Dmytro Mishkin 🇺🇦 @ducha_aiki Twitter profile

Pinned Tweet

Dmytro Mishkin 🇺🇦

2 years

❗️🇺🇦🇺🇦🇺🇦❗️ My friends have created Notion wiki with many ways how you can help #Ukraine - donation, petitions, housing, DDOS and many others #HelpUkraine Please, share.

9

175

302

Last Seen Profiles

@PARIS_sats

@bestjjaehever

@EbaChronicles

@MrJonCryer

@chaergy

@Kensuke_38

@shiratama232396

@pemuas_ayangmu

@ohdestinations_

@ERAUAthletics

@coltford

@burntout_shell

@aka_themamba

@_CTRLBOY

@_tsukuyxmi

@zafarsah1

@edalpash

@chaeyaloveb0t

@_Jolene_Jolene

@ImdadSoomro6

@KEI722796079329

@boujieeasnae

@Sukabinorstw11

@LCS__Football

@boobataebubble

@VVDAmstelveen

@chihamaru_desyO

@dodorimohere

@localpride62

@mimetic_ilinx

@aerithisbaby

@betsy_klein

@bachnation

@timfremaux

@cody_cardillo

@MxtrixM

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Dear friends. I realize that you followed me for the news in computer vision, image matching, machine learning. Not for political bullshit. But my country is in grave danger. It is not politics, it is life&death. I will continue to write about ML&CV. And also about 🇺🇦.

58

185

4K

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

I have submitted my PhD thesis, yay!

37

11

594

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Yesterday Facebook blackout shows why I am afraid of self-driving cars. Not because of deep learning: I believe we will solve robust perception at some point. But mega-centralization...one wrong update and millions of self-driving cars become a weapons of mass self-destruction.

24

82

535

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

If you are interested in doing PhD in computer vision and/or robotics, consider CTU in Prague. We are bad in self-PR, but good in science. Also, Prague is a great and affordable city to live in.

24

70

445

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

That is 🔥 Linformer: Self-Attention with Linear Complexity Sinong Wang, Belinda Li, Madian Khabsa, Han Fang, Hao Ma

6

104

379

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Torch.manual_seed(3407) is all you need: On the influence of random seeds in deep learning architectures for computer vision @david_picard tl;dr: one can easily get +0.2 pp accuracy on ImageNet just by changing random seed. And much more on CIFAR10

8

82

359

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

ReLU Fields: The Little Non-linearity That Could @AnimeshKarnewar , Tobias Ritschel, Oliver Wang, Niloy J. Mitra tl;dr: grid-based 3D representation + ReLU are almost NERFs, but much faster.

10

58

347

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

Fully Differentiable RANSAC Tong Wei, @yash_patel2307 , Jiri Matas, @majti89 tl;dr: another (now working better?) differentiable RANSAC with clever pretraining (KL-div for inlier probabilities, then regular losses). Eval is good :) 1/2

7

60

298

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Good practice, which I forget to do every other time. #OpenSource

4

33

288

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

PhD defence done! Thank you all for support and coming online for the presentation!

29

4

266

Dmytro Mishkin 🇺🇦

@ducha_aiki

6 months

It seems that RoMA by @Parskatt almost solved WxBS. I have added it to eval colab here

9

51

259

Dmytro Mishkin 🇺🇦

@ducha_aiki

10 months

“Diffusion models are general purpose learning machineries”, not only generation. Talk by @YasutakaFuruka1 #CVPR2023 1/

3

41

254

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Good news: Pytorch works on M1 GPU :) Bad news: not all Pytorch operations work there. I guess we should create a separate test case for device='mps' to evaluate that at @kornia_foss

PyTorch

@PyTorch

2 years

We’re excited to announce support for GPU-accelerated PyTorch training on Mac! Now you can take advantage of Apple silicon GPUs to perform ML workflows like prototyping and fine-tuning. Learn more:

79

711

3K

6

41

248

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

SaaS: Schmidhubering as a Service

3

13

241

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

We are happy to introduce our paper, rejected from #CVPR2020 "Image Matching across Wide Baselines: From Paper to Practice". It has so many messages, that I will create separate tweet-thread for each part.

Kwang Moo Yi

@kwangmoo_yi

4 years

We are hosting a challenge/workshop at #CVPR2020 . Please see for more info. Here's also a paper related to it that you WON'T see at CVPR ;) TLDR: SIFT Keypoint+HardNet is still the best!

