Kaggle Landmark Recognition Challenge in PyTorch

Method

We use an attention based model similar to DELF, but we do not use PCA to reduce dimensionality. The model uses pretrained ResNet50 weights on ImageNet

Files:

• clean_df.py
  • Script to clean the dataframe by checking every file is there and openable, deleting all the entries that do not exist
• data_utils.py
  • Contains various functions and classes for loading data among other stuff
• download_all.py
  • Script to download all of the data
• models.py
  • Contains the PyTorch code for our model
• split_data.py
  • Script to split the trainset into a train and validation set, keeping class ratios as balanced as we can
• submit-NN.py
  • Script to generate the submission
• train-NN.py
  • Script to train the model
• train_utils.py
  • Contains various useful utility functions for training

References

H. Noh, A. Araujo, J. Sim, T. Weyand, B. Han, "Large-Scale Image Retrieval with Attentive Deep Local Features", Proc. ICCV'17