Automatic identification of individual yaks in in-the-wild images using part-based convolutional networks with self-supervised learning

Yaks (Bos grunniens) are the most important domestic animals for people living at high altitudes. In order to implement precise livestock management for yaks, it is of significant importance to automatically identify, keep track of, and monitor yaks. Traditional animal identification methods such as ear tags, tattoos, and RFID based methods suffer from problems like animal infection, high maintenance cost, inefficiency or sensor failure. Existing biometric-based identification methods for livestock such as muzzle prints, iris patterns, and retinal vascular patterns mostly require that animals are under control, either technically or physically, and are thus costly to deploy especially for yaks which are loosely raised on the grassland pastures and migrate with the seasons. In this paper, we propose a novel method for identifying individual yaks in in-the-wild images captured under unconstrained conditions. We utilize the part-based convolutional network (PCN) to obtain discriminative part-level feature representations. To further enhance the feature discriminativeness and alleviate the impact of small amount of yak image data, we implement self-supervised learning strategy by proposing random erasure and region-visibility prediction (RERP) as an auxiliary learning task. Experiments performed on the YakReID-103 dataset demonstrate that (i) when left and right side views of yaks are treated separately, the Rank-1 accuracy and mAP achieved by the proposed method with SEResNet50 backbone are up to 97.57% and 76.30%, which significantly advance the state-of-the-art, and (ii) when generalizing to different views, the proposed method with ViT backbone again obtains the best results compared with the counterpart methods.