Skip to content

Latest commit

 

History

History
71 lines (52 loc) · 2.68 KB

README_FD.md

File metadata and controls

71 lines (52 loc) · 2.68 KB
Raw

Face detection.

Data preparation

The training procedure can be done using data in LMDB format. To launch training or evaluation at the WiderFace dataset, download it from the source, extract images and annotations into <DATA_DIR> folder and use the provided scripts to convert original annotations to LMDB format.

Create LMDB files

To create LMDB files go to the '$CAFFE_ROOT/python/lmdb_utils/' directory and run the following scripts:

  1. Run docker in interactive session with mounted directory with WIDER dataset
nvidia-docker run --rm -it --user=$(id -u) -v <DATA_DIR>:/data ttcf bash
  1. Convert original annotation to xml format for both train and val subsets:
python3 $CAFFE_ROOT/python/lmdb_utils/wider_to_xml.py /data /data/WIDER_train/images/ /data/wider_face_split/wider_face_train_bbx_gt.txt train
python3 $CAFFE_ROOT/python/lmdb_utils/wider_to_xml.py /data /data/WIDER_val/images/ /data/wider_face_split/wider_face_val_bbx_gt.txt val
  1. Convert xml annotations to set of xml files per image:
python3 $CAFFE_ROOT/python/lmdb_utils/xml_to_ssd.py --ssd_path /data --xml_path_train /data/wider_train.xml --xml_path_val /data/wider_val.xml
  1. Run bash script to create LMDB:
bash $CAFFE_ROOT/python/lmdb_utils/create_wider_lmdb.sh
  1. Close docker session by ctrl+D and check that you have lmdb files in <DATA_DIR>.

Face detection training

On next stage we should train the Face Detection model. To do this follow next steps:

cd ./models
python3 train.py --model face_detection \                          # name of model
                --weights face-detection-retail-0044.caffemodel \  # initialize weights from 'init_weights' directory
                --data_dir <DATA_DIR> \                            # path to directory with dataset
                --work_dir <WORK_DIR> \                            # directory to collect file from training process
                --gpu <GPU_ID>

Face Detection model evaluation

To evaluate the quality of trained Face Detection model on your test data you can use provided scripts.

python3 evaluate.py --type fd \
    --dir <WORK_DIR>/face_detection/<EXPERIMENT_NUM> \
    --data_dir <DATA_DIR> \
    --annotation wider_val.xml \
    --iter <ITERATION_NUM>

Export to IR format

python3 mo_convert.py --name face_detection \
    --dir <WORK_DIR>/face_detection/<EXPERIMENT_NUM> \
    --iter <ITERATION_NUM> \
    --data_type FP32

Face Detection demo

You can use this demo to view how resulting model performs.