This repository is based on A Large-Scale Car Dataset for Fine-Grained Categorization and Verification

The website

Github Gist

(The dataset is not open. You have to follow their instructions to download it)

See summary.md to know the summary of the paper

Experiment Flow:

There are 6 experiments as you can see in the picture above:

1. pre-trained ImageNet model -> Web model (the model is provided by the author in the Github Gist)
2. Web model -> Web best model
3. pre-trained ImageNet model -> SV model
4. Web best model -> SV model
5. pre-trained ImageNet model -> Mix(Web+SV) as 431 class
6. pre-trained ImageNet model -> Mix(Web+SV) as 531 class

The last two experiments are different because the intersection of Web and SV are only 181 classes, not all 281 classes in SV

So 431 class is for those who have intersection, 531 are their union

See each readme.md file in each exp folder

You have to collect all images into a dir. The example code is cptodir.py .

Then you have to resize the images and crop them. The example code is prep.sh .