This repository contains my upgraded version of using YoloV3 with OpenCV DNN to detect objects by training a YoloV4 model to recognize and detect 4 specific classes of traffic signs :
- Traffic lights.
- Speedlimit.
- Crosswalk.
- Stop signs.
The photos for this dataset were taken by Erik Mclean (big thanks to him).
This dataset contains 877 images of 4 distinct classes for the purpose of road sign detection. Bounding box annotations were provided in the PASCAL VOC format, I transformed the data in YoloV4 format to make it easier for this project (using roboflow), you can download my YOLOv4 format version of the dataset here.
Each images contained in the training, test and validation sets are labelled using the YoloV4 Format i.e :
- One row per object
- Each row is
confidenceProbability, x_center, y_center, width, heigh, classProbability... - If there is more than one class, each row will be
confidenceProbability, x_center, y_center, width, heigh, (1st)classProbability, (2nd)classProbability...(k)classProbability.
- Cmake
- OpenCV 4.0 or higher
- CUDA 11.0 or higher
- cuDNN 7.0 or higher
- Python3 or higher
- Visual Studio 15 and higher
- Windows OS (I haven't tried on other OS)
- GPU with CudaCompiler
- Lot of patience (but worth it)
To make custom training with YoloV4, we'll need first to setup our working environment. To make it easier for people to use, I developped some utilities python files allowing the automatisation of this process. So basically, all you'll need to do it's to launch the console (Windows Shell) and execute the commands in the Python console (it's not that hard I promise you).
I'll present two ways to do custom training model for this particular detection task : (Local version) and (Easier/cloud version). The first one will need to do every steps of installation directly in your code folder (locally), the second one is much easier because everything is done in Colab.
Before all, you'll first need to install and setup OpenCV with a GPU back-end, as we will need OpenCV DNN GPU-enabled module : this is mandatory.
-
If you want to train your own model for this detection task you can do it as following :
After the OpenCV installation in the root folder containing yourtraffic-signs-detectionfolder, you'll have to install and compile the darknet YoloV4 algorithm in your workspace by doing these setups tutorial : if you want to use cmake, and this one if you want to use vcpkg. All the steps to setup and train your custom model are very well explained in AlexeyAB repository so I will not going in details here. Thebackupfolder generated will store your trained weights. -
If you want to use my pre-trained weights you can simply open a
Windows Command Prompt, go to thetraffic-signs-detectionfolder and type the following command :python3 download_weights.py.
It will call this python file and launches the instructions inside. This command will download automatically the weights provided here and put them in theweightsfolder directly. You'll be able to manipulate the weights to do everything you want with them.
You can also use this Google Colab notebook if you do not have enough computational ressources to perform the training. The backup folder will contain your trained weights, you'll juste need to collect them after enough training of your model.
(Local version) : I'm providing the dataset here, if you want to do the splitting, parsing of the train, validation and test sets and extracting the files manually.
/!\ In case of laziness /!\ (I know this feeling quite well), I did "two little scripts" : download_sets.py and setup_sets.py to automate the downloading, extracting and splitting into training, validation and test sets.
If you want to use my scripts, just open Windows Command Prompt, go to the traffic-signs-detection folder and type the following commands by order :
python3 download_sets.py: it will download the datasets and extract them in thetraffic-signs-detectionfolder.python3 setup_sets.py: this one will split and copy the images and labels in the proper folders (more infos here).
After doing all these steps, you're ready to train your custom model using the downloaded pre-trained YoloV4 weights !
(Local version) : Now, you can start to train our custom model. In order to see if your darknet installation is working on Windows. To do that, you'll just have to launch a Windows PowerShell Console, go to the traffic-signs-detection folder and type the following command :
.\darknet.exe detector train ../data/obj.data ../cfg/[CONFIG].cfg ../weights/[WEIGHTS].weights -map
Needlessly to tell you to replace CONFIG and WEIGHTS by the own names you gave to these files (more infos here).
To test your custom model, this is the same steps to do than training but instead you have to enter this command :
.\darknet.exe detector test ../data/obj.data ../cfg/[CONFIG].cfg ../weights/[WEIGHTS].weights -map
(Again) needlessly to tell you to replace CONFIG and WEIGHTS by the own names you gave to these files (more infos here).
Using the weights provided in this repository, the model achieved an accuracy mAp @0.5 of 95.13% (it tooks ~ 2000 iterations) on classification and detection of the 4 different traffic road signs classes.
In case or you want to test the custom model using Jupyter Notebooks, I'm providing two notebooks for images testings and for videos testings.
The inputs folder contains images and videos for testing your custom model, feel free to add more images/videos you're finding relevant ! The results folder contains the outputs of our custom model when we input images and videos from the inputs folder.
Using the notebooks for images, we can show some of the results :
This work project is the continuation of my first project combining YoloV3 and OpenCV DNN, I wanted to push furthermore and train the model only on specific categories for this particular detection task.
Thank you Joseph Redmon and Ali Farhadi for their incredible work on YOLO object detector : YOLOv3: An Incremental Improvement, Redmon, Joseph and Farhadi, Ali ; arXiv, 2018.
Thank you AlexeyAB for your repository : YOLOv4 improvement
Thank you Andrew Mvd for your dataset : Kaggle profile