A lightweight domain-specific appearance embedding for Turkey re-identification in dense barn environments

Individual-level monitoring of poultry in commercial barns requires identity-consistent tracking, which in turn depends on appearance features that can discriminate between visually homogeneous animals. Generic appearance models used in trackers such as DeepSORT are trained on human re-identification data and transfer poorly to livestock, where individuals lack the distinctive clothing and texture cues that drive pedestrian re-identification. In this work we present a lightweight domain-specific Siamese feature extractor for turkey re-identification, designed through a systematic, ablation-driven refinement process and trained with a combined batch-hard triplet and identity-classification objective. The final architecture combines pre-activation residual blocks, squeeze-and-excitation channel attention, and a BNNeck embedding head, and contains only 714K parameters — approximately 3.9 × fewer than the 2.8M-parameter mars-small128 model used by default in DeepSORT. On an open-set evaluation of 16 unseen turkey identities, the proposed embedding raises ROC AUC from 0.890 to 0.943, classification accuracy from 0.849 to 0.911, and the distribution separability index from 1.14 to 1.71, while reducing model storage by 3.8 × and inference latency relative to the baseline. A t-SNE projection of the learned space shows tight, well-separated clusters for the majority of identities, with residual overlap confined to the most visually similar individuals. When integrated into a DeepSORT tracking pipeline and evaluated on three commercial turkey-barn video sequences, the domain-specific embedding yields consistent gains in identity-related tracking metrics (IDF1 81.3 → 83.2%, association accuracy 68.8 → 69.9%). The results demonstrate that careful, domain-specific architectural design can substantially improve re-identification of visually homogeneous animals while simultaneously reducing computational cost. Within the scope of this controlled evaluation, the compact embedding is therefore well suited to resource-constrained edge hardware for continuous precision-livestock monitoring; large-scale field validation remains future work.