torch, torchvision, PIL, scipy, matplotlib
The CompCar dataste is provided here http://mmlab.ie.cuhk.edu.hk/datasets/comp_cars/index.html
It's inherited from the work of "Linjie Yang, Ping Luo, Chen Change Loy, Xiaoou Tang. A Large-Scale Car Dataset for Fine-Grained Categorization and Verification, In Computer Vision and Pattern Recognition (CVPR), 2015."
It's ecpteced to download and upzip the whole dataset under PROJECT_ROOT/Dataset.
Meanwhile, there is also a file called "full_model_name_updated.txt" under PROJECT_ROOT?Dataset, while contains 136725 lines. Each line will map a image file's directory (used to load this image) to the actual car model label.
Directly use the implementation in torchvision for resnet50.
The pretrained model should be downloaded here, and put into PROJECT_ROOT/Code/checkpoint2/10_epoch.tar
All the codes are push onto this repository, but the final one with full annotations is PROJECT_ROOT/Code/CarClassificationDemo2.ipynb
All the other codes are not for demonstration purpose.
(1) Download and extract the dataset here.
(2) Download this project. And put the dataset just downloaded into PROJECT_ROOT/Dataset/. An image in thid dataset should look like this CompCarDemo/Dataset/data/image/1/1101/2011/07b90decb92ba6.jpg
(3) This project is built with jupyter notebook and several scientific packages of python. So it's highly recommended to use anaconda to manage all these required packages in one place. Futhermore, miniconda is a much more lighweighted version, which takes much smaller space than Anaconda. It could be downloaded here.
(4) After installing the anaconda, the command 'conda' should be available in the terminal. The discription below is provided for Mac/Linux command line interface, simply because I'm unfamiliar with Windows termimal.
conda create --name CompCarDemoEnv python=3.6 #create a new environment in case the default python version is 2.x
conda activate CompCarDemoEnv #enter the new environment
conda install jupyter #for installing jupyter notebook
#use this line if you have a CUDA accelerated GPU
conda install pytorch torchvision cudatoolkit=10.2 -c pytorch
#use this line if not
conda install pytorch torchvision
conda install PIL,scipy,matplotlib(5) Download the pretrained model here, and put it like CompCarDemo/Code/checkpoint2/10_epoch.tar
(6) It's time to see the demo.
jupyter notebookThen open the file CompCarDemo/Code/CarClassificationDemo2.ipynb in the browser interface. Ready to go.
Add anotehr model in Code/CarClassification3.ipynb which is able to handle both car make and car model.
The pretrained model is provided here, and should be put like Code/checkpoint4/10_epoch.tar