|
Lekha Revankar I'm a 3rd year PhD student at Cornell University, where I work on computer vision and machine learning advised by Prof. Bharath Hariharan and Prof. Kavita Bala Previously, I did my M.S. and B.S at the University of Maryland, College Park under the supervision of Prof. Ming Lin. Currently I am taking a year to do research full time at Planet in San Francisco! Email / Resume / Google Scholar / GitHub |
|
ResearchOverall I work on multi-modal foundation models for visual understanding. I'm currently building models that can understand and reason about temporal events through visual, textual and temporal information. I have experience in creating large-scale multi-modal systems and working with unstructured data. I'm particularly interested in the interdisciplinary applications of these techniques, collaborating with researchers in archaeology, grape pathology, and other domains to address real-world challenges. |
Select PublicationsPlease check out my Google Scholar for a complete list of publications. |
|
MONITRS: Monitoring with Intermittent High-Resolution Sensing
Shreelekha Revankar, Utkarsh Mall, Cheng Perng Phoo, Kavita Bala, Bharath Hariharan NeurIPS D&B, 2025 Spotlight paper / code / project page We introduce MONITRS, pairing temporal satellite imagery of ~10,000 natural disasters with natural language descriptions from news articles. Fine-tuned models achieve 88.69% accuracy in disaster classification, outperforming state-of-the-art remote sensing models by over 38%, establishing a new benchmark for automated disaster monitoring. |
|
Scale-Aware Recognition in Satellite Images under Resource Constraints
Shreelekha Revankar, Cheng Perng Phoo, Utkarsh Mall, Bharath Hariharan, Kavita Bala ICLR, 2025 paper / code / project page We introduce a new approach to scale-aware recognition in satellite imagery under resource constraints. Our approach allows one to accurately detect various concepts using a fixed budget of HR imagery, outperforming entirely HR baselines by more than 26% mAP in zero-shot techniques, and more than 8% mAP in supervised techniques using 5× fewer HR images. |
News
 
|
|
Website template inspired by Jon Barron |