Aldey/tripleS-Member-Recognition

tripleS Member Recognition Model

This repository contains face recognition models specifically trained to identify members of the K-pop girl group tripleS. The models are built using deep learning techniques with InsightFace embeddings and can classify faces of all 24 tripleS members with varying levels of precision and recall depending on the model version.

Model Versions

This repository includes multiple model versions, each trained with different confidence thresholds to provide various trade-offs between precision and recall:

Base Model (Model-base/)

• Description: Base classifier model trained only on the tripleS Member Objekt dataset without any fine-tuning or external data
• Training samples: 3,480 samples
• Fine-tuned: No
• Use case: Baseline performance reference

Version 0.4 (Model-0.4/)

• Confidence threshold: ≥ 0.4
• Description: Trained using more inclusive data, resulting in a model that may generalize better but risks misclassification
• Training samples: 13,951 samples
• Characteristics: Higher recall, lower precision
• Use case: When you want to catch as many faces as possible, even with some false positives

Version 0.5 (Model-0.5/)

• Confidence threshold: ≥ 0.5
• Description: Balanced training set offering a trade-off between recall and precision
• Training samples: 13,707 samples
• Characteristics: Balanced recall and precision
• Use case: General-purpose recognition with balanced performance

Version 0.6 (Model-0.6/)

• Confidence threshold: ≥ 0.6
• Description: Trained only on high-confidence samples, producing a stricter classifier with higher precision
• Training samples: 13,358 samples
• Characteristics: Higher precision, lower recall
• Use case: When accuracy is critical and you want to minimize false positives

Recognized Members

The model can identify all 24 tripleS members:

Korean Members: ChaeWon, ChaeYeon, DaHyun, HaYeon, HyeRin, JiWoo, JiYeon, JooBin, NaKyoung, SeoAh, SeoYeon, ShiOn, SoHyun, SooMin, YeonJi, YooYeon, YuBin

Japanese Members: Kaede, Kotone, Lynn, Mayu

Chinese Members: Nien, Sullin, Xinyu

Model Files Structure

Each model version contains:

• config.json: Model configuration and metadata
• face_model.pkl: Trained model weights
• face_model_names.json: Label mapping (versions 0.4-0.6 only)

Model Selection Guide

• Choose Base Model if you want the original performance without fine-tuning
• Choose Version 0.4 if you prioritize detecting all possible members (high recall)
• Choose Version 0.5 if you want balanced performance for general use cases
• Choose Version 0.6 if you prioritize accuracy and want to minimize false identifications