Alongside my doctoral research, I spent several years in the industry, learning to apply computer vision to a
wide range of real-world problems.
I was with
Mojow Autonomous Solutions
for over two years, researching ways to
automate farming operations,
including rock detection, implement folding, and autonomous navigation in a self-driving tractor.
I spent another year at
ACAMP,
working on
human and animal detection,
and chain-link fence damage detection for an
autonomous security ATV.
I also did a 6-month internship at
ISL Adapt,
where I worked on
vehicle and pedestrian tracking
for road traffic analysis to improve the design of intersections and roundabouts.
My research interests include all aspects of computer vision and machine learning, though I am particularly
interested in the application of deep learning for visual recognition tasks like
object detection, segmentation
and tracking.
Projects
Nucleus segmentation and
classification in histopathology images: (Jan 2026 - Present): Supervisors:
Gilbert
Bigras,
Cross Cancer Institute,
Nilanjan Ray,
University of Alberta
Adapting
QuPath,
CVAT,
and
FiftyOne
to work together for efficient interactive semi-automated annotation of
histopathology images with pathologist-in-the-loop
End-to-End Multi-Object Tracking with Sequence Prediction: (July 2020 - Oct 2023): Supervisor: Nilanjan Ray, University of Alberta
Adapting a couple of dense video captioning methods for multi-object tracking
Creating a language of trajectories so that a sentence corresponds to a trajectory and a
multi-sentence caption corresponds to the set of all trajectories in the video clip
Deep MDP: (May 2019 - April 2021): Supervisor: Nilanjan Ray, University of Alberta
Attempted to improve the performance of the MDP system and remove its heavy dependency on
handcrafted features and heuristics by incorporating deep learning
Created an elegant abstraction to generalize the MDP framework to incorporate deep learning-based
trackers, feature extractors and classifiers
Added support for cell tracking in microscopy images
Fence Damage Detection with Laser Imaging: (September 2020 - March 2021): Supervisor: Ken Brizel, ACAMP
Developed and analysed algorithms to detect small holes in fences using laser line detection based
fence mask extraction and computer vision based hole detection
Animal and Human Detection with Deep Learning : (May-August 2018, January-April 2019): Supervisors:
Nilanjan Ray, University of Alberta,
Nehla
Ghouaiel, ACAMP
Trained and tested 6 deep CNN based object detectors - RFCN, SSD, YOLO(1,2,3) and 3 variants of
Faster RCNN - Inceptionv4, ResNet101, NAS - for detecting humans and 8 different types of animals
Trained and tested several instance segmentation methods like Mask RCNN and Sharpmask to generate
realistic augmented data
Used Tensorflow, Darknet and Pytorch for implementation.
Created a ROS compatible application ready for deployment on an autonomous ATV.
Developed efficient ways to get compressed video data from the ATV to the GPU server.
Helped develop a fence damage detection algorithm using laser based line detection and unsupervised
learning.
Vehicle and Pedestrian Tracking for Road Traffic Analysis : (March 2017-Jan 2018):
Supervisor: Nilanjan Ray, University of
Alberta, Douglas Hallett,
ISL Adapt
Created a system for tracking vehicles and pedestrians in real time traffic videos of roundabouts
and intersections captured from UAV and pole mounted cameras.
Created a highly efficient python implementation of the MDP online multi object tracker
using opencv, cuda and numpy to obtain a tenfold speed increase over the Matlab version and make it
suitable for real time deployment.
Adapted an open source PyQt based graphical tool to perform integrated detection, tracking and semi
automated labelling.
Modular Tracking Framework: A Unified Approach to Registration based Tracking : (Jan 2015-Feb 2017):
Supervisor: Martin Jagersand,
University of Alberta
Devised a novel way to study registration based trackers by decomposing them into three constituent
sub modules: appearance model, search method and state space model
Empirically tested different combinations of sub modules to produce novel and interesting
observations and insights missing in each method's original paper due to limited testing.
Proposed two new appearance models and several new composite search methods to achieve a new state
of the art in registration tracking.
Conducted experiments using four large datasets: TMT, UCSB, LinTrack and PAMI - with over
100,000 frames in all to ensure their statistical significance.
Created an open source tracking framework that can be used to reproduce all results and, owing to
its efficient C++ implementation, also address practical tracking requirements.
Using Deep Learning techniques for 3D Object Recognition : (Jan 2014-April 2014): Supervisor: Dale Schuurmans, University
of Alberta
Developed and tested several variants of the 6 hidden layer LeNet-5 architecture for classification
of handwritten digits and 3D real world objects represented by the MNIST and NORB datasets
respectively.
Used Matlab deep learning toolbox for implementation.
Multi Object Recognition in an Indoor Environment : (July 2012-Nov 2012): Supervisor: U.S. Tiwary, IIIT
Allahabad
Implemented an object recognition system based on a hierarchical model of visual processing in the
human visual cortex
System learns both invariance to object transformations and selectivity towards specific object
features in an incremental manner, across several layers
Matlab was used for implementation and a custom created database of several common indoor objects
was used for testing the system
Object Based Change Detection in Multi Temporal High Resolution Satellite Images : (July 2011-Nov
2011): Supervisor:
Anupam Agrawal,
IIIT Allahabad
Analyzed, debugged and documented an existing work based on multi feature fusion based change
detection and also extended it with a hierarchical clustering based approach
The results of these were compared with each other as well as with existing pixel based methods
The system was tested on Google Earth images of IIIT-A taken in 2004 and 2008 and was found to give
high accuracies of over 95% with both methodologies
