Despite recent advances in video segmentation, many opportunities remain toimprove it using a variety of low and mid-level visual cues. We proposeimprovements to the leading streaming graph-based hierarchical videosegmentation (streamGBH) method based on early and mid level visual processing.The extensive experimental analysis of our approach validates the improvementof hierarchical supervoxel representation by incorporating motion and colorwith effective filtering. We also pose and illuminate some open questionstowards intermediate level video analysis as further extension to streamGBH. Weexploit the supervoxels as an initialization towards estimation of dominantaffine motion regions, followed by merging of such motion regions in order tohierarchically segment a video in a novel motion-segmentation framework whichaims at subsequent applications such as foreground recognition.
展开▼
