# scrfd

**Repository Path**: nextvpu2016/scrfd

## Basic Information

- **Project Name**: scrfd
- **Description**: 源自insightface的scrfd
- **Primary Language**: Python
- **License**: GPL-3.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2023-03-16
- **Last Updated**: 2023-03-20

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

## Installation

Please refer to [mmdetection](https://github.com/open-mmlab/mmdetection/blob/master/docs/en/get_started.md#installation) for installation.
 
  1. Install [mmcv](https://github.com/open-mmlab/mmcv). (mmcv-full==1.2.6 and 1.3.3 was tested)
  2. Install build requirements and then install mmdet.
       ```
       pip install -r requirements/build.txt
       pip install -v -e .  # or "python setup.py develop"
       ```

## Data preparation

### WIDERFace:
  1. Download WIDERFace datasets and put it under `data/retinaface`.
  2. Download annotation files from [gdrive](https://drive.google.com/file/d/1UW3KoApOhusyqSHX96yEDRYiNkd3Iv3Z/view?usp=sharing) and put them under `data/retinaface/`
 
   ```
     data/retinaface/
         train/
             images/
             labelv2.txt
         val/
             images/
             labelv2.txt
             gt/
                 *.mat
             
   ```
 

#### Annotation Format 

*please refer to labelv2.txt for detail*

For each image:
  ```
  # <image_path> image_width image_height
  bbox_x1 bbox_y1 bbox_x2 bbox_y2 (<keypoint,3>*N)
  ...
  ...
  # <image_path> image_width image_height
  bbox_x1 bbox_y1 bbox_x2 bbox_y2 (<keypoint,3>*N)
  ...
  ...
  ```
Keypoints can be ignored if there is bbox annotation only.


## Training

Example training command, with 4 GPUs:
```
CUDA_VISIBLE_DEVICES="0,1,2,3" PORT=29701 bash ./tools/dist_train.sh ./configs/scrfd/scrfd_1g.py 4
```
if use int8mdoel, dont use GN layer.
need change '\configs\scrfd_RGB.py'  datapath:
        ```
        train=dict(
                type='RetinaFaceDataset',
                ann_file='xxxxxx/data/retinaface/train_label/labelv2.txt',
                img_prefix='xxxxxx/data/retinaface/train/images/',
                ...
        val=dict(
                type='RetinaFaceDataset',
                ann_file='xxxxxx/data/retinaface/val/widerface.txt',
                img_prefix='xxxxxx/data/retinaface/val/images/',
                ...
        test=dict(
                type='RetinaFaceDataset',
                ann_file='xxxxxx/data/retinaface/val/widerface.txt',
                img_prefix='xxxxxx/data/retinaface/val/images/',
                ...
        ```
want use ResNetV1e:
```
bash ./tools/train_RGB_bn.sh
```
want use MobileNetV1:
need change '\configs\scrfd_RGB3H.py'  datapath:
```
bash ./tools/train_RGB3H.sh
```
## WIDERFace Evaluation

We use a pure python evaluation script without Matlab.

```
GPU=0
GROUP=scrfd
TASK=scrfd_2.5g
CUDA_VISIBLE_DEVICES="$GPU" python -u tools/test_widerface.py ./configs/"$GROUP"/"$TASK".py ./work_dirs/"$TASK"/model.pth --mode 0 --out wouts
```


## Pretrained-Models

|      Name      | Easy  | Medium | Hard  | FLOPs | Params(M) | Infer(ms) | Link                                                         |
| :------------: | ----- | ------ | ----- | ----- | --------- | --------- | ------------------------------------------------------------ |
|   SCRFD_500M   | 90.57 | 88.12  | 68.51 | 500M  | 0.57      | 3.6       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyYWxScdiTITY4TQ?e=DjXof9) |
|    SCRFD_1G    | 92.38 | 90.57  | 74.80 | 1G    | 0.64      | 4.1       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyPVLI44ahNBsOMR?e=esPrBL) |
|   SCRFD_2.5G   | 93.78 | 92.16  | 77.87 | 2.5G  | 0.67      | 4.2       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyTIXnzB1ujPq4th?e=5t1VNv) |
|   SCRFD_10G    | 95.16 | 93.87  | 83.05 | 10G   | 3.86      | 4.9       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyUKwTiwXv2kaa8o?e=umfepO) |
|   SCRFD_34G    | 96.06 | 94.92  | 85.29 | 34G   | 9.80      | 11.7      | [download](https://1drv.ms/u/s!AswpsDO2toNKqyKZwFebVlmlOvzz?e=V2rqUy) |
| SCRFD_500M_KPS | 90.97 | 88.44  | 69.49 | 500M  | 0.57      | 3.6      | [download](https://1drv.ms/u/s!AswpsDO2toNKri_NDM0GIkPpkE2f?e=JkebJo) |
| SCRFD_2.5G_KPS | 93.80 | 92.02  | 77.13 | 2.5G  | 0.82      | 4.3       | [download](https://1drv.ms/u/s!AswpsDO2toNKqyGlhxnCg3smyQqX?e=A6Hufm) |
| SCRFD_10G_KPS  | 95.40 | 94.01  | 82.80 | 10G   | 4.23      | 5.0       | [download](https://1drv.ms/u/s!AswpsDO2toNKqycsF19UbaCWaLWx?e=F6i5Vm) |

mAP, FLOPs and inference latency are all evaluated on VGA resolution.
``_KPS`` means the model includes 5 keypoints prediction.

## Convert to ONNX

Please refer to `tools/scrfd2onnx.py`

Generated onnx model can accept dynamic input as default.

You can also set specific input shape by pass ``--shape 640 640``, then output onnx model can be optimized by onnx-simplifier.


## Inference

Please refer to `tools/scrfd.py` which uses onnxruntime to do inference.