Abstract
Teacher attire plays a significant role in classroom dynamics, influencing student cognition, trust, and the overall learning atmosphere. However, automated assessment of professional dress has been hindered by a lack of specialized datasets. This paper introduces the Teachers’ Classroom Dress Assessment (TCDA) dataset and the Teachers’ Dress Assessment Method (TDAM). By leveraging deep learning for multi-label attribute recognition and a rule-based scoring system, the proposed method provides a quantitative framework for evaluating teacher professionalism in educational settings.
Introduction
In educational environments, a teacher's appearance is more than a personal choice; it is a component of the professional atmosphere that impacts student engagement. While schools often implement dress codes, manual monitoring is subjective and labor-intensive. Existing pedestrian attribute recognition datasets, typically sourced from surveillance or retail contexts, do not capture the specific requirements of classroom professionalism. To address this, we present a novel dataset and an automated scoring pipeline designed specifically for the educational context.
The TCDA Dataset
The Teachers’ Classroom Dress Assessment (TCDA) dataset consists of 11,879 samples collected from diverse real-world educational scenarios, including national teaching competitions and middle school classrooms across nine subjects. Each sample is annotated with 25 specific clothing attributes categorized as positive, negative, or neutral.
The dataset includes four viewpoints (front, back, left, and right) to ensure robustness against occlusions common in classrooms, such as lecterns or desks. To protect privacy, faces were obscured using RetinaFace while preserving clothing details for analysis.
Methodology: The Teachers’ Dress Assessment Method (TDAM)
The TDAM framework operates in three distinct stages:
- Feature Extraction: A ResNet50 backbone extracts high-level visual features from multi-angle images.
- Multi-Label Classification: The model predicts 25 attributes (e.g., sleeveless, high collar, short trousers) using binary classifiers. A weighted binary cross-entropy loss is used to manage the inherent class imbalance in attire frequency.
- Quantitative Scoring: Attributes are mapped to a professionalism score. Starting from a base score of 10, the system adds 2 points for "positive" attributes (e.g., long trousers) and subtracts 2 points for "negative" attributes (e.g., hats).
Results and Discussion
Benchmarking various architectures revealed that ResNet50 and TResNet-M provide the best trade-off between accuracy and computational efficiency. The ResNet50-based TDAM achieved a Mean Absolute Error (MAE) of 5.4579 and a Root Mean Square Error (RMSE) of 6.2199. A key finding was the importance of the confidence threshold; a threshold of 0.5 was found to be optimal for balancing false positives and model stability during training.
Conclusion
This work establishes a baseline for automated teacher dress assessment. By providing both a specialized dataset and a scoring methodology, it enables schools to regulate professional standards more objectively and efficiently. Future research will focus on temporal analysis in video streams to account for movement and exploring cultural variations in professional attire norms across different regions.
Reference: Fang, M., Liu, Q., et al. (2024). A Teacher Classroom Dress Assessment Method Based on a New Assessment Dataset. Proceedings of the 33rd International Joint Conference on Artificial Intelligence (IJCAI-24).
Summary updated