JerCCows

Multi-camera cattle dataset made publicly available upon manuscript acceptance of Early Fusion Multi-View Aggregation for Multi-Camera Cattle Tracking

This is the JerCCows dataset from "Early Fusion Multi-Vew Aggregation for Multi-Camera Cattle Tracking"

Abstract
Intensive dairy farming operations require automated monitoring solutions to efficiently manage large herds across expansive areas. However, existing approaches face significant limitations. Single-camera systems provide insufficient coverage due to blind spots and reduced spatial resolution at greater distances. Current multi-camera tracking methods depend on detecting animals in individual views before cross-camera association, and are often restricted to specific barns or breeds. We introduce BEVine (bird's eye view for bovine tracking), a novel open-source multi-camera tracking framework that performs early multi-view aggregation by detecting animals directly in a unified bird's eye view (BEV) representation. Our algorithm achieves robust performance across two distinct farm datasets with varying camera configurations and cattle breeds: our JerCCows dataset (8 cameras, Jersey cattle) and the publicly available MmCows dataset (4 cameras, Holstein cattle), achieving multi-object tracking accuracies of 84.6% and 85.7%, respectively. Our practical visual localisation pipeline generates BEV ground-truth positions from time-synchronised multi-camera footage, supported by a web-based user interface for annotation refinement. We introduce a multi-sequence training protocol that prevents scene-specific overfitting of the temporal BEV feature cache, and two complementary architectural extensions: per-camera image auxiliary supervision providing explicit foot-point and bounding box geometry, and a differentiable calibration refinement module that learns per-camera extrinsic corrections end-to-end. These results establish early fusion BEV tracking as a viable and scalable solution for precision livestock farming across diverse agricultural settings. The code is available at https://github.com/MahejabeenNidhi/BEVine \