RESEARCH PROJECTS

Prof. Joycee Mekie | IIT Gandhinagar • Aug 2025 – Present

• Performed system-level quantization of depth estimation models (UniDepth-CNN, UniDepth-ViT, MASt3R) across 11 numerical formats including FP, Posit, and FixedPosit.
• Evaluated accuracy–efficiency trade-offs using SILog, AbsRel, and RMSE metrics, analyzing convergence under low-precision training.
• Benchmarked training and inference across five hardware accelerators (Eyeriss, Simba, EnX3D, Gemmini, AiM), demonstrating up to 4×–30× energy savings.

Prof. Himanshu Shekhar | IIT Gandhinagar • Jan 2026 – Present

• Developed a hybrid two-stage ULM pipeline combining neural microbubble localization with adaptive temporal tracking for sparse, overlapping, and noisy acquisitions.
• Built a physics-guided simulation engine for synthetic vascular structures and microbubble trajectories to enable scalable supervised training.
• Achieved super-resolution vascular reconstruction at reduced frame rates, preserving vessel continuity while lowering computational cost.

Prof. Nithin George | IIT Gandhinagar • Jan 2025 – Present

• Devised a Nearest Kronecker Product (NKP)-based adaptive GSC beamformer to reduce computational complexity in large microphone arrays.
• Achieved speedup in LMS/RLS updates by decomposing weight matrices into low-rank Kronecker factors.
• Demonstrated improved interference suppression with reduced execution time at the Undergraduate Research Showcase 2025.

Prof. Joycee Mekie | IIT Gandhinagar • Dec 2026 – Present

• Applied knowledge distillation on UniDepth-ViT to transfer performance from high-precision teacher models to low-precision student variants.
• Designed a reinforcement learning-based framework to assign optimal numerical formats to different model components (encoder, decoder, attention blocks).
• Explored mixed-precision configurations (e.g., FP8 encoder, FixedPosit16 decoder), demonstrating improved trade-offs over uniform quantization.

Prof. Himanshu Shekhar | IIT Gandhinagar • Feb 2026 – Present

• Developed OmniStrain-Net, an unsupervised deep learning framework integrating Vision Transformers (ViT) and Implicit Neural Representations (INR) for continuous tissue displacement tracking.
• Implemented a Physics-Informed Neural Network (PINN) loss function to enforce biomechanical constraints and ensure anatomical consistency.
• Currently benchmarking against SOTA models including ReUSENet, MUSSE-Net, and SMURF.

Prof. Shanmuganathan Raman | IIT Gandhinagar • Dec 2024 – Aug 2025

• Formulated a pipeline to perform inverse rendering of outdoor scenes from a single RGB image by estimating lighting (SH coefficients) using VQGAN-based feature extraction.
• Implemented 2D and 3D Gaussian Splatting for high-fidelity depth map and surface normal reconstruction.
• Evaluated geometry and lighting accuracy across diverse real-world outdoor datasets.

Prof. Shanmuganathan Raman | IIT Gandhinagar • Aug 2024 – Dec 2024

• Enhanced IBRNet’s robustness to large baseline variations using CLIP embeddings for generalized, high-quality feature representation.
• Designed a WaveNet-inspired encoder to compress CLIP’s 768-dimensional embeddings to 32 dimensions while retaining spatial context.
• Fine-tuned the encoder-decoder pipeline with pretrained weights, improving interpolation accuracy and multi-view rendering quality.