batch-face 1.5.0

Batch Face Preprocessing for Modern Research

Batch Face for Modern Research

🚧Documentation under construction, check tests folder for more details. 🚧

This repo provides the out-of-box face detection and face alignment with batch input support and enables real-time application on CPU.

Features

1. Batch input support for faster data processing.
2. Smart API.
3. Ultrafast with inference runtime acceleration.
4. Automatically download pre-trained weights.
5. Minimal dependencies.
6. Unleash the power of GPU for batch processing.

Requirements

• Linux, Windows or macOS
• Python 3.5+ (it may work with other versions too)
• opencv-python
• PyTorch (>=1.0)
• ONNX (optional)

While not required, for optimal performance it is highly recommended to run the code using a CUDA enabled GPU.

Install

The easiest way to install it is using pip:

pip install git+https://github.com/elliottzheng/batch-face.git@master

No extra setup needs, most of the pretrained weights will be downloaded automatically.

If you have trouble install from source, you can try install from PyPI:

pip install batch-face

the PyPI version is not guaranteed to be the latest version, but we will try to keep it up to date.

Usage

You can clone the repo and run tests like this

python -m tests.camera

Face Detection

Detect face and five landmarks on single image

import cv2
from batch_face import RetinaFace

detector = RetinaFace(gpu_id=0)
img = cv2.imread("examples/obama.jpg")
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

max_size = 1080 # if the image's max size is larger than 1080, it will be resized to 1080, -1 means no resize
resize = 1 # resize the image to speed up detection, default is 1, no resize
threshold = 0.95 # confidence threshold

# now we recommand to specify return_dict=True to get the result in a more readable way
faces = detector(img, threshold=threshold, resize=resize, max_size=max_size, return_dict=True)
face = faces[0]
box = face['box']
kps = face['kps']
score = face['score']

# the old way to get the result
faces = detector(img, threshold=threshold, resize=resize, max_size=max_size)
box, kps, score = faces[0]

Running on CPU/GPU

In order to specify the device (GPU or CPU) on which the code will run one can explicitly pass the device id.

from batch_face import RetinaFace
# 0 means using GPU with id 0 for inference
# default -1: means using cpu for inference
fp16 = True # use fp16 to speed up detection and save GPU memory

detector = RetinaFace(gpu_id=0, fp16=True)

	GPU(GTX 1080TI,batch size=1)	GPU(GTX 1080TI，batch size=750)	CPU(Intel(R) Core(TM) i7-7800X CPU @ 3.50GHz)
FPS	44.02405810720893	96.64058005582535	15.452635835550483
SPF	0.022714852809906007	0.010347620010375976	0.0647138786315918

Batch input for faster detection

Detector with CUDA process batch input faster than the same amount of single input.

import cv2
from batch_face import RetinaFace

detector = RetinaFace()
img= cv2.imread('examples/obama.jpg')
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

max_size = 1080 # if the image's max size is larger than 1080, it will be resized to 1080, -1 means no resize
resize = 1 # resize the image to speed up detection, default is 1, no resize
resize_device = 'cpu' # resize on cpu or gpu, default is gpu
threshold = 0.95 # confidence threshold for detection
batch_size = 100 # batch size for detection, the larger the faster but more memory consuming, default is -1, which means batch_size = number of input images
batch_images = [img,img] # pseudo batch input

all_faces = detector(batch_images, threshold=threshold, resize=resize, max_size=max_size, batch_size=batch_size) 
faces = all_faces[0] # the first input image's detection result
box, kps, score = faces[0] # the first face's detection result

Note: All the input images must of the same size, for input images with different size, please use detector.pseudo_batch_detect.

Face Alignment

face alignment on single image

from batch_face import drawLandmark_multiple, LandmarkPredictor, RetinaFace

predictor = LandmarkPredictor(0)
detector = RetinaFace(0)

imgname = "examples/obama.jpg"
img = cv2.imread(imgname)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

faces = detector(img)

if len(faces) == 0:
    print("NO face is detected!")
    exit(-1)

# the first input for the predictor is a list of face boxes. [[x1,y1,x2,y2]]
results = predictor(faces, img, from_fd=True) # from_fd=True to convert results from our detection results to simple boxes

for face, landmarks in zip(faces, results):
    img = drawLandmark_multiple(img, face[0], landmarks)

Head Pose Estimation

Head pose estimation on video

from batch_face import RetinaFace, SixDRep, draw_landmarks, load_frames_rgb, Timer

vis = True
gpu_id = 0
batch_size = 100
threshold = 0.95
detector = RetinaFace(gpu_id=gpu_id)
head_pose_estimator = SixDRep(gpu_id=gpu_id)
video_file = 'examples/ross.mp4'
frames = load_frames_rgb(video_file) # simple wrapper to load video frames with opencv and convert to RGB, 0~255, UInt8, HWC
print(f'Loaded {len(frames)} frames')
print('image size:', frames[0].shape)
# it might take longer time to detect since is first time to run the model
all_faces = detector(frames, batch_size=batch_size, return_dict=True, threshold=threshold, resize=0.5)
head_poses = head_pose_estimator(all_faces, frames, batch_size=batch_size, update_dict=True, input_face_type='dict')
# the head pose will be updated in the all_faces dict
out_frames = []
for faces, frame in zip(all_faces, frames):
    for face in faces:
        head_pose_estimator.plot_pose_cube(frame, face['box'], **face['head_pose'])
    out_frames.append(frame)

if vis:
    import imageio
    out_file = 'examples/head_pose.mp4'
    imageio.mimsave(out_file, out_frames, fps=8)

check out the result video here you can run the script python -m tests.video_head_pose to see the result.

References

• Face Detection Network and pretrained model are from biubug6/Pytorch_Retinaface
• Face Alignment Network and pretrained model are from cunjian/pytorch_face_landmark
• Face Reconstruction Network and pretrained model are from cleardusk/3DDFA
• Head Pose Estimation Network and pretrained model are from jahongir7174/SixDRepNet