Adapted from linxiaohui/mtcnn-opencv. Modifications include uses of onnx runtime as inference backend and provide a raw output API. Maybe this package should be a fork but I have already had a forked version to address another problem, so I made a new package.
MTCNN Face Detector using ONNX-runtime OpenCV, no reqiurement for tensorflow/pytorch.
Select one method from below:
pip install mtcnn-onnxruntime
: Use existing onnxruntime version in environment to run, if no onnxruntime is in the environment,opencv
will be used as backend.pip install mtcnn-onnxruntime[cpu]
: Installmtcnn-onnxruntime
withonnxruntime
pip install mtcnn-onnxruntime[gpu]
: Installmtcnn-onnxruntime
withonnxruntime-gpu
import cv2
from mtcnn_ort import MTCNN
detector = MTCNN()
test_pic = "t.jpg"
image = cv2.cvtColor(cv2.imread(test_pic), cv2.COLOR_BGR2RGB)
result = detector.detect_faces(image)
# Result is an array with all the bounding boxes detected. Show the first.
print(result)
"""
[{'box': [60, 0, 314, 356],
'confidence': 0.9993509650230408,
'keypoints': {'left_eye': (136, 71),
'right_eye': (289, 58),
'nose': (218, 148),
'mouth_left': (162, 243),
'mouth_right': (290, 228)}}]
"""
detector.detect_faces_raw(image)
"""
(array([[ 60.58798278, -66.81823712, 374.15868253, 356.04121107,
0.99935097]]),
array([[136.35648 ],
[289.0994 ],
[218.10023 ],
[162.28156 ],
[290.98242 ],
[ 71.76702 ],
[ 58.487453],
[148.75732 ],
[243.27672 ],
[228.3274 ]], dtype=float32))
"""
Illustration:
import cv2
if len(result) > 0:
bounding_box = result[0]["box"]
keypoints = result[0]['keypoints']
cv2.rectangle(image,
(bounding_box[0], bounding_box[1]),
(bounding_box[0] + bounding_box[2], bounding_box[1] + bounding_box[3]),
(0,155,255),
2)
cv2.circle(image,(keypoints['left_eye']), 2, (0,155,255), 2)
cv2.circle(image,(keypoints['right_eye']), 2, (0,155,255), 2)
cv2.circle(image,(keypoints['nose']), 2, (0,155,255), 2)
cv2.circle(image,(keypoints['mouth_left']), 2, (0,155,255), 2)
cv2.circle(image,(keypoints['mouth_right']), 2, (0,155,255), 2)
cv2.imwrite("result.jpg", cv2.cvtColor(image, cv2.COLOR_RGB2BGR))
# Generate labeled images
with open(test_pic, "rb") as fp:
marked_data = detector.mark_faces(fp.read())
with open("marked.jpg", "wb") as fp:
fp.write(marked_data)
Warped patch (then face recognition SOTA ArcFace) can consume it (otherwise, if one just use bounding box, what some models such as UltraNet can only make, the performance will significantly compromised.).
from skimage import transform as trans
import numpy as np
image = cv2.cvtColor(cv2.imread(test_pic), cv2.COLOR_BGR2RGB)
src = np.array([
[30.2946, 51.6963],
[65.5318, 51.5014],
[48.0252, 71.7366],
[33.5493, 92.3655],
[62.7299, 92.2041]], dtype=np.float32)
src[:, 0] += 8.0
landmark5 = detector.detect_faces_raw(image)[1].reshape(2, 5).T
tform = trans.SimilarityTransform()
tform.estimate(landmark5, src)
M = tform.params[0:2, :]
img = cv2.warpAffine(image, M, (112, 112),
borderValue=0.0)
cv2.imwrite("warped.jpg", cv2.cvtColor(img, cv2.COLOR_RGB2BGR))