Although the real world is dynamic, many techniques used to image/capture it are fundamentally sequential in nature. For example, capturing a high-dynamic range (HDR) image of a scene (which contains a wide range of illumination) without special hardware involves taking a set of sequential images at different exposures, each one measuring a small range of illumination. However, this approach has problems when reconstructing the HDR image of dynamic scenes with moving subjects, since the individual frames are not registered correctly. Problems like this appear in many research areas, from medical imaging to computer vision.
In this project, the PI and his team are developing a common framework that addresses artifacts from motion for a variety of different applications. Their key insight is that patch-based optimization can be used to handle motion inconsistencies without explicitly solving the challenging problem of accurate motion estimation. This produces results that are reconstructed from different inputs in a consistent manner without motion artifacts. The PI is exploring how this framework can be applied to several important research areas, from high-quality imaging to computer vision applications such as the reconstruction of dynamic scenes.
Improved capture of dynamic scenes has the potential to transform the way certain imaging procedures (such as medical imaging) are done. Scientific results of this work are disseminated through technical publications at top venues in the graphics/vision communities, and the PI plans to release the algorithms developed online. Finally, this project involves high school and under-represented students into the research effort.