LdDl / road-anomaly-detection-train

Training object detector for road accidents detection

Training object detector for road accidents detection

Table of Contents

• About
• Pretrained models
• Usage
• References

About

It is just bunch of scripts to train road accidents detector for my pet-project written in Rust programming languages: https://github.com/LdDl/road-anomaly-detection

There are two scripts in this repository:

• download.py to download dataset of interest;
• train.py to run training process; (w.i.p)

Pretrained models

If you want just to download pretrained models here are links:

• YOLOv8 nano - Best weights (ONNX), Best weights (Pytorch) Last weights (Pytorch) Parameters:

  • Image size: 608x608
  • Batch size: 16
  • Epochs: 300
  • Cache images: yes
  Training results (images, CSV)

  Training and validation batches examples are here

  Results . Corresponding CSV is here

  Confusion matrix

  Normalized confusion matrix

  F1 Curve

  P Curve

  R Curve

  PR Curve

  Labels

  Labels correlogram

• YOLOv8 small - Best weights (ONNX), Best weights (Pytorch) Last weights (Pytorch)

  Parameters:

  • Image size: 608x608
  • Batch size: 16
  • Epochs: 300
  • Cache images: yes
  Training results (images, CSV)

  Training and validation batches examples are here

  Results . Corresponding CSV is here

  Confusion matrix

  Normalized confusion matrix

  F1 Curve

  P Curve

  R Curve

  PR Curve

  Labels

  Labels correlogram

• YOLOv8 medium - @todo train

• YOLOv8 large - @todo train

• YOLOv8 extra large - Best weights (ONNX), Best weights (Pytorch) Last weights (Pytorch)

  Parameters:

  • Image size: 608x608
  • Batch size: 16
  • Epochs: 300
  • Cache images: yes
  Training results (images, CSV)

  Training and validation batches examples are here

  Results . Corresponding CSV is here

  Confusion matrix

  Normalized confusion matrix

  F1 Curve

  P Curve

  R Curve

  PR Curve

  Labels

  Labels correlogram

Usage

• Clone the repository and navigate to root folder:

  git clone https://github.com/LdDl/road-anomaly-detection-train.git
  cd road-anomaly-detection-train

• Install dependencies

  pip3 install -r requirements.txt

• Navigate to selected dataset. In this case the link is:

  https://universe.roboflow.com/accident-detection-ffdrf/accident-detection-8dvh5
  

  Click Download button:

  

• Navigate to Terminal tab and get dataset ID and unique key to download it.

  

• Run download.py script

  export DATASET_ID=YOUR-DATASET-ID
  export ROBOFLOW_KEY=YOUR-ACCOUNT-KEY
  
  python3 download.py --dataset_id $DATASET_ID --key $ROBOFLOW_KEY --output dataset.zip

  You can adjust classes if you need to in lines 119 and 124:

  • Undefined classes would be marked as (max class ID + 1).
  • Warning: Re-labeled annotations would be stored in /train/labels, /test/labels and /valid/labels. Source labels would be stored in /train/labels_source, /test/labels_source and /valid/labels_source respectively.

• Run train.py script

  python3 train.py --cache_images t --model_size n --image_size 608 --yaml_path extracted_dataset --batch_size 16 --epochs 300

  When training is done you can extract both ONNX and Pytorch weights from run directory which would be generated during training process.

References

• Developers of YOLOv8 - https://github.com/ultralytics/ultralytics. If you are aware of some original papers for YOLOv8 architecture, please contact me to mention it in this README.
• Dataset source https://universe.roboflow.com/accident-detection-ffdrf/accident-detection-8dvh5