COCO-N, Cityscapes-N, COCO-WAN

本研究引入了三个新的数据集：COCO-N、Cityscapes-N和COCO-WAN，用于评估和增强实例分割模型对标签噪声的鲁棒性。COCO-N和Cityscapes-N模拟了不同程度的标签噪声，包括类噪声和空间噪声，以评估模型在复杂任务中的表现。COCO-WAN则专注于使用基础模型和弱监督来模拟半自动化标注工具及其噪声标签。这些数据集的创建旨在解决实际数据集中常见的标注不准确问题，特别是在自动驾驶和医学图像分析等领域的应用中，精确的物体边界至关重要。通过这些数据集，研究者可以更好地理解和改进模型在面对真实世界噪声时的性能。

This study introduces three novel datasets: COCO-N, Cityscapes-N, and COCO-WAN, which are designed to evaluate and enhance the robustness of instance segmentation models against label noise. COCO-N and Cityscapes-N simulate label noise at varying levels, including class noise and spatial noise, to assess model performance on complex tasks. COCO-WAN, by contrast, focuses on simulating semi-automated annotation tools and their noisy labels using foundation models and weak supervision. The creation of these datasets aims to address the common issue of inaccurate annotations in real-world datasets, particularly in applications such as autonomous driving and medical image analysis, where precise object boundaries are critical. Using these datasets, researchers can gain a deeper understanding of and enhance model performance when confronted with real-world noise.