Tasking on the Natural Statistics of Pictures and Videos

摘要

I will discuss a variety of topics related to the statistics of pictures and videos of the real world, how they relate to visual perception, and most importantly how they can be used to accomplish perceptually relevant picture-processing and video-processing tasks. Underlying my talk is the thesis that pictures and videos of the real world obey lawful statistical behavior that can be modeled. These models supply useful statistical priors that can be used to define or regularize the solutions to a variety of visual problems. I will address the application of these models to such visual tasks as visual quality assessment, efficient video data delivery in rate-adaptive network environments, face detection in difficult environments, and depth estimation from a single image. I will describe the ongoing work in LIVE in these areas and pose some general problems to be solved in the future.