Minimum barrier superpixel segmentation

Superpixel Segmentation by Forgetting Geodesic Distance

“…LSC [20] and SNIC [21] also follow SLIC's framework to balance the color and space distance between cluster centers and pixels for their local neighbors. MBS [28] introduces minimum barrier distance (MBD) to evaluate the similarity between pixel and seeds. Although this method utilizes the difference between the pixel's maximum intensity and minimum intensity to evaluate similarity, it tends to suffer from undersegmentation in the case of weak boundary regions.…”

“…However, due to the Euclidean distance computation, they cannot accurately segment object parts, such as pillars. Meanwhile, the weak roof boundary results in low boundary adherence for MBS [28] and FLIC [29]. Our method calculates the geodesic distance to focus on the object boundary and extract the pillar boundaries.…”

“…NC [11] and TP [15] are time consuming with respect to their polynomial time complexity and multiple iterations. MBS [28], FLIC [29] and SNIC [21] achieve the top three results.…”

“…Meanwhile, computational efficiency becomes its bottleneck due to measuring the geodesic distance and iterative updating. Recently, Zhao et al and Hu et al propose two superpixel segmentation methods based on raster scanning, i.e., MBS [28] and FLIC [29], to improve computation efficiency. The former uses a coarseto-fine hierarchical scheme to estimate superpixel centers.…”

“…To obtain an object and comprehensive comparison, we test our method by three different metric measures, which are the common quality indications in existing works. Meanwhile, we compare our method with nine state-of-the-art methods: FH [12], SLIC [16], TP [15], NC [11], LSC [20], RPS [24], SNIC [21], MBS [28] and FLIC [29]. We execute segmentation based on publicly available codes provided by authors.…”