Color recovery in scattering media
With Derya Akkaynak I have been working out when color recovery in scattering media is a well-posed problem at all. Our first preprint gives conditions that bound the uncertainty in recovered color. A follow-up preprint drops the assumption of a perfect hyperspectral camera with known spectral sensitivities, and gives an idealized model in which the uncertainty tends to zero as image resolution grows and the scene provides enough depth variation. It also states conditions under which the required segmentation into regions of constant inherent radiance can be determined algorithmically. These are sufficient conditions in an idealized setting, not a working restoration system.
Camera spectral response
Spectral sensitivity describes what a camera actually measures, and it is normally recovered in a lab. In Spectral Sensitivity Estimation Without a Camera we estimate it from public color-calibration data instead, without access to the camera. The method combines numerical optimization with an autoencoder trained on 51 measured sensitivity functions, evaluated by leave-one-out validation on those functions. The project released reusable software, a citable dataset, and predicted sensitivities for more than 1,000 camera models.
Fast exact algorithms
My work in algorithms is about exact computation with proved cost. With coauthors I developed a list decoder for fully general algebraic-geometry codes with the best known complexity, quasi-linear encoding and unencoding for structured families such as Hermitian and norm-trace codes, and deterministic algorithms for Popov and Hermite normal forms of rectangular polynomial matrices that match the previously best randomized Popov bound and improve earlier Hermite bounds. A related line of work gives quasi-linear bivariate multipoint evaluation and interpolation for generic point sets, with performance that degrades smoothly away from genericity and an application to modular composition.
Product and research software
Before my PhD I worked as a developer. At RailCOMPLETE I built C# modeling software for railway engineers, implementing product features from domain requirements in a Scrum team with code review. At Schlumberger I developed and tested C++ modeling tools for geoscientists. That period set the habits I still bring to research code: version control, automated tests, review, and results a colleague can reproduce.
