Vasiliy Karasev email github
I am a senior software engineer on the computer vision team at Zoox.

I mostly live in New York City.
Before graduating from UCLA in Fall 2015, I was a PhD student in Electrical Engineering, where I worked with Prof. Stefano Soatto. My interests are in ''value of information'' problems, optimization, and their applications in computer vision. During my studies I had the opportunity to visit Honda Research Institute and Samsung Display labs.
Prior to that (in reverse chronological order) I worked with Prof. Jin Hyung Lee (UCLA EE), Prof. Avideh Zakhor (UCB EECS), and Prof. Martin Banks (UCB Vision Science).

email: karasev00 at gmail dot com


Intent-Aware Long-Term Prediction of Pedestrian Motion V. Karasev, A. Ayvaci, B. Heisele, and S. Soatto. In ICRA 2016. [pdf] [video]

Forecasting what pedestrians intend to do is easier if they behave rationally. We show how this assumption simplifies motion prediction in the assisted/autonomous driving setting.

Causal Video Object Segmentation from Persistence of Occlusions B. Taylor, V. Karasev, and S. Soatto. In CVPR 2015. [pdf] [project page]

We show how to exploit occlusions to discover salient objects in video.

Active Frame, Location, and Detector Selection for Automated and Manual Video Annotation V. Karasev, A. Ravichandran, and S. Soatto. In CVPR 2014. [pdf] [project page]

How to choose where and which detectors to run (or m-turks to query), if we can run only a few of them? We answer this question using the ''information gathering'' framework, and show results on semantic video segmentation.

Controlled Recognition Bounds for Visual Learning and Exploration V. Karasev, A. Chiuso, and S. Soatto. In NIPS 2012. [pdf]

We show how to (greedily) search for an unknown object, under occlusions, quantization-scale, and uncertain measurements.

Compressed sensing enabled ultra-high resolution optogenetic functional magnetic resonance imaging (ofMRI) J.H. Lee, J. Li, V. Karasev. In Society for Neuroscience Meeting, 2011. [pdf]

Compressed sensing in dynamic MRI is normally used to improve temporal resolution. Here we used it to improve spatial resolution. Reconstruction used the simplest possible TV regularization.
This work was also a part of my MS project.

Temporal presentation protocols in stereoscopic displays: Flicker visibility, perceived motion, and perceived depth D. Hoffman, V.Karasev, and M. Banks. In Journal of the Society for Information Display, 2011. [pdf]

We studied how different 3D display presentation methods affect flicker, motion artifacts, and errors in perceived depth.