Spatiotemporal segmentation for stereoscopic video

摘要

We propose an efficient and practical spatiotemporal segmentation method for stereoscopic videos. In our framework, an improved graph-based video segmentation method is firstly applied to group pixels with similar color, depth and motion into spatiotemporal superpixels. We then construct a graph over these superpixels and use spectral clustering method to obtain further segmentation results. This approach divides the segmentation into two processes, which are able to generate spatially and temporally coherent segmentations for videos at a fast speed. Additionally the information of depth and motion, which not only guides the connections of pixels in different frames and views but also measures the similarity, improves the quality of segmentation. Experimental results have demonstrated our method is competitive with the state-of-art methods.