Vishal Chandra

I'm an ECE master's student at the University of Michigan, where my focus area is computer vision. I received my BSE in computer engineering also from Michigan. Here, I work at the Lab for Progress on hardware-accelerated DeepRL and on 3D object reconstruction. I'm also an advisor at Atlas Digital Consulting and a member of Michigan Investment Group.

This past summer, I was a research intern at MIT Lincoln Laboratory focusing on unsupervised representation learning for object re-identification across large video databases.

In my free time, I enjoy learning to fly planes, collecting watches, and taking film photos.

Email  /  Resume  /  Github

Research

My main research interest is in deep learning for computer vision. I'm also interested in graph learning, inverse problems, and scientific machine learning. As a computer engineer, I'm also always learning about efficient and alternative hardware for machine inference.

Training a DiGCN to fit black-box node rankings on directed graphs, then applying multi-scale Grad-CAM to visualize the learned decomposition of the black-box ranking function.

Formulated methods for 2D surface comparison for use in vision model loss function, based on rank correlation and ideas from differential geometry. (image source)

Generating the modulation transfer function (MTF) of an imager from various calibration targets, with derivations and survey of MTF literature. Presented to large R&D audiences.

Simulating and sculpting a class of toroidal ring linkages called kaleidocycles. Deformation of their basis units create interesting designs on the whole with both artistic and mechanical functionality.

Other Projects

Some of my other related projects on vision, AI, and hardware.

Low-cost, open-source, esp32-based camera system for tracking 3D markers in real-time. The goal was to make motion caption more accessible to small labs and game studios. Prototyped in Python, converted to Rust & C for fast data reconstruction and trasmission.

PDENets is a collection of Physics-Informed Neural Networks (PINN) and Fourier Neural Operators (FNO) trained to solve key families of partial differential equations including the heat equation and Helmholtz equation. This project was a survey of methods to better understand the field of AI for Science.

A survey of three methods for tomographic image reconstruction: filtered backprojection (FBP), algebraic reconstruction technique (ART), and neural network-based reconstruction. We step through each method, compare their performance, and ablate some of the key hyperparameters.

A simple SVM for predicting the outcomes of the 2022 March Madness tournament. Team vectors comprised of some team statistics and six expert power rankings including Pomeroy, Sagarin, Massey, and Colley. From Colley's method, I began to explore graph analytics applied to sports. (image source)

For the Michigan Daily

Template from Jon Barron