Guha Balakrishnan

Face




Email:
balakg@mit.edu

Address:
32-D466 Vassar St.
Cambridge, MA
02139

About Me

Hello! I am a postdoctoral researcher in Bill Freeman’s group at MIT. I am interested in developing algorithms that learn or reason through image synthesis. Data-driven synthesis techniques have rapidly improved in recent years, and are now capable of simulating realistic faces, outdoor scenes, and even videos of human actions. I want to leverage such techniques to make models that can (1) learn by simulating hypothetical visual scenarios, and (2) causally analyze the decision processes of other vision algorithms.

I am also driven by important vision problems in applied fields, particularly medicine. Examples of my medical-related work include non-invasive vitals signs measurement from videos, and fast learning-based registration of medical images. Please take a look below at some of my projects.

I completed my Master’s (2013) and Ph.D. (2018) degrees in the EECS department at MIT, advised by John Guttag and Frédo Durand. My Ph.D. dissertation was titled Analyzing and Synthesizing Deformations in Image Datasets. My Master's dissertation focused on analyzing vitals signs from subtle body motions in video recordings.

I completed my B.S.E. (2011) degrees in Computer Science and Computer Engineering at the University of Michigan, Ann Arbor.

Some Projects By Category

Image/Video Synthesis


Counterfactual faces
Framework using learning-based image synthesis methods and human annotators to measure causal effects of semantic attributes on analysis systems.

Towards Causal Benchmarking of Vision Systems using Counterfactual Image Synthesis.
G. Balakrishnan, P. Perona (Amazon Web Services)
(current work, stay tuned for paper!)
Art
Learning-based approach that synthesizes plausible painting timelapses given only a finished painting.

Painting Many Pasts: Synthesizing Time Lapse Videos of Paintings.
A. Zhao, G. Balakrishnan, K.M. Lewis, F. Durand, J. Guttag, A. V. Dalca
CVPR 2020
vdp-teaser
Learning-based approach that synthesizes plausible images/videos from projected, lower-dimensional observations.
Visual Deprojection: Probabilistic Recovery of Collapsed Dimensions in Images and Videos.
G. Balakrishnan, A. V. Dalca, A. Zhao, J. Guttag, F. Durand,
W.T. Freeman.
ICCV 2019
paper

Press:
mit news, engadget, petapixel, science daily, techspot

pose teaser
A modular neural network design that synthesizes high-quality images of human poses in new poses.

Synthesizing Images of Humans in Unseen Poses.
G. Balakrishnan, A. Zhao, A. V. Dalca, F. Durand, J. Guttag.
CVPR, 2018 (oral)
paper, code

Press:
csail news
videodiff-teaser
A fast, robust method to compare the spatiotemporal differences between a pair of videos depicting similar motions.

Video diff: highlighting differences between similar actions in videos.
G. Balakrishnan,  F. Durand, J. Guttag.
Siggraph Asia, 2015 (oral)
project page

Medical Image Analysis


voxelmorph
We introduce Voxelmorph, an unsupervised method that registers medical images as accurately as the state-of-the-art, but orders of magnitude more quickly.

An Unsupervised Learning Model for Deformable Medical Image Registration.
G. Balakrishnan, A. Zhao, M. R. Sabuncu, J. Guttag, A. V. Dalca.
CVPR, 2018
paper, code

Press:
mit news, nvidia, innovatorsmag, extremetech, semiengineering, physicsworld, electronics360, radiologybusiness, futurelab3d, techexplorist, healthimaging, acceleratingbiz, bionengineeringtoday
tmi-teaser
Presents unsupervised and semi-supervised versions of VoxelMorph, along with substantial experimental analysis.

VoxelMorph: A Learning Framework for Deformable Medical Image Registration.
G. Balakrishnan, A. Zhao, M.R. Sabuncu, J. Guttag, A. V. Dalca.
IEEE TMI: Transactions in Medical Imaging, 2019
paper
augmentation
A trainable augmentation method that learns independent models of spatial and appearance transforms, and uses them to synthesize new training examples.

Data augmentation using learned transforms for one-shot medical image segmentation.
A. Zhao, G. Balakrishnan, F. Durand, J. Guttag, A. V. Dalca.
CVPR, 2019 (oral)
paper, code
miccai
A probabilistic, diffeomorphic version of VoxelMorph with the same accuracy and speed as the original.

Unsupervised Learning for Fast Probabilistic Diffeomorphic Registration.
A. V. Dalca, G. Balakrishnan, J. Guttag, and M. R. Sabuncu.
MICCAI, 2018 (oral, best paper runner-up)
paper, code
media-gm
A thorough theoretical framework for probabilistic VoxelMorph, an extension for surface registration, and extensive experiments.

Unsupervised Learning of Probabilistic Diffeomorphic Registration for Images and Surfaces.
A.V. Dalca, G. Balakrishnan, J. Guttag, M.R. Sabuncu.
Medical Image Analysis (MedIA), 2019
paper

Healthcare


pulse teaser
Method that extracts accurate pulse signals from video recordings by measuring subtle head movements.

Detecting Pulse from Head Motions in Video.
G. Balakrishnan,  F. Durand, J. Guttag.
CVPR, 2013.
project page

Press:
mit news, phys.org, newscientist, cnet, fox news, sciencenews, kurzweilai
ataxia
Method that uses pose estimation and motion features to predict ratings of neurological movement disorder severity from videos of motor exams.

A Video-Based Method for Objectively Rating Ataxia.
R. Jaroensri, A. Zhao, G. Balakrishnan, D. Lo, J. Schmahmann, J. Guttag, F. Durand.
MLHC, 2017.
paper
eeg
Method that reduces the number of time-series data examples needed to accurately train epileptic seizure detectors.
Scalable personalization of long-term physiological monitoring: Active learning methodologies for epileptic seizure onset detection.
G. Balakrishnan, Z. Syed.
AISTATS, 2012.
paper