|
Some of my recent research topics
1. Learning-driven deformable
models
|
|
I work on topology
independent solutions for segmenting objects with texture patterns of any
scale, using geometric models and active contours driven by
discriminative pixel and region classification, such as the Conditional
Random Fields (CRFs). I integrate region and edge information as image
driven terms, whereas the probabilistic shape and internal (smoothness)
terms use representations similar to the level-set based methods. The
model evolution is solved as a MAP problem, where the target conditional
probability is decomposed into internal (local shape smoothness) terms
and image-driven terms.
In my
approach
(i) I
integrate learning-based classification with deformable models in a
tightly coupled framework,
(ii) I handle
local feature variations by updating the model interior statistics and
processing at different spatial scales,
(iii) I achieve
the independence from the image topology, and
(iv) I tackle
multi-modal segmentation, by incorporating features from different
modalities. For example, for the segmentation of brain tumors I use
Magnetic Resonance (MR) and Chemical Shift Imaging (CSI) data: I
integrate the MR intensity information the registered CSI spectral
information of the tissue, to evolve the deformable model based on the
appearance and the chemical properties of the tissue.
|
  
Textured object segmentation in clutter
Segmentation of the tear (film and menisci)
from Optical Coherence Tomography (OCT)
 
Segmentation of the Geographic Atrophy in
dry Age-related Macular Degeneration from OCT
 
Segmentation of the Geographic Atrophy in dry
Age-related
Macular Degeneration from Autofluorescence images
 
Segmentation of retinal tumor in the mouse model of
retinoblastoma
 
Brain tumor segmentation from MR and spectroscopy
|
|
2. Microscopic imaging and
applications in physiology, neuroscience, molecular and cellular biology
|
|
I have been working on computational problems in biology and
neuroscience, including segmentation, tracking and representation of structures
and features of interest, as well as recognition of patterns. The model
organisms I have been working on are C.
elegans, Drosophila and mice, in collaboration with researchers from
the respective domains.
|
Simultaneous tracking and locomotion classification of
C. elegans for modeling the phenotype expression of ion channel mutations
Detection of the foetal mouse brain stem regions that
contribute to the respiration rhythms from (a) the brain stem functional
(fluorescence) imaging and (b) the respiration rhythms, as observed from
the electrical recordings from the spinal cord
|
|
3. Object recognition in
cluttered environments
|
|
I have been working an
active learning approach for visual multiple object class recognition,
using a graphical model for discriminative learning. I have named this
graphical model “collaborative” Conditional Random Field, because it
infers class posteriors in instances of occlusion and missing information
by assessing the joint appearance and geometric assortment of neighboring
sites. The model can handle scenes containing multiple classes and
multiple objects inherently, while using the confidence of its
predictions to enforce label uniformity in areas where evidence supports
similarity. This method uses classification uncertainty to dynamically
select new training
samples to retrain the discriminative formulations used in the
CRF.
In Situ
Ichthyoplankton Imaging System (ISIIS)
I apply the above framework to a marine biology problem, namely, the
recognition of specimens in ichthyoplankton imagery.
ISIIS is capable of imaging large
water volumes with the goal of quantifying even rare plankton in situ.
ISIIS produces very high resolution imagery for extended periods of times
necessitating automated data analysis and recognition. Since the
identification and quantification of a large number of organisms is
required, I am developing fully
automated software for detection and recognition of organisms of interest, based on the collaborative CRF. This
framework aims at (i) the detection of all organisms of interest
automatically, directly from the raw data, while filtering out noise and
out-of-focus instances, (ii) the extraction and modeling of the appearance of each
segmented organism, and (iii) the
fully automated recognition of all the detected organisms simultaneously,
using appearance and topology information. What differentiates this work
from existing approaches (e.g. SIPPER) is that it is capable to recognize
simultaneously all existing organisms of interest without any human interaction, from the data
acquisition to the recognition stage.
|
Simultaneous multiple object recognition in clutter
Data acquired with ISIIS vehicle (blue
boxes=ground-truth
regions; red boxes=detected candidate specimens)
ISIIS-from the raw data to the specimens recognition:
over- or under-segmented regions are recognized efficiently, while the
segmented false positives are successfully rejected
|
|
4. Human motion analysis and
recognition
|
|
3D Hand tracking
from monocular sequences and American Sign Language (ASL) recognition.
I have been in a close collaboration with C. Neidle (http://www.bu.edu/asllrp/carol.html),
and D. Metaxas (http://www.cs.rutgers.edu/~dnm), Rutgers University,
working on the automated recognition of American Sign Language (ASL) from
3D visual cues. Our work on sign recognition is
distinguished from previous approaches in the following aspects.
(a)
The use of learning-based coupling
between temporal 3D hand tracking and static hand configuration
estimation. There are generally two major approaches in hand tracking:
(i) temporal methods that use both temporal and static information from
the input sequence, and (ii) static methods, which handle each frame
separately, using only static information and prior knowledge from
mapping 3D hand configurations to 2D image features. Temporal methods
provide high accuracy and low complexity, exploiting the continuity
constraints over time, but they often lose track and cannot recover
easily, due to the hands’ fast articulated motions. Static approaches do
not suffer from error accumulation over time, giving independent solutions
at each instance, but their accuracy depends on the generality of the
prior knowledge they utilize. To track the hand articulations in 3D, we have been developing a data-driven
dynamic coupling between static and temporal methods, which overcomes the
limitations of the two aforementioned techniques.
(b)
The use of domain knowledge in sign
recognition process. Despite the fact that no prior approaches to ASL
recognition have taken into account linguistic properties of signing, an
understanding of certain linguistic categorizations of signs, such as the
fundamental differences in the composition of distinct types of signs, is
essential to any machine learning-based recognition strategy. In our
framework, we are developing a parallel
learning framework that consists of two levels: at the first level, we
aim at recognizing linguistically important movements and hand shape
features, such as fingerspelling,
and at the second level we aim at recognizing the sign in the specific
linguistic category. This bi-level formulation tackles a crucial issue in
signing recognition, namely it links linguistic properties with observed
movement features.
Our
research takes advantage of a large annotated corpus that is collected
from native ASL signers. Annotations are carried out using SignStream(TM),
a program developed by the Dr. Neidle’s team for linguistic analysis of
visual language data.
|
|
|