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

Gaussian Splatting Based 3D Tomato Fruit Counting and Multi-Stage Classification from Monocular Video

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

Citation:  2026 ASABE Annual International Meeting  2600666.(doi:10.13031/aim.202600666)
Authors:   Yan-An Chen, Guan-Hua Chen, Ya-Ping Lin, Shih-Fang Chen
Keywords:   Fruit counting, 3D reconstruction, 3D Gaussian Splatting, deep learning, precision agriculture.

Abstract. Tomato breeding programs require comprehensive phenotypic data including fruit count and maturity progression across growth stages to evaluate yield potential and select superior genotypes. However, manual monitoring of these traits is prohibitively labor-intensive at field scale, while existing 2D imaging approaches struggle with severe occlusion and depth ambiguity in dense canopy environments. To enable scalable, continuous phenotypic monitoring, this study developed an automated framework that reconstructs 3D field scenes from monocular video for multi-stage fruit detection and counting. A dataset of more than 1,200 tomato plant images containing over 26,000 annotated fruit instances was collected across three maturity stages: green, turning-pink, and harvest stages. The image dataset was used to train a Mask2Former segmentation model. The segmentation masks were subsequently integrated with a surface-regularized 3D Gaussian Splatting (3DGS) framework, with monocular depth priors supervising geometric consistency. Fruit instances were then isolated and enumerated using the multi-layer perceptron (MLP) decoder head and Density-Based Spatial Clustering of Applications with Noise (DBSCAN) clustering algorithm. The segmentation model achieved a mean Average Precision (mAP50) of 86%. The 3DGS reconstruction attained Peak Signal-to-Noise Ratio (PSNR) values of 21.41 dB and 22.64 dB, and Structural Similarity Index Measure (SSIM) values of 0.81 and 0.80, for the early and late growth stage datasets, respectively. The framework yielded overall fruit counting accuracies of 0.86 and 0.82 across two evaluation datasets. Future work will incorporate morphological fruit sizing and floral-to-fruit correspondence tracking to support comprehensive yield forecasting for precision breeding programs.

(Download PDF)    (Export to EndNotes)