Accurate phenotypic measurement of pigs is critical for optimizing breeding strategies, enhancing production efficiency, and ensuring animal welfare in modern intensive livestock farming. Existing 3D reconstruction approaches remain constrained by pose variation, occlusion, and high hardware costs. Achieving accurate body measurements in uncontrolled and freely moving environments remains particularly difficult. To address these issues, this study proposes Pig3D, a pig-specific deformable statistical model, and a novel two-stage multi-view 3D reconstruction framework based on RGB image. In the first stage, a deep neural network, Pig3DNet, is designed to estimate pig shape and pose parameters within a normalized 3D space. The network does not rely on 3D supervision but instead through 2D projection constraints from keypoints and segmentation masks. A geometry-driven view quality and fusion (GVQF) module is proposed to evaluate the geometric reliability of each view, enhance multi-view feature fusion, and adaptively emphasize lateral views that provide clearer structural information. In the second stage, the predicted shape and root pose parameters from the first stage are combined with calibrated camera parameters to perform scale recovery optimisation, enabling accurate 3D reconstruction of pigs in real-world coordinates. Experimental results show that the proposed method achieves IoU of 80.19%, PCK@0.15 of 96.21%, and RMSE of 6.93 pixels on the Pig3d-MV dataset. The five measured parameters include body length (BL), chest width (CW), hip height (HH), abdominal girth (AG) and chest girth (CG), BL and HH achieve the highest accuracy, with MAE and MAPE below 2.4 cm and 5.74%, respectively. The proposed framework offers promising potential for practical deployment in precision livestock farming.
Xiang et al. (Mon,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: