Maarten Perneel joins CV4Animals Workshop 2023

Dynamic multi-pose, multi-viewpoint re-identification of Holstein-Friesian cattle
Holstein-Friesian dairy cattle have distinct, unique coat patterns, offering opportunities to identify animals in video captures obtained during experiments requiring behavioural analysis. However, for applications in practice, a re-identification algorithm has to deal with images of freely moving animals, which is a complex task due to the high asymmetry of cattle coat patterns. Therefore, we introduce a re-identification algorithm that combines pose-estimation with embedding generation to handle this complexity. Our methodology requires only a single embedding network to be trained and obtains a rank-3 accuracy of 95.0%, offering new perspectives to perform tag-free behavioural research in dairy cattle.