eden500/COCO-N

Noisy-Labels-Instance-Segmentation数据集是一个用于图像分割任务的数据集，特别关注于标签噪声的基准测试。数据集支持通过命令行工具添加不同类型的噪声到注释中，包括随机噪声、定位噪声、近似噪声和类别翻转噪声。原始注释需要符合COCO格式。数据集的使用需要遵循Apache 2.0许可证，并且建议使用具有适当许可证的数据集，如MS-COCO/LVIS和Cityscapes。The Noisy-Labels-Instance-Segmentation dataset is designed for image segmentation tasks, specifically focusing on benchmarking label noise. The dataset supports adding various types of noise to annotations through command-line tools, including random noise, localization noise, approximation noise, and class flip noise. The original annotations must be in COCO format. The dataset is released under the Apache 2.0 license and is recommended to be used with properly licensed datasets such as MS-COCO/LVIS and Cityscapes.

原始信息汇总

Noisy-Labels-Instance-Segmentation

数据集概述

• 任务类别: 图像分割
• 语言: 英语
• 许可证: Apache 2.0

数据集使用

• 原始标注格式: COCO 格式

• 运行基准: bash python noise_annotations.py /path/to/annotations --benchmark {easy, medium, hard} --seed 1

• 自定义噪声方法: bash python noise_annotations.py /path/to/annotations --method_name method_name --corruption_values [{rand: [scale_proportion, kernel_size],localization: [scale_proportion, std_dev], approximation: [scale_proportion, tolerance], flip_class: percent_class_noise}]

引用

plaintext @misc{grad2024benchmarkinglabelnoiseinstance, title={Benchmarking Label Noise in Instance Segmentation: Spatial Noise Matters}, author={Eden Grad and Moshe Kimhi and Lion Halika and Chaim Baskin}, year={2024}, eprint={2406.10891}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2406.10891}, }

许可证

• 项目许可证: Apache 2.0
• 数据集许可证: 使用 MS-COCO/LVIS 和 Cityscapes