Click on the above “ Xiaobai studies vision ”, Optional plus " Star standard " or “ Roof placement ”

 Heavy dry goods , First time delivery 

distance YOLO v4 Launch , It's over 5 More than a month .YOLO Frame adoption C Language as the underlying code , This is for the customary Python For the researchers , It's really a little unfriendly . Therefore, there are many online based on various in-depth learning frameworks YOLO Duplicate version . In recent days, , There are researchers GitHub Based on PyTorch Of YOLOv4.

From this year 4 month YOLOv4 After the release of , For this target detection framework , Perhaps the most frequently asked question is ：
「 Have any classmates reappeared YOLOv4 Of , Can you talk about it 」.
Because of the original YOLO Use C Programming with language , This alone makes many students flinch . There are many online TF/Keras and Caffe Etc , But many projects only give code , There is no given model in COCO、PASCAL VOC Training results on the dataset .

In recent days, , There are researchers GitHub Open source a project on ： be based on PyTorch Deep learning framework YOLOv4 Duplicate version , This version is based on YOLOv4 The implementation given by the author AlexeyAB/darknet, And in PASCAL VOC、COCO And custom datasets .

Project address ：https://github.com/argusswift/YOLOv4-PyTorch

in addition to , The project also adds some useful Attention method , And implemented mobilenetv2-YOLOV4 and mobilenetv3-YOLOV4.

attentive YOLOv4

The project adds some attention methods to the backbone , Such as SEnet、CBAM.

SEnet (CVPR 2017)

CBAM (CVPR 2018)

mobilenet YOLOv4

The research also achieved mobilenetv2-YOLOV4 and mobilenetv3-YOLOV4（ Just change config/yolov4_config.py Medium MODEL_TYPE that will do ）.

The following table shows mobilenetv2-YOLOV4 Performance results of ：

Now let's look at the details and requirements of the project .

Environmental requirements

• Nvida GeForce RTX 2080TI

• CUDA10.0

• CUDNN7.0

• windows or linux System

• python 3.6

characteristic

• DO-Conv (https://arxiv.org/abs/2006.12030) (torch>=1.2)

• Attention

• fp_16 training

• Mish

• Custom data

• Data Augment (RandomHorizontalFlip, RandomCrop, RandomAffine, Resize)

• Multi-scale Training (320 to 640)

• focal loss

• CIOU

• Label smooth

• Mixup

• cosine lr

Install dependencies

Run the script to install dependencies . You need to provide conda The installation path （ for example ~/anaconda3） And created conda Name of environment （ Here is YOLOv4-PyTorch）.

pip3 install -r requirements.txt --user

It should be noted that ： The installation script is already Ubuntu 18.04 and Window 10 Tested on the system . If something goes wrong , Please check the detailed installation instructions ：https://github.com/argusswift/YOLOv4-PyTorch/blob/master/INSTALL.md. 

preparation

1. git Copy YOLOv4 library

The first step in preparation is to copy YOLOv4.

git clone github.com/argusswift/YOLOv4-PyTorch.git

Then update the configuration file 「config/yolov4_config.py」 in 「PROJECT_PATH」.

2. Data set preparation

The project is ready Pascal VOC and MSCOCO 2017 Data sets . among PascalVOC The dataset includes VOC 2012_trainval、VOC 2007_trainval and VOC2007_test,MSCOCO 2017 The dataset includes train2017_img、train2017_ann、val2017_img、val2017_ann、test2017_img、test2017_list.

PascalVOC Data set download command ：

# Download the data.cd $HOME/data
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2012/VOCtrainval_11-May-2012.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtrainval_06-Nov-2007.tar
wget http://host.robots.ox.ac.uk/pascal/VOC/voc2007/VOCtest_06-Nov-2007.tar# Extract the data.
tar -xvf VOCtrainval_11-May-2012.tar
tar -xvf VOCtrainval_06-Nov-2007.tar
tar -xvf VOCtest_06-Nov-2007.tar

MSCOCO 2017 Data set download command ：

#step1: download the following data and annotation
 2017 Train images [118K/18GB]
 2017 Val images [5K/1GB]
 2017 Test images [41K/6GB]
 2017 Train/Val annotations [241MB]
 #step2: arrange the data to the following structure
 COCO
 ---train
 ---test
 ---val
 ---annotations

After downloading the data set , You need to do the following ：

• Put the dataset into the directory , to update config/yolov4_config.py Medium DATA_PATH Parameters .

• （ about COCO Data sets ） Use coco_to_voc.py take COCO The data type is converted to VOC data type .

• Transform data format ： Use utils/voc.py or utils/coco.py take pascal voc *.xml Format （ or COCO *.json Format ） Convert to *.txt Format （Image_path xmin0,ymin0,xmax0,ymax0,class0 xmin1,ymin1,xmax1,ymax1,class1 ...）.

