ECE Seminar Series: Vishwanath Saragadam

About

Flyer: Beyond Large-Scale RGB: Harnessing Priors and Physics Across Data-Scarce Regimes

Abstract: The recent revolution in computer vision has largely been propelled by the availability of massive, annotated RGB datasets. However, in domains such as infrared and other wavelengths, such large-scale labeled data is virtually nonexistent. As the amount of data shrinks, purely data-driven paradigms break down, necessitating a shift toward strong domain priors and physical models. In this talk, I will explore three distinct regimes to address these challenges. First, when no training data is available, we rely on the implicit biases of neural networks to perform tasks directly from single observations. Second, when only a small amount of data is available, we leverage natural image statistics to regularize learning and improve generalization. Third, when models are well trained but noise is correlated, we incorporate the physics of image formation to restore robustness and accuracy. Together, these strategies provide a framework for expanding computer vision beyond the limits of RGB by bridging data scarcity through physics and structural priors.