The accuracy of yolov7 in cracking down on counterfeits, not all papers are authentic

Pay attention to the parallel stars

Never get lost

Institute of computer vision

official account ID｜ComputerVisionGzq

Study Group ｜ Scan the code to get the join mode on the homepage

Computer Vision Institute column

author ：Edison_G

Recently we shared Yolov6 and Yolov7 Two new frameworks , But many students hope that we can really interpret the code in detail , today “ Institute of computer vision ” Just... First Yolov7 Make a real experimental comparison of the performance of .

Before “ Institute of computer vision ” Shared yolo The latest two versions of the series ：

Yolov7： The latest and fastest real-time detection framework , The most detailed analysis and explanation （ With source code ）

Explode ！YOLOv6 The fast and accurate target detection framework is open source （ With source code download ）

Rep-PAN chart

Today we are based on Yolov7 Open source code , Realized some experiments in his paper .

stay MS COCO The results on the dataset of are as follows ：

First, in the coco On validation set , Verify the official yolov7.pt, In fact, it uses L edition （ similar YOLOv5-L,YOLOX-L,PPYOLOE-L）, give the result as follows ：

Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.51206
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.69730
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.55521
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.35247
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.55937
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.66693
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.38453
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.63765
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.68772
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.53766
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.73549
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.83868

notes ： The following changes require bs=1 To verify :

python3.7 test.py --data data/coco.yaml --img 640 --batch 1 --conf 0.001 --iou 0.65 --device 0 --weights yolov7.pt --name yolov7_640_val

• Change data preprocessing letterbox Of auto=True, Because the actual prediction is a fixed size , Only to 640 Size

It is obvious from the above results mAP by 50.8, It fell off 0.4.

• eval When NMS Yes, there is trick Of ,multi_label Indicates whether a box can be assigned two categories , But when we actually deploy, one box corresponds to one category , So put it again multi_label Set to False

The test result seems to fall again 0.2....

such YOLO There is no comparison between the series of speed measurements nms Time consuming , So for the sake of accuracy, we can nms A great article . Into the nms Before max_nm It's set to 30000, In actual deployment , Get into nms Of Tensor If it's very big and time-consuming , Generally, it will not be set to this large ,1000 enough , Change max_nms=1000. As well as max_det=300 Represents the maximum number of boxes in each diagram , It's really necessary 300 So much ? cocoapi Although the evaluation tool is max_det=100, But change 300 It will really rise .

max_nms=30000,max_det=300 This operation is not only eval The process slows down , Generate json It's slower when it's time to go , And if it's a time when you haven't trained well in the early stage of training eval, It's going to be slow .

 Why training while eval I feel very fast when I'm here ？
 reason ： While training eval And take the weight alone eval, From data processing to evaluation tools are not a set of logic ！

Add and change max_nms from 30000 To 1000,mAP Although it didn't fall , but recall falling ：

Add and change max_det from 300 To 100,mAP It's falling again 0.1：

According to the above 3 Click after operation ,X edition mAP Only 52.1 了 , It fell off 0.8 A little bit .

python3.7 test.py --data data/coco.yaml --img 640 --batch 1 --conf 0.001 --iou 0.65 --device 3 --weights yolov7x.pt --name yolov7_640_val

But the actual comparison yolov7 and yolov6, The result will indeed be some qualitative improvement .

Yolov7

Yolov6

Congratulations first 2022LPL The focus of the regular season in summer ,RNG 2-1 Overturn and defeat EDG, Take down LPL The victory of the Derby .

 THE END 

Please contact the official account for authorization.

The learning group of computer vision research institute is waiting for you to join ！

We created “ Computer vision society ” Knowledge planet has more than two years , It has also been recognized by many students , Recently, we started the operation of knowledge planet . We Regular meeting Push practical content to share with you , Students on the planet can Ask questions at any time , Be ready to ask for it , We will give timely reply and corresponding reply .

ABOUT

Institute of computer vision

The Institute of computer vision is mainly involved in the field of deep learning , Mainly devoted to face detection 、 Face recognition , Multi target detection 、 Target tracking 、 Image segmentation and other research directions . The Research Institute will continue to share the latest paper algorithm new framework , The difference of our reform this time is , We need to focus on ” Research “. After that, we will share the practice process for the corresponding fields , Let us really experience the real scene of getting rid of the theory , Develop the habit of hands-on programming and brain thinking ！

VX：2311123606

Previous recommendation  

• Yolov7： The latest and fastest real-time detection framework , The most detailed analysis and explanation （ With source code ）

• SSD7 | Embedded friendly target detection network , Product landing

• Face technology ： Photos of unclear people are restored into high-quality and high-definition image frames （ With source code download ）

• “ object detection ”+“ Visual understanding ” Realize the understanding and translation of the input image （ With source code ）

• object detection | Target size prediction based on statistical adaptive linear regression
  

• Accuracy improvement method ： The adaptive Tokens Efficient vision Transformer frame （ Open source ）
  

• ONNX elementary analysis ： How to accelerate the engineering of deep learning algorithm ？

• Explode ！YOLOv6 The fast and accurate target detection framework is open source （ With source code download ）

• FastestDet： Than yolov5 faster ！ stronger ！ Newly designed ultra real time Anchor-free Target detection algorithm （ With source code download ）
  

• At present, the target detection model with the highest accuracy, the fastest efficiency and the smallest storage （ Download the source code ）