YOLOv3-tiny-FaceDetection-by-PaddlePaddle

Note: If there's any error like "No module named 'xxx'", please use command "pip install xxx" to repair.

Feel free to add my QQ: 793729558 to discuss with me.
Also you can add the QQ group: 647303915 to discuss together.

1. Import the module

from imageSolver import *
from detection import *

2. How to load the data

trainData = WIDER(imgPath='xxx/images',
                  imgBboxGt='xxx/wider_face_train_bbx_gt.txt'
                  maxBoxNum=256)

imgPath is your img root path;
imgBboxGt is your "..._bbx_gt.txt" path;
maxBoxNum is the maximum number of faces in each picture;

The storage structure of image data is based on WIDER data.
If you want to use your own data, please keep storage structure consistent or rewrite the data utils class.

3. How to train your model

First you need to create a YOLOv3_tiny object.

model = YOLOv3_tiny(use_cuda=True)

use_cuda is whether to use GPU;

Then you need to build the model structure.

model.build(boxNum=64)

boxNum is the maximum number of faces in each picture which should be consistent with WIDER's maxBoxNum.

Next you can train your model.

model.train(imgClass=trainData, epoches=10, batchSize=64)

imgClass is your img data class object.
batchSize is the number of data used for each train step;
epoch is the total iteration number of your training data;

And the log will be print like follows:

...
After iterations 7400: loss = 52.402  mAP: 0.478  51.425 images//s  Estimated remaining time: 57995.182s
After iterations 7500: loss = 32.706  mAP: 0.575  51.565 images//s  Estimated remaining time: 57775.328s
After iterations 7600: loss = 17.977  mAP: 0.511  51.344 images//s  Estimated remaining time: 57961.978s
...

Finally you need to save your model for future use.

model.save('./infer_model')

First parameter is the path of model saved.

Ok, I know you are too lazy to train your own model. Also you can use my trained model in '/infer_model'. The model is trained on WIDER train data and score mAP = 0.779 on WIDER val data.

4. How to use your model to detect face

First you need to create a Detector object.

detector = Detector(modelPath='./infer_model',USE_CUDA=False)

modelPath is your model path;
USE_CUDA is whether to use GPU;

Then you can do face detection by call it.

imgs,bboxes_pre = detector(imgList=['imgs/1.jpg','imgs/2.png'],
                          confidence_threshold=0.5,nms_threshold=0.3)

imgList is your image path list; confidence_threshold is confidence threshold in non-maximum suppression;
nms_threshold is nms threshold in non-maximum suppression;

It will return a list of each image and its detection result.
Also you can plot it and save result:

for i,(img,bbox_pre) in enumerate(zip(imgs,bboxes_pre)):
    draw_bbox(img, bbox_pre, savePath=f'imgs/{i+1}_out.png')

savePath is your image save path;

The results are shown below: