-[face_recognition_sface_2021sep.onnx](./face_recognition_sface_2021sep.onnx) is converted from the model from https://github.com/zhongyy/SFace thanks to [Chengrui Wang](https://github.com/crywang).
- Support 5-landmark warpping for now (2021sep)
Results of accuracy evaluation with [tools/eval](../../tools/eval).
| Models | Accuracy |
|-------------|----------|
| SFace | 0.9940 |
| SFace quant | 0.9932 |
\*: 'quant' stands for 'quantized'.
## Demo
***NOTE***: This demo uses [../face_detection_yunet](../face_detection_yunet) as face detector, which supports 5-landmark detection for now (2021sep).
...
...
@@ -17,6 +27,7 @@ Run the following command to try the demo:
The script is modified based on [evaluation of InsightFace](https://github.com/deepinsight/insightface/blob/f92bf1e48470fdd567e003f196f8ff70461f7a20/src/eval/lfw.py).
This evaluation uses [YuNet](../../models/face_detection_yunet) as face detector. The structure of the face bounding boxes saved in [lfw_face_bboxes.npy](../eval/datasets/lfw_face_bboxes.npy) is shown below.
Each row represents the bounding box of the main face that will be used in each image.
`x1, y1, w, h` are the top-left coordinates, width and height of the face bounding box, `{x, y}_{re, le, nt, rcm, lcm}` stands for the coordinates of right eye, left eye, nose tip, the right corner and left corner of the mouth respectively. Data type of this numpy array is `np.float32`.
### Prepare data
Please visit http://vis-www.cs.umass.edu/lfw to download the LFW [all images](http://vis-www.cs.umass.edu/lfw/lfw.tgz)(needs to be decompressed) and [pairs.txt](http://vis-www.cs.umass.edu/lfw/pairs.txt)(needs to be placed in the `view2` folder). Organize files as follow: