faceboxes

introduction

A tensorflow implement faceboxes. And some changes has been made in RDCL module, to achieve a better performance，and runs faster:

1. input size is 512 (1024 in the paper), then the first conv stride is 2, kernel size 7x7x12.
2. replace the first maxpool by conv 3x3x24 stride 2
3. replace the second 5x5 stride2 conv and maxpool by two 3x3 stride 2 conv
4. anchor based sample is used in data augmentaion.

codes like below

    with tf.name_scope('RDCL'):
     net = slim.conv2d(net_in, 12, [7, 7], stride=2,activation_fn=tf.nn.relu, scope='init_conv1')
     net = slim.conv2d(net, 24, [3, 3], stride=2, activation_fn=tf.nn.crelu, scope='init_conv2')
    
     net = slim.conv2d(net, 32, [3, 3], stride=2,activation_fn=tf.nn.relu,scope='conv1x1_before1')
     net = slim.conv2d(net, 64, [3, 3], stride=2, activation_fn=tf.nn.crelu, scope='conv1x1_before2')
     
     return net

I want to name it faceboxes++ ,if u don't mind

Pretrained model can be download from:

• baidu disk (code eb6b )

• google drive

Evaluation result on fddb

fddb
0.961

Speed: it runs over 70FPS on cpu (i7-8700K), 30FPS (i5-7200U), 140fps on gpu (2080ti) with fixed input size 512, tf1.14, multi thread. And i think the input size, the time consume and the performance is very appropriate for application :)

Hope the codes can help you, and i am struggling with the new tf api, contact me if u have any question, 2120140200@mail.nankai.edu.cn .

requirment

• tensorflow1.14

• tensorpack (for data provider)

• opencv

• python 3.6

useage

train

1. download widerface data from http://shuoyang1213.me/WIDERFACE/ and release the WIDER_train, WIDER_val and wider_face_split into ./WIDER,

2. download fddb， and release FDDB-folds into ./FDDB ， 2002,2003 into ./FDDB/img

3. then run python prepare_data.pyit will produce train.txt and val.txt

   (if u like train u own data, u should prepare the data like this: ...../9_Press_Conference_Press_Conference_9_659.jpg| 483(xmin),195(ymin),735(xmax),543(ymax),1(class) ...... one line for one pic, caution! class should start from 1, 0 means bg)

4. 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

5. after training ,convert to pb file: python tools/auto_freeze.py

finetune

1. (if u like train u own data, u should prepare the data like this: ...../9_Press_Conference_Press_Conference_9_659.jpg| 483(xmin),195(ymin),735(xmax),543(ymax),1(class) ...... one line for one pic, caution! class should start from 1, 0 means bg)

2. set config.MODEL.pretrained_model='your model pretrained', and config.MODEL.continue_train=True in train_config.py

3. python train.py

evaluation

    python test/fddb.py [--model [TRAINED_MODEL]] [--data_dir [DATA_DIR]]
                          [--split_dir [SPLIT_DIR]] [--result [RESULT_DIR]]
    --model              Path of the saved model,default ./model/detector.pb
    --data_dir           Path of fddb all images
    --split_dir          Path of fddb folds
    --result             Path to save fddb results

example python model_eval/fddb.py --model model/detector.pb --data_dir 'FDDB/img/' --split_dir FDDB/FDDB-folds/ --result 'result/'

visualization

1. If u get a trained model, run python tools/auto_freeze.py, it will read the checkpoint file in ./model, and produce detector.pb.

2. python vis.py --img_dir 'your images dir ,by default it detect pics with .jpg

3. or use a camera: python vis.py --cam_id 0

You can check the code in vis.py to make it runable, it's simple.

reference

FaceBoxes: A CPU Real-time Face Detector with High Accuracy