Human-in-the-loop generative pipeline for synthetic classroom training data with domain-expert validation¶

Strategy · All strategies
Evidence · no claims cited

Description¶

A reusable recipe for building synthetic training data for classroom understanding: consult teachers to select action classes, run a preliminary study of video generators to pick suitable models, generate uniform and scene assets with generative image models in a human-in-the-loop process, design prompts specifying CCTV camera placement, viewing angle, and subject scale, then apply multi-reviewer quality inspection and domain-expert validation. The authors state "the proposed pipeline provides a blueprint that researchers and practitioners can adapt to generate synthetic data for classroom understanding and related applications."

Design Implications¶

Context¶

Requirements¶

• Each generated video is retained only if three independent reviewers agree on visual realism, camera properties, motion realism, and action semantics
• Generated uniform assets are manually verified against corresponding real-world uniforms before inclusion

Constraints¶

• Generated classroom videos must be used carefully to avoid unrealistic behavior, privacy leakage, and biased incident representations
• Real student images were anonymized and used only as references for generating non-identifiable uniform assets

Target Learners¶

• preschool students and teachers in Singaporean partner preschools

Target Learning Goals¶

• creating realistic, privacy-sensitive training data for classroom incident recognition

Related Strategies¶

• Involve domain experts throughout the AI safety evaluation pipeline, especially for defining unsafe content
• Center co-design with teachers and students throughout AI tool R&D, using human-in-the-loop workflows as a primary risk mitigation

Examples¶

-

Key Sources¶

• Paritosh Parmar, Landy Lan, Hong Yang, Chen Yi, Chiat Pin Tay. (2026). Robust and Efficient Motion Reasoning for Privacy-Aware Classroom Incident Recognition. arXiv preprint. https://arxiv.org/abs/2608.05115