ASABE Technical Library - Abstract

Click on “Download PDF” for the PDF version or on the title for the HTML version.
If you are not an ASABE member or if your employer has not arranged for access to the full-text, view access and subscription options

VF-FruitRecon: A Vision Foundation Model-based Pipeline for Occlusion-Robust Fruit 3D Reconstruction and Size Estimation

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2026 ASABE Annual International Meeting  2601079.(doi:10.13031/aim.202601079)
Authors:   Xingjian Li, Mr., Pengyao Xie, Dr., Weilong He, Mr., Nicholas Kaczmer, Mr., Neil Mattson, Dr., Lirong Xiang, Dr.
Keywords:   fruit size estimation, foundation models, 3D reconstruction, dense canopy

Abstract. Accurate fruit size estimation in the field is critical for yield prediction and precision crop management, yet remains challenging due to severe occlusions and complex plant geometries. Traditional approaches typically rely on two-dimensional (2D) pixel measurements, geometric assumptions, or multi-view pipelines, which often struggle under dense canopy conditions, suffer from scale ambiguity, or require extensive data collection and calibration. These limitations hinder robust size estimation when only partial fruit surfaces are visible or when depth sensing is noisy. We propose VF-FruitRecon, a framework that leverages Vision Foundation Models (VFMs) to robustly reconstruct on-plant tomatoes from stereo image pairs. Our approach integrates Grounded-SAM, FoundationStereo, and SAM-3D, for instance segmentation, stereo depth estimation, and three-dimensional (3D) reconstruction to recover the complete geometry of individual fruits with scale consistency. Our approach was evaluated on both photorealistic synthetic data and real-world greenhouse datasets, and compared it against traditional 2D baselines and RANSAC-based sphere fitting. Experimental results demonstrate that VFM-driven 3D reconstruction substantially improves accuracy and robustness to depth inaccuracies in complex simulated or field environments. Specifically, the framework with the Swin-T backbone achieves a Pearson correlation coefficient (r) of 0.58 with a mean absolute error (MAE) of 0.72 cm in high-occlusion scenarios, and reaches r = 0.81 (MAE = 0.34 cm) in low-occlusion conditions, outperforming baseline models for reliable fruit size estimation under dense canopy and sensing constraints. By bridging foundation-model priors with precision agriculture, VF-FruitRecon establishes an annotation-efficient pathway toward robust fruit reconstruction and size estimation under challenging real-world occlusion conditions.

(Download PDF)    (Export to EndNotes)