Superpixel via coarse-to-fine boundary shift

摘要

K-means is used by numerous superpixel algorithms, such as SLIC, MSLIC and LSC, because of its simplicity and efficiency. Yet those k-means based algorithm failed to perform well on connectivity and accuracy. In this paper, we propose a coarse-to-fine boundary shift strategy (CFBS) as a replacement of k-means. The CFBS solves the superpixel segmentation problem by shifting boundries rather than clustering pixels. In other words, it can be defined as a special k-means algorithm optimized for superpixel segmentation. By replacing k-means with CFBS, SLIC and LSC are upgraded to NeoSLIC and NeoLSC. Experiments show that NeoSLIC and NeoLSC outperform SLIC and LSC in accuracy and efficiency respectively, and NeoSLIC and NeoLSC alleviate dis-connectivity. In addition, experiments also show that CFBS achieves great improvements on semantic segmentation, class segmentation and segmented flow.