AutoCattlogger: Track-assisted, automatic, instant cattle identity cataloger for recognition and beyond

We create an automatic Holstein cattle cataloging system called the AutoCattlogger to address the dynamic needs of dairy farms with frequently changing herds. This system instantly registers the identities of cattle by tracking them passing underneath a camera. The AutoCattlogger saves the identities of each individual cow in a predefined, interpretable representation space as barcodes, which allows for direct addition or deletion of cow identities without expensive model retraining. Thus, this system only requires a farmer to make a new cow walk under a camera for its identity to be instantly learned. Additionally, it is also capable of one-shot learning from single images of cows when videos are not available, has built-in illumination calibration functionality, and is robust to varying cow poses. The AutoCattlogger supports AutoCattleID, which is a cattle identification functionality that yields consistently high Track-level Top-1 recognition accuracy of above 90% even after a year on unannotated, unsegmented dairy videos. This demonstrates the robustness of our system to long-term changes in cow sizes and scene lighting conditions, which is a hard problem for Deep Metric Learning based recognition methods. The AutoCattlogger, with its tracks and identities, can thus serve as the foundation for many cattle video analytics tasks such as video retrieval, weight estimation, and primarily, cattle recognition. • AutoCattlogger tracks cows under cameras and learns their identities as cowbarcodes. • Barcode feature space allows direct and instant addition/deletion of cow identities. • Barcode features make AutoCattlogger also good at one-shot learning. • AutoCattlogger calibrates for scene lighting from reflections on fully black cows. • Barcode based cattle recognition is robust to long-term changes in cow appearance.