Superpixels and supervoxels in an energy optimization framework

摘要: Many methods for object recognition, segmentation, etc., rely on a tessellation of an image into "superpixels". A superpixel is patch which better aligned with intensity edges than rectangular patch. Superpixels can be extracted any segmentation algorithm, however, most them produce highly irregular superpixels, widely varying sizes and shapes. more regular space may desired. We formulate the partitioning problem in energy minimization framework, optimize graph cuts. Our function explicitly encourages superpixels. explore variations basic energy, allow trade-off between less but accurate boundaries or efficiency. advantage over previous work computational efficiency, principled optimization, applicability to 3D "supervoxel" segmentation. achieve high boundary recall images spatial coherence video. also show that compact superpixels improve accuracy simple application salient