Figure 1
Many methods for object recognition, segmentation, etc., rely on tessellation of an image into ``superpixels". A superpixel is an image patch which is better aligned with intensity edges than a rectangular patch. Superpixels can be extracted with any segmentation algorithm, however, most of them produce highly irregular superpixels, with widely varying sizes and shapes. A more regular space tessellation may be desired. We formulate the superpixel partitioning problem in an energy minimization framework, and optimize with graph cuts. Our energy function explicitly encourages regular superpixels. We explore variations of the basic energy, which allow a trade-off between a less regular tessellation but more accurate boundaries or better efficiency. Our advantage over previous work is computational efficiency, principled optimization, and applicability to 3D ``supervoxel" segmentation. We achieve high boundary recall on 2D images and spatial coherence on video. We also show that compact superpixels improve accuracy on a simple application of salient object segmentation.
 |  |  |  |
Code For Unix (thanks to Alvaro Collet)
Our approach naturally extends to segmenting ``supervoxels" in 3D space. Segmentation of volumes into supervoxels can be useful, potentially, for medical image and for video processing. First we create a 3D volume by stacking the frames together, Figure 2. Analogously to the 2D case, we cover the 3D volume by overlapping 3D blocks. For clarity, in Figure 2 we show only a few non-overlapping blocks. The depth of a block can be different from its width and height. The larger the depth, the more temporal coherency is encouraged. Each block corresponds to a label.
| Results without temporal coherence |
| Results with temporal coherence |