SIBGRAPI 2018 - Real-Time Gender Detection in the Wild

This repo contains the source code to reproduce the results of our paper published in the SIBGRAPHI2018. Our code is projected to run on Linux environments.

More details @ http://luiszeni.com.br/gender_sib2018/

If you want an easy way to test and reproduce our method, check out our docker image @ Docker Hub: https://github.com/luiszeni/sibgrapi2018_gender_detection/blob/master/how_to_docker.md

This project is organized in the following manner:

sibgrapi2018_gender_detection
      |___code
            |___darknet [Darknet framework used to train and run our models]

            |___scripts [Scripts to download the datasets with our annotated data]

            |___visualization [Heatmap activation tool to visualize activations for each class of an trained model]

      |___data
            |___annotations [Set of annotations used to train our models]

            |___img [Testing images]
            |___vid [Testing videos]

Running the gender detector on darknet

1- Build the project (see the makefile flags if you are not using a GPU or openCV, more details at darknet site)

cd code/darknet
make

2- Download our pre-trained model:

wget http://inf.ufrgs.br/~lfazeni/sib2018_models/gender_detection_50voc_50celeb_darknet.weights

3- Run the demo

./darknet detector demo cfg/test_voc_only.data cfg/yoloGender.cfg gender_detection_50voc_50celeb_darknet.weights ../../data/vid/001.mp4

Training a gender detector model on darknet with celebA and PascalVoc

1- Download the datasets running the scripts to download

cd code/scripts
./get_celeba_with_gender_annotations.sh
./get_voc_with_gender_annotations.sh

2- download yolo v2 pre-trained mode and cfg:

cd ../../
cd code/darknet
wget https://raw.githubusercontent.com/pjreddie/darknet/master/cfg/yolov2-voc.cfg
wget https://pjreddie.com/media/files/yolov2.weights

3- get only the 29 first layers of the model

./darknet partial yolov2-voc.cfg yolov2.weights yolov2.weights.29 29

4- Train (this code trains our model using in each epoch 50% of image from each dataset)

mkdir backup
mkdir backup/gender_voc_50_celeb_50
./darknet detector train cfg/train_voc_50_celeb_50.data cfg/yoloGender.cfg yolov2.weights.29 29

Visualizing the heatmap activations

Dependencies: tensorflow, keras, opencv in python, numpy

1- Downloading the tensorflow's model

cd code/visualization
wget http://inf.ufrgs.br/~lfazeni/sib2018_models/gender_detection_50voc_50celeb_tensorflow.h5

2- Visualizing the heatmap of an image:

python heatmap_from_detection.py -m gender_detection_50voc_50celeb_tensorflow.h5  -i ../../data/img/000058.jpg -md 3

3- Visualizing the heatmap of an video:

python heatmap_from_detection.py -m gender_detection_50voc_50celeb_tensorflow.h5  -v ../../data/vid/001.mp4 -md 2

4- View all avaible options

python heatmap_from_detection.py --help