Automatic coarse-to-fine method for cattle body measurement based on improved GCN and 3D parametric model

Abstract

In the era of smart agriculture, automatic livestock body measurement plays a pivotal role in optimizing growth management and enhancing health monitoring. Accurate, automated, and non-contact methods for assessing morphological traits are vital for improving operational efficiency and minimizing human error. However, current 3D vision-based methods often suffer from limitations in accuracy and processing time, particularly when dealing with data with holes and noise. Moreover, given the lack of high-precision 3D scanning data for livestock, we observe that the existing parametric models are not expressive enough to represent extensive variations in shape and pose. To address these challenges, this paper proposes an innovative, automated two-stage coarse-to-fine method for cattle body measurement that integrates both parametric and non-parametric representations. Our approach initiates with a coarse estimation phase, where the alignment of the parametric model with the point cloud data is guided by shape priors, pose priors, and detected key points. Subsequently, to surpass the limited shape space of parametric models, the mesh refinement phase is introduced, predicting per-vertex non-parametric deformations via an encoder-decoder Graph Convolutional Network (GCN). Specifically, the downsampling and upsampling GCNs take as input a combination of point cloud-level global features and vertex-level local features to progressively extract non-linear variations in hierarchical meshes. Finally, morphological traits are measured using the reconstructed mesh. We measured 9 body dimensions of 100 cattle, with a mean absolute percentage error (MAPE) of 3.58%. Compared to previous body measurement methods, our approach demonstrates superior accuracy and speed. Further comprehensive evaluations and ablation studies provide strong validation for our scheme. The code and data can be accessed at https://github.com/luhexiao/BM-GCN.