A Fully Convolutional Network for Salient Object Detection

In this paper we address the task of salient object detection without requiring an explicit object class recognition. To this end, we propose a solution that exploits intermediate activations of a Fully Convolutional Neural Network previously trained for the recognition of 1,000 object classes, in order to gather generic object information at different levels of resolution. This is done by using both convolution and convolution-transpose layers, and combining their activations to generate a pixel-level salient object segmentation. Experiments are conducted on a standard benchmark that involves seven heterogeneous datasets. On average our solution outperforms the state of the art according to multiple evaluation measures.