Robust broiler pose estimation in diverse farming environments: A teacher–student test-time adaptation approach

Broiler pose estimation (BPE) is pivotal for behavioral analysis and health monitoring in poultry farming, with recent deep learning significantly enhancing its precision. Existing models critically depend on the independent and identically distributed (i.i.d.) assumption between training and testing data. However, this assumption is frequently violated in dynamic farming environments due to varying lighting conditions, weather changes, and heterogeneous barn layouts. For example, models trained on sunny-condition images may perform poorly in rainy environments, a challenge known in artificial intelligence as domain shift. To address this, we introduce a novel teacher–student test-time adaptation (T/S-TTA) for BPE. T/S-TTA uses multi-domain data augmentation (MDA) to simulate real-world variability, broadening input distribution for unpredictable broiler conditions. Augmented data informs a teacher–student architecture, where the teacher extracts resilient features from transformed inputs. These features guide the student model to predict target images. To ensure consistent performance across shifting domains, dynamic sample selection (DSS) evaluates sample quality through prediction confidence metrics. High-quality samples reinforce reliable features through positive learning, while low-quality samples are suppressed via negative learning to suppress misleading features. Then, the random restoration mechanism resets student parameters periodically to retain source knowledge. The adapted parameters acquire cross-domain generalization, significantly improving final pose estimation accuracy. Experimental validation uses standardized dataset for general performance evaluation and challenging farm-collected dataset for domain shift testing. The framework achieves 72.3% mAP@50–95, 77.6% mAR, 75.3% [email protected] accuracy and 11.2 px Avg. PE in cross-domain setups, establishing state-of-the-art performance for BPE under dynamic farming conditions. • First to identify domain shift in broiler pose estimation. • Proposed teacher–student test-time adaptation (T/S-TTA) framework. • Enables deployment-stage model optimization. • Significant performance gains in diverse farming environments. • 72.3% mAP@50–95 and 75.3% [email protected] under domain shifts.