have an api for colmap sparse outputs (if I'm not misreading)

GitHub - FangjinhuaWang/PatchmatchNet: Official code of PatchmatchNet (CVPR 2021 Oral)

Official code of PatchmatchNet (CVPR 2021 Oral). Contribute to FangjinhuaWang/PatchmatchNet development by creating an account on GitHub.

github.com

1

6

Johan Edstedt

@Parskatt

17 days

CNN > Transformer

2

0

6

Johan Edstedt

@Parskatt

18 days

Anyone visualized neural activations in @rerundotio yet? Would be super cool and similar to @karpathy s scifi short.

2

0

6

Johan Edstedt

@Parskatt

3 months

Big quality of life improvement

Rerun

@rerundotio

3 months

You can now switch to a First Person camera

2

7

0

1

6

Johan Edstedt

@Parskatt

3 months

RIP lucidrains era?

0

6

Johan Edstedt

@Parskatt

3 months

@CVPR heartbeat just went up 2x

2

0

6

Johan Edstedt

@Parskatt

25 days

@TimDarcet Theyre probably not contributing much to submission except putting their name on it 😄 There should be a cap at 10.

2

0

6

Johan Edstedt

@Parskatt

9 days

@ducha_aiki @guipotje @eucadar @andrefaraujo @renatojmsdh @ericksonrn_ Works quite well !

2

0

6

Johan Edstedt

@Parskatt

2 months

🤩

Rerun

@rerundotio

2 months

@Parskatt should be included in the next release!

0

1

3

0

5

Johan Edstedt

@Parskatt

3 months

Can we just agree that center of top-left pixel is [0.5,0.5] and not [0,0] and things like this wouldn't have to be done?

2

0

5

Johan Edstedt

@Parskatt

8 days

@janusch_patas Feel like splat people should really evaulate more on dl3dv, not just the simple MVS datasets.

0

5

Johan Edstedt

@Parskatt

5 months

Cool stuff

NAVER LABS Europe

@naverlabseurope

5 months

Check-out 📢DUSt3R📢 - a new 3D reconstruction model that works with no prior info on camera calibration nor viewpoint poses! Outperforms SoA monocular & multiview depth estimation & relative pose estimation. Paper, demo, videos (& soon code!) available

2

35

146

0

5

Johan Edstedt

@Parskatt

4 months

Join our lab! Exciting positions in machine learning for satellite imaging :)

Computer Vision Laboratory (CVL), Linköping

@CvlIsy

4 months

Join us at CVL! We now offer PhD Student opportunities, the research will be focused on machine learning for remote sensing. More info and application here:

0

4

7

0

5

Johan Edstedt

@Parskatt

3 months

@CVPR Is this Kaimings first ever paper? Pretty good start in that case.

0

5

Johan Edstedt

@Parskatt

3 months

@karpathy Thoughts on visual tokenizers? Current seems to be uniform patches with lin filters, can we do better?

0

5

Johan Edstedt

@Parskatt

1 year

@ducha_aiki @yash_patel2307 @majti89 They really did a lot of work here, if I'm not mistaken they seem to have ported the minimal solvers into pytorch? These are not trivial to implement. Perhaps something to put into @kornia_foss ?

1

0

5

Johan Edstedt

@Parskatt

8 months

Official colmap docker now updated, available with docker pull colmap/colmap

0

5

Johan Edstedt

@Parskatt

16 days

@Michael_J_Black Any advice for finding places that are fun to work at? Of course after working there I'll know; but are there some good questions to ask e.g. at interview?

2

0

5

Johan Edstedt

@Parskatt

1 month

RoMa warp*confidence

Federico Italiano

@FedeItaliano76

1 month

Danila Tkachenko's astounding ‘Restricted Areas’ series on the relics of Soviet utopianism Can you see beauty in it, or only decay?

36

477

4K

0

5

Johan Edstedt

@Parskatt

10 months

Pretty disappointed with torch.compile so far 😐 Compilation errors constantly, even for simple graphs. Seems they have a long way to go.

