• Navigate into this repository
• Execute following command: conda env create -f environment.yml
• Activate the environment: conda activate landmark-recognition

• Download one or more datasets from Datasets
  • It doesn't matter where you save them
  • Don't rename the downloaded csv. They should be named train or test.
• Navigate into this repository
• Activate the environment
• Execute following command: python scripts/download_dataset.py --name={FOLDER_NAME} --csv={PATH_TO_DOWNLOADED_CSV}
  • You have to download train and test images separately
  • Images are saved in ./data/{FOLDER_NAME}/{CSV_NAME}/{ID}.jpg
  • For testing purpose you can also download only the first N images with passing --num {N}

• ./data/ contains the datasets.
• ./evaluation/ contains the evaluation artifacts, such as evaluated metrics.
• ./tensorboard/ contains the tensorboard logs.
• ./log/ contains logs, such as logged stdout.
• ./experiments/ contains experiment scripts.
  • experiment scripts must be named after the following structure: exp_{ID}_{NAME}.py

• Dataset for challenge: https://www.kaggle.com/google/google-landmarks-dataset
• Bigger dataset: https://github.com/cvdfoundation/google-landmark
  • Resized to 256x256: http://storage.googleapis.com/landmark-recognition-2019/compressed/train-256.tar
  • Discussion: https://www.kaggle.com/c/landmark-recognition-2019/discussion/91770#latest-530831

Start experiments with python main.py {ID} {FLAGS} Checkpoints, evaluation artifacts, logs are stored in sub directories named after the experiment and passed flags.

Only for testing purpose.

Crop input image into 5 sub images and extract features on each one. Uses triplet loss. Does not work yet!