PhD Research
Recurrent Pattern Discovery: A Modern Deep-Learning Approach and its Applications.
​
I'm in the process of accurately defining the problem and formulating it. But in a nutshell, given a single-view image, I'm trying to discover recurring patterns in the image. The definition of "Recurring Patterns" (RP) is rather nuanced and that, probably, is the most challenging and intriguing aspect of the problem. An RP could be an object or a texture, could be in the foreground or in the background etc. The motivation behind this problem is that RPs are powerful in giving us much information about the 3D geometry of the scene. One can infer many things about the scene with just a single view. This link directs to the project page that has more information. Our paper can also be accessed here.
​
RECURRENT PATTERN DISCOVERY | DEEP LEARNING | COMPUTER VISION
Master's Research and Graduate Projects
Deep Learning object tracker for carotid artery wall motion
Motion estimation of the carotid wall is performed using a deep learning-based object tracker with a non-linear motion model. Extended Kalman Filter is designed to track the motion of the arterial wall accurately.
​
MOTION ESTIMATION | SIAMESE NETWORKS | EXTENDED KALMAN FILTER
Faster Search Algorithm for Speckle Tracking in Ultrasound Images
Exhaustive search is the most commonly used search technique for similarity matching in ultrasound images, which is slow and expensive. We propose to adopt a faster search algorithm called the Adaptive Rood Pattern Search (ARPS) along with sub-pixel accurate matching to enhance the performance.
​
MOTION ESTIMATION | ARPS | EXHAUSTIVE SEARCH
Augmented Reality Viewer
The project aimed at displaying artificial objects overlaid on images of a real 3D scene. Key aspects of the projects were, 3D reconstruction to recover point cloud, RANSAC-based plane fitting, creating and placing the virtual object in the scene, 3D-to-2D camera projection with Z-ordering
​
3D - RECONSTRUCTION | COLMAP | RANSAC | Z-ORDERING
Reinforced SiamFC
​
Limitations of SiamFC are overcome by incorporating Linear Kalman Filter and anchoring-based reference image update. The original tracker is made robust towards drastic changes of the reference object.
​
SIAMESE NETWORKS | LINEAR KALMAN FILTER | TEMPLATE UPDATE
Identification of Unit Lattice in Symmetric Wallpapers using Region Proposal
This project aimed at detecting and localizing unit lattices in symmetric images that do not contain foreground-background distinction. Key aspects of the project were, transfer learning, localization, NMS and t-SNE analysis
​
REGION PROPOSAL NETWORKS | SYMMETRY |
​
Undergrad Research
(This is a bit clumsy)
At PES Center for Intelligent Systems, my research was in machine learning, particularly on Artificial Neural Networks for sequential learning. I've explored ways to build systems to learn the pattern from raw signals. This process is derived from the principles of the human cognitive learning model. I've built systems with cognitive and meta-cognitive components for time series data to classify human behavioural attributes and detect peaks in complex signals such as PPG. In particular, I worked on the classification of signals in real-time applications – Human Activity Recognition and Peak Detection in Photoplethysomogram Signals. The uniqueness lies in that the signals are classified without any kind of feature extraction or complex pre-processing, unlike conventional methods. The system was further improved by adding a Meta-Cognitive component. Meta-Cognition is an integral part of the entire system that specifically obtains knowledge about what the network is actually learning. Further, the advantage of using ensemble networks is captured. Different signals are fed to different customizable networks. The algorithms are inspired by the Human Neural Processing System and Cognitive Learning.
​
Details
Research Fellow: @PES Center for Intelligent Systems, PES University
Working on machine learning problems under Prof. Koshy George,
Director, PES Center for Intelligent Systems, Dept. of Telecommunications, PES University.
Research Work:
1. Meta-Cognition for online sequential learning in Feedforward Neural Networks.
2. Online Peak Detection in PPG signals using a sequential learning algorithm.
3. Artificial Neural Networks for Detection and Localisation of Dynamite Fishing.
​
Research Project Intern: @Indian Institute of Science, DESE
Worked on Machine Learning problems and Embedded Systems under Prof. HS Jamadagni,
Dept. of Electronic Systems and Engineering.
​
Project 1: Power Line Communication
Provide better data rates from a single source for a large crowd, we split the single source of hotspot into ‘n’ number of sources. Power lines were used to communicate between the sources.
Project 2: Lab-on-Tab
Lab-on-Tab is a creative idea which uses mobile phones as a function generator and CRO. Simple circuits can be rigged up without actually going to labs.
​
Project 3: Dynamite fishing detection and localization.
The idea was to detect and localize a blast, if and where it occurred. Detection was done using artificial neural network and localization using RSSI based triangulation.
Project 4: Designing a creative and innovative game based on IoT built on Intel Edison board.
Project Intern: @Li2 – Innovations
Carried out projects on Embedded Systems.
I worked here on several basic robotic projects. Most of the projects involved interfacing different kinds of modules like Bluetooth, accelerometer, GSM, etc., with embedded boards like Arduino and MSP430. The projects included Bluetooth controlled bot, gesture-controlled bot, bot control using available speech recognition, maze solver, etc.