NanKai Sonar Image Dataset (NKSID)
收藏资源简介:
该数据集包含2617张来自8个类别的图像,标签显示自然长尾分布。数据收集发生在渤海湾,使用配备多波束前视声纳的遥控潜水器捕捉水下数据。为了减少目标间的干扰并便于定位,目标通过绳索连接到浮标,并悬挂在水面下约5-10米处。数据收集过程中,从不同视角、距离(2-15米)和频率(750kHz, 1.2MHz)捕捉每个目标的图像,以增强数据集的丰富性。
This dataset comprises 2,617 images from 8 categories, exhibiting a natural long-tail distribution in their labels. The data collection took place in the Bohai Bay, where a remotely operated vehicle equipped with a multibeam forward-looking sonar was utilized to capture underwater data. To minimize interference between targets and facilitate positioning, each target was tethered to a buoy and suspended approximately 5-10 meters below the water surface. During the data collection process, images of each target were captured from various angles, distances (2-15 meters), and frequencies (750kHz, 1.2MHz) to enhance the richness of the dataset.
NK-Sonar-Image-Dataset (NKSID) 概述
数据集基本信息
- 名称: NanKai Sonar Image Dataset (NKSID)
- 类别数: 8
- 图像数量: 2617
- 数据收集地点: Bohai Bay ($39^circ N 118^circ E$)
- 数据收集工具: 配备Oculus M750d多波束前视声呐的遥控潜水器(ROV)
- 图像特征: 目标通过绳索附着在浮标上,悬挂在水面下约5-10米,从不同角度、距离(2-15m)和频率(750kHz, 1.2MHz)捕捉
- 数据处理: 目标选择、预处理和标注
数据集使用
- 下载与解压: 从仓库直接下载并解压所有
.zip文件,每个类别的图像单独压缩 - 文件说明:
train_abs.txt: 包含每个图像的相对路径和标签kfold_train.txt和kfold_val.txt: 存储十折交叉验证的随机训练集/验证集分割,$n$表示样本索引,对应train_abs.txt中的第$n$行
示例应用
- 示例仓库: Jorwnpay/Sonar-OLTR (github.com),展示使用此数据集进行开放集长尾识别的示例
引用信息
- 论文引用: latex @article{jiao2024open, title={Open-set recognition with long-tail sonar images}, author={Jiao, Wenpei and Zhang, Jianlei and Zhang, Chunyan}, journal={Expert Systems with Applications}, pages={123495}, year={2024}, publisher={Elsevier} }




