EOImageNET - multiscale global dataset and DNN model for object detection on optical EO data

This package contains the EOImageNet dataset: a comprehensive multi-dataset training framework for object detection in remote sensing data (satellite, drone, aircraft images) that unifies 13 fully open-source remote sensing datasets together with the supplementary Tuatara dataset into a single standardized benchmark. Our approach addresses dataset heterogeneity through systematic preprocessing, annotation format unification, and a novel cross-labeling methodology that automatically annotates missing object classes across datasets with different taxonomies. The unified dataset encompasses 51,826 images with 810,551 annotated objects across 26 object classes, representing diverse geographic regions, imaging conditions, and object scales.This package uses following naming convention for dataset splits:- train.txt: Training set (41,208 images)- test.txt: Hyperparameter tuning set (10,316 images)- validation.txt: Final evaluation set (302 images)本套件包含EOImageNet数据集：一款面向遥感数据（卫星、无人机、航空器影像）目标检测任务的综合性多数据集训练框架，将13个完全开源的遥感数据集与补充的Tuatara数据集整合为单一标准化基准测试集。我们的方案通过系统性预处理、标注格式统一，以及一种全新的跨标注方法解决了数据集异质性问题，该方法可自动对采用不同分类体系的数据集的缺失目标类别进行标注。该统一数据集涵盖51826幅影像，包含26个目标类别下共计810551个标注目标，覆盖多样化的地理区域、成像条件与目标尺度。本套件针对数据集划分采用如下命名规范： - train.txt：训练集（41208幅影像） - test.txt：超参数调优集（10316幅影像） - validation.txt：最终评估集（302幅影像）