Information-Driven Imaging for Appearance Capture and Recognition

Jinwei Gu
Computer Scientist in the Center for Vision Technologies
SRI International
Wachman 1015D
Friday, November 15, 2013 - 11:00

One of the fundamental problems in imaging, computer vision, and visualization is the study of appearance. Despite of tremendous success, the popular data-driven approach is soon becoming incompetent for the study of many high-dimensional appearance phenomena, such as high-speed motion, light fields, and spectral images. In this talk, I argue that a better way to tackle high-dimensional appearance is the information driven approach. It has three key distinctions compared to the data-driven approach. (1) Physics-based appearance models and statistical priors of natural images should be actively incorporated to better constrain ill-posed problems. As an example, I will show how we designed algorithms to remove image artifacts caused by dirty lenses and thin occluders from videos. (2) Instead of passively recording 2D slices of appearance (which is both expensive and redundant), novel imaging systems can be designed to take coded, information-condensed projections, which can later be decoded computationally. In particular, I will describe how we designed coded exposure for CMOS image sensors for flexible space-time sampling and motion detection. (3) Low- level image acquisition should be directly connected with high-level tasks such as detection and recognition in order to measure low-dimensional, representative features. As an example, I will describe how we learned discriminative illumination for raw material classification.