3D deep learning identifies pigs using point cloud data during different growth stages of grow-finish pigs

Effective precision livestock farming relies on the ability to accurately identify individual animals. While ear tags offer a straightforward identification method, they are rarely used in grow-finish pigs due to the cost of the tag, labor demands to install the tags, and the resulting decrease in carcass value when the ear is damaged. Therefore, this study aims to identify pigs from top-down point clouds using a 3D deep learning algorithm. Three different time periods of grow-finish pigs were tested (early, mid-, and late-finishing). It was also tested for long-term identification across growth stages. PointNet architecture is used to categorically classify more than 20 individuals using approximately 800-point clouds per individual. The method was tested for identification in the short term, up to 10 days, and long-term throughout their time in the grow-finish pen. For the early stage, the model could achieve an overall macro-averaged F1 score of 0.72 the day after training data collection, but could not achieve consistent identification in the future 10 days. However, for the mid- and late-finishing stages, the model could achieve a consistent F1 score greater than 0.68 for short-term identification. Performance between early- and mid-finishing animals was poor (F1 score: 0.2), and even lower between early- and late-finishing animals (F1 score: 0.1), whereas classification between mid- and late-finishing animals performed notably better, with F1 scores exceeding 0.47 across all remaining days. Pigs can be identified after mid-growth using 3D deep learning; however, retraining with limited data between mid- and late-finishing stages leads to highly accurate identification.