Car Recognition

This repository is to do car recognition by fine-tuning ResNet-152 with Cars Dataset from Stanford.

Dependencies

Dataset

We use the Cars Dataset, which contains 16,185 images of 196 classes of cars. The data is split into 8,144 training images and 8,041 testing images, where each class has been split roughly in a 50-50 split.

You can get it from Cars Dataset:

$ cd Car-Recognition
$ wget http://imagenet.stanford.edu/internal/car196/cars_train.tgz
$ wget http://imagenet.stanford.edu/internal/car196/cars_test.tgz
$ wget --no-check-certificate https://ai.stanford.edu/~jkrause/cars/car_devkit.tgz

ImageNet Pretrained Models

Download ResNet-152 into models folder.

Usage

Data Pre-processing

Extract 8,144 training images, and split them by 80:20 rule (6,515 for training, 1,629 for validation):

$ python pre-process.py

Train

$ python train.py

If you want to visualize during training, run in your terminal:

$ tensorboard --logdir path_to_current_dir/logs

Analysis

Update "model_weights_path" in "utils.py" with your best model, and use 1,629 validation images for result analysis:

$ python analyze.py

Validation acc:

88.70%

Confusion matrix:

Test

$ python test.py

Submit predictions of test data set (8,041 testing images) at Cars Dataset, evaluation result:

Test acc:

88.88%

Demo

Download pre-trained model into "models" folder then run:

$ python demo.py --i [image_path]

If no argument, a sample image is used:

$ python demo.py
class_name: Lamborghini Reventon Coupe 2008
prob: 0.9999994

1	2	3	4

Hyundai Azera Sedan 2012, prob: 0.99	Hyundai Genesis Sedan 2012, prob: 0.9995	Cadillac Escalade EXT Crew Cab 2007, prob: 1.0	Lamborghini Gallardo LP 570-4 Superleggera 2012, prob: 1.0

BMW 1 Series Coupe 2012, prob: 0.9948	Suzuki Aerio Sedan 2007, prob: 0.9982	Ford Mustang Convertible 2007, prob: 1.0	BMW 1 Series Convertible 2012, prob: 1.0

Mitsubishi Lancer Sedan 2012, prob: 0.4401	Cadillac CTS-V Sedan 2012, prob: 0.9801	Chevrolet Traverse SUV 2012, prob: 0.9999	Bentley Continental GT Coupe 2012, prob: 0.9953

Nissan Juke Hatchback 2012, prob: 0.9935	Chevrolet TrailBlazer SS 2009, prob: 0.987	Hyundai Accent Sedan 2012, prob: 0.9826	Ford Fiesta Sedan 2012, prob: 0.6502

Acura TL Sedan 2012, prob: 0.9999	Aston Martin V8 Vantage Coupe 2012, prob: 0.5487	Infiniti G Coupe IPL 2012, prob: 0.2621	Ford F-150 Regular Cab 2012, prob: 0.9995

Docker for API

You can build and run the docker using the following process:

Cloning

git clone https://github.com/jqueguiner/car-classification.git car-classification

Building Docker

cd car-classification && docker build -t car-classification -f Dockerfile .

Running Docker

echo "http://$(curl ifconfig.io):5000" && docker run -p 5000:5000 -d car-classification

Calling the API for image detection

curl -X POST "http://MY_SUPER_API_IP:5000/detect" -H "accept: image/png" -H "Content-Type: application/json" -d '{"url":"https://i.ibb.co/Lzpp400/input.jpg"}'

fiberhead/car-classifier