Car Brand Recognizer

Table of Contents

• Car Brand Recognizer
• Introduction
• Usage
  • Dependencies
  • Installation
  • Running the program
    • Single Image
    • Multiple Images
  • Output
• Examples
• Details
  • Basic Structure
  • detector.py
  • classifier.py
  • drawer.py
• Dataset

Introduction

This is a Python-based tool for detecting cars and identifying their brands in images, utilizing deep-learning object detection and classification models

The models are trained to predict 16 common car brands:
Audi, BMW, Chevrolet, Ford, Honda, Hyundai, Jeep, Kia, Lexus, Mercedes-Benz, MINI, Nissan, Subaru, Tesla, Toyota, Volkswagen

I made this as a practice in deep learning, computer vision, and Python development.

Usage

Dependencies

pytorch 2.5+
torchvision 0.20
PIL 11.0
opencv-python 4.11
numpy 2.2.2

CUDA is optional, the code will automatically run the model on cpu if cuda isn't available.

Installation

git clone https://github.com/vincent8264/car-brand-recog.git

IMPORTANT: Due to GitHub's file size limits, the models must be downloaded separately.

You can download the models on google drive, and place the .pt files in the ./functions/models folder like this:

│── car_recog.py
│── functions
│  │── models
│  │  │── PUT_THE_MODELS_HERE.txt
│  │  │── model_ft.pt
│  │  │── model_tl.pt

Running the program

Inputs with .jpg .jpeg .png are supported, other file formats will be ignored.

Single image

python car_recog.py --input path/to/input_image.jpg --output path/to/output_folder

Multiple images in a folder

python car_recog.py --input path/to/input_folder --output path/to/output_folder

If --input is omitted, the program will scan the current directory for supported image files.
If --output is omitted, results will be saved in a new ./output folder in the current directory.

Output

After running the program, the output folder will contain annotated images with bounding boxes and labels, adding "detected" to the image name:

│── output folder
│  │── detected_input1.jpg
│  │── detected_input2.jpeg
│  │── detected_input3.png

Examples

More examples can be found in the samples folder along with the sources of the images.

Details

Basic structure

The system uses three models in two steps. The first step uses a Retinanet model with pre-trained COCO dataset parameters that detects cars in the input image. The second step predicts the brands of the cars detected in the image, using ensembling of the two other classification models, a fine-tuned CNN and a transfer-learned efficientnet model.

The following description explains each step in detail.

detector.py

Uses retinanet_resnet50_fpn_v2 to detect cars in the input image, for its balance between accuracy and computational needs. All objects that are detected as labels "car" and "truck" are considered as cars. This is because pick-ups, which most people would call them "cars", are in the "truck" class in the COCO dataset. Unfortunately, this causes other actual "trucks" like semis to be included.

The cars in the images then go through additional checks to filter out those that are:

• Too close to the border
• Too small
• Blocked by other cars

All the detected cars that satisfy these thresholds are then cropped, and send to the classifier.

classifier.py

The classification step uses two models to predict the brand of the cropped cars. The first model is a fine-tuned CNN which uses a VGG-like architecture with fewer units and layers. The second model is a transfer-learned model, with efficientnet_v2_s as the base model. The classification models both have about 80% accuracy (f1 score) on the validation set.

After prediction, the output probabilities of each class are ensembled. The default ensembling method takes the maximum probability among the following three:

1. probabilities of each class from model 1
2. probabilities of each class from model 2
3. element-wise product of 1. and 2.

This is similar to using the class with the highest probability, but if both models also think another class has a significant amount of chances, the final prediction will be that class instead.

drawer.py

Finally, in this part, the bounding boxes of cars are rendered, along with the predicted class and ensembled probabilities as labels. The default color for the boxes and labels are green for more visibility.

Dataset

Car images used to train the classification models are from three datasets, with car models mostly between 2005~2020. In other words, it might not perform well on newer (or older) car models.

1. Stanford cars dataset
   https://www.kaggle.com/datasets/jessicali9530/stanford-cars-dataset

2. The Car Connection Picture Dataset
   https://github.com/nicolas-gervais/predicting-car-price-from-scraped-data/tree/master/picture-scraper https://www.kaggle.com/datasets/prondeau/the-car-connection-picture-dataset/data

3. Copyright-free images downloaded myself
   pexels.com
   pixabay.com
   unsplash.com