Tanuj Sur

I'm a Data Scientist at Micron Technology in Hyderabad, India, where I focus on problems at the intersection of semiconductor industry and AI. Additionally, I am a computer vision researcher with publication at A and A* peer reviewed journals and conferences.

Previously, I have been fortunate to work under the guidance of Dr. Subhasis Chaudhury and Dr. Biplab Banerjee at the Indian Institute of Technology Bombay in the field of 3D Vision and Lifelong learning. Additionally, I have served as a research assistant at the Indian Institute of Technology Delhi under the supervision of Dr. Brejesh Lall working on Data-efficient learning and Biomedical Vision problems.

I did my MSc. in Data Science from Chennai Mathematical Institute and my BSc. in Statistics from St. Xavier's College (Autonomous), Kolkata (under Calcutta University).

Email  /  Bio  /  Scholar  /  Github

Research

I'm interested in computer vision problems related to egocentric vision, video understanding, 3D reconstruction and scene understanding. My research till now has been primarily in meta learning, incremental learning and data efficient learning. My research works have been highlighted below.

3D point cloud recognition and segmentation are vital for scene understanding and motion planning in autonomous driving and robotics. Traditional methods struggle with dynamic environments and new categories, highlighting the need for integrating few-shot learning and class incremental learning in 3D point cloud segmentation. Despite some existing approaches, there is a notable gap in research specifically focused on Few-Shot Class Incremental Learning (FSCIL) for this purpose.

Continual learning in deep neural networks faces challenges like catastrophic forgetting and issues with shifting data distributions. This paper addresses these challenges within the Unsupervised Incremental Domain Adaptation (UIDA) framework, where the initial source domain is labeled while subsequent domains are not. Existing methods often lack cross-domain generalization and adaptation. We propose UIDAPLE, a new approach that uses a unified prompt across all domains, leveraging the CLIP foundation model to eliminate the need for separate domain treatments.

The text discusses advancements in deep learning for image semantic segmentation but highlights the challenge of requiring large annotated datasets. It notes a shift towards Few-Shot Learning (FSL), particularly in medical domains where pixel-level annotations are costly. The proposed solution is Regularized Prototypical Neural Ordinary Differential Equation (R-PNODE), which utilizes Neural-ODEs alongside cluster and consistency losses to improve Few-Shot Segmentation (FSS) of organs.

Few-shot Learning (FSL) methods are increasingly used in data-scarce environments, particularly in the medical field where obtaining annotations is costly. However, Deep Neural Networks, especially in FSL, are vulnerable to adversarial attacks, which can critically affect clinical decisions. This paper presents a framework aimed at enhancing the adversarial robustness of few-shot segmentation models in the medical domain.

Mathematical expressions are essential not only for numerical calculations but also for enhancing clarity in discussions and documentation. They are present in various software engineering artifacts like source code, documentation, and bug reports. Research shows that these expressions impact the accuracy of commit message generation tools. To support future research, a dataset of bug reports with annotated mathematical expressions has been created and shared, along with a tool called MEDSEA to identify these expressions.

Projects

I have highlighted some of the projects that I have been a part of.

We curate a dataset comprising of egocentric driver attention maps in simulated evironments built upon CARLA and Unreal Engine. We use Meta's Eye Tracking ARIA glasses for registering user gaze.

Research Internships

Services

The code for this website is adopted from Jon Barron's website.