Key Issues for the Construction of Salient Object Datasets with Large-Scale Annotation

Salient object detection (SOD) has been extensively studied in recent decades, especially after the boom of convolutional neural networks (CNNs). To direct supervised CNN-based methods to its highest function for SOD, more challenging datasets with reasonable large-scale annotations have been proposed. However, due to a lack of verdict of defining multiple salient objects on images or sequences with complex natural scenes and objects, there are certain degrees of bias in current SOD datasets. Therefore, we survey the methods for salient object annotation and further conclude several key issues for the future SOD dataset construction. To the best of our knowledge, this is the first work that synthesizes all the existing salient object annotation methods.