A Images Dataset of Rural Road for Instance Segmentation in Northern Xinjiang

This dataset is a collection of 1,285 valid high-definition images from 2021 to 2023, which were manually captured using a monocular sports video camera, GoPro HERO9 (pixels of 3840 × 2160), in the rural areas of northern Xinjiang, extracted by frame extraction technique, and pre-processed to obtain 1,285 valid high-definition images. The data collection locations were mainly selected from the countryside, fields, and urban and rural roads in northern Xinjiang. According to the main operating time and driving area of agricultural machines, the scenes are mainly asphalt, concrete, gravel, and dirt roads; the time is daytime and dusk; and the recognized target object categories are mainly: vehicles (agricultural machines, cars, trucks, tricycles, bicycles, etc.), pedestrians, livestock, and other static trees, traffic signs, street lamps, fences, and walls. By studying datasets such as Cityscapes, Mapillary Vistas, BDD100K, etc., and also based on the category objects in the images, this data is categorized into a total of 40 categories, which contain 20 instance categories. This data uses the polygonal annotation tool in the CVAT image annotation tool to manually annotate all the images at a detailed pixel level, in which a total of 10062 instance objects are annotated, which can meet the training needs of mainstream deep learning image segmentation models. This dataset includes three data files, of which: (1) images.zip is the original image file containing 1285 high-definition images of rural roads in the northern region of Xinjiang; (2) mask-images.zip is 1285 mask image files generated in one-to-one correspondence with the original images; (3) annotations.zip is the JSON file generated by annotation本数据集包含2021至2023年间采集的1285张有效高清图像。所有图像均采用GoPro HERO9单目运动摄像机（分辨率3840×2160）手动拍摄于新疆北部乡村区域，经帧提取技术提取并预处理后，最终得到该1285张有效高清图像。 数据采集地点主要选自新疆北部的乡村、农田及城乡道路。结合农业机械的主要作业时段与行驶区域，采集场景涵盖沥青路、水泥路、碎石路与土路；采集时段包含日间与黄昏；识别的目标物体类别主要包括：车辆（农机、轿车、货车、三轮车、自行车等）、行人、牲畜，以及静态的树木、交通标志、路灯、围栏与墙体。 本数据集参考Cityscapes、Mapillary Vistas、BDD100K等公开数据集，并结合图像内的目标类别，将标注类别划分为共计40类，其中包含20个实例类别。本数据集采用CVAT图像标注工具中的多边形标注工具，对所有图像开展精细化像素级手动标注，总计标注得到10062个实例对象，可满足主流深度学习图像分割模型的训练需求。 本数据集包含三类数据文件，具体如下：(1) images.zip为原始图像文件，内含1285张新疆北部乡村道路高清图像；(2) mask-images.zip为与原始图像一一对应的1285张掩码图像文件；(3) annotations.zip为标注生成的JSON格式文件。