Efficient Multiscale Object-based Superpixel Framework

Abstract

Superpixel segmentation can be used as an intermediary step in many applications, often to improve object delineation and reduce computer workload. However, classical methods do not incorporate information about the desired object. Deep-learning-based approaches consider object information, but their delineation performance depends on data annotation. Additionally, the computational time of object-based methods is usually much higher than desired. In this work, we propose a novel superpixel framework, named Superpixels through Iterative CLEarcutting (SICLE), which exploits object information being able to generate a multiscale segmentation onthe-fly. SICLE starts off from seed oversampling and repeats optimal connectivity-based superpixel delineation and object-based seed removal until a desired number of superpixels is reached. It generalizes recent superpixel methods, surpassing them and other state-of-the-art approaches in efficiency and effectiveness according to multiple delineation metrics.

Supervoxel Approach in an Efficient Multiscale Object-Based Framework

“…Classic methods, like Iterative Spanning Forest (ISF), use pixel clustering to generate superpixels in a short span of time, but they fail to ensure the desired number of superpixels Considering that, an Object-based Dynamic Iterative Spanning Forest (ODISF) was created -using a method of (i) seed oversampling, (ii) superpixel generation, and (iii) object-based seed removal -to be more accurate and faster than the aforementioned frameworks and have a minimum influence of saliency errors [3]. Finally, upgrading the seed removal method of ODISF, SICLE was developed to achieve adequate delineation for all objects tested in Belém's [2] work, irrespective of map saliency errors.…”

“…For this reason, both arc-cost functions may present similar delineations, especially in object borders. Therefore, VI-SICLE uses of a root-based function, henceforth named ROOT [2].…”

“…In image superpixel segmentation, SICLE [2] has achieved impressive results in several data-sets. The Superpixel through Iterative Clearcutting (SICLE) is a generalization of two stateof-art methods that "starts off from oversampling and, through several iterations, generates superpixels from the seed set and remove a portion of irrelevant seeds to preserve the accurate object delineation from the previous iteration", being seed a chosen pixel on the image.…”

Novel Arc-Cost Functions and Seed Relevance Estimations for Compact and Accurate Superpixels

“…Another strategy consists of building end-to-end supervised [12,13,37] or unsupervised [14,15] networks for superpixel generation. However, although promising, both approaches demand more research [20,35] since they present: (a) high data dependency; (b) moderate delineation; and (c) high computational cost.…”

“…Another saliency-based framework, named Superpixels through Iterative CLEarcutting (SICLE) [20], generalizes both DISF and ODISF for computing superpixels more efficiently. SICLE is composed of three independent steps: (i) seed oversampling; (ii) path-based superpixel generation; and (iii) seed removal.…”