Abstract
Despite significant progress in computer vision for precision livestock farming, a critical challenge persists: the lack of effective long-term cattle re-identification datasets under real-world conditions, primarily due to substantial phenotypic changes. To address this challenge, we present the Beef Cattle dataset (BECA), a novel, large-scale dataset specifically designed to support long-term and diverse cattle recognition from dorsal views. BECA encompasses two sub-datasets: the Beef Cattle dataset for Diversity (BECA-D), containing 16,889 images from 5,661 beef cattle across multiple breeds, designed to capture visual diversity for recognition; and the Beef Cattle dataset for Long-term recognition (BECA-L), comprising 12,172 annotated images from 103 cattle tracked over a continuous period of up to five months-representing a notably long duration for cattle recognition datasets. Additionally, we provide annotation subsets for key pipeline tasks such as object detection and pose estimation, supporting the development of diverse models for long-term recognition. Collectively, BECA establishes a comprehensive benchmark for vision-based livestock management, facilitating research in cattle identification, behavior analysis, and welfare monitoring.