Car Retrieval System

By: Musa Izzanardi Wijanarko

This project is intended for completing technical test. This Car Retrieval System will be able to detect multiple car instances then classify/retrieve what is the type (eg. MPV, Sedan, Hatchback, etc) of the detected car.

This system consists of two models:

• Car Detection model, using DETR-ResNet50
• Car Body Type Classifier model, using MobilenetV3

The predicted result video can be seen on: ▶ this link

You can also access the technical reports documentation here to see the system architecture, training documentation, etc.

---

Getting started

This guide is for UNIX-based OS.

Requrements

• Python>=3.10.2
• All dependencies are in the requirements.txt
• Ideally a linux-based system, equipped with CUDA (check it with nvcc --version)

Setting up

To set this system up, you need to:

1. Clone the repo with git clone
2. go to the inside of the directory with cd car-retrieval system
3. (Optional) create a virtual environment venv with python3 -m venv venv
4. Activate it with source venv/bin/activate
5. Install the required dependencies pip install -r requirements.txt
6. You also need to setup a .env for these variables:

• HF_TOKEN: your huggingface credential
• WANDB_PROJECT: project name for tracking in Weight & Biases website
• HF_HOME: huggingface default download path (models, datasets, etc)
• TORCH_HOME: torch default download path
• VIDEO_GDRIVE_ID: the google drive id of test video

---

Car Detection Model

This object detection model was based on DETR-ResNet50 with 40M+ parameters, and was trained on the Indian Driving Dataset (Varma et al, 2018) 40K used for object detection task. We will use this dataset since it is resemble the Indonesian road condition.

The dataset has 9 labels and defined as such:

{
    "0": "traffic sign",
    "1": "motorcycle",
    "2": "car",
    "3": "rider",
    "4": "person",
    "5": "truck",
    "6": "autorickshaw",
    "7": "vehicle fallback",
    "8": "bus"
}

And we manually convert this dataset into to a CPPE-5-like (YOLO annotation) format and uploaded it into huggingface.
🤗 Huggingface Dataset Link

---

Car Classifier Model

The classifier model was based on MobileNetV3 available on torch architecture, and then trained on the Stanford Cars dataset introduced by Krause et al. (2015). This dataset does not contains car type, however, mayurmahukar create a mapping to 10 types of car, which are:

{
    0: 'Coupe',
    1: 'Sedan',
    2: 'Cab',
    3: 'Convertible',
    4: 'SUV',
    5: 'Minivan',
    6: 'Hatchback',
    7: 'Other',
    8: 'Van',
    9: 'Wagon'
 }