Cross-species knowledge sharing for improved animal activity recognition with limited labelled data

Deep learning dominates automated animal activity recognition (AAR) tasks due to high performance on large-scale labelled data. However, such data are often limited due to the laborious and time-consuming data labelling process, leading to weak feature representation ability and poor model performance. Existing works generally augment data sizes based on labelled data, exploit unlabelled data via semi-supervised learning, or pre-train models on large-scale datasets (e.g., ImageNet dataset). In this paper, we boost the capability of models trained on limited labelled data from a new perspective. Notably, there are increasing open-source datasets of various animal species in the field of wearable sensor-aided AAR. Different species may have similar movement patterns and thus possess shared features that can improve the model's feature representation ability. Nevertheless, data distributions across species are distinct due to their inherent discrepancy, bringing difficulty in learning generic representations. Therefore, we develop a Cross-species Knowledge Sharing network (CKS-Net), where convolutional layers capture species-shared features and a Species-specific Batch-Normalization (SBN) module is designed following each convolutional layer to avoid inter-species conflicts. The SBN module involves multiple BN layers that separately fit the distributions of different species. To verify our method's effectiveness, a case study of cattle behaviour recognition is conducted on data collected from six cattle via accelerometers. Two distinct public datasets from sheep and horses are used to provide shared knowledge. The results demonstrate that our method remarkably enhances the cattle behaviour classification performance with increments in precision, recall, F1-score, and accuracy of 14.27%, 0.22%, 8.71%, and 4.65%, respectively. © 2024 11th European Conference on Precision Livestock Farming. All rights reserved.