An Online Hierarchical Supervoxel Segmentation Algorithm Based On Uniform Entropy Slice

Supervoxel plays a significant role in video segmentation task in which moving objects in the video are detected and their external boundaries are obtained. This is due to the observation that hierarchies of supervoxel can process the video into a multiscale decomposition with more information for later analysis than other approaches using the concept of supervoxels to video segmentation. Most available supervoxel video segmentation methods require the whole voxels in the video to be loaded into memory first before processing can occur which is impossible even for medium sized videos containing 100 frames, of reasonable complex background and foreground of moving objects. The results of deploying these more traditional approaches show that they tend to under-segment at lower levels and over-segment at high levels, so it is a challenge to use the results to analyze the video, e.g., in detection of objects. In this paper, we present an algorithm to overcome these limitations. Our method, called StreamUES, is based on the Uniform Entropy Slice (UES). It not only can seek a selection of supervoxels by using the post hoc feature criterion such as optical flow, but also can work on streaming modem use a clip of frames at one time until the end of the video clip is reached. Our results indicate the StreamUES can work very well with a challenging dataset viz., VSB 100, which we used to evaluate the proposed StreamUES technique.