Dan Oneață

Picture of Dan Oneata

Hello and welcome to my web-page! My name is Dan and I'm a research scientist at the University Politehnica of Bucharest in the SpeeD lab. Previously, I have worked in the industry at Fordaq and Eloquentix. I have received my PhD in computer vision and machine learning from the Université Grenoble Alpes, where I was fortunate to work under the supervision of Cordelia Schmid and Jakob Verbeek. Before that I have received an MSc in artificial intelligence from the University of Edinburgh and I did my dissertation under Iain Murray. My research interests include general machine learning techniques and its applications to computer vision, speech, and natural language processing. I also enjoy reading about category theory and its applications.

You can find below some of my publications and presentations. The complete list of publications is available on my Google scholar profile.


Translation symbol
Conference paperpaper · code
Multilingual multimodal learning with machine translated text
Empirical Methods in Natural Language Processing, 2022
Chen Qiu, Dan Oneață, Stella Frank and Desmond Elliott.
Journal paperpaper
Keyword localisation in untranscribed speech using visually grounded speech models
IEEE Journal of Selected Topics in Signal Processing, 2022
Kayode Olaleye, Dan Oneață and Herman Kamper.
Conference paperpaper · slides
Improving Multimodal Speech Recognition by Data Augmentation and Speech Representations
IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2022
Dan Oneață and Horia Cucu.
Journal paperpaper
Multimodal speech recognition for unmanned aerial vehicles
Computers & Electrical Engineering, 2021
Dan Oneață and Horia Cucu.
Conference paperpaper · slides
An Evaluation of Word-level Confidence Estimation for End-to-end Automatic Speech Recognition
IEEE Spoken Language Technology, 2021
Dan Oneață, Alexandru Caranica, Adriana Stan and Horia Cucu.
Conference paperpaper
Data-filtering methods for Self-training of Automatic Speech Recognition Systems
IEEE Spoken Language Technology, 2021
Lucian Georgescu, Cristian Manolache, Dan Oneață, Horia Cucu and Corneliu Burileanu.
Conference paperpaper · slides · presentation · code
Revisiting SincNet: An Evaluation of Feature and Network Hyperparameters for Speaker Recognition
European Signal Processing Conference (EUSIPCO), 2020.
Dan Oneață, Lucian Georgescu, Horia Cucu, Dragoș Burileanu and Corneliu Burileanu.
Conference paperpaper · poster · webpage
Kite: Automatic speech recognition for unmanned aerial vehicles
Interspeech, 2019.
Dan Oneață and Horia Cucu.
Competition reportpaper · code
The Quo Vadis submission at Traffic4cast 2019
arXiv, 2019.
Dan Oneață, Cosmin George Alexandru, Marius Stănescu, Octavian Pascu, Alexandru Măgan, Adrian Postelnicu and Horia Cucu.
Optical flow stabilization
Journal paperpaper · code
A Robust and Efficient Video Representation for Action Recognition
International Journal of Computer Vision, Springer Verlag, 2015, pp.1-20.
Heng Wang, Dan Oneață, Jakob Verbeek and Cordelia Schmid.
Spatio temporal proposals
Conference paperpaper · poster · code
Spatio-Temporal Object Detection Proposals
European Conference on Computer Vision, 2014.
Dan Oneață, Jérôme Revaud, Jakob Verbeek and Cordelia Schmid.
Approximate normalization of Fisher vectors
Conference paperpaper
Efficient Action Localization with Approximately Normalized Fisher Vectors
IEEE Conference on Computer Vision and Pattern Recognition, 2014.
Dan Oneață, Jakob Verbeek and Cordelia Schmid.
Action recognition results
Conference paperpaper · poster · code
Action and Event Recognition with Fisher Vectors on a Compact Feature Set
IEEE International Conference on Computer Vision, 2013.
Dan Oneață, Jakob Verbeek and Cordelia Schmid.


University of Grenoble logo
Doctoral thesisthesis
Robust and Efficient Models for Action Recognition and Localization
Université Grenoble Alpes, 2015.
Dan Oneață
University of Edinburgh logo
Master's thesisthesis · slides
Fast Low-Rank Metric Learning
University of Edinburgh, 2011.
Dan Oneață
Politehnica University of Bucharest logo
Bachelor's thesis (in Romanian)thesis
Detecția de fețe utilizând metoda Viola-Jones
"Politechnica" University of Bucharest, 2010.
Dan Oneață


Bucharest CV logo
Paper presentationslides
Deep Image Prior
Bucharest CV reading group · Bucharest, Romania · February, 2018
Bucharest FP logo
Coding dojoslides · code
A simple Sudoku solver in Haskell
Bucharest FP meet-up · Bucharest, Romania · June, 2017
SpeeD lab logo
Short course ◇ lectures 1 · 2 · 3 · 4
An Introduction to Machine Learning
Course at SpeeD lab · Bucharest, Romania · June, 2016
Eloquentix logo
An Introduction to Machine Learning
Eloquentix annual meeting · Brașov, Romania · October, 2015
Meet up logo
Keynote format
A Tutorial on Action Recognition in Video
Grenoble Data Science meet-up · Grenoble, France · March, 2015
Thumos'14 logo
Talk ◇ slides 1 · 2
LEAR–INRIA Submission at Thumos’14
THUMOS'14 workshop · Zürich, Switzerland · September, 2014
NIST logo
TRECVid'13 workshop · Gaithersburg, USA · November, 2013
Graphical models
Paper presentationslides
A Thousand Frames in Just a Few Words
LEAR–XRCE reading group · Grenoble, France · August, 2013
Spectral learning
Paper presentationslides · extra
A Spectral Algorithm for Learning Hidden Markov Models
LEAR–XRCE reading group · Grenoble, France · March, 2013
Stochastic gradient with Langevin dynamics
Paper presentationslides
Posterior Sampling with Stochastic Gradients
LEAR reading group · Grenoble, France · July, 2012
Latent temporal model
Paper presentationslides
Learning Latent Temporal Structure for Complex Event Detection
LEAR reading group · Grenoble, France · August, 2012


being a collection of notes.
being a declarative drawing library for Python, developed with Sasha Rush. The library's desgin is inspired by Haskell's diagrams and Scala's doodle.
Vorbis: Multimodal speech recognition
being my post-doctoral project in which we attempt to improve automatic speech recognition by incorporating the visual context of what the speaker sees.
NHLA grading tool
being an user interface implemented in Javascript that allows NHLA grading of wooden boards. The tool was developed for the Neural Grader project at Fordaq.
Code gistlink
Fisher vectors
being a minimal Python implementation of the Fisher vectors method based on numpy and scikit-learn. For an application of Fisher vectors for action recognition in videos, check out the fv4a repository.
Code gistlink
being an exploration into applicative functors and a solution to a puzzle by Conor McBride. The snippet was highlighted as an example of functor-oriented programming. For more exercises on the functorial kit checkout this other exercise from Conor McBride and Jeremy Gibbons's paper, Calculating Functional Programs.
Diacritics restoration
being a machine learning challenge I set up for the second-year students. The goal was to add the missing diacritics marks for a given text in Romanian.
Haskell playground
being the place where I go down the rabbit hole into the Haskell world and I toy with free monoids, comondas, anamorphisms and others.
Algebra of programming
being an incomplete cheatsheet of equational properties in category theory to ease my way through the Algebra of Programming book.