Can the unique pattern of a cow's muzzle serve as a reliable biometric identifier, much like a fingerprint in humans? This question lies at the heart of new work published by the STRADE group in Pattern Recognition, one of the most respected journals in computer vision and machine learning.
Cattle identification based on muzzle patterns is a promising modality for livestock management and traceability, yet existing approaches often suffer from bias induced by the specifics of individual datasets and from being limited to a closed set of identities. In this work, we introduce a rigorous open-set verification benchmark designed to systematically mitigate this bias.
Method
We conducted an extensive ablation study evaluating the interplay between three factors: architecture (convolutional networks versus vision transformers), input resolution, and modern angular margin loss functions, specifically ArcFace, CosFace, and AdaFace. A total of twenty architectures were compared.
Results
The results revealed an interesting dependency between architecture and resolution. At lower resolution (256 × 192), keypoint-based alignment boosted the accuracy of convolutional networks up to 95.45% (VGG11-BN with ArcFace). At higher resolution (384 × 384), the Swin Transformer achieved the best result of 96.54%, but only with explicit alignment, whereas modern convolutional networks (e.g. MobileNetV3-L) reached a comparable accuracy of 96.14% even on unaligned crops. Sensitivity to alignment thus proves to be architecture-dependent: pronounced for the Swin Transformer, yet virtually absent for DeiT and modern convolutional networks.
Significance
The work shows that even with the limited amount of data available in this domain, transformers can effectively capture the distinctive features of the muzzle. The source code and pair lists will be released to support reproducibility of research in animal biometrics.
Tomáš Goldmann, Samuel Šimún, Filip Orság
The muzzle as a biometric feature: Cattle identification using unique muzzle patterns
Pattern Recognition, Elsevier
https://www.sciencedirect.com/science/article/pii/S0031320326013828
