FEEL (Framework for Emotion Evaluation) is a large-scale benchmarking framework for emotion recognition from physiological signals. It unifies evaluation across 19 publicly available datasets containing electrodermal activity (EDA) and photoplethysmography (PPG) signals, enabling standardized comparisons and robust generalization analysis. FEEL benchmarks 16 diverse architectures from traditional machine learning to deep and self-supervised models under both within-dataset and cross-dataset settings. The results highlight that contrastive language signal pretraining (CLSP) achieves state-of-the-art performance, and handcrafted feature-based architectures remain competitive in noisy, low-resource environments. By systematically analyzing generalization across datasets, devices, and labeling strategies, FEEL offers a unified foundation to advance emotion-aware and mental health technologies.

           DOI                                                                                                                                                                                                NeurIPS 2025

EEVR (Emotion Elicitation in Virtual Reality) is a novel dataset for emotion recognition tasks, combining physiological signals (EDA, PPG) from 360-degree VR videos with emotional descriptions from semi-structured qualitative interviews. Featuring data from 37 participants, it pairs text with signals, offering richer context for emotion classification. We also introduce the Contrastive Language Signal Pre-training (CLSP) algorithm, which enhances emotion recognition by learning from both signals and text. Our results show improved performance over baseline models and strong zero-shot transferability, demonstrating the effectiveness of contextualized learning.

           DOI                                                                                                                                                                                                NeurIPS 2024

        AnnoSense is a stakeholder-informed framework designed to improve everyday emotion data collection for AI systems. As emotion sensing moves from controlled labs to real-world environments, ensuring high-quality, accurately annotated data has become increasingly challenging. To address this, we engaged 119 stakeholders, including the public, mental health professionals, and emotion AI experts, through surveys, interviews, and focus groups. Their insights shaped 15 actionable guidelines that promote clarity, usability, and ethical practices in emotion annotation. Evaluated by 25 domain experts, AnnoSense offers a practical foundation for advancing emotion-aware AI and improving data collection in real-world contexts.                                                                                                                                                                                                                    

                                               DOI                                                                                                                                                                                                ACM IMWUT 2025

The proof-of-concept paper examines the growing interest among children in evaluating diverse interactive elements through a book on the Indian freedom struggle. I contributed to a research project on designing virtual and augmented reality tools. I used visual projection mapping and gesture tracking to build interactive learning experiences for children

Springer series                                                                                                                                                                                 ACM India-HCI 2023

“Studies in Computational Intelligence"