Computer vision makes animal identity a validation issue, not a camera feature

A computer-vision system can broaden individual cattle identification, but it cannot yet reliably replace established systems under real farm conditions. That is the central finding of a project published on 10 November by Meat & Livestock Australia (MLA). The prototype was intended to identify animals individually from visual features and complement conventional identifiers such as RFID tags, without manual handling.

On a controlled dataset, the system achieved 100% top-1 identification accuracy. On a second dataset designed to reflect practical farm conditions, performance fell to 62.5%. The gap is technically important: lighting, camera angle, movement, breed and environmental variation turn visual identification from a laboratory task into a robustness challenge.

MLA sees potential for the technology to support traceability and automation, as well as to link with vision applications for behaviour monitoring or weight estimation. It could also reduce animal handling and associated stress. The results, however, do not support presenting the system as an immediate RFID replacement: the report itself recommends integrating additional features, improving data fusion, analysing video sequences and testing across more breeds, viewpoints and lighting conditions.