for gpu and pytorch,tensorflow mixed version

dont pip install opencv and pyside

  conda create -n masking python=3.8
  conda activate masking
  pip install -r requirement_gpu.txt
  conda install -c anaconda cudatoolkit
  conda install -c conda-forge opencv
  conda install -c conda-forge pyside2
  pip install tensorflow=2.5

frame memory : replaced by deepSORT object track, initially used to mask previously detected area as the AI may fail to detect the plate in some frame

blur size : the strength of blurring , higher the value, stronger the effect

Detection threshold : as its name implies, the threshold of conf level

ROI ratio : scale up or down the detected bounding box (also currently deactivated in order to evaluate the real effect of AI)

inference size : unused params (the original wirter use it to lower the resolution of input frame in order to increase runtime)

credit:https://colab.research.google.com/drive/11GRCzo-yKzkntLK1PwuhE4RV9nGUYLhs#scrollTo=Fl7PsmikjCBW

source:https://github.com/theAIGuysCode/OIDv4_ToolKit for license plate, the maximum number of image for train and validation data is around 4000 and 500 respecitively

python main.py downloader --classes 'Vehicle registration plate' --type_csv train --limit 4000
python main.py downloader --classes 'Vehicle registration plate' --type_csv validation --limit 500
python convert_annotations.py
rm -r OID/Dataset/train/'Vehicle registration plate'/Label/
rm -r OID/Dataset/validation/'Vehicle registration plate'/Label/

source: flickr (https://www.flickr.com/photos/j3tourshongkong/) the current dataset contain 1000 local car plates

advantages

multiple car types, different color, free to use, well-organized drawback:

fixed lens, resolution, angle and location

Current composition:

**After data augmentation:(need more improvementssss) **

• brightness decrease
• mosaic
• shear angle (whole picture / inside bounding box)
• blur

for data augmentation, i just use roboflow to augment some of the pictures in dataset (free account limit) as a proof of work. You may use other tools to process all the images, including non-local part to generate a bigger data for AI trainning

for futhre data augementation:

• birgthness variation based on certain distribution
• mosaic +blur+ brightness
• other data augmentation techniques mentioned in yolov4 paper e.g. cutoff
• random combination of data augmentation techniques

please be reminded that after shearing angle, u need to check or recalculate the bounding box coordinate. If u are using roboflow to generate dataset, u need to convert the _annotation.txt file into individal txt file

tools: https://github.com/tzutalin/labelImg

tips: 1.read the instructions on the tool github 2.change the classes.txt before labelling 3.open autosave and deafult label option if u are just labelling license plate (save times) 4.choose YOLO format output before labelling

**If u are also new to darknet training, I recommed you to open a new folder and create all the files u need and paste them to darknet folder later

step1:define new cfg file(veryyyyyyyy important)

texts from tutorial page

I recommend having batch = 64 and subdivisions = 16 for ultimate results (Personal advise: if the training fail to start, reduce the batch and sudivision size as the CUDA may be out of memory) . If you run into any issues then up subdivisions to 32.Make the rest of the changes to the cfg based on how many classes you are training your detector on.

Note: I set my max_batches = 6000, steps = 4800, 5400, I changed the classes = 1 in the three YOLO layers and filters = 18 in the three convolutional layers before the YOLO layers.

How to Configure Your Variables:

width = 416

height = 416 (these can be any multiple of 32, 416 is standard, you can sometimes improve results by making value larger like 608 but will slow down training)

max_batches = (# of classes) * 2000 (but no less than 6000 so if you are training for 1, 2, or 3 classes it will be 6000, however detector for 5 classes would have max_batches=10000)

steps = (80% of max_batches), (90% of max_batches) (so if your max_batches = 10000, then steps = 8000, 9000)

filters = (# of classes + 5) * 3 (so if you are training for one class then your filters = 18, but if you are training for 4 classes then your filters = 27)

Optional: If you run into memory issues or find the training taking a super long time. In each of the three yolo layers in the cfg, change one line from random = 1 to random = 0 to speed up training but slightly reduce accuracy of model. Will also help save memory if you run into any memory issues.

Summary

• change width and height to 416
• change classes to 1
• change batches to 6000 (increase this if u want to train more epoches)
• search "yolo" in text editor, there should be 3 yolo layers in total, for each layer, change the filter into 18 <=$(1+5) * 3

Put the {name of model}.cfg to /darknet/cfg/ My cfg file name is yolov4-obj.cfg. you may check that for reference

step2: define obj.names and obj.data

obj.names : name list of classes

obj.data : files that show where the darknet should get data from

Put obj.names and obj.data into /darknet/data/

step3: generate train.txt and test.txt

noticed that the darknet author defined the folder name of training as "obj" and testing as "test". steps if u are using my code

#test_gen.py
import os

image_files = []
os.chdir(os.path.join("data", "test"))
for filename in os.listdir(os.getcwd()):
    if filename.endswith(".jpg"):
        image_files.append("data/test/" + filename)
os.chdir("..")
with open("test.txt", "w") as outfile:
    for image in image_files:
        outfile.write(image)
        outfile.write("\n")
    outfile.close()
os.chdir("..")

1. creat 2 folders named 'obj' and 'test'
2. put dataset for training into obj and dataset for testing into test
3. create a folder named 'data' and put 'obj' and 'test' into it
4. run test_gen.py and train_gen.py
5. Put the files inside data folder into /darknet/data/ file

step4: run AI train

./darknet detector train data/obj.data cfg/yolov4-obj.cfg /home/h06607/darknet/backup/yolov4-objv1_6000.weights -map -show_image -clear 

format: ./darknet detector mode(train/detect) cfg path initial weight file(at least use yolov4.conv.137) flags flags:

• map: plot loss and mAP@50 in a png file
• show_image: show the map in livetime
• clear : remove the epoch count in the initial weight files (e.g train with yolov4-objv1_6000.weights as starting point without clear flag will directly stop as max_epoch that i used is 6000 )

runtime

performance