CVClassroom

Classroom-based Structure for CV Training (based on Noisy Student Training)

• Multi-Teacher?
• Multi-Student?
• BOTH?

stay tuned!

Reproducibility notes

Most files in this project are in keras/tensorflow.

Part 1: Stanford Cars Dataset

1. download the zip file of the dataset here: [https://www.kaggle.com/datasets/jessicali9530/stanford-cars-dataset?resource=download] This is what the folder should look like after you extract:

2. Move the files into the stanford cars folder, as seen below:

In the end, this is what the folder should look like: Note that cars_annos.mat is unnecessary and can be removed.

Sources of files in the folder:

• The CSV files of some labels of the dataset were from this link: [https://github.com/BotechEngineering/StanfordCarsDatasetCSV]
• More labels of cars were downloaded (the file called "cars_test_annos_withlabels (1).mat") from this kaggle link: [https://www.kaggle.com/datasets/abdelrahmant11/standford-cars-dataset-meta]

Part 2: The Car Connection Picture Dataset (somewhat unlabelled)

1. Download the zip file of the dataset here: [https://www.kaggle.com/datasets/prondeau/the-car-connection-picture-dataset?resource=download] The folder should consist of only .png files Name the folder "imgs", and put the "imgs" folder under the car_connection_dataset folder in this repository.

Part 3: teacher models

For the transfer learning, we need a different structure of the stanford cars dataset. Run download_stanford_cars_dataset.py to download this other structure.