Understanding the three-dimensional social behaviors of freely moving large mammals is crucial for agriculture and life science. Existing methods for animal pose estimation often fail due to occlusions during close interactions. This study introduces MAMMAL (Multi-Animal Mesh Model Alignment), a system using an articulated surface mesh model to capture 3D surface motions of multiple animals, even with severe occlusions. MAMMAL enables quantitative analysis of pig locomotion, postures, animal-scene interactions, social interactions, and tail motions. Experiments on mice and dogs demonstrate MAMMAL's generalizability.

Liang An, Jilong Ren, Tao Yu, Tang Hai, Yichang Jia, Yebin Liu