|
CURRENT PROJECTS
|
Connections
between SIFT and Biological Vision
The
determination of dominant orientation at a given image location is
formulated as a decision-theoretic question. This leads to a novel measure for
the dominance of a given orientation, which is similar to that used by
SIFT. We then show that the new measure can be computed with a network that
implements the sequence of operations of the standard neurophysiological
model of V1. We verify that the network units exhibit trademark properties
of V1 neurons, such as cross-orientation suppression, sparseness,
independence, off-looking strategy etc. In this work, we establish a
connection between SIFT (and similar histogram based features that are
popular in computer vision) and biological vision. This connection provides
a justification for the success of SIFT-like features and tries to
establish the sequence of operations that are statistically optimal for
low-level image representation. It also shows that this optimal sequence of
operations is performed by the visual cortex.
|
|
Fully
Automated Design of Fast Object Detectors
We consider
the problem of learning a real-time object detector cascade with no manual
supervision and propose a two-step solution. A discriminant saliency based
procedure is used to automatically collect a high quality training set by
cleaning and cropping a dataset obtained from a web search. The second step
involved a new optimization procedure for automated cascade design that achieves
an optimal trade-off between classification accuracy and detection speed. Demos include an
interface that enables the user to train a real-time object detector for
any desired object, and a smartphone application that runs real-time
detectors designed using our system for 10 real world objects.
|
|
How
Language and Visual Saliency Jointly Guide Eye Movements in the Perception
of Complex Scenes
Integrating visual
and linguistic information e.g. combining the visual properties with a
spoken description of a scene, is an extremely important skill that humans
use almost constantly. Despite this, how linguistic information influences
the processing of complex visual input is still unclear. In this collaboratory
project, we use methods from psychology and linguistics as well as computer
vision in order to understand (1) How linguistic descriptions of scenes
drive eye movements (2) how the visual aspects of scenes interact with
language processing and (3) how we can derive computational models that
predict the interaction between eye movements and language.
|
|
Shape
Coding in Area V4
The ventral visual
pathway in primates is often modeled as a feedforward hierarchy that is
selective to increasingly complex patterns (from simple oriented edges to
objects) and is increasingly invariant. Specifically, it has been shown
from neural recordings of macaque monkeys that Area V4 is selective to
complex properties of boundaries such as orientation and curvature in
addition to being invariant to spatial translation. This project involves
the derivation of a computational model of area V4 as a natural extension
to our biologically plausible model of V1. We also investigate applications
to computer vision tasks.
|
PUBLICATIONS
A
biologically plausible network for the Computation of Orientation Dominance
Kritika Muralidharan and Nuno Vasconcelos
Proc. Neural Information Processing Systems(NIPS),
2010 (spotlight presentation)
On
the connections between SIFT and Biological Vision
Kritika Muralidharan and Nuno Vasconcelos
Frontiers in Systems Neuroscience, March 2010.
|
