A special target detection called You Only Looks Once

Abstract

This paper proposes a unique target detection called You Only Look Once (YOLO for short) and a recognition method. Unlike the previous idea of many classifiers with object detection functions, the object detection box is set as a spatially separated bounding box and the regression problem of related class probabilities is realized. The neural network model can directly scan the entire image during testing, predicting bounding boxes and class probabilities from the complete picture. At the same time, because the whole detection channel relies on a single neural network, it is more straightforward and concise when upgrading and updating. The unified architecture used in this paper is fast, with smaller versions of the YOLO model processing a staggering 155 frames per second. At the same time, it has also achieved excellent results in mAP. Compared with other detection systems, YOLOv3 has optimized many past problems, including positioning errors. At the same time, the probability of predicting false detections in the absence of false detections is small. Finally, like the YOLO base model, YOLOv3 may produce significant errors when processing abstract works of art and images with a large number of small objects. However, its actual results are still better than detection methods such as Region -Convolutional Neural Networks (R-CNN).