- Clone the repo
git clone git@github.com:davidygp/Pedestrian-Attribute-Recognition.git
cd Pedestrian-Attribute-Recognition/working_code - Install the required packages
pip install -r requirements.txt
- Get the raw data
- (The additional annotated .txt files are already in the folder "./Updated_Labels/")
- (The previous PETA.mat from https://github.com/dangweili/pedestrian-attribute-recognition-pytorch is already renamed it as "./PETA_old.mat")
- download the original PETA dataset (http://mmlab.ie.cuhk.edu.hk/projects/PETA.html), and place it as the folder "./PETA dataset"
- Run the .py script to process the PETA images and generate the new PETA.mat file in the "./data" folder
(Note: Ordering of the image names differs between Windows & Mac, to get the exact same IDs it should be run on Windows)
python ./process_updated_labels_n_images_v3.1.py
- Run the .py script to generate the dataset.pkl file in the "./data" folder
python ./format_peta.py
- Copy the "./data" folder to the main repo folder
mv ./data ../
- Run the training as required
cd ../
python ./train.py
(see config.py for examples)
python ./train.py --batchsize 128 --train_epoch 100
PETA dataset can be obtained from: http://mmlab.ie.cuhk.edu.hk/projects/PETA.html
PA100k dataset can be obtained from: https://drive.google.com/drive/folders/0B5_Ra3JsEOyOUlhKM0VPZ1ZWR2M
RAPv2 dataset can be obtained from: https://drive.google.com/file/d/1hoPIB5NJKf3YGMvLFZnIYG5JDcZTxHph/view
Codes are based on the repositories. (Thank you for your released code!):