SINCE 2022
HOVER Inc.
Computer Vision
- Wide Baseline Stereo
- Benchmarks & Honest Evaluation
- Classical CV Made Differentiable
- Eastern-European CV Infrastructure
COMPUTER VISION RESEARCHER · HOVER INC. · CTU PRAGUE
Dmytro Mishkin
Prague, Czechia
He married classical geometry to deep learning, then built the benchmarks that keep computer vision honest.
Dmytro does computer vision at HOVER and remains a researcher at CTU Prague, where he co-maintains Kornia, the differentiable 'OpenCV in PyTorch.' His PhD on wide baseline stereo — which won CTU's Dean's and Rector's Prizes — ran from crafted matchers like MODS and WxBS to learned ones like HardNet and AffNet. He also builds Ukrainian science: he co-founded the Szkocka research group, named for the 1930s Lviv mathematicians of the 'Scottish Book,' and takes his handle ducha.aiki from aikido.
- 6.3KCVPR 2020 tutorial views
- 5,400CVPR 2022 workshop views
- 268Medium followers
- 152Claps on his most-read essay
HAPPENING NOW
STANDING TODAY
SINCE 2015
Czech Technical University in Prague
Computer Vision Researcher
HOW THEY THINK4 BELIEFS
Four convictions run through the work, each stated in his own words.
THE IMPACT
Pull one out: the papers, the benchmarks, and the code that carries them.
EVERYTHING ELSE · 12 WORKS
THE TIMELINE
One problem since 2012 — matching photos the world says shouldn't match — crafted first, then learned, then measured honestly.
The whole record, newest first – tap any moment and Aster tells the story.
2022–PRESENT
Industry, joint matching, and asking how we even know the state of the art
Since 2022 Mishkin does computer vision at HOVER while staying a researcher at CTU Prague. With Daniel Barath and the ETH Zürich group he builds StereoGlue (ECCV 2024), which matches features and estimates geometry jointly, and the Large Scale Homography Benchmark (CVPR 2023). His talks turn to the meta-question — 'How to move the goalposts' (ECCV 2024) and 'What is the state-of-the-art in 2025 and how can we be sure?'
“I do research for curiosity and living.”
2019–2021
Benchmarks, Kornia, and building Ukrainian science
The integration years. Mishkin ships Kornia (WACV 2020) so classical and learned code can train together, and the 'From Paper to Practice' image-matching benchmark (IJCV 2020) that seeds the Image Matching Challenge. In parallel he pours energy into Ukraine — co-founding the EECVC conference, joining the Ministry AI committee, and finishing a prize-winning PhD in December 2021.
“expertise is required to decide, which should and, more importantly, should not be implemented.”
2015–2018
The Prague PhD: crafting and learning for local features
Under Prof. Jiří Matas at CTU's Center for Machine Perception, Mishkin builds the two halves of his thesis. On the learned side: LSUV initialization (ICLR 2016), HardNet's descriptor loss (NeurIPS 2017), and AffNet's affine shapes (ECCV 2018). On the community side, he co-founds the Szkocka NGO (2016), under which HardNet is listed as a group publication.
2011–2014
Kyiv: professor, machine perception, and a Bay-Area startup
Before the PhD, Mishkin taught AI and image recognition as an assistant professor at KPI in Kyiv, then joined CTU Prague's Center for Machine Perception on a research fellowship — where he started MODS, the matcher that would define his career. In 2014 he co-founded Clear Research in the Bay Area, shipping the same matching technology as product.