Reproduce “Beyond Specialization: Assessing the Capabilities of MLLMs in Age and Gender Estimation” on OpenResearch

About this paper

Maksim Kuprashevich, Grigorii Alekseenko, Irina Tolstykh

Researchers at SaluteDev's Layer Team demonstrated that a fine-tuned open-source Multimodal Large Language Model (MLLM), ShareGPT4V-7b, surpassed their state-of-the-art specialized model, MiVOLO384, for age and gender estimation, achieving an MAE of 3.52 compared to 3.65 on the LAGENDA benchmark. This work evaluates the ability of MLLMs to perform specific computer vision tasks against dedicated models.

Key insights

• Fine-tuning Multimodal Large Language Models (MLLMs) significantly boosts their performance, enabling them to surpass specialized models in accuracy for tasks like age estimation with relatively simple adaptation.
• General-purpose MLLMs exhibit strong generalization for age estimation (e.g., ChatGPT-4O) and generally robust gender identification, though OpenAI's ChatGPT models showed weaker gender classification compared to other MLLMs.
• While MLLMs can achieve higher accuracy when fine-tuned, their current computational cost is orders of magnitude greater than specialized models, making them suitable for accuracy-critical, non-real-time applications like data labeling.

Summary by alphaXiv.