1

6

34

7

50

236

Dmytro Mishkin 🇺🇦

@ducha_aiki

5 months

Bag of Image Patch Embedding Behind the Success of Self-Supervised Learning @Yubei_Chen , @AdrienBardes , @LiZengy , @ylecun tl;dr: bag of patches and compositional structure of the horse is all you need. (except where is Dinov2 comparison?) and

3

41

232

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 months

Deep Networks Always Grok and Here is Why @imtiazprio , @randall_balestr , @rbaraniuk tl;dr: early stopping is bad, instead you should train much longer than you thought to have more robust model. Bonus: BatchNorm hurts

6

39

231

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

Ouch. I haven't read this full story about the ChatGPT virtual machine. That is incredible.

8

38

224

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

My slides on "How to navigate ML research literature" for @ucu_apps Winter ML school. How to read papers? How to filter out? Where to get? What does peer review mean? Unfortunately, no part about re-implementing - this is covered by other speaker.

How_to_navigate_ML_research.pdf

drive.google.com

7

44

227

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

Wow, there is huge revolution on #kaggle happening right now. More and more competitions are "kernel-only-submisson". Which means you should upload your model to cloud and then all the inference is done in submission kernel with runtime limit. Farewell 100500 models ensembles!

3

37

225

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Tl;d: Pytorch team is working on M1 GPU support and plan to release alpha in 4 month.

Timothy Liu

@timothy_lkh_

2 years

It’s coming!

1

40

250

2

21

216

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

Faster AutoAugment - first paper, which uses @kornia_foss Just in case if anyone was wondering why one needs differentiable augmentation :)

2

63

206

Dmytro Mishkin 🇺🇦

@ducha_aiki

11 months

VanillaNet: the Power of Minimalism in Deep Learning Hanting Chen, Yunhe Wang, Jianyuan Guo, Dacheng Tao tl;dr: 4x4conv/4->n x {1x1conv->{seriesAct}->MaxPool2x2}. seriesAct = stack of BN(ReLU(BN(ReLU)))

6

46

205

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Dear colleagues. Could you please not hardcode "cuda" device in your example code? Some of us running your code on a laptop :)

8

9

204

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

ML theory twitter, here is a question: why classification works so much better that regression? Bonus question: is it true only for DL, or for XGBoost as well?

Eric Brachmann

@eric_brachmann

1 year

@ducha_aiki No clue, happy for pointers! I suspect there is some nice theory in the ML community. Stability of soft max? Robustness to multi-modality? Networks preferring over-parametrizations?

4

1

9

34

30

202

Dmytro Mishkin 🇺🇦

@ducha_aiki

5 years

How good are classic navigation algorithms vs fancy RL? Spoiler: much better Read our paper "Benchmarking Classic and Learned Navigation in Complex 3D Environments" Paper: Source code (most of it) :

GitHub - ducha-aiki/navigation-benchmark: Code for "Benchmarking Classic and Learned Navigation in...

Code for "Benchmarking Classic and Learned Navigation in Complex 3D Environments" paper - ducha-aiki/navigation-benchmark

github.com

5

61

201

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

NeRF-SLAM: Real-Time Dense Monocular SLAM with Neural Radiance Fields @RosinolToni , John J. Leonard, @lucacarlone1 tl;dr: DroidSLAM + Instant-NGP.

0

46

196

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

"RaspberryPI for mosquito neutralization by power laser" "The system can neutralize 2 mosquitos per sec and this result can be easily improved. "

10

34

196

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

Tormentor : pyTORch augMENTOR. New augmentation library based on the #pytorch and #kornia @kornia_foss

0

60

194

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

NeRN -- Learning Neural Representations for Neural Networks Maor Ashkenazi, Zohar Rimon, Ron Vainshtein, Shir Levi, Elad Richardson, Pinchas Mintz, Eran Treister tl;dr: NERFs to compress the CNN classifiers. Not hypernetworks, although seem similar

3

38

191

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Kubric: A scalable dataset generator 34 co-authors. tl;dr: (Blender+PyBullet)-based framework for the dataset generation, suitable for many tasks code:

5

29

191

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

Metric learning: cross-entropy vs. pairwise losses Tl;dr: paper derives that the most of metric learning losses are kind of equivalent to cross-entropy. And CE is not worse than them.

7

42

187

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Identical Initialization: A Universal Approach to Fast and Stable Training of Neural Networks Anon #ICLR2023 submission tl;dr: initialize your weights with (almost) identity.

