Abstract
Point cloud analysis is a crucial task in computer vision. Despite significant advances over the past decade, the developments in agricultural domain have faced challenges due to a scarcity of datasets. To facilitate 3D point cloud research in agriculture community, we introduce Crops3D, the diverse real-world dataset derived from authentic agricultural scenarios. Crops3D distinguishes itself through its unique properties: diversity, authenticity, and complexity. The dataset incorporates data from diverse point cloud acquisition methods, encompassing eight distinct crop types with 1,230 samples, authentically representing crops in the real-world. It stands as the pioneering dataset that comprehensively supports the three critical tasks in 3D crop phenotyping: instance segmentation of individual plants in agricultural settings, plant type perception, and plant organ segmentation. Additionally, the intricate crop structures in Crops3D exhibit higher complexity than available 3D public datasets, showcasing substantial self-occlusion and increased complexity as crops mature. We analyse diverse crop point cloud acquisition methods and evaluate multiple models' performance with the Crops3D dataset.