9 March 2026 in IEEE Access
SPSibi Chakravarthy Parivendan
Sibi Chakravarthy Parivendan
Pose estimation enables quantitative animal phenotyping, but progress is limited by sparse annotated datasets and domain variability in real-world farm environments. This study evaluates cross-species transfer learning by adapting a zebra pose model (ZebraPose) for 27-keypoint dairy cattle pose estimation. We systematically evaluated three training configurations: zebra-only pre-training, cattle-specific fine-tuning, and combined dataset training. While the combined training achieved strong benchmark performance (average precision [AP] = 0.86, average recall [AR] = 0.87, percentage correct keypoints [PCK] at 0.5 IoU = 0.869), generalization to unseen farm environments deteriorated significantly. Analysis reveals that synthetic pretraining data and anatomical similarity alone are insufficient for robust cross-farm deployment, despite comparable model architectures and training protocols. The persistent domain gap—characterized by environmental heterogeneity, lighting variation, and pose diversity—demonstrates that successful livestock phenotyping requires environment-aware model development and multi-site dataset diversity rather than single-species transfer learning approaches. These findings establish practical constraints on cross-species adaptation for animal pose estimation and inform future directions toward farm-adaptive phenotyping systems.