Captioning daily activity images in early childhood education: benchmark and algorithm
Multimedia Systems · 2026-11 · 32 (10)
DOI: 10.1007/s00530-026-02704-0 ↗Kurz gefasst
Fachartikel in der Zeitschrift Multimedia Systems, online erschienen am 23.10.2026. Verfasst von Sixing Li und 6 weiteren Autor:innen. Die Arbeit nennt 47 Literaturangaben. Gefördert von: the National Natural Science Foundation of China, 面向恶劣成像环境的鲁棒视觉智能感知技术.
Abstract (Originalsprache)
Early childhood education (ECE) produces large volumes of daily activity images that document classroom routines, play, peer interaction, and teacher-guided activities. Converting such multimedia records into educationally meaningful captions remains difficult because ECE scenes contain fine-grained activity areas, visually ambiguous teaching toys, and professional terminology that are underrepresented in generic captioning datasets. We present KinderMM-Cap, a multimedia captioning system built around ECAC, a controlled-access benchmark containing 251,954 real-world kindergarten images with expert-level captions, activity-area labels, and structured teaching-toy annotations. We introduce Teaching Toy Recognition Score (TTS) for evaluating professional toy-name accuracy and propose RSRS, a reward-conditional hybrid training framework that switches between reinforcement learning and supervised fine-tuning when sparse rewards cause zero-advantage groups. Based on ECAC and RSRS, KinderMM-Cap-3B achieves a TTS of 51.06 and an overall caption-quality score of 86.20, outperforming strong multimodal baselines. Precision analysis further shows that RSRS improves measured toy-recognition precision among the compared training strategies. The results suggest that child-sensitive multimedia captioning benefits from domain data, specialized evaluation, controlled data access, and reward-aware training.
Förderung
- the National Natural Science Foundation of China
- 面向恶劣成像环境的鲁棒视觉智能感知技术
Zitieren
APA
Li, S., Gu, Z., Zhang, Z., Pan, W., Li, B., Wang, Y., & Liu, H. (2026). Captioning daily activity images in early childhood education: benchmark and algorithm. Multimedia Systems, 32(10). https://doi.org/10.1007/s00530-026-02704-0
Vancouver
Li S, Gu Z, Zhang Z, Pan W, Li B, Wang Y, et al. Captioning daily activity images in early childhood education: benchmark and algorithm. Multimedia Systems. 2026;32(10). doi:10.1007/s00530-026-02704-0
BibTeX
@article{Li2026,
author = {Li, Sixing and Gu, Zhibin and Zhang, Ziqi and Pan, Weiguo and Li, Bing and Wang, Ying and Liu, Hongzhe},
title = {Captioning daily activity images in early childhood education: benchmark and algorithm},
journal = {Multimedia Systems},
year = {2026},
volume = {32},
number = {10},
doi = {10.1007/s00530-026-02704-0}
}
EUVIMED