Wild ActionFormer: Enhancing wildlife action recognition for 11 endangered species in Wolong

A novel video self-supervised learning-based network, Wild ActionFormer, achieves 95.09% Top-1 accuracy in recognizing actions of 11 endangered species in the Wolong Nature Reserve, marking a significant advancement over baseline models. This breakthrough leverages UniformerV2 as its backbone and incorporates differential dispersion regularization loss to enhance feature extraction robustness while mitigating interference. The Focal Loss reweighting strategy further optimizes performance for long-tail classes, addressing data imbalance biases. Validated on the LoTE-Animal dataset—comprising 10,000 videos of species like giant pandas and Sichuan snub-nosed monkeys—the study demonstrates the potential of automated action recognition in conservation research, offering precise insights into wildlife behavior without manual annotation.

Liu, D., Sun, Q., Hou, J., Zheng, B., Zhang, J., Li, D., Norton, T., & Ning, J. (2025). Wild ActionFormer: Enhancing wildlife action recognition for 11 endangered species in Wolong. Ecological Informatics, 89, 103148. https://doi.org/10.1016/j.ecoinf.2025.103148