Phenomics

Learning 3D Features from Dual-View Point Cloud for Estimation of Swine Subcutaneous Biological Traits

---

Abstract

Accurately estimating subcutaneous biological traits from external images remains a fundamental challenge in animal phenotyping. This study aimed to develop a dual-view point-cloud learning framework for non-invasive prediction of subcutaneous biological traits in live pigs using external surface morphology, with post-mortem measurements serving as reference phenotypes. A dual-view depth imaging system sequentially captured dorsal and lateral surface geometry of 625 pigs from four genetic groups (Large White, Landrace, Duroc, and Crossbred), generating 5,836 pairs of high-quality point clouds. Meanwhile, post-mortem measured phenotypic values of loin muscle area (LMA, 35-95 cm²) and backfat thickness (BFT, 5-55 mm) were obtained for all 625 pigs, providing a dataset with broad phenotypic variation. Based on this dataset, a Dual-stream Multi-scale Fusion Network (DSMF-Net) was developed and trained to predict LMA and BFT, and the predictions were compared with their corresponding post-mortem reference values. Model performance was evaluated using root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), coefficient of determination (R²), and Pearson correlation coefficient (r), and was compared against classic point-cloud networks (PointNet, PointNet++, and PointMLP) as well as traditional regression models based on manually derived body measurements. DSMF-Net achieved superior prediction accuracy for both traits, with an MAE of 5.38 cm² (R² = 0.74, r = 0.86) for LMA and an MAE of 3.40 mm (R² = 0.83, r = 0.91) for BFT, significantly outperforming all baseline methods. Ablation studies further confirmed that joint multi-trait prediction outperformed single-task models and validated the contribution of each architectural component. Attention visualization highlighted anatomically relevant dorsolumbar regions associated with both traits. These results demonstrate that integrating complementary geometric information from multiple viewpoints enables accurate prediction of subcutaneous biological traits from external morphology, establishing a scalable, non-contact phenotyping paradigm suitable for real farm conditions and with strong potential applications in precision livestock farming and animal breeding programs.

Keywords: 2026

How to Cite:

ding, x., li, q., mi, y. & zhang, z., (2026) “Learning 3D Features from Dual-View Point Cloud for Estimation of Swine Subcutaneous Biological Traits”, World Congress on Genetics Applied to Livestock Production Digital Archive 2026(1): 2283741. doi: https://doi.org/10.31274/wcgalp.23508

Rights: 1