Video Segmentation Tutorial
CVPR2014 Tutorial: Video Segmentation
Main Tutorial Material
Course Overview: Technology advances have made it quite simple to record a video nearly anytime, anywhere. As a result, we are drowning in video content---YouTube, for example, receives 72 hours of video uploaded every minute. In many applications, there is so much video content that a sufficient supply of human observers to manually tag or annotate the videos is unavailable. Furthermore, these unconstrained videos contain rich and variable content, are acquired from a broad set of devices, and are largely out of the scope of many known techniques; they hence present a rich problem-space for the next steps in computer vision research!In recent years, segmentation has emerged as a plausible first step in early video processing of unconstrained videos, without needing to make an assumption of a static background as earlier methods have. Video segmentation and over-segmentation, or more commonly supervoxel extraction, is a complementary early video processing step to the more traditional feature extraction, such as STIP and trajectories, and it extends the long history of image segmentation methods. This tutorial will survey and present the important models and algorithms for video segmentation. We will cover direct extensions of image segmentation methods through video-specific spatiotemporal and streaming methods. In addition to core methodological elements, the tutorial will also cover benchmark and evaluation of video segmentation as well as applications of video segmentation. Participants will be introduced to the details of these methods not only through traditional slide presentations but also example implementations through the LIBSVX library.
Course ScheduleThis is an estimate of the schedule...