TOTTORI UNIVERSITYPattern Recognition System Laboratory
Menu

RESEARCH

Pedestrian Age Recognition using Contrastive Learning by Automatically Generating Descriptions with High Inter-class Separability

Key Points

We use a vision–language model to automatically generate descriptions of the characteristics of each age class and align image and text features through contrastive learning. This improves pedestrian age recognition while reducing the trial and error involved in manually preparing descriptions.

Overview of pedestrian age recognition using automatically generated age descriptions and contrastive learning

Abstract

This research aims to identify a person's age group from a whole-body image. AI that handles both images and text can help, but the descriptions of each age group matter. Preparing descriptions that clearly distinguish the groups has required people to repeatedly write and revise them.

We use AI that answers short questions about images to automatically produce descriptions of physical characteristics and things people are wearing. By including more words that express differences between age groups, these descriptions provide useful clues for recognition. We then train the system to match images with descriptions of the same age group. This helps it learn visual clues that distinguish one group from another.

We evaluated the method on four public datasets: RAPv2.0, PETA, MSP60k, and PA100k. It achieved an average accuracy of 84.1%, which was 4.6 percentage points higher than the best method used for comparison. The method also removed the trial-and-error step of manually preparing the descriptions.

Publications

Back to research