license: cc-by-nc-4.0
base_model: ultralytics/yolo11n
library_name: ultralytics
metrics:
- precision
- recall
- mAP
pipeline_tag: object-detection
tags:
- face-detection
- age-estimation
- yolo11
- onnx
YOLO11 Nano - Age Classification
This is a lightweight face detection and age classification model based on the YOLO11 Nano architecture. It was fine-tuned for 300 epochs on a heavily curated dataset of approximately 75,000 face images to classify individuals into three broad age brackets: 0-14, 15-22, and 22+.
The training data was sourced and filtered from IMDB-Wiki, Adience, UTKFace, and FairFace. Consequently, this model is released under a CC-BY-NC-4.0 license and is intended strictly for non-commercial research.
Performance
The model achieves high precision despite its small size, making it suitable for edge devices or local privacy-preserving inference.
- mAP@50: ~89.5%
- mAP@50-95: ~86.5%
- Recall: ~87%
- Precision: ~83%
Confusion Matrix
The table below shows the normalized prediction results. The columns represent the actual age (True Labels), and the rows represent the model's predictions.
| True: 0-14 | True: 15-22 | True: 22+ | |
|---|---|---|---|
| Pred: 0-14 | 0.91 | 0.07 | 0.01 |
| Pred: 15-22 | 0.08 | 0.74 | 0.16 |
| Pred: 22+ | 0.01 | 0.19 | 0.82 |
The model is highly reliable at identifying children, with 91% accuracy and minimal confusion with adults (only 1% of children were predicted as 22+).
Limitations & Legal
Since this model was trained exclusively on real photos, performance may degrade on AI-generated faces, cartoons, or 3D renders. However, personal tests suggest it maintains reasonable detection capabilities on realistic AI generated content.
- Accuracy: This model is not 100% accurate. It uses a "buffer" class (15-22) to handle the ambiguity of young adulthood, but obviously it's not perfect.
- Critical Use: It should not be used for critical age verification or legal identity checks.
- Biometric Privacy: This model processes biometric features (faces). It is provided for research and local-inference contexts only and should not be deployed for mass surveillance or without considering local regulations (such as the GDPR guidelines in Europe or CCPA in the US).