Car Dekhliya

An AI tool that streamlines your car insurance process by detecting damages with the most accuracy.

Table of Contents

• Demo
• Background
• Installation
• Usage
• API
• TODO
• Contributing
• License

Demo

Watch the video demonstration: https://www.loom.com/share/c258aae54fcd494ab2c041cc9923b5f9

Background

Car damages can be a major safety hazard if not detected and repaired in time. However, detecting damages in cars can be a tedious and time-consuming task for humans. This is where machine learning can be of great help. By training a machine learning model on a dataset of car images with damages, we can develop an automated system for detecting damages in cars.

Installation

To install the necessary dependencies for this project, run the following command:

pip install -r requirements.txt

To use the car damage detection model, follow these steps:

• Provide the path to the image you want to test
• The script will output whether the car has damages or not

from prediction_engine import get_yolov5
model = get_yolov5(input_image)
print(model)

API

The POST methods in the `app.py` file:

"/obj-to-json"

This will return the predictions in a JSON format in the response body

{
  "result": [
    {
      "xmin": 330.4482116699,
      "ymin": 80.6117782593,
      "xmax": 649.321105957,
      "ymax": 470.5021972656,
      "confidence": 0.9025965333,
      "class": 15,
      "name": "rear-bumper-dent"
    }
  ]
}

"/obj-to-img"

This route will return the resulting image with bounding box around the detected damage area.

and finally,

"/severity"

This route returns the severity of the damage into 3 classes, minor, moderate, severe.

{
  "result": {
    "severity": "minor"
  }
}

TODO

• Percentage value of damage on a vechicle
• Generalize the model for any damaged object
• Add a ChatGPT powered chatbot that assists the customer throughout the process
• Fire UI

Contributing

We welcome contributions to this project. To contribute, please fork this repository, make your changes, and submit a pull request.