3

0

4

Johan Edstedt

@Parskatt

1 month

@SattlerTorsten An alternative to MVS I guess would be NERF/GS-based methods, but I'm not entirely convinced by the precision of those methods yet when you don't have a huge number of images (perhaps I'm wrong though).

1

0

5

Johan Edstedt

@Parskatt

5 months

What should I research?

K > 2 multi-view

28

Object centric matching

14

Scaling (datasets/models)

16

Other (suggest pls)

4

1

5

Johan Edstedt

@Parskatt

2 months

@ducha_aiki Please ban quaternions from 3DV

1

0

5

Johan Edstedt

@Parskatt

14 days

@ducha_aiki Random seed xd

1

0

4

Johan Edstedt

@Parskatt

5 months

@ducha_aiki @BokmanGeorg Thanks for sharing! Really excited for this work. We show that descriptions in different orientations are just a "matrix mult away" from eachother, when you train correctly. I think this has potential for other things too (e.g. skew, lighting, etc)

1

0

4

Johan Edstedt

@Parskatt

4 months

If you're interested in tracking, check out Jie's work where we investigated the potential of data augmentation for modern Transformer based trackers!

Ryohei Sasaki

@rsasaki0109

4 months

DATr PyTorch implementation of "Leveraging the Power of Data Augmentation for Transformer-based Tracking" (WACV2024)

0

19

87

0

1

4

Johan Edstedt

@Parskatt

30 days

@chrisoffner3d @arankomatsuzaki For depth I feel like the issue is that you could regress depth, log depth, disparity. These are non-lin of eachother and therefore the network would need to have seperate rep for all of them to be linsep.

1

0

4

Johan Edstedt

@Parskatt

27 days

@AlexStoken @ducha_aiki @BokmanGeorg @zhenjun_zhao Haha thanks, I spent way too much time on changes that didn't help at all. So might as well warn others. Loss functions for detectors are quite tricky to get right (and v2 is still not the right one).

0

4

Johan Edstedt

@Parskatt

3 months

@ducha_aiki @imtiazprio @randall_balestr @rbaraniuk They don't actually know why. I suggest a change of title: "Deep Networks Always Grok and We Don't Know Why" or "Deep Networks Always Grok and Here is What it is"

2

0

4

Johan Edstedt

@Parskatt

1 month

@SattlerTorsten I should note that patchmatch generally works well, but I feel like there should have been some development the last 15 years.

0

4

Johan Edstedt

@Parskatt

4 months

Well deserved :)

Accepted papers at TMLR

@TmlrPub

4 months

DINOv2: Learning Robust Visual Features without Supervision Maxime Oquab, Timothée Darcet, Théo Moutakanni et al.. Action editor: Abhishek Kumar. #supervised #visual #features

1

18

118

0

4

Johan Edstedt

@Parskatt

3 months

@LuLing26466911 @CVPR @PurdueCS Can you explain why this requires a non-commercial license? I understand the need for anonymization, but this shouldn't prevent commercial usage? It is also unclear to me what applies to model weights trained on this dataset? Do they inherent this license?

1

0

4

Johan Edstedt

@Parskatt

4 months

@rerundotio Rerun is really convenient to run :)

1

4

Johan Edstedt

@Parskatt

8 months

Pushed up-to-date colmap and hloc docker image to docker hub. Available under parskatt/colmap and parskatt/hloc

0

4

Johan Edstedt

@Parskatt

2 months

@gabriberton @ducha_aiki @AlexStoken @bcaputo_iit Ah, you might also be interested in Steerers @BokmanGeorg

GitHub - georg-bn/rotation-steerers: A steerer for D-dimensional keypoint descriptions is a DxD...

A steerer for D-dimensional keypoint descriptions is a DxD matrix that transforms the descriptions as if they were computed from a rotated image. - georg-bn/rotation-steerers

github.com

1

0

4

Johan Edstedt

@Parskatt

10 months

Better features, better matching, some new results:

0

4