Yolo v3 came out in 20 18, which is later than SSD and retinanet, and much faster than SSD and retinanet in the index of map0.5. In industrial applications, map0.5 has been fully utilized, and yolo v3 has a brief introduction, rich documents, and some series of variants such as tiny version. The most important thing is that it is fast, faster than SSD and retinanet.

YOLOv3 is not put forward to solve any problems, and the whole paper is actually a technical report. YOLOv3 has made some minor improvements on the basis of YOLOv2. The article is not long, and the core idea is similar to YOLOv2 and YOLO9000.

Model improvement:

Bounding box prediction: The anchor box is used to predict the bounding box for the positioning task, and YOLOv3 uses logistic regression to predict an objectivity score for each bounding box, which is based on the overlap between the predicted box and the object. If the overlap of one frame is higher than other frames, its score is 1, and those frames that are not the best frames and have an overlap greater than a certain threshold (0.5) are ignored.