Age-Gender-Classification

This repo is created to describe the project 'age-gender classification' conducted from May 30 to Jun 4. Also, it is used for finding and predicting gender and age of human from a video. My goal is 93 % acc for gender and MAE for 5.2 for age can be achieved after 50 epochs of training.

Requirement

• pip3 install --upgrade opencv-python, imutils, skimage
• pip3 install retina for face detection

Usage

Training

1. Firstly, cd process_imdb-wiki, run IMDB-WIKI_dataset.ipynb for filtered dataset IMDB-WIKI

2. Run train_test_split.py for create train.csv and test.csv file. Also, Using AFAD Dataset for finetune on AsianFace. Combine both of the datasets, get train_combinedataset.csv and test_combinedataset.csv for training and evaluation

3. Run train.py

Real-time Prediction

Run inference.py, put you video in this file and get the result.

Details:

• Train a model base on ResNet18,
  • The latest conv feature is put into 2 discriminative branches: gender(2 neuron male-female) , age(10 neuron for bin-age normalized (0-9), each bin(10 neuron represent (0-9)))
  • Modify model can be found on model.py
• Test :
  • Using default retina (external python library) for face detection, I would recommend you train another model for face detection to get better result and inference.