Video Deconfounding: Hearing-Aid Inspired Video Enhancement

摘要

We introduce a set of techniques for selective amplification of video subjects that are largely hidden from view by confounding light paths, such as daylight reflections from tinted automobile windows, reflections from windows and screens when imaging outdoor scenes at night from indoors, and reflected light from fluid surfaces. In these situations, the subject of interest is represented by only a small fraction of the light being captured at each pixel. We show that enhancement approaches commonly employed for selective amplification of speech in audio applications can serve as the basis for selective amplification of hidden objects in video streams.