| layout | page |
|---|---|
| title | Valentin Peretroukhin |
| head_title | Home |
| subtitle | Fan of Robots & Humans |
From August 2020 to July 2021, I was a postdoctoral research associate in the Robust Robotics Group with Prof. Nicholas Roy at MIT CSAIL. I did my doctoral work in the STARS Lab (Space & Terrestrial Autonomous Robotic Systems) under the supervision of Prof. Jonathan Kelly. My Ph.D. dissertation was entitled Learned Improvements to the Visual Egomotion Pipeline.
I am currently...
- learning a ton about cryo-EM!
Recently...
- advising the organizing committee for the third annual debates workshop at ICRA 2021.
- I was web chair for Conference on Robot Learning (CoRL 2020).
- I co-chaired the 2020 RSS Pioneers workshop;
- I co-organized the Power On And Go (POGO) RSS (2020) workshop;
- my RSS work titled A Smooth Representation of SO(3) for Deep Rotation Learning with Uncertainty was awarded the Best Student Paper award, thank you to my incredible co-authors; (Official Page, arXiv preprint, YouTube)
I love reading poetry, philosophy and literature (especially Weil, Chesterton, Schopenhauer, Blake, Thomas, McCarthy, and Tolstoy), playing sports (especially soccer, hockey, tennis, and basketball), hiking & portaging, and all things general aviation (I received my Canadian private pilot's license in 2018).
{::nomarkdown}
Using a multi-head structure, we build a deep network that can regress uncertain estimates of rotation using quaternion averaging and noise injection. This work is currently published in the CVPR Workshop on Uncertainty in Visual Learning as a workshop paper.
<div class="unit half">
<div class="project">
<h4 class="project-title"><a href="https://github.com/utiasSTARS/dpc-net"><i class="fa fa-github" aria-hidden="true"></i> DPC-Net: Deep Pose Correction</a></h4>
<img src='/assets/img/projects/dpc.png' class='project-img'>
<p>Deep Pose Correction Network (DPC-Net) is a deep learning approach to achieving accurate, robust visual localization by correcting existing geometric, probabilistic estimators. It's been <a href="https://ieeexplore.ieee.org/document/8125095"><i class="fa fa-file-text" aria-hidden="true"></i> published</a> in Robotics and Automation Letters (and presented at ICRA 2018).</p>
</div>
</div>
Sun-BCNN (Bayesian Convolutional Neural Network) infers the full 3D sun direction from single RGB images (which can then be used to improve localization).
PROBE, and its follow-up, PROBE-GK (Generalized Kernels) is a heteroscedastic covariance model for image keypoints in visual odometry. Notably, PROBE-GK presents a novel Bayesian formulation for deriving a robust cost function for sparse keypoint reprojection errors.
Below you'll find a recent medley of things I really enjoy and love recommending to friends.
- Farther Away Jonathan Franzen on David Foster Wallace and loneliness.
- My Family's Slave Alex Tizon on Lola, a woman who helped raise him.
- Lear, Tolstoy and the Fool George Orwell on why Tolstoy wasn't a fan of Shakespeare.
- A Confession - Leo Tolstoy
- Sapiens: A Brief History of Humankind - Yuval Noah Harari
- Frankenstein; or, The Modern Prometheus - Mary Shelley
- Blood Meridian or the Evening Redness in the West - Cormac McCarthy
- The Ones Who Walk Away from Omelas - Ursula K. Le Guin
- Fallen Leaves: Last Words on Life, Love, War, and God - Will Durant