Cow identification network trained with similarity learning

Cow identification is a key phase in automated processing of cow video footage for behavioural analysis. Previous cow identification works have achieved up to 97.01% accuracy on 45 cows and 94.7% on datasets containing up to 200 cows. This paper presents new results from applying similarity learning to a cow identification Convolutional Neural Network on a group of 537 cows. Our method achieves identification accuracy of up to 99.3% and generalizes well to new cows, eliminating the need for retraining every time a new cow is added to the heard