One-stage : YOLO series, SSD. No filtering, fast, suitable for real-time, average accuracy.
Two-stage : RCNN, faster-RCNN, after screening (RPN region proposal network, network for extracting candidate frames), the speed is average and the accuracy is high.
mAP: comprehensive measure of detection effect,
FPS: speed
IOU: intersection and union ratio Intersection
: intersection of real box and predicted box
Union: union of real box and predicted box
TF indicates whether the prediction is correct: PN indicates predicted label 
accuracy (Accuracy) = (TP +TN) / (TP+FP+TN+FN) =70%
Precision : Precision represents the accuracy of prediction in positive sample results.
P=TP / (TP+FP) = 40/(40+20) = 66.7%
Recall : measures the classifier's ability to identify positive examples.
R=TP / (TP+FN) = 40/(40+10) = 80%
Confidence threshold: 0.9 is the possibility of the detection frame detecting an object or a face.
PR curve: Precision-Recall curve
AP: PR curve area (evaluation index)
mAP: average accuracy value of each category
ground-truth bounding box (GT): real box (real target position of the picture)
Predicted bounding box (bBox): prediction frame.