Skip to content

Latest commit

 

History

History
123 lines (88 loc) · 5.84 KB

README.md

File metadata and controls

123 lines (88 loc) · 5.84 KB

CATI (Common Automotive Tasks on Images)

"CATI" is a tool to easily integrate and test deep learning automotive detection models. It is based on our paper "CATI – An Open-Source Framework to Evaluate Attacks on Cameras for Automated Vehicles".

Core features include:

  • Automotive Object Detection: Specifically trained version of YOLOv10 for the task of automotive object detection. It was trained on the COCO dataset with the following object types: "Person," "Car," "Bicycle," "Motorcycle," "Bus," "Truck," and "Traffic Light."
  • Traffic Sign Detection: Specifically trained version of YOLOv10 for the task of traffic sign detection. It was trained on a subset of the Mapillary Traffic Sign Dataset including 51 different German traffic signs.
  • Modular Framework: Its modular architecture allows the integration of further automotive machine learning models.

drawing

Requirements

CATI itself is built using default Python libraries, but a conda installation is required to create environments and install requirements for the different models.

For the installation we refer to the original homepage: https://docs.anaconda.com/miniconda/install/ Tested with Miniconda versions 24.9.2, 24.7.1, and 24.3.0.

Setup

  1. Clone the repository
  2. Install CATI with
python setup.py

This will create the needed environments and requirements for the two integrated models.

Hint: If you have problems with conda, try running the setup in the "base" conda environment (conda activate base).

Usage & Examples

General Syntax:

usage: cati.py [-h] --input INPUT --output OUTPUT --model
                    {yolov10_object_detection,yolov10_traffic_sign_detection}
                    [{yolov10_object_detection,yolov10_traffic_sign_detection} ...]
                    [--conf CONF] [--save_images] [--cuda]

List of parameters:

  • --input: Filepath to the folder containing the images to analyze.
  • --output: Filepath to the folder that will contain the results (will be created if not existing).
  • --model: Space-separated list of models to execute.
  • (Optional) --conf: Confidence threshold value between 0.0 and 1.0. Default: 0.5
  • (Optional) --save_images: If given, save the images that include the detected objects
  • (Optional) --cuda: Use CUDA to execute the models. This requires a compatible hardware and driver.

Examples:

  1. Change the confidence cutoff value:
python cati.py --input /test_images --output /test_output --model yolov10_object_detection --conf 0.4
  1. Save the images with marked detections (flag):
python cati.py --input /test_images --output /test_output --model yolov10_object_detection --save_images
  1. Use cuda for detection (flag):
python cati.py --input /test_images --output /test_output --model yolov10_object_detection --cuda
  1. Run inference on multiple models:
python cati.py --input /test_images --output /test_output --model yolov10_object_detection yolov10_traffic_sign_detection

Output of Automotive Object Detection

Object Detection

Output of Traffic Sign Detection

Traffic Sign Detection

Output

By default the output will only contain a json-file for each image that will list all detected objects. There will be one sub-directory for each image in the output folder containing the images for each selected model (if enabled) and the json-file with all detected objects.

Model Addition

General syntax:

model_addition.py [-h] --model_path MODEL_PATH --inference_path INFERENCE_PATH --weights_path WEIGHTS_PATH
                         --requirements_path REQUIREMENTS_PATH --python_version PYTHON_VERSION --name NAME

List of parameters:

  • --model_path: Filepath to the trained custom model.
  • --inference_path: Filepath to the python script that runs the inference of the custom model.
  • --weights_path: Filepath to the trained model (typ. "PT" file).
  • --requirements_path: List of Python requirements in the requirements file format.
  • --python_version: Specification of the exact Python version
  • --name: Name of the model (avoid using spaces)

Example

python model_addition.py --model_path /yolov10 --inference_path /yolov10/inference.py --weights_path pretrained_weights/yolov10m.pt --requirements_path yolov10_requirements.txt --python_version 3.10 --name yolov10_object_detection

Important to note is that the file for the --inference_path (the file that initalizes the model and will load in the images etc.) has to be inside the --model_path to ensure that the paths can be correctly set for the interface.

After calling model_addition.py, all files should are moved to the respective directories, an environment with the necessary dependencies is created and the paths are configured such that cati.py includes the newly added model now in its selection.

Please also make sure that the --inference_path file, will accept all the parameters that can be used in cati.py to ensure correct functionality.

Acknowledgements

Datasets:

  1. bdd100k (https://doc.bdd100k.com/)
  2. German Traffic Sign Detection Benchmark (GTSDB) (https://benchmark.ini.rub.de/gtsdb_dataset.html)
  3. Mappillary Traffic Sign Dataset (https://www.mapillary.com/dataset/trafficsign)
  4. Common Objects in Context (COCO) (https://cocodataset.org/#home)

Models:

  1. YOLOv10 (https://github.com/THU-MIG/yolov10)