Hu et al. propose minimum barrier superpixels (MBS) using a minimum barrier distance transform for generating superpixels. Instead of computing the distance transform explicitly, however, an iterative raster-scan approach is employed. In particular, the image is iteratively scanned in raster scan and reverse raster scan order; based on seed pixel, each pixel is assigned a label based on the minimum barrier distance defined along the path between seed pixel and current pixel. To obtain compact superpixels, the distance transform is augmented by a compactness term (i.e. considering the length of the path). The whole process is applied in a hierarchical manner; starting with a subsampled image, the image is divided into superpixels. The original seeds are then updated based on the computed superpixels and projected into the next-higher resolution. I refer to the paper for details beyond this very rough description.
Figure 1: Qualitative results in comparison to other superpixel algorithms.
In experiments, the authors compare against other superpixel algorithms. Figure 1 (see the paper for references) shows qualitative results on the BSDS500 [1] dataset. Quantitatively, they argue that their approach performs equally well as ERS does but is significantly faster than the other approaches. The code can be found on GitHub: YinlinHu/MBS.
[1] P. Arbelaez, M. Maire, C. Fowlkes and J. Malik. Contour Detection and Hierarchical Image Segmentation. IEEE TPAMI, Vol. 33, No. 5, pp. 898-916, May 2011.
What is your opinion on this article? Let me know your thoughts on Twitter @davidstutz92 or LinkedIn in/davidstutz92.
Hu et al. propose minimum barrier superpixels (MBS) using a minimum barrier distance transform for generating superpixels. Instead of computing the distance transform explicitly, however, an iterative raster-scan approach is employed. In particular, the image is iteratively scanned in raster scan and reverse raster scan order; based on seed pixel, each pixel is assigned a label based on the minimum barrier distance defined along the path between seed pixel and current pixel. To obtain compact superpixels, the distance transform is augmented by a compactness term (i.e. considering the length of the path). The whole process is applied in a hierarchical manner; starting with a subsampled image, the image is divided into superpixels. The original seeds are then updated based on the computed superpixels and projected into the next-higher resolution. I refer to the paper for details beyond this very rough description.
Figure 1: Qualitative results in comparison to other superpixel algorithms.
In experiments, the authors compare against other superpixel algorithms. Figure 1 (see the paper for references) shows qualitative results on the BSDS500 [1] dataset. Quantitatively, they argue that their approach performs equally well as ERS does but is significantly faster than the other approaches. The code can be found on GitHub: YinlinHu/MBS.