SyneticAI/ApplesM5-Dataset

ApplesM5是一个合成的苹果检测基准数据集，通过程序内容生成和基于物理的渲染（PBR）创建，为农业AI提供了一个干净、高度泛化的训练信号。该数据集证明了仅使用合成数据训练的模型比仅使用真实世界数据训练的模型具有更好的泛化能力，mAP50-95指标提高了+34.24%。数据集结构针对YOLO模型训练进行了优化，分为训练集和验证集，其中训练集包含超过10,000张合成图像和标签，验证集包含约300张来自外部果园的真实世界图像样本。

ApplesM5 is a synthetic apple detection benchmark dataset created through procedural content generation and physically-based rendering (PBR) to provide a clean, highly generalized training signal for robust agricultural AI. The dataset demonstrates that models trained exclusively on synthetic data achieve superior generalization compared to models trained solely on real-world data, with a +34.24% increase in mAP50-95. The dataset structure is optimized for YOLO model training, split into a training set with over 10,000 synthetic images and labels, and a validation set with about 300 real-world image samples from external orchards.