Abstract: Deep neural network (DNN) has become an indispensable tool for most applications. Yet, its superior performance relies on correctly annotated large-scale datasets. In image segmentation, acquiring high-quality annotations is time-consuming and error-prone. Inaccurate annotations introduce noise to the training labels and will significantly impact DNNs' performance due to their strong memorization power. Therefore, it is important to develop methodology that can improve the robustness of segmentation models against noisy annotations.

We explore existing literature on learning-with-label-noise in both classification and segmentation scenarios. Then we focus on training a segmentation model in the presence of noisy annotations. In the segmentation context, existing methods mostly ignore the spatial correlation of the annotation noise. By explicitly modeling spatial correlation, we propose a method to correct segmentation label noise during training. Preliminary results show that our method has better performance compared with existing methods.