A tensorflow implement dsfd, and there is something different with the origin paper.

It‘s a ssd-like object detect framework, but slightly different, combines lots of tricks for face detection, such as dual-shot, dense anchor match, FPN and so on.

now it is mainly optimised about face detection, and borrows tons of code from tensorpack

it achieves 0.982 on FDDB, not tested on widerface.

ps, the code maybe not that clear, please be patience, and i am still working on it, and forgive me for my poor english :)

tensorflow1.12

tensorpack for data provider

opencv

python 3.6

if u like to show the anchor stratergy, u could simply run :

python anchor/utils.py

it will draw the anchor one by one,

if u like to know how the data augmentation works, run :

python data/augmentor/augmentation.py

u should prepare the data like this:

...../9_Press_Conference_Press_Conference_9_659.jpg| 483,195,735,543,1

one line for one pic

caution! class should start from 1, 0 means bg

download widerface data from http://shuoyang1213.me/WIDERFACE/

and release the WIDER_train, WIDER_val and wider_face_split into ./WIDER, or somewhere u like,then run

python prepare_wider_data.py it will produce train.txt and val.txt

1.download the imagenet pretrained resnet50 model from http://download.tensorflow.org/models/resnet_v1_50_2016_08_28.tar.gz

release it in the root dir, as in train_config.py set config.MODEL.pretrained_model='resnet_v1_50.ckpt',config.MODEL.continue_train=False

2.but if u want to train from scratch set config.MODEL.pretrained_model=None,

3.if recover from a complet pretrained model set config.MODEL.pretrained_model='yourmodel.ckpt',config.MODEL.continue_train=True

then, run:

python train.py

and if u want to check the data when training, u could set vis in train_config.py as True

 71 class DataFromList(RNGDataFlow):
 72     """ Wrap a list of datapoints to a DataFlow"""
 73 
 74     def __init__(self, lst, shuffle=True):
 75         """
 76         Args:
 77             lst (list): input list. Each element is a datapoint.
 78             shuffle (bool): shuffle data.
 79         """
 80         super(DataFromList, self).__init__()
 81         self.lst = lst
 82         self.shuffle = shuffle
 83     
 84     #def __len__(self):
 85     #    return len(self.lst)
 86 
 87     def __iter__(self):
 88         if not self.shuffle:
 89             for k in self.lst:
 90                 yield k
 91         else:
 92             while True:
 93                 idxs = np.arange(len(self.lst))
 94                 self.rng.shuffle(idxs)
 95                 for k in idxs:
 96                     yield self.lst[k]

(caution: i dont know where the demo picture coms from, if u think it's a tort, i would like to delet it)

python vis.py

u can check th code in vis.py to make it runable, it's simple.

download a pretrained model(detector.pb) from https://pan.baidu.com/s/1SRMoJIcqHRoVydl2XIZ3lA (code m3bg) put it in to './model/detector.pb'