Active learning with sampling by joint global-local uncertainty for salient object detection

The training of the SOD model relies on abundant annotated data, which needs laborious and expensive manual labeling. The generated pseudo-labels for reducing the annotation of the salient object will inevitably introduce noise, which will degrade the performance of the model and cannot fully represent the ground truth of manual labeling. To address this issue, we propose a novel active sampling strategy for salient object detection. The method is made up of two parts: a prediction module and an active learning module. The prediction module predicts the saliency of the image and provides the saliency prediction map for the active learning module. Then, the active learning module measures the global uncertainty and local uncertainty of the prediction map, aiming to select the most informative samples for the model. The selected samples are manually annotated and added to the training set to retrain the prediction model. Experimental results on DUTS dataset indicate that the amount of data can be reduced by 48.3% with competitive performance compared with the state-of-the-art SOD model.