RS-STOD遥感影像小目标检测数据集

本项目中使用的数据集专为遥感图像中的小目标检测任务设计，涵盖了小型车辆、大型车辆、飞机、船只、储油罐等5种常见的小尺寸目标类型，本数据集包含2315张影像，以及50854个人工标注的实例，平均尺寸为13.4像素，适合小目标检测模型的训练与评估，特别是在密集分布、低对比度和遮挡情况下的小目标识别场景，其中，超微小目标定义为 bbox 最长边小于 16 像素。本数据集包含两种标注格式：YOLO格式与COCO格式，分别放置在对应名称的文件夹中，其中coco格式进行了train与val的划分，YOLO格式未进行划分。

The dataset utilized in this project is specifically tailored for small object detection tasks in remote sensing imagery. It encompasses five common small-sized object categories: small vehicles, large vehicles, airplanes, ships, and oil storage tanks. The dataset comprises 2315 images and 50854 manually annotated instances, with an average bounding box size of 13.4 pixels. It is well-suited for training and evaluating small object detection models, particularly in scenarios involving small object recognition under dense distribution, low contrast, and occlusion conditions. Ultra-small objects are defined as those whose longest side of the bounding box (bbox) is less than 16 pixels. The dataset offers two annotation formats: YOLO format and COCO format, which are stored in folders named after the respective formats. The COCO format has been split into training (train) and validation (val) subsets, whereas the YOLO format has not undergone such a split.

搜集汇总

背景与挑战

背景概述

RS-STOD是一个包含2315张遥感影像和50854个标注实例的小目标检测数据集，涵盖5种常见小尺寸目标类型，平均尺寸13.4像素，支持YOLO和COCO两种标注格式，适用于密集分布、低对比度和遮挡场景下的小目标识别任务。