PrimateFace: A Foundational Resource for Generalizable Cross-Species Facial Analysis

Felipe Parodi, Konrad Kording

Keywords: face detection, face pose, cross-species, primatology, gaze, dataset

TL;DR: PrimateFace, a cross-species dataset & models that show pre-training on taxonomically diverse data creates generalizable models for animal behavior that generalizes from primates to humans, enabling scalable, multi-species biosignal analysis.

Abstract: The analysis of facial kinematics, a critical class of behavioral biosignals, is fundamental to understanding social cognition. Progress has been hampered by the lack of large-scale, diverse datasets needed to train robust and generalizable models for face detection and facial landmark estimation (FLE). We introduce PrimateFace, a foundational resource designed to overcome this bottleneck. It consists of a large-scale dataset of over 260,000 images spanning more than 60 primate genera, and a suite of pretrained models. Models trained on PrimateFace achieve high cross-species performance, from tarsiers to gorillas, and demonstrate remarkable generalization to human data, validating the benefits of pre-training on taxonomically diverse data. We showcase how PrimateFace serves as an essential front-end for diverse downstream applications, including quantifying social gaze in human infants, enabling multimodal analysis of vocalizations, and powering the data-driven discovery of behavioral repertoires. PrimateFace provides a standardized platform for extracting and analyzing behavioral biosignals, empowering scalable, data-driven studies of behavior.

Submission Number: 82