Introduction

this repository is the implementation of MTCNN in MXnet

• core: core routines for MTCNN training and testing.
• tools: utilities for training and testing
• data: Refer to Data Folder Structure for dataset reference. Usually dataset contains images and imglists
• model: Folder to save training symbol and model
• prepare_data: scripts for generating training data for pnet, rnet and onet

Useful information

You're required to modify mxnet/src/regression_output-inl.h according to mxnet_diff.patch before using the code for training.

• Dataset format The images used for training are stored in ./data/dataset_name/images/ The annotation file is placed in ./data/dataset_name/imglists/

  • For training: Each line of the annotation file states a training sample.
    The format is: [path to image] [cls_label] [bbox_label]
    cls_label: 1 for positive, 0 for negative, -1 for part face.
    bbox_label are the offset of x1, y1, x2, y2, calculated by (xgt(ygt) - x(y)) / width(height)
    An example would be 12/positive/28 1 -0.05 0.11 -0.05 -0.11.
    Note that all the strings are seperated by space.

  • For testing: Similar to training but only path-to-image is needed.

• Data Folder Structure (suppose root is data)

  cache (created by imdb)
  -- name + image set + gt_roidb
  -- results (created by detection and evaluation)
  mtcnn # contains images and anno for training mtcnn
  -- images
  ---- 12 (images of size 12 x 12, used by pnet)
  ---- 24 (images of size 24 x 24, used by rnet)
  ---- 48 (images of size 48 x 48, used by onet)
  -- imglists 
  ---- train_12.txt
  ---- train_24.txt
  ---- train_48.txt
  custom (datasets for testing) 
  -- images
  -- imglists
  ---- image_set.txt

• Scripts to generate training data(from wider face dataset)

  • run wider_annotations/transform.m (or transform.py) to get the annotation file of the format we need.
  • gen_pnet_data.py: obtain training samples for pnet
  • gen_hard_example.py: prepare hard examples. you can set test_mode to "pnet" to get training data for rnet, or set test_mode to "rnet" to get training data for onet.
  • gen_imglist.py: ramdom sample images generated by gen_pnet_data.py or gen_hard_example.py to form training set.

Results

License

MIT LICENSE

Reference

Kaipeng Zhang, Zhanpeng Zhang, Zhifeng Li, Yu Qiao , " Joint Face Detection and Alignment using Multi-task Cascaded Convolutional Networks," IEEE Signal Processing Letter