Dmytro Mishkin

Computer vision and deep learning consultant.

I save your time and money by advising directions that are more suitable, from a performance, sustainability, and business perspective. While the current frameworks allow good software engineers to replace domain experts in computer vision, when it comes to implementation, expertise is required to decide, which should and, more importantly, should not be implemented. My consulting agreement template is [here]

Recent consulting projects


Since 2019
AI Ukraine

Serve as volunteer Member of the Expert Committee on Artificial Intelligence at Ministry of Digital Transformation Of Ukraine. Areas of responsibility: science and education.

2019 — 2020

Co-founder of the Eastern European Computer Vision Conference. We running the largest computer vision conference in Eastern Europe to connect industrial and academic worlds.

EECVC
Since 2016
Szkocka

Have started Ukrainian Research Group "Szkocka" is an initiative to promote and advance Ukrainian science. It is a platform for cooperation between researchers and supervisors on doing high-quality academic research. It is non-government organization free from bureaucracy and regulations, it exists due to free cooperation and donations.

We supervise and fund students, PhD students and volunteers research in computer vision and machine learning area. Group is named after Lviv mathematician community in 1930s

Since 2015

Started my PhD at CTU in Prague under supervision of Prof. Jiri Matas to deepen and expand my expertise in computer vision and machine learning. My research is mostly devoted to wide baseline stereo and local features: the workhorse of 3D reconstuction, SLAM and image retrieval. During the study, I have done reserch internship at Intel Labs Munich, where I studied classical and learning-based navigation algorithms.

I am the maintainer of the open source Kornia library — OpenCV in PyTorch.

CTU in Prague
2014 — 2017
Clear Research

Co-founder and CTO of Clear Research. Team under my supervision have developed a mobile visual commerce system for madora.co app. In particular:

  • bags and shoes recommendation engine, based on actual photos of things user like, or already have;
  • proprietary deep learning powered similarity search engine, based on user tap on camera photo;
  • proprietary algorithm, which discovers potential items to sell, based on photo and description from supplier web-page;
  • proprietary automatic image adjustment algorithm, so all things we sell, have standardized view and and the photos are of desired quality, even the items are from different suppliers
2012 — 2013

Worked as visiting researcher at Center of Machine perception at CTU in Prague. I have developed MODS - the state-of-the-art method for the wide baseline stereo matching under the extreme viewpoint change.

CTU in Prague
2011 — 2014
NTUU KPI

Was an Assistant Professor at National Technical University of Ukraine "KPI". I taught masters and undergraduate courses:

  • Image recognition
  • Satellite imagery processing
  • Microcontroller systems

Invited talks

2020
14.06.2020. Seattle, USA. CVPR 2020. Tutorial "RANSAC in 2020".

Talk "Benchmarking Robust Estimation Methods"

15.06.2020. Seattle, USA. CVPR 2020. Tutorial "Local features: from SIFT to differentiable methods".

Talk "Local features: from paper to practice".

2019
13.12.2019. Barcelona, Spain. Computer Vision Center of Autonomous University of Barcelona.

"Crafting and learning for image matching".

26.09.2019. Tallinn, Estonia. Veriff Computer Vision Meet-up.

"Deep-learned vs Handcrafted navigation".

06.07.2019. Odesa, Ukraine. Eastern European Conference on Computer Vision.

"Crafting and learning for image matching".

2018
29.03.2018. Prague, Czech Republic. Institute of Informatics, Academy of Science.

Convolutional neural networks from basics to the recent advances


Interviews


Selected publications

2021
Efficient Initial Pose-graph Generation for Global SfM

Authors: Daniel Barath, Dmytro Mishkin, Ivan Eichhardt, Ilia Shipachev, Jiri Matas

CVPR 2021.[pdf],[bib]

SfM-paper logo
2020
Image Matching across Wide Baselines: From Paper to Practice

Authors: Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, Eduard Trulls

IJCV 2020.[preprint],[bib], [sources], [web-site]

Image Matching Benchmark
Kornia: an open source differentiable computer vision library for PyTorch

Authors: Edgar Riba, Dmytro Mishkin, Daniel Ponsa, Ethan Rublee, Gary Bradski

WACV 2020.[pdf],[bib], [sources], [docs]

Kornia logo
2019
Benchmarking Classic and Learned Navigation in Complex 3D Environments

Authors: D. Mishkin, A. Dosovitskiy, V. Koltun

arXiv 2019.[preprint],[bib], [sources], [video], [web-site]

Navigation benchmark
2018
DeblurGAN: Blind Motion Deblurring Using Conditional Adversarial Networks

Authors: O.Kupyn, V.Budzan, M.Mykhailych, D. Mishkin, J.Matas

CVPR 2018.[preprint],[bib], [sources]

DeblurGAN
Repeatability Is Not Enough: Learning Affine Regions via Discriminability

Authors: D. Mishkin, F.Radenovic, J.Matas

ECCV 2018.[preprint],[bib], [sources]

AffNet
2017
Working hard to know your neighbor's margins: Local descriptor learning loss

Authors: A.Mishchuk, D. Mishkin, F.Radenovic, J.Matas

NeurIPS 2017. [preprint],[bib], [sources], [poster][slides]

HardNet
Systematic evaluation of convolution neural network advances on the Imagenet

Authors: D. Mishkin, N.Sergievskiy, J.Matas

CVIU 2017.[paper],[preprint],[bib], [sources], [slides]

Systematic CNN evaluation
2016
All you need is a good init

Authors: D. Mishkin and J.Matas

ICLR, 2016. [pdf], [bib], [poster]

LSUV logo
2015
MODS: Fast and Robust Method for Two-View Matching

Authors: D. Mishkin, J.Matas, M.Perdoch

CVIU 2015.[preprint],[paper],[bib], [sources]

MODS
WxBS: Wide Baseline Stereo Generalizations

Authors: D. Mishkin, M. Perdoch, J.Matas, K. Lenc

BMVC, 2015. [pdf], [bib], [poster]

WxBS

Teaching

2021
Intellectual property of AI companies

Joint lecture with LawLawLaw about how AI works and which additional intellectual property AI company generates compared to the non-AI information technology company. Invited by WIPO

WIPO seminar
Computer Vision Methods

Lectures on correspondence methods for computer vision: local feature detectors, descriptors and matching. [MPV course] at CTU in Prague. Slides and videos are available at the course page.

Computer Vision Methods course 2021
2020
Local features in computer vision

Private 3 days workshop about modern local detectors, descriptor for correspondence search.
For University of Ostrava Institute for Research and Applications of Fuzzy Modeling

WBS workshop 2020
Computer Vision Methods

I have re-designed and taught the practical part of the [MPV course] at CTU in Prague.

Computer Vision Methods course 2020
2018
Visual object tracking course

For Winter School at Ukrainian Catholic University. [Course link]

Visual object tracking course logo
2016
Intro into Deep Learning course

At Kyivstar Big Data School

Deep Learning course logo

Contact

The best way to contact me is email: ducha.aiki@gmail.com
I also write about computer vision and science on:
twitter
my new blog
medium (old blog)
blog devoted to wide baseline stereo.

In my free time I am playing kaggle (Kaggle Master), chess and practicing aikido. Yes, that's where my nickname comes from.