cv

MASc

University of Waterloo

Systems Design Engineering

• Thesis: Whale detection from Aerial images

B.Tech + M.Tech

Indian institute of technology (IIT) Kharagpur

Industrial Engineering with micro specialization in Optimization theory and Applications

• Thesis: Gym exercises posture correction using Computer Vision

Research Intern

Philips Innovation Campus, Philips India Limited

• Enhanced the quality of MRI image annotations by manually revising over 184 images, achieving higher inter-annotator agreement.
• Improved deep learning model performance by 45% through realistic data augmentation for 3D Brain MRI images using TorchIO.
• Boosted software quality with unit testing and debugging, conducted a literature survey on active learning algorithms, and developed a customizable PyTorch framework for models and datasets.

Remote research Intern

under Prof. Panagiotis Kalnis, Computer Science | King Abdullah University of Science and Technology (KAUST), Saudi Arabia

• Simulated a federated learning environment in PyTorch to train models on decentralized edge devices without data exchange, ensuring data privacy.
• Reduced communication bottlenecks by implementing lossy compression schemes on local models during client-server communications.
• Implemented and benchmarked the state-of-the-art Compressed Loopless Gradient Descent (L2GD) algorithm on real datasets.

Foundation model for SAR imagery

• Developed a foundation model using Masked autoencoders on SAR images using large amounts of SAR data

AutoIce Competition

• Developed a multi-task UNet model which won the first place in the Autoice competition hosted by the European Space Agency (ESA)
• We were awarded €3000
• This work was published in the Cryosphere journal

Pixel level segmentation from Region level labels

• Developed a novel loss function that facilitates the training of pixel-level segmentation models using only region-level labels, bypassing the need for labor-intensive pixel-wise annotation
• This method outperforms the Auto ICE competition model and has significantly better accuracy(12% average) for minority classes
• This work was published at ICLR2024 in ML4RS

Whale detection from Large Aerial Images

• Developed an innovative training pipeline tailored for small object detection, achieving an impressive 90% precision and recall rate
• Shipped the model into production using Labelstudio with a support for online learning. This framework is currently used by the DFO (Fisheries and Oceans Canada)

Region-Level Labels in Ice Charts Can Produce Pixel-Level Segmentation for Sea Ice Types

In Proceedings of the ICLR 2024 Machine Learning for Remote Sensing (ML4RS) Workshop

We introduce a novel loss function to enable learning from regional labels to produce pixel-level segmentations. Leveraging the AI4Arctic Sea Ice Challenge Dataset, our method outperforms the fully supervised U-Net benchmark, the top solution of the AutoIce challenge

MMSeaIce: a collection of techniques for improving sea ice mapping with a multi-task model

in The Cryosphere

In this work we outline the solution that helped us win the AutoIce competition organised by European Space Agengy (ESA)

Simulating Realistic MRI variations to Improve Deep Learning model and visual explanations using GradCAM.

in IEEE CCEM

To overcome limited data availability, we implement realistic data augmentation to generate synthetic 3D volumetric data and Grad-CAM to visualize the model's prediction

International Master’s Award of Excellence (IMAE)

University of Waterloo

Awarded CA$12500 for academic excellence

Lalit Chugh and Hira Ahuja International Graduate Scholarship

Lalit Chugh and Hira Ahuja, UWaterloo

Awarded CA$10000 for academic excellence

Inspire Scholarship

Department of Science and Technology, Govt. of India