Collection of face detection algorithms and One-Shot Face recognition

Face detection

Detectors

• CenterFace
  • Model (7.3Mb) already included in this repository
    • Trained on WIDER FACE
• DSFD
  • Implementation from this repository
  • Model (WIDERFace_DSFD_RES152.pth - 481Mb) needs to be downloaded from the original repository
    • Trained on WIDER FACE
    • Should be placed under dsdf/weights
• FaceBoxes
  • PyTorch implementation is used in this project
  • Model (FaceBoxesProd.pth - 4.1Mb) needs to be downloaded from PyTorch implementation version
    • Trained on WIDER FACE
    • Should be placed under face_boxes/weights
• Haarcascade from OpenCV
  • Model (1MB) already included in this repository
• MTCNN
  • Implementation from this repository
• RetinaFace
  • ResNet50 model (112Mb) pretrained on WIDER FACE
  • Automatically downloaded with first use (the model is placed @ $USER$\.insightface\models\)
• S3FD
  • Model (sfd_face.pth - 90Mb) needs to be downloaded from the repository
    • Trained on WIDER FACE
    • Should be placed under S3FD/weights
• YoloFace
  • Model (yolov3-wider_16000.weights - 246Mb) needs to be downladed from the repository
    • YOLOv3 model trained on WIDER FACE
    • Should be placed under yoloface/model-weights

Accuracy

Reported by the authors

• Results on validation set of WIDER FACE:

Model	Easy Set	Medium Set	Hard Set
CenterFace	93.5	92.4	87.5
DSFD	96.6	95.7	90.4
FaceBoxes	-	-	-
Haarcascade	-	-	-
MTCNN	-	-	-
RetinaFace	96.9	96.1	91.8
S3FD	93.7	92.4	85.2
YoloFace	-	-	-

• Results on test set of WIDER FACE:

Model	Easy Set	Medium Set	Hard Set
CenterFace	93.2	92.1	87.3
DSFD	96.0	95.3	90.0
FaceBoxes	-	-	-
Haarcascade	-	-	-
MTCNN	-	-	-
RetinaFace	96.3	95.6	91.4
S3FD	92.8	91.3	84.0
YoloFace	-	-	-

Face recognition

InsightFace/ArcFace recognition model is used to preform face recognition. Faces are saved in a list of recognized faces once they are recognized as a new face. A face is recognized as a new face if none of the other recognized faces doesn't achieve higher similarity than FACE_CONF_THRESHOLD. Face recognition can be easily switched on by using retina_face detector and setting retina_face.Recognition = True.

Performance with current settings used to detect faces of handball players

Model	Seconds per frame
CenterFace	1
DSFD	48
FaceBoxes	0.3
Haarcascade	1
MTCNN	4
RetinaFace	90
S3FD	20
YoloFace	10

More

State-of-the-art methods for Face Detection on WIDER Face (Hard) dataset

Papers with code - Face Detectors