RGB Image Dataset of Urochloa Hybrids for High-Throughput Phenotyping and Artificial Intelligence Applications

This dataset represents an extended version of a previous work, accessible at this link: https://doi.org/10.7910/DVN/U0KL6Y. An additional 139 images and a total of 24,983 new annotations have been included. Combined with the original dataset, a total of 394 images with 47,323 annotations are now available. This new dataset differs from the previous one in several key ways, primarily in the conditions and types of images captured, as well as in the expanded annotations. In the initial release, lighting conditions were carefully controlled to standardize histogram distribution across all images. The images were also captured at a fixed distance and exclusively in a nadir (top-down) view, using a single sensor in a single geographic location. For this updated dataset, variability was prioritized across all aspects. Images were taken in multiple geographic locations, including Palmira, Colombia, and Ocozocoautla de Espinosa, Mexico. Different sensors were used, including a professional Nikon D5600 camera, smartphones (such as the Realme C53 and Oppo Reno 11), and even a Phantom 4 Pro V2 drone. The capture distance varied from 1 to 3 meters, resulting in images with differing spatial resolutions. Additionally, several capture angles were employed: no longer just nadir views but also oblique and frontal angles. Raceme density per plant was also increased. In the original dataset, the plant with the highest raceme count had 851 racemes. In the updated dataset, raceme counts reach as high as 1,586 in a similar area (~1m²), nearly doubling the count. This increase leads to a much higher degree of raceme overlap. This expanded dataset is expected to provide significant benefits for deep learning applications. The enhanced variability supports the development of more robust deep learning models, better suited to handle real-world diversity and complexity.

本数据集为先前研究工作的扩展版本，可通过以下链接获取：https://doi.org/10.7910/DVN/U0KL6Y。新增139张图像及总计24,983条注释；结合原始数据集，目前共包含394张图像与47,323条注释。该更新数据集与先前版本的核心差异体现在图像捕获条件、图像类型及扩展注释三个方面。在初始版本中，光照条件经过严格控制，以实现所有图像直方图分布（histogram distribution）的标准化；图像捕获采用固定距离，且仅使用单一传感器在单一地理位置获取天底视角（nadir view，即自上而下视角）的图像。而本更新数据集则优先考虑各维度的变异性：图像采集于多个地理位置，包括哥伦比亚的Palmira与墨西哥的Ocozocoautla de Espinosa；使用多种传感器，涵盖专业尼康D5600相机、智能手机（如Realme C53及Oppo Reno 11）以及Phantom 4 Pro V2无人机；捕获距离范围为1至3米，导致图像空间分辨率（spatial resolution）存在差异；此外，捕获角度不再局限于天底视角，还包括倾斜视角与正面视角。每株植物的总状花序（raceme）密度亦有所提升：原始数据集中单株最高总状花序数为851，而更新数据集在相似区域（约1平方米）内的单株最高总状花序数达1586，几乎翻倍，这显著增加了总状花序的重叠程度。该扩展数据集有望为深度学习（deep learning）应用带来显著收益。增强的变异性可支持开发更鲁棒的深度学习模型，使其更能适应现实世界的多样性与复杂性。