Options
2014
Master Thesis
Titel
FastStreamGBH: A fast streaming method for hierarchical spatio-temporal video segmentation
Abstract
Online video material has increased dramatically in recent years and so does research interest in video processing tasks such as video retrieval, summarization, event detection or video classification. Video segmentation describes a useful prerequisite for many of those high-level tasks since it helps to structure and reduce the huge amount of data in videos to be processed. However, current state-of-the-art video segmentation approaches show high computational and memory costs and often assume that the entire video clip fits into memory. Streaming approaches tackle this problem but produce only moderate results. In this thesis, FastStreamGBH, a hierarchical streaming algorithm for video segmentation based on the efficient graph-based approach of Xu et al., called StreamGBH, is proposed. F astStreamGBH incorporates an efficient spatio-temporal superpixel approach that enables a significant faster runtime than StreamGBH and makes use of an effective feature extension based on dense optical flow and textures. Benchmark evaluation shows that it competes with current state-of-the-art results that consider complete videos at once.
ThesisNote
Magdeburg, Univ., Master Thesis, 2014
Verlagsort
Magdeburg