Luengo et al. introduce SMURFS – superpixels from multi-scale refinement of super-regions. The main idea is to alternatingly refine a regualr grid by merging superpixels into super-superpixels and splitting these super-superpixels again. This idea is illustrated in Figure 1. In particular, the merging of superpixels follows the approach by Felzenswalb and Huttenlocher [1] (like FH superpixels) with an added size constraint on the superpixels. Superpixels are merged until a rough average size of the super-superpixels is obtained. Then, these super-superpixels are splitted using an MRF formulation – this can be done in parallel for each region individually. The MRF takes into account the cluster centers (which are initially chosen randomly within the super-superpixels) in the unary term and employs a simple Pott's model as pairwise term. When not optimized individually, graph cuts can be employed.
Figure 1: Illustration of the employed optimization scheme.
In experiments, the authors consider different versions of the proposed method. For example, in a total of 10 iterations, we can split super-superpixels always into $K = 2$ or $K = 5$ superpixels. It is also not clear how to choose the corresponding seeds; the authors use random sampling or run $k$-means to identify seeds (which seems a bit weird as $k$-means would already produce superpixels, so no need to apply their scheme). Figure 2 shows qualitative results; I won't discuss quantitative results, though.
Figure 2: Qualitative results.
[1] Pedro F. Felzenszwalb, Daniel P. Huttenlocher. Efficient Graph-Based Image Segmentation. International Journal of Computer Vision 59(2): 167-181 (2004).
What is your opinion on this article? Let me know your thoughts on Twitter @davidstutz92 or LinkedIn in/davidstutz92.
Luengo et al. introduce SMURFS – superpixels from multi-scale refinement of super-regions. The main idea is to alternatingly refine a regualr grid by merging superpixels into super-superpixels and splitting these super-superpixels again. This idea is illustrated in Figure 1. In particular, the merging of superpixels follows the approach by Felzenswalb and Huttenlocher [1] (like FH superpixels) with an added size constraint on the superpixels. Superpixels are merged until a rough average size of the super-superpixels is obtained. Then, these super-superpixels are splitted using an MRF formulation – this can be done in parallel for each region individually. The MRF takes into account the cluster centers (which are initially chosen randomly within the super-superpixels) in the unary term and employs a simple Pott's model as pairwise term. When not optimized individually, graph cuts can be employed.
Figure 1: Illustration of the employed optimization scheme.
In experiments, the authors consider different versions of the proposed method. For example, in a total of 10 iterations, we can split super-superpixels always into $K = 2$ or $K = 5$ superpixels. It is also not clear how to choose the corresponding seeds; the authors use random sampling or run $k$-means to identify seeds (which seems a bit weird as $k$-means would already produce superpixels, so no need to apply their scheme). Figure 2 shows qualitative results; I won't discuss quantitative results, though.
Figure 2: Qualitative results.