Evaluating transfer learning strategies for improving dairy cattle body weight prediction in a small farm using depth-image and point-cloud data.

Computer vision-based approaches have emerged as automated, non-invasive, and scalable tools for monitoring dairy cattle, thereby supporting effective management, health assessment, and productivity. Although transfer learning is commonly employed in studies predicting body weight from images, its effectiveness and optimal fine-tuning strategies remain poorly understood in livestock applications, particularly beyond the straightforward use of pretrained ImageNet or Common Objects in Context weights. In addition, while both depth images and 3-dimensional point-cloud data have been explored for body weight prediction, direct, head-to-head comparisons of these 2 modalities in dairy cattle are limited. Therefore, the objectives of this study were to 1) evaluate whether transfer learning from a large farm can enhance body weight prediction performance on a small farm with limited data, and 2) compare the predictive performance of depth-image- and point-cloud-based approaches under 3 experimental designs. Top-view depth images and point-cloud data were collected from 1,201, 215, and 58 cows at large, medium, and small dairy farms, respectively. Four deep learning models were evaluated: ConvNeXt and MobileViT for depth images, and PointNet and DGCNN for point clouds. We found that transfer learning consistently improved body weight prediction on the small-farm test set compared with single-source learning and achieved performance comparable to, or greater than, that obtained through joint learning. These results indicate that pretrained representations can generalize effectively across farms with differing imaging conditions and dairy cattle populations. No consistent performance difference was observed between depth-image- and point-cloud-based modeling pipelines under the data acquisition and processing conditions evaluated in this study. Taken together, our findings suggest that transfer learning provides a practical alternative to cross-farm data sharing for small-farm prediction scenarios, where direct sharing of raw data may be constrained by logistical, data-governance, or policy considerations, as it requires access only to pretrained model weights rather than raw data.