The Heat Based Rump Descriptor for Identification of Very Thin Goats in Dairy Farms

In this paper we address the problem of identifying goats in dairy farms with sub-optimal values of Body Condition Score (BCS). The BCS conveys information on whether an animal is fat or thin and its identification has a strong economic impact as very thin animals have poorer milk production and associated health problems. Albeit its important implications, not only there is no automated way of assessing the BCS in dairy farms, but current available techniques require specially trained personnel. However, the recently introduced Pictorial Scale for BCS assessment in dairy goats shows that the rump region has several visual cues strongly correlated with animal’s BCS values. In this paper we move towards the automatic assessment of BCS by developing a descriptor for rump’s 3D surfaces, collected by an RGB-D camera. The use of 3D surfaces as the basis for identification is fundamental, as it allows data collection without requiring animal handling to ensure careful alignment between camera and animal. However, the identification of the rump region in the 3D surfaces is very difficult, which leads to a large variability in the type of surfaces associated with the same BCS value. The descriptor we here introduce, the Heat Based Rump Descriptor (HBRD), uses diffusion geometry concepts to seamlessly handle the difficulty in defining a rump region and the resulting large variability of shapes. We test our descriptor in a dataset of 32 dairy goats and show that our descriptor is able to effectively cluster all the very thin animals.