1、 Original paper

link , Extraction code ：mk96

2、 The thought of thesis

3、 Thesis translation

     YOLO 9000 Better 、 faster 、 stronger

Abstract

      We have introduced an advanced 、 In real time 、 Can detect more than 9000 Type of target detection system -YOLO 9000. First, we put forward some suggestions for YOLO The improvement of this detection algorithm , There are some new things in these improvements , We also refer to some previous work . The improved model YOLO v2 stay Pascal voc and COCO The performance of these standard testing tasks is the best . A novel multi-scale training method is used to make the same YOLO v2 The model can run on different scales , And a good balance between speed and accuracy . With 67fps The detection speed of ,YOLO v2 stay voc Obtained on dataset 76.8 Of mAP. With 40fps The detection speed of ,YOLO v2 To obtain the 78.6 Of mAP, Than some advanced algorithms, such as resnet Of RCNN and SSD Good performance , And significantly faster . Finally, we propose a method of joint training target detection and classification . In this way , We are at the same time COCO Target detection and ImageNet Simultaneous training on image classification data set YOLO v2. This kind of joint training makes YOLO v2 It can detect the categories of detection data that are not marked . We are ImageNet Verify our method in the detection task . Even though 200 There are only 44 Detection data of categories ,YOLO v2 stay ImageNet The detection verification set obtained 19.7 Of mAP. In not belonging to COCO Of 156 Among categories ,YOLO v2 To obtain the 16 Of mAP. however ,YOLO The categories that can be detected are not just 200 Kind of , It can detect more than 9000 Different target categories . also , He is still real-time detection .

1、 introduction

      A general purpose target detection system should be fast 、 Accurate and able to recognize a large number of objects . Since the introduction of neural networks , The detection system becomes more and more fast and accurate . however , A large number of detection methods can still detect only a small number of objects .
      Compared with classified or data sets , The current target detection data set is limited . The most common detection data sets contain thousands to hundreds of thousands of images with hundreds of labels . The classified dataset has millions of images in tens of thousands or hundreds of thousands of categories .
      We hope the detection can reach the classification level . however , Detect the picture annotation in the task , Far more precious than sorting or labeling ( Tags are usually provided free of charge by users ). So in the near future , We are unlikely to see the detection data set reach the size of the classification data set .
      We propose a new method to use the large amount of classified data we already have , And use it to expand the detection range of our current target detection system . Our method uses a hierarchical view of target classification , It allows us to combine different data sets .
      We also propose a joint training algorithm , This algorithm enables us to train our target detector on the detected and classified data . This algorithm uses the labeled detection images to learn to locate objects accurately , At the same time, the classification image is used to improve his vocabulary and robustness .
      Using this method, we trained YOLOv2 This is real-time , Can detect more than 9000 A target detection system for different types of objects . First of all we have YOLO Improved on the basis of YOLO v2 This advanced real-time detector . Then we use data set merging method and joint training algorithm , stay ImagNet Of 9000 Categories and COCO The model is trained on the detection data set .

picture 1：YOLO v2 It can detect a large number of object categories in real time .

2、 Better

     YOLO Compared with the advanced target detection system, there are a lot of problems .YOLO And Fast RCNN The error comparison analysis shows that YOLO There are a lot of positioning errors .YOLO And based on region proposal Compared with ,recall Relatively low . So we want to maintain the classification accuracy , Focus on improving recall And precise positioning .
      Computer vision