6.yolov4-tiny

yolov4-tiny official website: https://github.com/AlexeyAB/darknet

Source code: https://github.com/bubbliiiiing/yolov4-tiny-tf2

6.1. Introduction

Release time node

• 2020.04: YOLOv4 officially released

• 2020.06: YOLOv4-Tiny officially released

YOLOv4-Tiny performance on COCO: 40.2% AP50, 371 FPS (GTX 1080 Ti) Whether it is AP or FPS performance, it is a huge improvement compared to YOLOv3-Tiny, Pelee, and CSP. As shown below:

Comparison of YOLOv4 and YOLOv4-Tiny detection results, source network

YOLOv4 test results

xxxxxxxxxx
Done! Loaded 162 layers from weights-file
data/dog.jpg: Predicted in 27.039000 milli-seconds.
Bicycle: 92%
dog: 98%
truck: 92%
Potted plants: 33%

YOLOv4-Tiny test results

xxxxxxxxxx
Done! Loaded 38 layers from weights-file
data/dog.jpg: Predicted in 2.609000 milli-seconds.
Bicycle: 29%
dog: 72%
truck: 82%
cars: 46%

We can see that the detection accuracy of Yolov4-tiny has declined, but Yolov4-tiny has obvious advantages in terms of time consumption: Yolov4-tiny detection takes only 2.6 milliseconds, while Yolov4 detection takes 27 milliseconds, which is faster. More than 10 times!

6.2. Use

xxxxxxxxxx
#Raspberry Pi 5 master needs to enter docker first, please perform this step
#If running the script into docker fails, please refer to ROS/07, Docker tutorial
~/run_docker.sh

xxxxxxxxxx
roslaunch yahboomcar_yolov4_tiny yolodetect.launch display:=true

• [display] parameter: whether to enable the visual interface.

Supports real-time monitoring of web pages, such as:

xxxxxxxxxx
192.168.2.89:8080

View node information

xxxxxxxxxx
rqt_graph

Print detection information

xxxxxxxxxx
rostopic echo /DetectMsg

Print as follows

xxxxxxxxxx
data:
   -
     frame_id: "person"
     stamp:
       secs: 1646128857
       nsecs: 530594825
     scores: 0.609634816647
     ptx: 109.685585022
     pty: -2.94450759888
     distw: 374.364135742
     Disth: 236.672561646
     centerx: 187.182067871
     centery: 118.336280823

• frame_id: identification name.

• scores: recognition scores.

• ptx, pty: Coordinates of the upper left corner of the identification box.

• distw, disth: the width and height of the identification box.

• centerx, centery: identify the center.

6.3. Folder structure

xxxxxxxxxx
yolov4-tiny-tf2
├── font # Store font package
│ └── Block_Simplified.TTF
├── garbage_data # Data set
│ ├── GetData.py # Get the data set code
│ ├── image # Target source file
│ ├── JPEGImages # Data set pictures (as many as possible)
│ ├── texture # Background image (as many as possible)
│ └── train.txt # Label file corresponding to the data set image
├── img # Store test images
│ └── 1.jpg
├── logs # Store test logs and final training model last1.h5.
├── model_data #Storage pre-trained model (weight file)
│ ├── coco.txt
│ ├── garbage.h5
│ ├── garbage.txt # Custom label file (corresponding to the target source file)
│ ├── yolo_anchors.txt
│ ├── yolov4_tiny_weights_coco.h5 # Weight file
│ └── yolov4_tiny_weights_voc.h5 # Weight file
├── predict_img.py # Image detection code
├── predict_video.py # Video detection code
├── README.md
├── train.py # Training model code
├── utils # library files
└── yolo_nets # Network structure library file

The concept of anchor box was introduced in the YOLO-v2 version, which greatly increased the performance of target detection. The essence of anchor is the reverse of the SPP (spatial pyramid pooling) idea, and what does SPP itself do? It is to combine different sizes The input is resized to become the output of the same size, so the reverse of SPP is to reverse the output of the same size to get the input of different sizes.