4

26

186

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

LERF: Language Embedded Radiance Fields Justin Kerr, Chung Min Kim, @Ken_Goldberg , @akanazawa , Matthew Tancik tl;dr: if NERF is trained to predict DINO+CLIP features in addition to RGB and depth, you can use it later for volumentric segmentation.

2

41

187

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

My PhD defence will be on Thursday Dec 16, 14:00 CET. Available online at this MS Teams link:

27

8

185

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Wow, #CVPR2023 reviewing will be not on CMT, but on OpenReview!

5

23

184

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 months

CLIP Can Understand Depth Dunam Kim, Seokju Lee tl;dr: learn input embedding for the monodepth, then use it for the input -> not sota, but quite good, unlike previous CLIP based depth.

0

24

183

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

One could learn more about good software engineering from this 1 hour talk than from a university course, tbh. Incredible presentation of nbdev, jupyter notebook and live coding environments by @jeremyphoward

3

20

185

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

We announce "Affine Correspondences (AC) and their Applications in Practice" #CVPR2022 tutorial by @majti89 @prittjam , me and Levente Hajder. ✅ AC applications: rectification, surface normals, fast pose ✅ How to get AC @CVPR P.S. Dolls and photos by Olha Mishkina.

3

29

182

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

Image Matching Challenge 2023 starts NOW! Task: 3D reconstructions from 10-100 images Entry Deadline: June 6, 2023. Prize Money: $50,000 #IMC2023 #CVPR2023 @CVPR

Image Matching Challenge 2023

Reconstruct 3D scenes from 2D images

www.kaggle.com

3

41

176

Dmytro Mishkin 🇺🇦

@ducha_aiki

6 months

If you are into image matching and want to get feeling for different algorithms - try this amazing webdemo. It has tons of modern local features, as well as many version of RANSAC (from OpenCV)

2

31

177

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Global Pooling, More than Meets the Eye: Position Information is Encoded Channel-Wise in CNNs Authors: all tagged + Neil D. B. Bruce tl;dr: CNNs with global average pooling encode the position in channel order. Nicely formulated and tested hypothesis

7

40

177

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

ArXiving before submission : - helps professional identity building; -protects against idea re-discovery/theft/gate-keeping - facilitates open research distribution - reduces inequality. our refined statement, with @amy_tabb & Jiri Matas, now on arXiv

ArXiving Before Submission Helps Everyone

We claim, and present evidence, that allowing arXiv publication before a conference or journal submission benefits researchers, especially early career, as well as the whole scientific community....

arxiv.org

10

40

175

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 months

Short guide for modern local features: - if you are GPU and CPU-rich, use RoMA - if you are allergic to learned matching, use DeDoDe. - otherwise use ALIKED+LightGlue.

7

27

174

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

Wow, fastai is on the front-page of top repositories selected for the github.archive project: together with Linux, Ruby, Postgres, jquery. Congrats!

0

31

173

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

This comment gives me some hope. Context: M1 GPU support in PyTorch.

6

12

168

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

OpenGlue: Open Source Graph Neural Net Based Pipeline for Image Matching @OstapViniavskyi , @DobkoMaria @ducha_aiki @dobosevych tl;dr: MIT license impl+additional geom input. Caveat: same as SGMNet, fails to match official ML perf

5

34

164

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 month

Track Everything Everywhere Fast and Robustly Yunzhou Song, Jiahui Lei, Ziyun Wang @LingjieLiu1 @KostasPenn tldr: track in (optimized)3D and use DINOv2 for long-term matching.

1

28

166

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Can Vision Transformers Learn without Natural Images? It turns out that one can pretrain vision networks solely on the synthetic images - "formula supervision", fractal-based.

1

35

161

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

@CSProfKGD Here is my opinion

3

4

163

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

I am preparing for my lectures on How to navigate CV/ML literature, how to read papers in an efficient way, re-implement then, etc. Which issues in this area do you personally have? Questions? Whar is the most misleading advice you ever get? Please write in comments below

13

8

160

Dmytro Mishkin 🇺🇦

@ducha_aiki

10 months

Enjoy Image Matching Challenge 2023 recap: tl;dr: - SfM is not solved - global desc similarity is hard - orientation invariance for SuperGlue is easy - PixSfM is good idea, but need follow-ups - KeyNetAffNet @kornia_foss rocks #IMC2023 #CVPR2023 @CVPR

