Salient Object Ranking: Saliency model on relativity learning and evaluation metric on triple accuracy

Abstract

Salient object ranking (SOR) aims to evaluate the saliency level of each object in an image, which is crucial for the advancement of downstream tasks. The human visual system distinguishes the saliency levels of different targets in a scene by comprehensively utilizing multiple saliency cues. To mimic this comprehensive evaluation behavior, the SOR task needs to consider both the objects’ intrinsic information and their relative information within the entire image. However, existing methods still struggle to obtain relative information effectively, which tend to focus too much on specific objects while ignoring their relativity. To address these issues, this paper proposes a Salient Object Ranking method based on Relativity Learning (RLSOR), which integrates multiple saliency cues to learn the relative information among objects. RLSOR consists of three main modules: the Top-down Guided Salience Regulation module (TGSR), the Global–Local Cooperative Perception module (GLCP), and the Semantic-guided Edge Enhancement module (SEE). At the same time, this paper proposes a Triple-Accuracy Evaluation (TAE) metric for the SOR task, which can evaluate the segmentation accuracy, relative ranking accuracy, and absolute ranking accuracy in one metric. Experimental results show that RLSOR significantly enhances SOR performance, and the proposed SOR evaluation metric can better meets human subjective perceptions.