A Teacher Classroom Dress Assessment Method Based on a New Assessment Dataset

Ming Fang, Qi Liu, Yunpeng Zhou, Xinning Du, Qiwen Liang, Shuhua Liu

TL;DR

Researchers have introduced the first-of-its-kind Teachers’ Classroom Dress Assessment (TCDA) dataset and a corresponding automated scoring method (TDAM). By leveraging computer vision to identify specific clothing attributes, the system can objectively evaluate whether a teacher's attire meets professional standards, aiming to enhance the classroom learning environment.

What Problem Does This Paper Address?

In the field of education, a teacher’s professional image significantly impacts student perception, trust, and the overall learning atmosphere. While many institutions have dress codes, enforcing them currently relies on manual, subjective, and labor-intensive "spot checks."

From a technical standpoint, while "pedestrian attribute recognition" (identifying if someone is wearing a hat, glasses, etc.) is a well-studied area in surveillance and retail, it hasn't been applied to the nuances of a classroom. Existing datasets focus on outdoor streets or malls, which don't capture the specific professional requirements or the environmental constraints (like being partially hidden behind a lectern) of a school setting. The authors aim to bridge this gap by providing both the data and a baseline model for automated teacher attire assessment.

Methods

The authors propose the Teachers’ Dress Assessment Method (TDAM), which operates in two primary stages:

  1. Attribute Recognition: The system uses a ResNet50 backbone as a feature extractor. It treats the problem as a multi-label classification task, predicting the presence or absence of 25 specific attributes (e.g., "high collar," "short skirt," "without hat"). To handle the inherent imbalance in how often certain clothes appear, they employ a weighted binary cross-entropy loss function.
  2. Score Generation: Instead of a "black box" score, the model uses a transparent, rule-based generator. It starts with a base score of 10 points. It then adds 2 points for "positive" professional attributes (like long trousers or a high collar) and subtracts 2 points for "negative" attributes (like hats or sleeveless tops).

To protect teacher privacy, the researchers integrated RetinaFace to automatically detect and encode faces in the video clips before processing.

Datasets

The core contribution is the Teachers’ Classroom Dress Assessment (TCDA) dataset:

  • Scale: 11,879 teacher dress samples.
  • Content: 2,354 video files collected from real-world scenarios (National Teacher Competitions, middle school daily recordings) and 1,295 simulated "negative example" videos recorded by volunteers.
  • Annotations: Each sample is labeled with 25 attributes across four viewpoints (front, back, left, right) to account for classroom occlusions like desks or lecterns.
  • Diversity: The data covers nine different high school subjects to ensure the model isn't biased toward the typical style of, say, a PE teacher versus a Math teacher.

Key Insights and Ideas

  • Attribute-Based Scoring vs. Direct Regression: Rather than training a model to guess a "score" directly from an image, the authors break the task down into discrete attributes. This makes the AI's decision-making interpretable—an administrator can see exactly why a score was low (e.g., "detected sleeveless top").
  • The "Classroom Perspective": Most pedestrian datasets assume a clear view of the person. TCDA specifically includes four viewpoints because teachers are frequently obscured by classroom furniture.
  • Defining "Professionalism" via Data: The paper explicitly categorizes clothing into "positive," "negative," and "other" (neutral) attributes based on existing educational professional norms, effectively codifying a subjective social standard into a machine-learning objective.

Key Results

The authors tested several baseline architectures, including various ResNet versions, TResNet, and Vision Transformers (ViT):

  • Optimal Performance: ResNet50 emerged as the best balance between accuracy and computational efficiency, achieving a Mean Absolute Error (MAE) of 5.4579 and a Root Mean Square Error (RMSE) of 6.2199.
  • Model Stability: The experiments showed that increasing model depth (e.g., moving to ResNet101) actually led to overfitting, likely due to the specific, constrained nature of the classroom environment compared to massive general-purpose datasets.
  • Thresholding: They found that a confidence threshold of 0.5 for attribute detection provided the most stable results, preventing an influx of "false positives" that would unfairly penalize teachers.

Limitations and Future Work

  • Cultural Specificity: The definitions of "positive" and "negative" attire are based on specific professional norms (likely aligned with Chinese educational standards, given the authors' affiliation). These standards vary significantly across different cultures and countries.
  • Occlusion Challenges: While the dataset includes multiple viewpoints, extreme occlusion (e.g., a teacher sitting behind a desk for an entire clip) remains a challenge for attribute recognition.
  • Dynamic Attire: The current model evaluates static snapshots or short clips. Future work could involve temporal analysis to see how a teacher's "professional score" might change during active movements, like writing on a high chalkboard.

Summary updated

BibTeX
@inproceedings{fang2024teacher,
  author = {Ming Fang and Qi Liu and Yunpeng Zhou and Xinning Du and Qiwen Liang and Shuhua Liu},
  title = {{A Teacher Classroom Dress Assessment Method Based on a New Assessment Dataset}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure