Proceedings of the Annual Meeting of the Cognitive Science Society

Wearable head-mounted cameras offer a window into infant's social world, yet manual coding remains a computational bottleneck. To address this limitation, we evaluate a scalable, open-source pipeline for automated facial expression estimation using the EfficientNet-B0 architecture on naturalistic, egocentric videos. Despite non-canonical viewpoints, poor lighting and motion artifacts, the model achieves performance (F1 = 0.543) comparable to state-of-the-art benchmarks. We demonstrate representational alignment (r = 0.76) between model confusion patterns and human perception, suggesting that the model reflects the adult social-perceptual structure. While categorical accuracy varies, the model's distributional probabilities capture the graded nature of affect. We also identify a critical divergence in neutral expression processing, where model and human ambiguity dissociate (r = -0.26). These findings support automated facial expression estimation as a viable approach for quantifying infants' natural social input, while highlighting category-specific limitations. Together they represent a step toward a "big data" science of early social development.