Data and Model Uncertainty Aware Salient Object Detection

초록

In general, salient object detection (SOD) datasets have ambiguity due to annotation accuracy and human subjectivity in determining saliency. Since this data uncertainty causes inaccurate prediction, many techniques tackling data uncertainty have been proposed so far. Previous works estimated data uncertainty in terms of predictive inaccuracy and adjusted the learning contribution so that a given model can focus more on specific data. However, inaccurate predictions can occur due to not only data uncertainty but also model uncertainty in which the model does not fully explain the data. As a result, a region that is inaccurately predicted due to model uncertainty is considered a region with high data uncertainty, resulting in insufficient learning. To solve this problem, we propose a novel uncertainty-aware learning scheme where model uncertainty is decomposed from prediction uncertainty and it is minimized. Also, we propose a refinement method to further improve performance by correcting the prediction result using data uncertainty in the inference step. The proposed uncertainty-aware method excludes data uncertainty from learning step and inference step more effectively, making the model more accurately detect salient object(s). The experimental results prove that the proposed method achieves state-of-the-art performance on several SOD datasets and qualitatively detects salient objects more accurately than the prior arts. The code will be uploaded on Github.

키워드