Abstract:
Thermal cameras don’t just capture heat maps—they see a mix of emitted and reflected infrared radiation. In this talk, I’ll show how we can computationally disentangle these signals to enable better interpretation of scenes from thermal data. I’ll begin with a dual-band imaging system that leverages differences in spectral emissivity to separate emitted radiation (from heat transport) and reflected radiation (from light transport). This allows us to estimate both material emissivity and dynamic temperature in real-world settings–like hot liquid in glass cups with people moving in the background–where traditional assumptions for thermography fail. Then, we ask a fundamental question: can we infer an object’s shape by just observing its heat flow? I’ll show how passive heating from everyday light bulbs induces conductive heat flow that encodes surface geometry, even for transparent or translucent objects. By modeling this heat flow and solving for shape, we achieve millimeter-accurate 3D reconstructions without any assumptions on lighting, optical property or the shape. Together, these techniques transform the thermal spectrum into a rich source of visual cues, opening new frontiers for perception in challenging environments where RGB falls short.

Committee:
Srinivasa Narasimhan (Advisor)
Aswin C Sankaranarayanan
Ioannis Gkioulekas
Benjamin Attal