malte-soe / landmark_recognition_challenge

Landmark Recognition Challenge

Preparation

Create Environment

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

Download Datasets

• 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}

Directories

• ./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

Datasets

• 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

Experiments

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

00 Test

Only for testing purpose.

01 ResNet50 trained with triplet loss

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