3

32

158

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 month

We are launching Structured Semantic 3D Reconstruction (S23DR) Challenge at @CVPR @structGeom workshop on @huggingface ! Prizes: 1: $10,000 2: $7,000 3: $5,000 Deadline: June 4, 2024 Task: point cloud+segmentations+monodepth ->roof wireframe. 1/2

2

36

160

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

So true :)

Colin Raffel

@colinraffel

4 years

Hot take: Mathiness [1] is like an adversarial patch [2] for ML conference reviewers: Mathiness causes a reviewer to classify the paper as "accept" regardless of whether the math is useful/valid and the paper is any good. [3] Fig. 6 has some empirical evidence of this. (refs ⬇️)

14

82

510

1

35

159

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 month

Image Matching Challenge 2024 is live on #Kaggle ! Prize fund: $50k. Entry deadline: May 28 2024. Goal: reconstruct camera poses given images. Catch: many nuisance factors: illumination, nature, season, rotation, doppelgängers etc. Examples from train(easy) #IMC2024 @CVPR 1/3

3

37

157

Dmytro Mishkin 🇺🇦

@ducha_aiki

5 years

Our 7th place solution to Whale competition. Great thanks to @Ana_Geneva and Igor for team, @radekosmulski for starter kit and @fastdotai for great library.

3

32

154

Dmytro Mishkin 🇺🇦

@ducha_aiki

5 years

Just to make clear: Kornia () is a differentiable OpenCV in Pytorch, ambitious project led by @edgarriba . It has filters, geometry, local features, losses. And much more in plans.

Dmytro Mishkin 🇺🇦

@ducha_aiki

5 years

@michalwols What exactly, THFeat, library? kornia is much more than local features, it is more like "pytorch differentiable OpenCV".

1

2

6

2

33

150

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Image as Set of Points #ICLR2023 anon submission tl;dr: SuperPixel meet ViT

1

33

151

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 months

Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, Hengshuang Zhao tl;dr: DINOv2 meets MIDAS and ZoeDepth and unlabeled images with strong augmentation like CutMix.

2

26

150

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

SLAM people, I have couple of questions. I understand that ORBSLAM(1,2,3) is a great and well engineered piece of software. But it is a paper from 2015. Many things have changed, e.g. hardware, we have Full HD (GPU)-SIFT in realtime now. Why don't we have SuperSLAM at least?

21

18

150

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

DeepMatcher: A Deep Transformer-based Network for Robust and Accurate Local Feature Matching Tao Xie, Kun Dai, Ke Wang, Ruifeng Li, Lijun Zhao tl;dr: include position embeddings everywhere,not only beginning, better "fine" refinement -> better LoFTR 1/2

7

26

147

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Personal announcement: I am joining @HOVER3D to work with @giridhar89 on the WxBS application in real life -- house 3D reconstruction from several photos. I am also keeping my CTU in Prague affiliation and will continue to work on #IMC2022 and @kornia_foss .

20

1

147

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Tune It or Don’t Use It: Benchmarking Data-Efficient Image Classification tl;dr: When carefully tuned (hyperparams), cross entropy works as good, or better than specialized "low-data" losses/methods. P.S. Why am I not surprised? #ICCV2021

2

25

143

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

The map of occupied Ukrainian territory. The enemy is not as fast as they dreamed. But they are here.

4

16

135

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 month

Benchmarking Object Detectors with COCO: A New Path Forward Shweta Singh, Aayan Yadav, Jitesh Jain, Humphrey Shi, Justin Johnson, Karan Desai tl;dr: SAM-refined masks for MS-CoCo -> re-evaluated benchmark -> all methods score higher.

2

25

135

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Self-Attention Between Datapoints: Going Beyond Individual Input-Output Pairs in Deep Learning @janundnik , Neil Band, @clarelyle , @AidanNGomez , @tom_rainforth , @yaringal tl;dr: deep learning makes kNN 2.0 by modeling interactions between data points.

1

30

135

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Exploring Simple Siamese Representation Learning Xinlei Chen, Kaiming He Tl;dr: Stop-grad + siamese similarity is all you need for semi-supervised learning (in agreement with recent FROST paper by @lnsmith613 )

0

28

130

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Brief reminder to myself and #IMC2021 #CVPR2020 paper reviewers. Our paper now has 80 citations and become one of the standard image matching benchmarks. Ofc, weak reject :)

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

We are happy to introduce our paper, rejected from #CVPR2020 "Image Matching across Wide Baselines: From Paper to Practice". It has so many messages, that I will create separate tweet-thread for each part.

