ABox-PC: Adaptive box-guided completion of single-view tomato fruit point clouds for improved greenhouse phenotyping

Accurate fruit-level phenotyping of tomatoes is important for quantitative analysis under greenhouse conditions. However, occlusion and limited viewpoints often produce incomplete single-view fruit point clouds, reducing the reliability of downstream geometric measurements. This study developed ABox-PC, an adaptive box-guided point cloud completion network, and integrated it into a completion-to-phenotyping pipeline. The network combines an encoder–decoder architecture, adversarial supervision, and adaptive box-guided local spatial modeling to recover missing regions within incomplete single-view visible-surface representations. Completed point clouds were subsequently used to estimate horizontal diameter, vertical diameter, and weight through an axis-weighted oriented bounding box method and a density-based geometric formulation. On paired incomplete–reference samples derived from relatively complete single-view fruits, ABox-PC achieved a squared L2 Chamfer distance of 784.76±91.08 mm 2 . For the same 109 naturally occluded fruits, completion increased the R² values for horizontal and vertical diameter estimation from 0.47 and 0.57 to 0.79 and 0.78, respectively, with corresponding RMSE values of 5.32 and 5.07 mm. Weight estimation after completion achieved an R² of 0.83, an RMSE of 29.77g, and a MAPE of 10.26%. These results indicate that visible-surface completion can provide more reliable geometric inputs for downstream phenotypic estimation under the evaluated conditions.