Research Projects

The rapid adoption of drone-based aerial imagery in applications such as urban planning, disaster management, and traffic analysis highlights the growing demand for robust AI models. However, developing domain-generalized models for drone-based imagery in the Indian context poses unique challenges due to the country’s diverse landscapes, infrastructures, and socio-structural complexities. From unplanned urban layouts and vast rural farmlands to hilly regions, India’s variability makes it difficult for models trained on foreign datasets to perform reliably. Our primary objective is to develop AI models capable of adapting seamlessly to India’s varied environments without retraining. To achieve this, we will build on existing annotated datasets such as VisDrone and synthetic datasets like SkyScenes, which provide a useful starting point. However, these alone are insufficient to address India’s unique complexities. To fill this gap, we will collect a large-scale drone-based aerial imagery dataset representing diverse Indian settings, with a small, carefully annotated subset for supervised learning. To enhance model robustness, we will develop novel augmentation techniques specifically tailored to Indian contexts, such as Inter-Domain Object Randomization (IDOR), which replaces semantic objects like vehicles and buildings across datasets while preserving contextual integrity. This will introduce controlled variability, enabling models to learn domain-invariant features. Additionally, we will leverage a semi-supervised learning (SSL) framework to utilize large volumes of unlabeled data, aligning predictions from weak and strong augmentations to improve performance while minimizing reliance on labeled data. By integrating IDOR, SSL, and data from diverse sources, we aim to create models that effectively address domain shifts and perform consistently across different environments. This project is expected to contribute significantly to aerial computer vision research by advancing domain generalization and semi-supervised learning methodologies. It will enable AI models to adapt to India’s complex and heterogeneous settings, ensuring reliable applications such as traffic monitoring, disaster response, and infrastructure assessment. By addressing these challenges, the project will promote smarter decision-making, resource optimization, and sustainable development in critical sectors, while laying the groundwork for further advancements in drone-based AI solutions tailored to diverse geographies.