Skip to content

End-to-end face detection, cropping, norm estimation, and landmark detection in a single onnx model

License

Notifications You must be signed in to change notification settings

atksh/onnx-facial-lmk-detector

Repository files navigation

onnx-facial-lmk-detector

End-to-end face detection, cropping, norm estimation, and landmark detection in a single onnx model, model.onnx.

Demo

You can try this model at the following link. Thanks for hysts.

Code

See src.

Example

example

import onnxruntime as ort
import cv2

sess = ort.InferenceSession("model.onnx")
img = cv2.imread("input.jpg")

scores, bboxes, keypoints, aligned_imgs, landmarks, affine_matrices = sess.run(None, {"input": img})
# float32 int64 int64 uint8 int64 float32
# (N,) (N, 4) (N, 5, 2) (N, 224, 224, 3) (N, 106, 2) (N, 2, 3)

This model requires onnxruntime>=1.11.

How does it work?

This is simply a merged model of the following underlying models with some pre- and post-processing.

Underlying models

model reference
face detection SCRFD_10G_KPS https://github.com/deepinsight/insightface/tree/master/detection/scrfd#pretrained-models
landmark detection 2d106det https://github.com/deepinsight/insightface/blob/master/alignment/coordinate_reg/README.md#pretrained-models

Pre- and Post-Processing

Implemented the following processing by PyTorch and exported to ONNX.

  • Input transform:

    • Resize and pad to (1920, 1920)
    • BGR to RGB conversion
    • Transpose (H, W, C) to (C, H, W)
  • (Face Detection)

  • Post-processing of face detection

    • Predicted bounding boxes and Confidence Score Processing
    • NMS (ONNX Operator)
  • Norm estimation and face cropping

    • Estimate the norm and apply an affine transformation to each face.
    • Crop the faces and resize them to (192, 192).
  • (Landmark Detection)

  • Perform post-processing for landmark detection.

    • Process the predicted landmarks and apply the inverse affine transform to each face.

Note

Please check with the model provider regarding the license for your use.

This model includes the work that is distributed in the Apache License 2.0.

About

End-to-end face detection, cropping, norm estimation, and landmark detection in a single onnx model

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages