Shivarth Rai

I am a B.Tech Computer Science undergraduate at Manipal Institute of Technology. Currently, I am a research intern at the Ubiquitous Lab, IIT BHU, where I am working under the guidance of Prof. Hari Prabhat Gupta on Low Power Wide Area Networks (LPWANs).

Previously, I interned as a Machine Learning Engineer at Kenvue, where I architected an end-to-end machine learning pipeline for cloud cost forecasting. Before that, I was a Project Intern at Samsung R&D Institute Bangalore through the PRISM program, where I developed a custom quantization methodology for image segmentation models.

My research primarily focuses on computer vision and computer networks. I have presented my work on Mamba-based dual domain learning for underwater image enhancement at the AAAI 2026 AI for Environmental Science Workshop, and my research on enhancing hazy wildlife imagery at the CVPR 2025 Computer Vision for Animals Workshop.

Email  /  GitHub  /  Google Scholar  /  LinkedIn  /  CV

Research

I'm interested in computer vision, machine learning, optimization and computer networks.

Hero-Mamba: Mamba-based Dual Domain Learning for Underwater Image Enhancement

Hero-Mamba enhances degraded underwater images using a novel Mamba-based network. It features parallel spatial-spectral dual-domain processing to decouple degradation factors, and a physics-guided ColorFusion block for high-fidelity color restoration.

Enhancing Hazy Wildlife Imagery: AnimalHaze3k and IncepDehazeGan

We introduce the AnimalHaze3k dataset and IncepDehazeGan, a novel generative architecture combining inception blocks with residual skip connections. Achieving state-of-the-art dehazing performance, our model improves downstream YOLOv11 wildlife detection mAP by 112%, providing ecologists reliable tools for robust population monitoring.

PCOS Diagnosis: Data-Driven Predictions and Workflow Enhancement

We developed a machine learning framework for PCOS diagnosis. By pairing Mutual Information feature selection with a Random Forest classifier, our model achieved 96% accuracy and 98% AUC ROC. We applied SHAP explainable AI to ensure transparent, clinically interpretable predictions.

Projects

Multi-Modal Price Prediction Framework

A multi-modal price prediction network utilizing Qwen 2.5-1.5B for feature enhancement, CLIP for joint image-text embedding generation and ensemble of MLP, RF and XGBoost for price prediction.

Course and Program Outcome Tracking and Analysis

Developed for the School of Computing, this MERN-stack platform digitizes Outcome-Based Education workflows, replacing manual Excel tracking. Key technical contributions include automated attainment calculations, role-based access, and bulk data integration.

Satellite Image Dehazing

A Generative Adversarial Network utilizing Inception blocks and multi-layer feature fusion for satellite image dehazing. It achieves state-of-the-art results on Hazelk and RICE datasets. Grad-CAM XAI results validate effective task localization.

Design and source code from Leonid Keselman's website