3. Download the weight file

1）darknet Pre training weights ：yolov4（https://drive.google.com/file/d/1cewMfusmPjYWbrnuJRuKhPMwRe_b9PaT/view）.

2）Mobilenet Pre training weights ：

mobilenetv2：（https://pan.baidu.com/share/init?surl=sjixK2L9L0YgQnvfDuVTJQ, Extraction code ：args）;

mobilenetv3：（https://pan.baidu.com/share/init?surl=75wKejULuM0ZD05b9iSftg, Extraction code ：args）.

3） Create in the root directory weight Folder , Put the downloaded weight file into weight / Under the table of contents .

4） Training in config/yolov4_config.py Set in MODEL_TYPE.

4. Convert to a custom dataset （ Training based on custom datasets ）

1） Put the pictures of the custom dataset into JPEGImages Folder , Put the comment file in Annotations Folder .

2） Use xml_to_txt.py File write the list of training and testing files ImageSets/Main/*.txt.

3） Transform data format ： Use utils/voc.py or utils/coco.py take pascal voc *.xml Format （ or COCO *.json Format ） Convert to *.txt Format （Image_path xmin0,ymin0,xmax0,ymax0,class0 xmin1,ymin1,xmax1,ymax1,class1 ...）.

Training

Run the following command to start training , For details, see config / yolov4_config.py. When training, you should DATA_TYPE Set to VOC or COCO.

CUDA_VISIBLE_DEVICES=0 nohup python -u train.py  --weight_path weight/yolov4.weights --gpu_id 0 > nohup.log 2>&1 &

It also supports resume Training , add to --resume, Use the following command to automatically load last.pt.

CUDA_VISIBLE_DEVICES=0 nohup python -u train.py  --weight_path weight/last.pt --gpu_id 0 > nohup.log 2>&1 &

testing

Modify the detection image path ：DATA_TEST=/path/to/your/test_data# your own images.

for VOC dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_voc.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval --mode det
for COCO dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_coco.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval --mode det

The result can be in output / View in , As shown below ：

assessment （Pascal VOC Data sets ）

Modify the evaluation dataset path ：DATA_PATH=/path/to/your/test_data # your own images

for VOC dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_voc.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval -

 

assessment （COCO Data sets ）

Modify the evaluation dataset path ：DATA_PATH=/path/to/your/test_data # your own images

CUDA_VISIBLE_DEVICES=0 python3 eval_coco.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval --mode val

type=bbox
Running per image evaluation...      DONE (t=0.34s).
Accumulating evaluation results...   DONE (t=0.08s).
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.438 
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.607 
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.469 
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.253 
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.486 
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.567 
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.342 
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.571 
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.632 
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.458 
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.691 
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.790

Visualization heat map

stay val_voc.py Set in showatt=Ture, The network can output the heat map .

for VOC dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_voc.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval
for COCO dataset:
CUDA_VISIBLE_DEVICES=0 python3 eval_coco.py --weight_path weight/best.pt --gpu_id 0 --visiual $DATA_TEST --eval

stay output / You can view the heat map , As shown below ：

The good news ！

Xiaobai learns visual knowledge about the planet

Open to the outside world

 download 1：OpenCV-Contrib Chinese version of extension module 

 stay 「 Xiaobai studies vision 」 Official account back office reply ： Extension module Chinese course , You can download the first copy of the whole network OpenCV Extension module tutorial Chinese version , Cover expansion module installation 、SFM Algorithm 、 Stereo vision 、 Target tracking 、 Biological vision 、 Super resolution processing and other more than 20 chapters .


 download 2：Python Visual combat project 52 speak 
 stay 「 Xiaobai studies vision 」 Official account back office reply ：Python Visual combat project , You can download, including image segmentation 、 Mask detection 、 Lane line detection 、 Vehicle count 、 Add Eyeliner 、 License plate recognition 、 Character recognition 、 Emotional tests 、 Text content extraction 、 Face recognition, etc 31 A visual combat project , Help fast school computer vision .


 download 3：OpenCV Actual project 20 speak 
 stay 「 Xiaobai studies vision 」 Official account back office reply ：OpenCV Actual project 20 speak , You can download the 20 Based on OpenCV Realization 20 A real project , Realization OpenCV Learn advanced .


 Communication group 

 Welcome to join the official account reader group to communicate with your colleagues , There are SLAM、 3 d visual 、 sensor 、 Autopilot 、 Computational photography 、 testing 、 Division 、 distinguish 、 Medical imaging 、GAN、 Wechat groups such as algorithm competition （ It will be subdivided gradually in the future ）, Please scan the following micro signal clustering , remarks ：” nickname + School / company + Research direction “, for example ：” Zhang San  +  Shanghai Jiaotong University  +  Vision SLAM“. Please note... According to the format , Otherwise, it will not pass . After successful addition, they will be invited to relevant wechat groups according to the research direction . Please do not send ads in the group , Or you'll be invited out , Thanks for your understanding ~