Realistic monkey body animation reveals an uncanny valley in macaque body perception.

Social interactions are essential for survival in primates, relying on both facial expressions and body signals. The accurate characterization of these signals is critical for understanding the neurocomputational mechanisms underlying social communication. While previous work has focused on recognizing monkey behavior, a causal and direct manipulation of individual cues strongly benefits from believable, dynamic body avatars-analogous to those successfully developed for faces. Creating lifelike monkey avatars with realistic body motion, however, is challenging. Acquiring sufficiently accurate movement data for animation with marker-based motion capture is impractical, and markerless tracking methods require extensive manual labeling. To address this, we developed MacAction, a realistic macaque body avatar animated from multi-camera markerless tracking data. Our method reconstructs accurate trajectories for a large number of keypoints, as required for the 3D animation of realistic body models. The entire time course of individual actions is captured using only two labeled keyframes per second, with performance further validated on a large-scale human multi-view dataset. We assessed the animation quality of our dynamic avatar in a free-viewing experiment with eight macaque observers for single macaque actions, where fixation behavior was indistinguishable between our animations and matched real videos. Moreover, by systematically varying the realism of the avatar, we found an uncanny valley effect in macaque body perception, similar to that previously described in both humans and macaque faces. These findings support the commonalities of social vision across primate species, providing a foundation for controlled experiments aimed at clarifying the detailed neurocomputational mechanisms of social body perception in primates.