This repository contains code used achieve 1st place in the Google Landmark Recognition/ Retrieval 2021 competition which was hosted on kaggle (https://www.kaggle.com/c/landmark-recognition-2021 , https://www.kaggle.com/c/landmark-retrieval-2021)

To derive the solution the following models were trained using 8xV100 NVIDIA GPU with distributed data parallel (DDP). The current repository only consits of dataset and model architecture as well as hyperparameters in form of config files, but lacks precise training and inference routine.

refer to https://github.com/haqishen/Google-Landmark-Recognition-2020-3rd-Place-Solution

refer to https://github.com/bestfitting/instance_level_recognition

The solution is summarized in the paper Efficient large-scale image retrieval with deep feature orthogonality and Hybrid-Swin-Transformers which is available under https://arxiv.org/abs/2110.03786

@misc{henkel2021efficient,
      title={Efficient large-scale image retrieval with deep feature orthogonality and Hybrid-Swin-Transformers}, 
      author={Christof Henkel},
      year={2021},
      eprint={2110.03786},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

include training/ inference script