Wei Jun ; Wang Shuhui ; Huang Qingming

Abstract

Most of existing salient object detection models have achieved great progressby aggregating multi-level features extracted from convolutional neuralnetworks. However, because of the different receptive fields of differentconvolutional layers, there exists big differences between features generatedby these layers. Common feature fusion strategies (addition or concatenation)ignore these differences and may cause suboptimal solutions. In this paper, wepropose the F3Net to solve above problem, which mainly consists of crossfeature module (CFM) and cascaded feedback decoder (CFD) trained by minimizinga new pixel position aware loss (PPA). Specifically, CFM aims to selectivelyaggregate multi-level features. Different from addition and concatenation, CFMadaptively selects complementary components from input features before fusion,which can effectively avoid introducing too much redundant information that maydestroy the original features. Besides, CFD adopts a multi-stage feedbackmechanism, where features closed to supervision will be introduced to theoutput of previous layers to supplement them and eliminate the differencesbetween features. These refined features will go through multiple similariterations before generating the final saliency maps. Furthermore, differentfrom binary cross entropy, the proposed PPA loss doesn't treat pixels equally,which can synthesize the local structure information of a pixel to guide thenetwork to focus more on local details. Hard pixels from boundaries orerror-prone parts will be given more attention to emphasize their importance.F3Net is able to segment salient object regions accurately and provide clearlocal details. Comprehensive experiments on five benchmark datasets demonstratethat F3Net outperforms state-of-the-art approaches on six evaluation metrics.