Abstract
3D computer vision offers powerful tools for efficient agricultural analysis by enabling automated extraction of quantitative traits using point clouds of field scenes. In the context of wheat, accurate analyses of yield-related traits such as spike count, plant density and canopy structure at plot-level and organ morphology such as length, width and volume of wheat heads are essential for high throughput phenotyping. Although recent advances in 3D imaging and point-cloud learning have enabled more detailed analysis of plant architecture, the development of robust wheat phenotyping methods remains limited by the lack of annotated 3D datasets for instance-level segmentation. Existing wheat point-cloud resources primarily focus on part-level or semantic segmentation, where points are assigned to broad categories such as wheat heads, leaves and stems. However, study and analysis of wheat head morphology requires instance-level annotations.
In this work, we introduce Wheat3DIS, a multi-source benchmark for 3D wheat head instance segmentation. The benchmark combines in-field wheat point clouds from various sources such as a laser-scanner, a camera array and an iphone. The benchmark provides diverse 3D representations across acquisition modalities, point densities, and reconstruction characteristics. Each scene is annotated with instance-level labels for wheat heads to support the evaluation of wheat organ separation, spike counting, and structure-aware trait extraction. Wheat3DIS provides standardized data splits, annotation protocols, evaluation metrics, and baseline results for representative 3D segmentation architectures. By focusing specifically on wheat and by moving beyond part segmentation toward instance-level understanding, Wheat3DIS establishes a reproducible benchmark for studying 3D wheat perception, cross-sensor generalization, and robust phenotyping under realistic geometric complexity.