7

50

236

1

14

133

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 months

The premise of DeDoDe paper: you don't need learned matcher, mutual nearest neighbor is enough. LightGlue: hold my beer DeDoDe: soon in @kornia_foss thanks to @Parskatt . DeDoDe-LightGlue - trained by @kornia_foss

3

18

131

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 month

FeatUp: A Model-Agnostic Framework for Features at Any Resolution @xkungfu , Mark Hamilton, Laura Brandt, Axel Feldman, Zhoutong Zhang, William T. Freeman tl;dr: in title. Upscale DINOv2, anyone? #ICLR2024

4

30

129

Dmytro Mishkin 🇺🇦

@ducha_aiki

5 months

GS-SLAM: Dense Visual SLAM with 3D Gaussian Splatting Chi Yan, Delin Qu, Dong Wang, Dan Xu, Zhigang Wang, Bin Zhao, Xuelong Li tl;dr: NERF-SLAM, now with Gaussian splatting. 8.34 FPS, of course from RGB-D, not just monocular.

2

17

129

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

Out-Of-Distribution Detection Is Not All You Need @jorisguerin , Kevin Delmas, @raulsferreira , Jérémie Guiochet tl;dr: ability to predict if model can make mistake >> good OOD performance.

3

22

125

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Right now, when civilians in Mariupol are being killed by Russian bombs, @IEEESpectrum spreads Russian propaganda in their recent article. WTF? @CVPR , we are IEEE/CVF conf. Can we have a public stance regarding war and spreading Russian propaganda?

8

35

128

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 months

Image matching folks, I know you couldn’t celebrate Christmas without knowing that Image Matching Workshop 2024 has been accepted to #CVPR2024 :) Kudos to Eduard, @kwangmoo_yi @fabio_bellavia , Jiři, @lcmorelli3 @3DOMFBK Fabio @CVPR

3

17

128

Dmytro Mishkin 🇺🇦

@ducha_aiki

5 years

I wrote a starter pack for @kaggle Landmark Retrieval 2019 competition. Since the competition data is too huge to iterate quickly, I have used SfM120k dataset from GeM paper. All the validation, fast NN-search code is already there. Library is @fastdotai

GitHub - ducha-aiki/google-retrieval-challenge-2019-fastai-starter: fast.ai starter kit for Google...

fast.ai starter kit for Google Landmark Retrieval 2019 challenge - ducha-aiki/google-retrieval-challenge-2019-fastai-starter

github.com

1

30

124

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Pros of @weights_biases : everything Cons: now you can do loss watching even from your phone :(

8

9

127

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Convolutional Hough Matching Networks for Robust and Efficient Visual Correspondence tl;dr: 6D (x-y-scale1, x-y-scale2) correlation ->scale max-pool -> refine -> softargmax -> correspondences. Juhong Min, Seungwook Kim, Minsu Cho

3

25

126

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Dear computer vision colleagues, I am really sorry that I am not working against social media ban at #CVPR now. Once we done with Russia and the war they started in Ukraine, we will resume our call for open science.

3

127

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Laugh is a weapon too. Today I will be posting news in meme format, please share :)

3

13

124

Dmytro Mishkin 🇺🇦

@ducha_aiki

13 days

Probing the 3D Awareness of Visual Foundation Models @_mbanani , Amit Raj, @kmaninis , Abhishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, @jcjohnss , @jampani_varun tl;dr: DINOv2 rules, but read paper. SD is good for semantic cores

2

26

125

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 months

Rethinking Inductive Biases for Surface Normal Estimation @BaeGwangbin , @AjdDavison tl;dr: fighting bitter lesson with geometry for better normals. Code is available, but non-commercial

2

20

125

Dmytro Mishkin 🇺🇦

@ducha_aiki

10 months

Detector-Free Structure from Motion Xingyi He, @JiamingSuen , Yifan Wang, Sida Peng, @qixing_huang , Hujun Bao, @XiaoweiZhou5 tl;dr: LoFTR-like->SfM -> lightweight PixSfM + BA interleaved. #IMC2023 1st place

2

30

121

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

Designing BERT for convolutional networks: sparse and hierarchical masked modeling Keyu Tian, Yi Jiang, Qishuai Diao, Chen Lin, Liwei Wang, Zehuan Yuan tl;dr: create image, which looks to CNN same, as transformers -> MIM starts working

4

27

119

Dmytro Mishkin 🇺🇦

@ducha_aiki

10 months

You think can Segment Anything? Try segmenting this trees then! #CVPR2023

1

17

120

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Everyone in Ukraine understands that and thus not agree to just "peace". It means that in 5-10 years we will be completely destroyed by reformed Russian army. We need to win. Don't wish us "may this end soon". Wish us victory.

