From Pixels to Prompts: Evaluating ChatGPT-4o in Face Recognition, Age Estimation, and Gender Classification

Abstract

The rapid development of multimodal large language models (MLLMs) has opened new possibilities for semantic reasoning over images, yet their capabilities in face understanding remain underdeveloped.This article presents a comprehensive evaluation of ChatGPT-4o's performance in age estimation, gender classification, and identity verification in two challenging datasets: the In-the-Wild Celebrity Children (ITWCC) dataset, containing 7, 990 images of children aged 6-17, and a Surgery Face dataset consisting of paired preoperative and postoperative images of pediatric patients.Tailored "AI-generated image" prompts were used to bypass built-in safeguards.The results show that ChatGPT-4o outperformed conventional face recognition models, achieving a mean absolute error (MAE) of 1.8 years for age estimation, with 82% of predictions within 2 years.It demonstrated 96% gender classification accuracy (F 1 = 0.96) and a 100% true match rate in identity verification for longitudinal pairs, compared to DeepFace 67%.Furthermore, ChatGPT-4o inferred identity in 95% of the cases for surgical pairs, while Oriented FAST and Rotated BRIEF (ORB) feature matching averaged 48 key points.These findings highlight the potential of MLLMs to surpass traditional CNN-based approaches, offering robust, interpretable, and rationale-rich outputs for biometric tasks, although limitations remain in handling extreme facial transformations.

Ähnliche Arbeiten