Andrew Gallagher
Research Interests
- Computer vision
- Social Media
- Machine Learning
- Images of People
Selected Research Projects
What's in a Name: First Names as Facial Attributes
The idea is to guess what first name a person might have, just by looking at his/her face. To do this, we built classifiers that try to distinguish between 100 different common first names (about half for each gender). Then, when a new face is encountered, the classifiers indicate the likelihood that the person has a given first name.
The underlying idea is that first names are not given to babies at random, but are influenced by many factors (e.g., gender, current popular names, ethnicity, socio-economic status) that are also correlated somewhat with facial appearance.
Jigsaw Puzzles with Pieces of Unknown Orientation
This paper decsribes advances for computationally solving jigsaw puzzles. Pieces from multiple puzzles can be mixed together, and the algorithm can still solve the puzzles.
Jointly
Estimating Demographics and Height with a Calibrated Camera
In this paper, we propose that measuring people and estimating demographics are
problems that should be solved jointly. We use anthropometric statistics (i.e.
statistics on measurements of the human body) with a calibrated camera to
provide accurate height measurements and infer gender and age too.
Understanding
Images of Groups of People
One common image is the "Group Shot" where several family members or
friends appear together in an image. In this paper, we study the spatial
juxtaposition of faces in an image. We introduce features that encapsulate the
structure of the group. These features provide useful context for demographic
recognition, find images of groups eating, and even locate the horizon!
Estimating
Age, Gender, and Identity using First Name Priors
People have an amazing ability to interpret still images of people. We innately
put our lifetime of experience in social situations to use for interpreting
images. We want to give computer vision algorithms this same social
context. In this paper, we merge demographic data related
to first names in the United States with gender and age classifiers
to identify people in images based on their names alone.
Clothing Cosegmentation
for Recognizing People
Consumer photographers usually capture multiple images during a single event.
Because people's clothing does not generally change during an event, clothing
can help identify people (sometimes more reliably than faces!) The
challenge is to achieve accurate clothing segmentation. In this paper we
explore the relationship between clothing segmentation and person recognition.
Also, we introduce the Gallagher Collection
Person Dataset for other researchers to use.
Image
Authentication by Detecting Traces of Demosaicing
How can we distinguish real images (from a camera)
from fakes (either computer-generated or constructed by compositing parts of
other images)? Images from a digital camera contain traces of demosaicing (the process used to construct a full-color
image from an image sensor with a color filter array). We demonstrate an
elegant approach for locally detecting demosaicing.
We show the best yet performance at distinguishing real photographs from
computer graphics on an established test set. And we show accurate localization
of forged image regions in tampered images.
Decision Making in NFL Football
One of my interests is in probabilistic decision-making in sports. In NFL football, coaches have some hard decisions to make. For example, on a fourth down, should the team go for it, punt the ball, or attempt a field goal. Here, I analyze a dataset of 430,022 plays to answer that question. Basically, coaches are too conservative with their decisions, costing real points that they could have had. Note this is a draft version.
MathTime: A Free Arithmetic Game for Windows
I made a game for practicing math problems where a photo is hidden by a set of math problems. Each time a problem is solved, it reveals a part of the photo.