Hierarchical Superpixel Segmentation by Searching Seeds

Abstract

Most superpixel methods generate a fixed number of superpixels, which is inconvenient for tasks requiring multi-scale image representations. In contrast, hierarchical methods provide efficient hierarchical superpixel segmentation, but may ignore preserving fine structure. To address these limitations, this paper presents Hierarchical Superpixel Segmentation by Searching Seeds (HSSS). As a seed-initiated method, HSSS develops a content-sensitive seed selection strategy by evaluating the local homogeneity of pixels/superpixels to improve the ability to capture details. Then, based on the foundation superpixel result, a superpixel-wise seed-searching scheme and an arc-pruning strategy are introduced in HSSS to guide the construction of the superpixel hierarchy, while delicate structures are preserved. To quantitatively assess fine-structure-preserving performance, an evaluation metric is introduced, termed Instance Boundary Recall (IBR). Experimental results on both conventional and newly introduced metrics show that HSSS achieves competitive overall performance, while showing particular advantages in preserving fine structures.