Noisy Label Rectification Using Weighted Fusion for Deep Unsupervised Salient Object Detection
Xiaohui Dong, Yi Liu · 2022
Salient Object Detection (SOD) based on deep neural networks (DNN) depends on pixel-level human annotations for training, which consumes much labor. Alternatively, we use pseudo-ground-truth labels generated from multiple handcrafted saliency methods to substitute for expensive ground-truth human annotation for deep salient object detection training. To remedy the noise of pseudo labels, we use a weighted fusion strategy to integrate several hand-crafted methods at the initial stage and rectify pseudo-ground-truth at the label update stage.