University of Chicago
Sept 2022 - Dec 2023: Master's Program in Computer Science - Data Analytics
I work on high performance C++ code
where every microsecond matters
I am a CS Master’s student at the University of Chicago specializing in Data Analytics. Before joining UChicago, I worked as a low-latency trading systems developer, improving upon and expanding the trading infrastructure at AlphaGrep Securities. Prior to this, I graduated from the Indian Institute of Technology Delhi (IIT Delhi) with a dual degree (B.Tech + M.Tech) in Computer Science and Engineering. I am enthusiastic about computer vision and machine learning applications. I have worked on multiple research projects across IIT Delhi, Nanyang Technological University and Aalto University.
My research work has ranged from developing adversarial attacks to fool state-of-the-art object tracking networks to improving stereo odometry models for autonomous vehicles. I have also contributed to our fight against the ongoing pandemic by being one of the first few teams in India to develop a chest X-ray-based Covid-19 detection model - back when RT-PCR tests were in extremely short supply.
I have also been a teaching assistant for graduate and undergraduate-level courses at IIT Delhi. During my time at the university, I got the opportunity to visit 13 countries which helped me develop the ability to work in a cross-cultural global workforce.
Sept 2022 - Dec 2023: Master's Program in Computer Science - Data Analytics
July 2022 - Sept 2022: Associate - Infra Developer
July 2021 - June 2022: Senior Analyst - Infra Developer
July 2020 - June 2021: Analyst - Infra Developer
Jan 2019 - July 2020: Computer Vision Lab
Jan 2020 - July 2020: Programming Languages Course COL226
July 2019 - Dec 2019: Computer Vision Course COL780
May 2019 - July 2019: Quantitative Research Intern
May 2018 - July 2018: Research Assistant (Computer Vision)
Aug 2017 - Dec 2017: Semester Exchange Student - EMACS Department
May 2017 - July 2017: Research Assistant (Computer Vision)
July 2015 - July 2020: Bachelor's and Master's in Computer Science
With the advancements in deep learning models for computer vision applications, many adversarial attacks have been developed for image classification tasks. However, there are few instances of such attacks in the domain of video streams. We developed a new semi-targeted attack that can cause the failure of state-of-the-art SiamMask object tracking network.
Given any video sequence, my pipeline generated and superimposed adversarial noise, which is not visible to the human eye but causes the object tracker to drift away in the direction of our choice.
Due to the dearth of RT-PCR testing kits during the early stage of the pandemic, we developed a chest X-ray based COVID-19 detection model to prioritize the selection of patients for receiving RT-PCR tests and triaging in isolation wards.
On the publicly available covid-chestxray-dataset (Cohen et al., 2020) we were able to significantly improve upon the best performing Covid-Net model (Wang et al., 2020)
"The Elephant in the Room" (Rosenfelt et al., 2018) gives a good picture of how deep neural networks are susceptible to failure due to their context bias. Our project improved upon the traditional Lucas Kanade (LK) object tracking algorithm by learning new algebraic operations that are better suited for the task at hand.
As an aside, I improved upon the performance of standard LK by defining a masked region of attention based on semantic segmentation.
I designed a model to incorporate semantic segmentation class information to improve egomotion estimates of vehicle trajectory when using stereo-mounted RGB cameras.
I also incorporated this semantic model into feature-based VISO2 and direct image alignment based Stereo Direct Sparse Odometry models to improve translation error in 6 out of 11 KITTI VO/SLAM sequences.
Localization can be very challenging in an indoor environment where GPS signals are unreliable and we have textureless scenes, occlusions and repetitive structures.
We developed a model to compute 6 DoF pose based on an online RGB input along with an offline input database of image-pose pairs. I also created the challenging University dataset to improve upon the shortfalls of the 7Scenes dataset.
A. Mangal, S. Kalia, H. Rajgopal, K. Rangarajan, V. Namboodiri, S. Banerjee & C. Arora (2020)
CovidAID: COVID-19 Detection Using Chest X-Ray
arXiv preprint arXiv:2004.09803
[Link]
Z. Laskar, I. Melekhov, S. Kalia, & J. Kannala (2017)
Camera Relocalization by Computing Pairwise Relative Poses Using Convolutional Neural Network (oral)
IEEE International Conference on Computer Vision (ICCV) workshop: Geometry Meets Deep Learning; Venice
[Link]