Pig localization using computer vision

Finding a pig in a large herd is tedious as pigs are housed in large groups and look-alike for humans. This project aims to solve the problem of finding pigs in a shed using computer vision. The problem is split up into acquiring an image of pigs, detecting, identifying and locating pigs. The focus is on the identification of pigs. A method similar to human face recognition is proposed where new pigs can be added without retraining. A dataset of 105 individual pigs is collected to train the network and evaluate the method. The evaluation shows that ten pigs seen during training of the embedding network can be identified with an accuracy of 87.8%. Identifying pigs not seen during training of the embedding network can be identified with an accuracy of 55.5%. There are still limitations when detecting pigs in another perspective or later in time than added to the identification system.