Install

Development and Testing were done on Windows 10

Requirements: Python>=3.6.0 is required for all
PyTorch>=1.7 is required for YOLOv5
Git>=2.33.0 is required to fetch the YOLOv5
YOLOv5>=6 is required for training and running wheel_detector
LabelImg is required for labeling images for training

Installing PyTorch

Development was done using Nvidia CUDA 11.3

pip3 install torch==1.10.0+cu113 torchvision==0.11.1+cu113 torchaudio===0.10.0+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

Installing YOLOv5

git clone https://github.com/ultralytics/yolov5/tree/v6.0
cd yolov5
pip install -r requirements.txt

Installing LabelImg

pip3 install labelImg

or automatic using

.\install.cmd

Directory Structure

wheel_detector
 ┣ data
    ┗ images
    ┗ labels
 ┣ data_split          # generated on the fly, contians train/validate
 ┣ test
 ┣ wheel_detector      # generated on the fly, contains wheel_model from trainng)
 ┣ classes.txt
 ┣ detect.cmd
 ┣ install.cmd
 ┣ README.md
 ┣ split.cmd
 ┣ split_data.py
 ┣ wheel_dataset.yaml   # defines the location of the data for YOLOv5
 ┣ wheel_detector.pt    # final model, can be used as base for deployment training
 ┗ wheel_detector.yaml  # model defintion, defines the layers of the model

Detect

Run command:

detect.cmd is a windows batch script with all the parameters by default .\output is where the output image results with labels

.\detect.cmd [image_directory_path] [detection_run_name]

example:

.\detect.cmd .\test test_run1

after running command the output results goes in \output directory with same folder name as specified detection_run_name

Split Data

Randomly splitting the data

Run command:

images and labels expected to be inside \data directory

.\split.cmd

after running .\split.cmd the \data_split gets generated with train and validate folders

Train

Before training make sure you run .\split.cmd to split the data and generate data_spit\train and data_split\validate

Run command:

default values for small dataset:

img_input_size = 512
num_batch = 4
num_epoch = 1000

.\train.cmd [img_input_size] [num_batch] [num_epochs]

example:

.\train.cmd 512 4 1000

Label Images

classes.txt contains the class name of the detector, in this case it't 'wheel'

make sure classes.txt in the root and inside /data/labels directory

Run command:

labelImg .\data\images\ classes.txt .\data\labels\

to make things easy, make sure you select auto save mode option under view menu

Paper

Lighting and Rotation Invariant Real-time Detector Based on YOLOv5: Vehicle Wheel Detector https://arxiv.org/abs/2305.17785

Overview Presentation

Detection Demo