Face Detector trained by SSD by WIDER dataset

By Danh Doan

Demo video available here: [link]

• Install MXNet framework and GluonCV toolkit

  • For CPU only:

    pip install mxnet gluoncv

  • For GPUs

    pip install mxnet-cu90 gluoncv

    Change to match with CUDA version. mxnet-cu100 if CUDA 10.0 is installed

• Obtain train and val datasets from WIDER [link] Create a symbol link to WIDER directory

  ln -s /path/to/WIDER

• Move all image file to JPEGImages folder to easily manipulate

  mkdir JPEGImages

Follow lst-rec-prep.ipynb to create List and RecordIO file - they are the dataset format developed by MXNet. To support the training procedure, it is preferred to utilize data in binary format rather than raw images as well as the annotations.

If you are not familiar with this process, refer this tutorial [link]

Follow step 0 to step 2 to first create .lst file. By performing these steps, 2 file train.lst and val.lst are generated.

After obtaining the .lst files. Start generating .rec by using built-in feature from MXNet

python /path/to/incubator-mxnet/tools/im2rec.py train.lst WIDER/JPEGImages/ --pass-through --pack-label

python /path/to/incubator-mxnet/tools/im2rec.py val.lst WIDER/JPEGImages/ --pass-through --pack-label

It's gonna take a few seconds to create record files for train and val datasets. After finishing this step, 4 files are created:

• train.idx
• train.rec
• val.idx
• val.rec

Move all generated files to datasets directory

mkdir datasets

mv *.lst *.idx *.rec datasets

After the preparation process is successfully done. The training procedure can be started by the following command:

python3 train_ssd.py

This is the full training program originally from MXNet [link].

In this program, face detector is trained by fine-tuning technique and the object detection pretrained model is trained from VOC dataset.

import gluoncv as gcv

TARGET_CLASSES = ['person']

net = gcv.model_zoo.get_model('ssd_300_vgg16_atrous_custom', classes=TARGET_CLASSES, pretrained_base=False, transfer='voc')

Checkpoints are saved every 10 epochs and the best model with the best accuracy in val set is also stored and updated after every single epoch.

After processing the training procedure, 3 files are obtained:

• ssd_300_vgg16_atrous_voc_best_map.log
• ssd_300_vgg16_atrous_voc_best.params
• ssd_300_vgg16_atrous_voc_train.log

test-face-detector.ipynbis the notebook developed to test the performance of trained Face detection model

Results from my work:

Test with video: python3 test-ssdface-vid.py -v /path/to/video -s -r -n video_name

Check the argument parser in altusi/helper/funcs.py to make use of the arguments