Download the following datasets
Data folder → Corresponding dataset [link]
cars → Stanford cars http://ai.stanford.edu/~jkrause/cars/car_dataset.html
deepfashion → Consumer-to-shop https://drive.google.com/drive/folders/0B7EVK8r0v71pRXllRUdQcC1zTHc?resourcekey=0-YgTkHTdQH_KN0VcXr9k_jQ
fashion200k → Fashion200k https://www.kaggle.com/datasets/mayukh18/fashion200k-dataset
food-101 → Food-101 https://www.kaggle.com/datasets/srujanesanakarra/food101
gld → Google landmarks https://github.com/cvdfoundation/google-landmark
gldv2 → Google landmarks https://github.com/cvdfoundation/google-landmark
prod → Products10k https://products-10k.github.io/challenge.html#downloads
products → Stanford online products https://github.com/rksltnl/Deep-Metric-Learning-CVPR16
rp2k → https://www.pinlandata.com/rp2k_dataset/
ss → iMaterialist Challenge (Furniture) https://www.kaggle.com/competitions/imaterialist-challenge-furniture-2018/data
storefronts →Kaggle dataset https://www.kaggle.com/datasets/kerrit/storefront-146
in-shop → In-shop https://drive.google.com/drive/folders/0B7EVK8r0v71pVDZFQXRsMDZCX1E?resourcekey=0-4R4v6zl4CWhHTsUGOsTstw
met → MET artwork dataset http://cmp.felk.cvut.cz/met/
130k_kaggle → 130k public dataset from kaggle https://www.kaggle.com/datasets/rhtsingh/130k-images-512x512-universal-image-embeddings
!!!! For gld, gldv2 and ss we provide the dataset link that have been uploaded to kaggle https://www.kaggle.com/datasets/socratis/modified-datasets to avoid any confusion.
Place them into comp_dir/data
There is a train.csv which contains the paths, the original labels, the encoded labels, the set (train or valid) and the category of labels (apparel, landmark, food etc.)
The train.csv file contains all the images that we used for training. By downloading each dataset and placing them to the corresponding folder, the path should remain the same.
In case there are problems for the data placement, we have also kept the original id from each source dataset in the column ‘orig_label’ preceded by the folder’s or category’s name and the original names. For example, the label fashion200k_90037639 refers to the original id 90037639 of the source dataset. The only exceptions are the Google Landmarks dataset, which uses the original id, and the Food-101 and Storefronts datasets, which utilize the folder’s name.
#using conda
conda create -n <environment-name> --file req.txt
#using pip
pip install -r requirements.txt
cd comp_dir
python train.py