Awesome

Face Detection

PyTorch implementations of various face detection algorithms (last updated on 2019-08-03).

Usage Example

import cv2, random
from detectors import DSFD
from utils import draw_bboxes, crop_thumbnail, draw_bbox

# load detector with device(cpu or cuda)
DET = DSFD(device='cpu')

# load image in RGB
img = cv2.imread('bts.jpg')
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)

# get bboxes with some confidence in scales for image pyramid
bboxes = DET.detect_faces(img, conf_th=0.9, scales=[0.5, 1])

# and draw bboxes on your image
img_bboxed = draw_bboxes(img, bboxes, fill=0.2, thickness=3)

# or crop thumbnail of someone
i = random.randrange(0, len(bboxes))
img_thumb, bbox_thumb = crop_thumbnail(img, bboxes[i], padding=1, size=100)

# you can use 'bbox_thumb' as bbox in thumbnail-coordinate system.
img_thumb_bboxed = draw_bbox(img_thumb, bbox_thumb)

Weight Files

MTCNN

./detectors/mtcnn/weights/pnet.npy
./detectors/mtcnn/weights/rnet.npy
./detectors/mtcnn/weights/onet.npy

FaceBoxes

./detectors/faceboxes/weights/FaceBoxes.pth

Tiny Face (trained on WIDER FACE)

./detectors/tinyface/weights/checkpoint_50.pth

PyramidBox (trained on WIDER FACE)

./detectors/pyramidbox/weights/pyramidbox_120000_99.02.pth

S3FD (trained on WIDER FACE)

./detectors/s3fd/weights/sfd_face.pth

DSFD (trained on WIDER FACE)

./detectors/dsfd/weights/dsfd_vgg_0.880.pth

Demo 01 : detect

python demo_detect.py

Note that it shows bounding boxes only for default scale image without image pyramid. Number of bounding boxes ─ not detected faces ─ and minimum box sizes are as follows:

	MTCNN	FaceBoxes	Tiny face	PyramidBox	S3FD	DSFD
# of boxes	210	28	686	512	475	528
minimum box size	71	1136	138	109	108	113
minimum box size	1046	27901	36912	24059	26711	15814

Demo 02 : crop

python demo_crop.py

Face Size

Minimum and maximum lengths of detected boxes are as follows:

	MTCNN	Tiny face	S3FD	DSFD
min length (pixel)	000.0	000.0	000.0	000.0
max length (pixel)	000.0	000.0	000.0	000.0