Mikhail Breslav
Objective
14 Buswell Street Apt 611 – Boston, MA
H (973) 508 0126 • B breslav@bu.edu • Í www.breslav.org
United States Citizen
To do research and development in computer vision and machine learning in a collaborative setting with the goal of
advancing a meaningful application.
Education
Boston University
Ph.D Candidate in Computer Science (Expected: June 2016). Advisor: Margrit Betke
Boston, MA
8/2010–current
The Pennsylvania State University
M.S Electrical Engineering, GPA 3.80/4.00
University Park, PA
8/2008–6/2010
B.S Electrical Engineering, GPA 3.60/4.00
8/2004–5/2008
Research Experience
Image & Video Computing Group, Boston University
Research Assistant
Boston, MA
8/2010–current
{ Designed, implemented, and evaluated novel computer vision and machine learning based algorithms for the problem of 2D and 3D
pose estimation in multi-view datasets.
{ Published and presented research work at leading peer-reviewed computer vision conferences.
Deep Learning
{ Examining how well pretrained Convolutional Neural Networks (CNNs) can represent images of Moths.
{ Experimenting with adaptation of CNNs for the problem of pose regression.
Discovering Useful Parts for Pose Estimation in Sparsely Annotated Datasets
{ Demonstrated that traditional part based models, like Pictorial Structures, can be improved by leveraging parts that are automatically
discovered from unannotated regions of training images. Useful parts are discovered by clustering patches and subsequently scoring how
predictive the cluster is of landmarks of interest. Votes for landmark positions are integrated and used to obtain more accurate appearance
likelihood terms.
{ Experimentally validated that this approach outperforms both traditional part based models and existing works from Biology on landmark
localization of Hawkmoths in high resolution visible light video.
{ Gained experience with features including: HOG, SIFT, SIFT variants based on BOW and Spatial Pyramid BOW, clustering algorithms:
greedy, k-means, affinity propagation, discriminative clustering, and classifiers: LDA, SVM.
3D Pose Estimation of Bats in the Wild
{ Demonstrated for the first time in the literature that 3D pose estimates of bats flying in the wild can be generated automatically,
even in very low resolution thermal infrared video data.
{ Formulated 3D pose estimation over time as a Markov Random Field (MRF) optimization problem. The 3D pose at any time
is constrained by an appearance term, multi-view geometry term, and a temporal smoothness term. The appearance term was
obtained by first developing a 3D model of the Mexican free-tailed bat in Blender and using it to capture how a bat’s appearance
relates to its 3D pose.
Camera Geometry
{ Helped to setup multiple thermal infrared cameras for a bat colony capture held in Texas. Gained experience with traditional
camera calibration approaches and have implemented routines to: estimate the fundamental and essential matrices from point
correspondences, decompose the fundamental and essential matrices into pairs of camera matrices, and triangulate 2D points to
obtain 3D reconstructions.
Image Processing Lab, The Pennsylvania State University
Research Assistant
University Park, PA
8/2008-6/2010
{ Handled integration of a new Endobronchial Ultrasound System (EBUS) into the lab. After setting up the unit and learning to
operate it, I designed novel underwater experiments for evaluating the 3D reconstruction of 2D ultrasound images. I prototyped
and evaluated several 3D voxel based reconstruction algorithms.
Work Experience
Nokia HERE
Intern
Berkeley, CA
5/2014–8/2014
{ Used C++ to implement and evaluate geometric algorithms for generating depth maps from large (≈ 107 ) lidar-based point clouds.
Custom modules were developed which built on top of the Point Cloud Library (PCL). Delivered a presentation showing the
feasibility of a commercial application that uses the generated depth maps.
MIT Lincoln Laboratory
Intern
{ Used GNU Radio to implement an end to end communications system over software defined radios
Lexington, MA
6/2010–8/2010
Lockheed Martin
Intern
Owego, NY
6/2008–8/2008
{ Wrote algorithms to integrate elevation data into a large Matlab-based signal location system
Computer skills
Programming: C++, Matlab, Python, C, Java, C#, LaTeX, GCC
Libraries: OpenCV, PCL, Boost, OpenGL, MEX
OS: Windows, Mac, Linux, Unix
3D Graphics: Blender, Unity3D
Publications
M. Breslav, T. L. Hedrick, S. Sclaroff, M. Betke. "Discovering Useful Parts for Pose Estimation in Sparsely Annotated
Datasets". WACV 2016.
M. Breslav, N. W. Fuller, S. Sclaroff, M. Betke. "3D Pose Estimation of Bats in the Wild". WACV 2014.
M. Breslav, N. W. Fuller, M. Betke. "Vision System for Wing Beat Analysis of Bats in the Wild". ICPR workshop,
November 2012.
X. Zang, M. Breslav, W. E. Higgins. "3D Segmentation and Reconstruction of Endobronchial Ultrasound". SPIE Medical
Imaging, 2013.
Breslav, Mikhail. 2010 "3D Reconstruction of 2D Endobronchial Ultrasound". Master’s Thesis, The Pennsylvania State
University Department of Electrical Engineering.
Teaching Experience
Boston University Teaching Fellow
CS112 - Intro to Computer Science 2 - Spring 2016 and Fall 2014
CS585 - Image and Video Computing - Fall 2015 and Fall 2012
CS237 - Probability in Computing - Fall 2013
CS108 - Application Programming in Python - Fall 2011
CS132 - Geometric Algorithms - Fall 2010
Penn State Teaching Assistant
EE453 - Digital Signal Processing - Spring 2010
EE455 - Image Processing - Fall 2009
CSE486 - Computer Vision - Spring 2009 and Fall 2008
Professional Activities
Memberships
Student Member of the Institute of Electrical and Electronics Engineers (IEEE)
Peer Review
IEEE Conference on Computer Vision and Pattern Recognition (with Advisor)
Languages
English: Native Speaker Russian: Fluent Spanish: Basic
Awards and Honors
William L. and Barbara A. Keefauver Scholarship, The Pennsylvania State University
2009
Inducted to Eta Kappa Nu Electrical and Computer Engineering Honor Society, The Pennsylvania State University
2008