Expand, Pool and Confine: Reliably Detaching Salient Objects From the Background

The majority of existing salient object detection (SOD) methods are fundamentally “one-shot” solutions, and the purpose of model training is to continuously improve the precision of instantly detaching the objects from the background at a glance. This paradigm may lead to the potential damage of edge details. To address this challenge, we rethink SOD from a probabilistic perspective and propose a multi-stage SOD network (EPC-Net) that integrates expanding, pooling, and confining. Firstly, we introduce a confidence region expanding technique to lock the reliable detection regions, thereby obtaining foreground probability maps for each stage. Then we propose a joint probability pooling to convert them into the joint probability maps, effectively preserving the edge details by filtering out edge-sensitive pixels of the foreground probability maps. Additionally, it can also filter out isolated background pixels. Finally, we design a compact background perceiving transformer block (CBPTer) with integrated foreground confine attention (FCA), aiming to accentuate foreground objects within trustworthy regions of joint probability maps. Experimental results show that the proposed model outperforms recent SOTA methods on five SOD benchmarks.