Kamil Galeev

@kamilkazani

2 years

The best formula of institutional evolution is: 1. Scare them 2. Don't finish them It skyrockets the chance that they evolve. Right now the regime is very scared. So they're working fervently on integration with China and "deescalation" will buy them time they need desperately

25

328

2K

0

26

118

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Efficient Initial Pose-graph Generation for Global SfM Our new paper without deep learning invoved! Tl;dr: Bag of tricks for 7x speed-up camera pose graph generation (29 hours for 402 130 image pairs vs 202 hours originally) Details in the thread 1/6

7

25

119

Dmytro Mishkin 🇺🇦

@ducha_aiki

3 years

Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions tl;dr: Pyramid makes everything better.

2

28

120

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

A short survey+benchmark of SuperPoint family: - SuperPoint, - eric-yyjau-superpoint, - Reinforced SP by @eric_brachmann et al - KP2D - LANet. tl;dr: original is mostly better then the rest except LANet. They also differ in keypoints they detect.

Un-SuperPoint family: who are they?

does it worth to know them?

ducha-aiki.github.io

11

27

120

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

Other people: release arXiv paper and empty repo. Some FB guys release as code with "as described in our paper [TODO link]" Screenshots in repo look very promising: Better than SuperGlue on #IMC2022

5

22

116

Dmytro Mishkin 🇺🇦

@ducha_aiki

5 years

I have summed up all cool and not so cool things about @fastdotai library from first-time-lazy-enough to not-watch-lessons user point of view

Fast.ai library: 1st impression

Recently I have tried fast.ai deep learning library and would like to share my experience.

towardsdatascience.com

2

31

119

Dmytro Mishkin 🇺🇦

@ducha_aiki

10 months

PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustment Jianyuan Wang @chrirupp , @davnov134 tl;dr:PoseRegression network + refinement via camera pose diffusion. Uses SuperGlue to get the correspondences guiding the diffusion

3

25

116

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 months

So far: - Our evaluation metrics become a common in many papers, - Popularity of better RANSACs, etc. - more awareness of the proper evaluation. - 300 citations. - IJCV journal paper is not much different from CVPR submission. Reviewers, are you happy with the rejection?)

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

We are happy to introduce our paper, rejected from #CVPR2020 "Image Matching across Wide Baselines: From Paper to Practice". It has so many messages, that I will create separate tweet-thread for each part.

7

50

236

2

118

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

Panoptic Neural Fields: A Semantic Object-Aware Neural Scene Representation Abhijit Kundu et 8 al. tl;dr: Make NERF predict not only RGB, but also semseg, depth, instance seg. 1 MLP for background, several for dynamic obj. It proposes learned init. 1/2

2

27

119

Dmytro Mishkin 🇺🇦

@ducha_aiki

4 years

I am struggling with finding any sense in the PhD thesis as information unit form-factor. I see use-cases,value,advantages of: - blogpost - paper - textbook - course - reference book. - source code. I see no advantages and use-cases (except PhD req satisfaction) for the thesis.

23

14

117

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

Deep Learning for Camera Calibration and Beyond: A Survey Kang Liao, Lang Nie, Shujuan Huang, Chunyu Lin, Jing Zhang, Yao Zhao, Moncef Gabbouj, Dacheng Tao tl;dr: well, it is a good survey :) updated at :

2

32

116

Dmytro Mishkin 🇺🇦

@ducha_aiki

2 years

2nd Place Solution to Facebook AI Image Similarity Challenge Matching Track SeungKee Jeon tl;dr: horizontally cat two images -> feed into (SimCLR-pretrained) ViT -> classifier (match/non-match). As simple as is. Comp:

6

24

114

Dmytro Mishkin 🇺🇦

@ducha_aiki

1 year

@vboykis What are you all talking about? That is clear linear regression problem, they are typically using it for job interviews…

1

3

111

Dmytro Mishkin 🇺🇦

@ducha_aiki

6 months

Zero123++: a Single Image to Consistent Multi-view Diffusion Base Model Ruoxi Shi, Hansheng Chen, Zhuoyang Zhang, Minghua Liu, Chao Xu, Xinyue Wei,Linghao Chen,Chong Zeng, Hao Su tl;dr: generate 2x3 image views, change noise schedule, reference attention

4

23

113