Rajat Sahay

I am a final year Master's student at RIT, majoring in Data Science.

I am currently working as a Graduate Research Assistant at the RIT Center for Human Aware AI under the supervision of Prof. Andreas Savakis. My work involves understanding representation learning in visual models through the lens of unsupervised domain adaptation.

I have had the pleasure of conducting research alongside researchers and scientists across the world. Prior to this, I have been associated with the Juno Mission at NASA JPL, IRIS-HEP at Princeton, the LIVIA Lab at ETS Montreal, ULR France, CamCann Smart Systems, and IIT Indore.

Email  /  CV  /  Google Scholar  /  Blog  /  Twitter  /  Github  /  LinkedIn

I'm interested in computer vision, machine learning, perceptual interfaces, and image processing. Much of my research is about inferring the physical world (shape, motion, color, light, etc) from images and videos. I also like to focus on interdisciplinary research, applying vision-based methods to innovatiely tackle problems in different fields. More recently, I have been focusing on problems at the intersection of computer vision and AI fairness, ethics and algorithmic biases as well as their applications in the real world.

We formulate Single Object Tracking (SOT) as an online incremental learning problem and develop a new method for dynamic sample selection and memory replay, preventing template corruption.

We propose a model adaptation method for Siamese trackers. We employ a generative model to produce a synthetic template from the object search regions of several previous frames, rather than directly using the tracker output.

We explore graph segmentation as a precursor to a larger deep learning pipelines, improving accuracy and efficiency of downstream tasks by using the underlying geometric structure of data. We apply our research to track particles generated during proton-proton collisions in the Large Hadron Collider (LHC).

We enhance Prototypical Networks with additional self-supervised modules to learn Euclidian embeddings and classify handwritten Urdu characters using a minimal number of examples.

We leverage colorization models to semantically attack images generating low-frequency adversarial perturbations while changing colors in ambigous areas. We were able to demonstrate that this method was successful in getting Vision Transformers (ViTs) to misclassify image labels.

Leveraged optical flow methods to map trajectories of random, fast-moving particles in constrained environments. Used statistical methods like Lucas-Kanade Optical Flow combined with Hungarian Algorithm to track and predict feature frames based on inputs by previous frames.

[ Fun Website ] [ Template ]