Google Landmark Recognition / Retrieval 2021 1st place

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)

Models

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.

model image size stride data private score recognition public score recognition private score retrieval public score retrieval
DOLG-EfficientNet-B5 768 2 GLDv2x 0.476 0.497 0.478 0.464
DOLG-EfficientNet-B6 768 2 GLDv2x 0.476 0.479 0.474 0.454
DOLG-EfficientNet-B7 448 1 GLDv2x 0.465 0.484 0.470 0.458
EfficientNet-B3-Swin-Base-224 896 2 GLDv2x 0.462 0.487 0.481 0.454
EfficientNet-B5-Swin-Base-224 448 1 GLDv2x 0.462 0.482 0.476 0.443
EfficientNet-B6-Swin-Base-384 384 1 GLDv2x 0.467 0.492 0.487 0.462
EfficientNet-B3 768 2 GLDv2 0.463 0.487
EfficientNet-B6 512 2 GLDv2 0.470 0.484 0.454 0.441
EfficientNet-B5 704 2 GLDv2x 0.459 0.428
Ensemble Recognition 0.513 0.534
Ensemble Retrieval 0.537 0.518

DOLG models

DOLG-EfficientNet-B5

DOLG-EfficientNet-B6

DOLG-EfficientNet-B7

Hybrid-Swin-Transformers

EfficientNet-B3-Swin-Base-224

EfficientNet-B5-Swin-Base-224

EfficientNet-B6-Swin-Base-384

Last years solutions

EfficientNet-B3 & EfficientNet-B6

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

EfficientNet-B5

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

Paper

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

Citing

BibTeX

@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}
}

ToDos

include training/ inference script