Comparisons of Full Body and Facial Dog Identification

Free-roaming dog population survey is very important in the veterinary research area. Modern technologies can promote the automation process of dog surveys. This paper investigates two networks, which are Facenet and Dlib for dog identification. These two networks predict embedded vectors which are a reduced dimension vector of images containing distance information that is aligned with dog identities. The full body and face images of dogs are fed into the networks. Our experiments show that facial images provide more discriminative embedded vector than full body images. The average accuracy of both networks with face and full body images are 91.43% and 87.32%, respectively. In overall, both networks provide good accuracy and promising results for dog recognition applications.