Benchmarking deep models on salient object detection

Huajun Zhou, Yang Lin, Lingxiao Yang, Jianhuang Lai, Xiaohua Xie

Abstract: Highlights•A benchmark to provide a fair and comprehensive comparison between SOD models.•Aged U-shape methods perform surprisingly well using modern optimization strategy.•Training strategy significantly contribute to the improvement of SOD methods.•A more generalized loss based on the analysis on existing losses for SOD models.•SOD models are sensitive to objectness shifting between training and testing data.