Cross-Species Behavioral Analysis via Multidimensional Knowledge Transfer

Computer vision tools can provide objective, reproducible, and scalable measurements for understanding animal behaviors, which is a fundamental challenge in cross-species research. Existing approaches primarily emphasize body detection and tracking, with limited advancement in integrative behavioral pattern analysis due to sparse annotations and substantial variability in species' physical and cognitive traits. In this article, we conduct two pilot studies for cross-species behavioral analysis within primates using multidimensional transfer learning. First, prior knowledge of physiological similarities between humans and monkeys is introduced to adjust the representation learning for macaque facial expression classification under limited labeled data. Second, we improve our approach by leveraging the foundation model and self-supervised strategy to mitigate appearance variations across primate species for downstream marmoset engagement level estimation. Experiments on public and private cross-primates datasets demonstrate the effectiveness of our framework, highlighting the potential of knowledge transfer to bridge the gap between species-specific behavioral analysis and generalizable models.