A Multisize Superpixel Approach for Salient Object Detection Based on Multivariate Normal Distribution Estimation

Abstract

This paper presents a new method for salient object detection based on a sophisticated appearance comparison of multisize superpixels. Those superpixels are modeled by multivariate normal distributions in CIE-Lab color space, which are estimated from the pixels they comprise. This fitting facilitates an efficient application of the Wasserstein distance on the Euclidean norm ( [Formula: see text]) to measure perceptual similarity between elements. Saliency is computed in two ways. On the one hand, we compute global saliency by probabilistically grouping visually similar superpixels into clusters and rate their compactness. On the other hand, we use the same distance measure to determine local center-surround contrasts between superpixels. Then, an innovative locally constrained random walk technique that considers local similarity between elements balances the saliency ratings inside probable objects and background. The results of our experiments show the robustness and efficiency of our approach against 11 recently published state-of-the-art saliency detection methods on five widely used benchmark data sets.

Salient object detection via spectral matting

“…In literature, the most suitable way of capturing this cue is to cluster colors of an image and model each color as a component of a Gaussian Mixture Model (GMM) [1,30,8]. We use GMM based representation of the colors in the image in this work,…”

“…Recently a plethora of techniques have been proposed that attempt to extract salient regions by computation of region-based contrast (in terms of primitive features) in a local or global fashion using either single or multiple scales [8,9,10,11,12,13].…”

“…The saliency of a region is then computed as a contrast with respect to any primitive feature, either by comparing a region with its surrounding superpixels [14], by computing the global contrast [15] or using a combination of both local and global contrast [8]. Such models ignore important object properties such as object boundaries, shapes and composition 1 , which results in inappropriate annotation 2 and non-uniform saliency assignment 3 inside object con- 1 We refer to the property of objects being composed of multiple parts as object composition.…”

“…3 Non-uniform saliency assignment means that considerably different saliency values have been assigned to different regions of the salient object. [8], that employs multivariate normal distribution termed as MND here, and the work of Yan et al [9] that proposed hierarchical saliency detection, termed as HS in this work.…”

“…This is achieved through handcrafted integration procedures in some models. Zhu et al [8] and Margolin et al [15] used element-wise multiplication to perform feature fusion. Gazit et al [18] employed the harmonic mean of feature maps to compute the final saliency.…”