6.4. Environmental requirements

The factory image is already configured and no installation is required.

xxxxxxxxxx
tensorflow-gpu==2.2.0
lxml
matplotlib
pandas
Pillow
scikit-learn
seaborn
tqd
imgaug

Installation example

xxxxxxxxxx
pip install imgaug

6.5. Customized training data set

6.5.1. Create data set

Method 1: Take some photos first, use the annotation tool to mark the targets on each photo, create a new [train.txt] file under the [garbage_data] folder, and write the target information in a specific format.

Method 2: Put background images (as many as possible) in the [garbage_data/texture] folder, modify the [GetData.py] code as needed, and execute [GetData.py] to generate a data set (as many as possible).

The names of the pictures and label files must correspond. The label format in the [train.txt] file is as follows:

xxxxxxxxxx
./garbage_data/JPEGImages/0.jpg 113,163,293,298,9 # Image path y, x, y + w, x + h ,label

Take method 2 as an example.

Open the [GetData.py] file

xxxxxxxxxx
sudo vim GetData.py

Modify the total number of generated data sets and fill it in as needed. [More], too few data sets will lead to unsatisfactory training results.

xxxxxxxxxx
img_total=10000

Run the [GetData.py] file to obtain the data set

xxxxxxxxxx
python GetData.py

6.5.2. Add weight file

There are good weight files (pre-trained models) [yolov4_tiny_weights_coco.h5] and [yolov4_tiny_weights_voc.h5] provided under the [model_data] file. Choose one of the two, and recommend coco’s weight file.

If you need the latest weight file, just search it on Baidu and download it.

6.5.3. Create label file

Be careful not to use Chinese tags and there should be no spaces in the folder!

For example: garbage.txt

xxxxxxxxxx
Zip_top_can
Old_school_bag
Newspaper
Book
Toilet_paper
... ...

6.5.4. Modify the train.py file

Modify according to your own needs by referring to the comments.

xxxxxxxxxx
# label position
annotation_path = 'garbage_data/train.txt'
# Get the location of classes and anchor
classes_path = 'model_data/garbage.txt'
anchors_path = 'model_data/yolo_anchors.txt'
# Location of pre-trained model
weights_path = 'model_data/yolov4_tiny_weights_coco.h5'
# Get classes and anchor
class_names = get_classes(classes_path)
anchors = get_anchors(anchors_path)
#How many categories are there in total?
num_classes = len(class_names)
num_anchors = len(anchors)
# The location where the trained model is saved
log_dir = 'logs/'
# Enter the image size. If the video memory is large, 608x608 can be used.
input_shape = (416,416)
#Initial epoch value
Init_epoch = 0
# Freeze the epoch value of training
Freeze_epoch = 50
# The size of Batch_size indicates how much data is fed each time. If there is OOM or insufficient video memory, please adjust it smaller.
batch_size = 16
# Maximum learning rate
learning_rate_base = 1e-3
#Total epoch value
Epoch = 100

Follow the above process, and after the operation is completed, directly run the [train.py] file for training.

xxxxxxxxxx
python3train.py

6.5.5. Model detection

• Modify the yolov4-tiny-tf2/utils/yolo.py file

xxxxxxxxxx
class YOLO(object):
     _defaults = {
         # Used to detect and train the model path.
         "model_path": 'model_data/garbage.h5',
         # yolo model parameter anchors path
         "anchors_path": 'model_data/yolo_anchors.txt',
         # Custom label file path
         "classes_path": 'model_data/garbage.txt',
         "score" : 0.5,
         "iou" : 0.3,
         "eager" : False,
         # The default is 416x416 (image size)
         "model_image_size" : (416, 416)
     }
... ...
#Font package path
self.font_path = 'font/Block_Simplified.TTF'

• Image detection

xxxxxxxxxx
python3 predict_img.py

During this period, you need to manually enter the images that need to be detected, as shown below:

• Video detection

python3 predict_video.py