Head–Body–Tail Segmentation of Poultry 3D Point Clouds Under Different Annotation Budgets: A Comparison of PCA-Based Rules and PointNet++-Based Methods

Abstract

Head–body–tail segmentation of poultry 3D point clouds is an underexplored task in precision poultry phenotyping. This study constructs a segmentation-oriented evaluation setting to examine how supervision source and annotation budget affect the relative performance of a principal component analysis (PCA)-based geometric rule and PointNet++-based segmentation methods. The evaluation setting was derived from a previously published single-view poultry point-cloud dataset originally collected for body-weight prediction, to which point-wise anatomical annotations, PCA-derived pseudo-labels, fixed data partitions, and a 500-sample manually annotated test set were added for the present segmentation task. The results show that the PCA Rule achieved the best performance under pseudo-label supervision, indicating its value as a strong low-cost baseline when no human-annotated training labels are available. With increasing annotation budgets, learning-based methods progressively surpassed the rule-based baseline, while the performance gain became marginal beyond 1000 annotated samples. A cost-performance analysis further shows that PCA-based geometric partitioning provides a highly efficient solution in low annotation scenarios, whereas learning-based methods become more advantageous when moderate human supervision is available. As auxiliary application validation, the weight-regression experiments showed that anatomical decomposition did not outperform the complete whole-body representation under the present data and model settings. These findings provide empirical evidence on the relative behavior of geometric and learning-based segmentation methods under different annotation conditions and establish a reproducible evaluation basis for subsequent poultry point-cloud part segmentation studies.