PISA: Pixelwise Image Saliency by Aggregating Complementary Appearance\n Contrast Measures with Edge-Preserving Coherence

Abstract

Driven by recent vision and graphics applications such as image segmentation\nand object recognition, computing pixel-accurate saliency values to uniformly\nhighlight foreground objects becomes increasingly important. In this paper, we\npropose a unified framework called PISA, which stands for Pixelwise Image\nSaliency Aggregating various bottom-up cues and priors. It generates spatially\ncoherent yet detail-preserving, pixel-accurate and fine-grained saliency, and\novercomes the limitations of previous methods which use homogeneous\nsuperpixel-based and color only treatment. PISA aggregates multiple saliency\ncues in a global context such as complementary color and structure contrast\nmeasures with their spatial priors in the image domain. The saliency confidence\nis further jointly modeled with a neighborhood consistence constraint into an\nenergy minimization formulation, in which each pixel will be evaluated with\nmultiple hypothetical saliency levels. Instead of using global discrete\noptimization methods, we employ the cost-volume filtering technique to solve\nour formulation, assigning the saliency levels smoothly while preserving the\nedge-aware structure details. In addition, a faster version of PISA is\ndeveloped using a gradient-driven image sub-sampling strategy to greatly\nimprove the runtime efficiency while keeping comparable detection accuracy.\nExtensive experiments on a number of public datasets suggest that PISA\nconvincingly outperforms other state-of-the-art approaches. In addition, with\nthis work we also create a new dataset containing $800$ commodity images for\nevaluating saliency detection. The dataset and source code of PISA can be\ndownloaded at http://vision.sysu.edu.cn/project/PISA/\n