Zero-Shot Cross Sensor Generalisation for Non-Invasive Canine Core

Rectal thermometry is the clinical gold standard for canine body temperature measurement, but it is invasive and unsuitable for continuous monitoring. In this letter, we present a thermal-imaging pipeline for non-invasive canine core temperature estimation that combines facial thermal imagery with ambient temperature measurements and generalises across heterogeneous thermal sensors. A model trained on high-resolution FLIR Boson data was evaluated in a zero-shot manner on previously unseen dogs recorded with a low-cost FLIR Lepton sensor under different environmental conditions. The proposed system achieved an RMSE of 0.52°C without retraining, with a slight improvement to 0.51°C after temporal smoothing. These results show that models trained on high-fidelity thermal data can be deployed on low-cost embedded thermal cameras for scalable animal health monitoring.

Here we provide the data used for this study to help boost further research in this area.