RetinaFace_Pytorch

Reimplement RetinaFace with Pytorch.

Installation

Clone and install requirements
$ git clone https://github.com/supernotman/RetinaFace_Pytorch.git
$ cd RetinaFace_Pytorch/
$ sudo pip install -r requirements.txt

Pytorch version 1.1.0+ and torchvision 0.3.0+ are needed.

Data

  1. Download widerface dataset

  2. Download annotations (face bounding boxes & five facial landmarks) from baidu cloud or dropbox

  3. Organise the dataset directory as follows:

  widerface/
    train/
      images/
      label.txt
    val/
      images/
      label.txt
    test/
      images/
      label.txt

Train

$ train.py [-h] [data_path DATA_PATH] [--batch BATCH]
                [--epochs EPOCHS]
                [--shuffle SHUFFLE] [img_size IMG_SIZE]
                [--verbose VERBOSE] [--save_step SAVE_STEP]
                [--eval_step EVAL_STEP]
                [--save_path SAVE_PATH]
                [--depth DEPTH]

Example

For multi-gpus training, run:

$ CUDA_VISIBLE_DEVICES=0,1,2,3,4,5 python train.py --data_path /widerface --batch 32 --save_path ./out

Training log

---- [Epoch 39/200, Batch 400/403] ----
+----------------+-----------------------+
| loss name      | value                 |
+----------------+-----------------------+
| total_loss     | 0.09969855844974518   |
| classification | 0.09288528561592102   |
| bbox           | 0.0034053439740091562 |
| landmarks      | 0.003407923271879554  |
+----------------+-----------------------+
-------- RetinaFace Pytorch --------
Evaluating epoch 39
Recall: 0.7432201780921814
Precision: 0.906913273261629
Pretrained model

You can download the model from baidu cloud or dropbox

Detect

$ python detect.py --model_path out/model_epoch_200.pt --image_path 4.jpg

Todo:

  • [ ] Wider Face mAP calculation
  • [ ] Deformable Convolution
  • [ ] More models support
  • [ ] Random crop and color distortion
  • [ ] Graph Convolution
  • [ ] Bug fix

GitHub