Papers & full reports
Multi-Modal Intelligent Student Monitoring Model Based on Attention and Lightweight Modules
Xiaowei Feng, sole author. This is the available publisher-formatted manuscript; the verified record is conference acceptance. A published proceedings version or DOI has not yet been confirmed for this page.
The starting point
Concern about school-based sexual violence led me to explore whether expression recognition and pose estimation could contribute to auxiliary warnings in one-to-one teacher–student meetings. The idea also raised questions about privacy and students’ ability to question a system.
Developing and evaluating the model
I adjusted attention and feature-filtering mechanisms and compared experimental results. The expression classifier achieved 91.35% accuracy on RAF-DB. The paper also examines pose estimation.
The paper, Multi-Modal Intelligent Student Monitoring Model Based on Attention and Lightweight Modules, was accepted by CIMSP 2026. I am its sole author and received methodological guidance during the research.
What this result establishes
Expression classification on a public benchmark dataset.
Reliable identification of abuse in real meetings, or effective protection of privacy in deployment.
A classification error rate cannot be converted into a rate of missed assaults. An expression alone cannot explain someone’s situation or replace listening to their account.
Questions beyond accuracy
- Who can access monitoring information, and who must respond to an alert?
- Can a student refuse monitoring or challenge an interpretation?
- Will a request for help be taken seriously even when the system gives no warning?
These questions are part of why I want to study gender, institutions, and power alongside technical methods. Improving a model and evaluating the conditions for using it are both necessary work.