Skills

Computer Vision

90%

Machine Learning / Deep Learning

90%

Perception pipeline for Autonomous Vehicles

80%

C/C++

90%

Python

100%

Tensorflow

90%

Pytorch

90%

git

90%

Big Data System (Hadoop Spark, GraphX, Distributed Tensorflow)

90%

Coursework

 
 
 
 
 

At University of Wisconsin-Madison

Sep 2018 – May 2020
  • CS838 Learning Methods in Computer Vision
  • CS861 Theoretical Foundations of Machine Learning
  • CS760 Machine Learning
  • CS726 Non-linear Optimization
  • CS761 Mathematical Foundations of Machine Learning(Audit)
  • CS744 Big Data Systems
  • CS537 Introduction to Operating Systems
 
 
 
 
 

At National Institute of Technology Karnataka, Surathkal

Jul 2012 – May 2016
Relevant Coursework -

  • Computer Programming
  • Digital Signal Processing
  • Data structures and Algorithms
  • Applications of Signal Processing to Image and Video (CT Image Reconstruction Algorithms)
  • Digital Processing of Speech and Audio Signals
  • Linear Algebra
  • Advanced Calculus
  • Advanced Computer Architecture
  • Control Systems
  • Microprocessors
  • Numerical Methods
  • Probability, Random Variables and Random Processes

Experience

 
 
 
 
 

Perception (Computer Vision) Intern

ARGO AI

Jun 2019 – Aug 2019 Greater Pittsburgh Area

Developed a Deep Learning based 3D object detection pipeline using LiDAR data.

Key Features :

  • Developed point-cloud features using Graph Convolutions to improve detection of objects far from the Autonomous Vehicle.
  • Worked on reducing false positive detection rate.
  • Worked on improving performance of proposed detection algorithm on rare classes.
  • Proposed an objective function to mitigate the effects of correlated point-cloud sweeps (non-independently and identically distributed data) on the detection algorithm.
 
 
 
 
 

Graduate Student Researcher

University of Wisconsin-Madison

Sep 2018 – May 2019 Madison, Wisconsin

Automated Analysis of 3D CT images of Brain using Deep Learning.

Key Features :

  • Developed and implemented Deep Learning Models for detecting Intracranial Hemorrhage and its subtypes using 3D CT images of brain.
  • The network architecture is capable of utilizing any form of CT data - Strongly labelled (labels available for each slice), Weakly labelled (labels available for each 3D scan), and Unlabelled.

  • Code

    PDF Report


 
 
 
 
 

Firmware Engineer

Sandisk, a Western Digital Brand

Jul 2016 – Aug 2018 Bangalore, INDIA
Responsibilities Include :

  • Design, Development and Validation of Firmware for USB flash drives.
 
 
 
 
 

Summer Intern

Signal Processing Lab, National Aerospace Laboratories

May 2015 – Jul 2018 Bangalore, INDIA
Projects :

  • Active Noise Control : Implemented FxLMS algorithm on a microcontroller (based on TI TMS320C6748) unit which controlled noise control headphones.
  • Digital Video Watermarking for metadata embedding : Developed and implemented an algorithm for embedding metadata directly in compressed MPEG-4 video bit-stream.

Projects

 
 
 
 
 

Robust 3D Object Detection for Autonomous Vehicles using Sensor Fusion

PDF Report

Code

Sept 2019 – Dec 2019
In this work, we explore the viability of augmenting the true point cloud (data obtained from lidar sensor) with pseudo point-cloud data generated from a combination of monocular image and range view of true lidar data. In particular, we explore if such an augmentation technique aids the downstream task of 3D Object Detection.
 
 
 
 
 

SerFer: Serverless Inference of Machine Learning Models

PDF Report

Code

Jan 2019 – May 2019
SerFer is a framework for serving Deep Learning queries built on top on Serverless Computing Platform (AWS Lambda). It distributes the computation across several lambda functions by splitting both the Deep Learning Model, as well as the input query (Model Parallelism and Data Parallelism).
 
 
 
 
 

Word Embeddings for Fine Grained Sentiment Analysis

PDF

Sep 2018 – Nov 2018
Current methods to learn word embeddings result in similar representations for words with different connotations and hence, sentiment recognition systems relying on them perform poorly. To overcome this, I proposed a new approach to learn word representations, which improves the performance of sentiment recognition systems that rely on word vector representations.
 
 
 
 
 

Reconfigurable Architecture for Face Detection

PDF

Aug 2015 – Apr 2016
  • Developed a custom Face Detection Model suitable for hardware implementation using Viola-Jones Face Detection Framework. This model was trained and validated on MATLAB.
  • Designed and developed hardware accelerator for face detection using Xilinx Zedboard.
  • Languages and Tools used - Xilinx Vivado Design suite, C++, VHDL, and OpenCV.
 
 
 
 
 

Algebraic Reconstruction using SART and Total Variation De-noising

PDF

Jan 2015 – Apr 2015
  • Developed and implemented an improved version of SART Algorithm (simultaneous algebraic reconstruction technique, a computerized tomography algorithm) that uses total variation denoising to reduce noise levels in the images reconstructed from limited CT projection data.

Contact

  • divatekodand@wisc.edu
  • Madison, Wisconsin, USA