Skip to content

nguyenit67/AI-Hackathon-Face-Analysis-2024

Repository files navigation

Guideline

This repo includes 3 main parts: using deepface to output age, gender, race, emotion and 2 models training from scratch are masked detection model and skintone detection model. All code are trained and evaluated on Kaggle with all provided data uploaded there.

DeepFace

Running experiment.ipynb notebook on Kaggle with public and private test data to get multiple backend detectors output for age, gender, race, emotion.

Skintone detection

In folder skinton_detection, run the skintone_analysis.ipynb notebook to train and predict skintone feature for public and private test data. A model skin_classify_150epoch.keras will be saved to Kaggle working directory /kaggle/working after training, which can be used for inference and fine-tuning later.

Masked detection

In folder masked_detection, run the masked_model_training.ipynb to train the masked detection model. A model named my_model.keras will be saved to Kaggle working directory /kaggle/working.

About Dataset

  • Total images: 15000 images (jpg format) including real images and synthetic images
  • Number of images with more than 1 face: 109 images (0.7%)
  • File annotation format:
  • CSV file
  • 15310 rows
  • Main attributes:
    • bbox: face bounding box of format (x, y, w, h) IOU (AVG precision IOU threshold từ 0.5 - 0.95)
    • age: 20-30s, 40-50s, Kid, Senior, Baby, Teenager - accuracy (*)
    • race: Caucasian, Mongoloid, Negroid - accuracy (*)
    • masked: Unmasked, Masked - accuracy
    • skintone: Mid-light, Light, Mid-dark, Dark - accuracy (*)
    • emotion: Neutral, Happiness, Anger, Surprise, Fear, Sadness, Disgust - accuracy (*)
    • gender: Male, Female - accuracy
  • Các thuộc tính khó: Độ tuổi, Cảm xúc, Tông màu da, Chủng tộc sẽ được đánh giá cao hơn các thuộc tính khác

Reference

Masked detection:

- https://medium.com/cloudnloud/building-a-face-mask-detector-using-python-and-opencv-2654e28d8d76

Skintone detection:

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published