SDCNet: Size Divide and Conquer Network for Salient Object Detection

Senbo Yan, Xiaowen Song, Chuer Yu; Proceedings of the Asian Conference on Computer Vision (ACCV), 2020

The fully convolutional neural network (FCN) based methods achieve great performances in salient object detection (SOD). However, most existing methods have difficulty in detecting small or large objects. To solve this problem, we propose Size Divide and Conquer Network (SDCNet) which learning the features of salient objects of different sizes separately for better detection. Specifically, SDCNet contains two main aspects: (1)We calculate the proportion of objects in the image (with the ground truth of pixel-level) and train a size inference module to predict the size of salient objects. (2) We novelly propose a Multi-channel Size Divide Module (MSDM) to learning the features of salient objects with different sizes, respectively. In detail, we employ MSDM following each block of the backbone network and use different channels to extract features of salient objects within different size range at various resolutions. Unlike coupling additional features, we encode the network based on the idea of divide and conquer for different data distributions, and learn the features of salient objects of different sizes specifically. The experimental results show that SDCNet outperforms 14 state-of-the-art methods on five benchmark datasets without using other auxiliary techniques.