The 1st Place Solution of the Google Landmark 2019 Retrieval Challenge and the 3rd Place Solution of the Recognition Challenge.

Our solution has been published! You can check from: Large-scale Landmark Retrieval/Recognition under a Noisy and Diverse Dataset

You can reproduce our environments using Dockerfile provided here https://github.com/lyakaap/Landmark2019-1st-and-3rd-Place-Solution/blob/master/docker/Dockerfile

• Google Landmark Dataset v1 (GLD-v1): https://www.kaggle.com/google/google-landmarks-dataset
• Google Landmark Dataset v2 (GLD-v2): https://github.com/cvdfoundation/google-landmark
• Clean version of the v2 (Newly Released!): https://www.kaggle.com/confirm/cleaned-subsets-of-google-landmarks-v2/kernels

Dataset statistics:

(You can skip this procedure to generate a cleaned subset. Pre-computed files are available on kaggle dataset.)

Run scripts/prepare_cleaned_subset.sh for cleaning the GLD-v2 dataset. The cleaning code requires DELF library (install instructions).

Prepare FishNet pretrained checkpoints first.

1. Download the checkpoint of FishNet-150 from https://www.dropbox.com/s/ajy9p6f97y45f1r/fishnet150_ckpt_welltrained.tar?dl=0
2. Place the checkpoint src/FishNet/checkpoints/
3. Execute src/FishNet/fix_checkpoint.py

Following commands are for reproducing our results.

cd ./experiments/
bash donwload_train.sh # download data
bash setup.sh  # setup data to ready training
bash reproduce.sh  # train models and predict for reproducing

python submit_retrieval.py  # for retrieval challenge
python submit_recognition.py  # for retrieval challenge

Model training and inference are done in experiments/ directory.

# train models by various parameter settings with 4 gpus (each training is done with 2 gpus).
python vX.py tuning -d 0,1,2,3 --n-gpu 2

# predict
python vX.py predict -m vX/epX.pth -d 0
# predict with multiple gpus
python vX.py multigpu-predict -m vX/epX.pth --scale L2 --ms -b 32 -d 0,1

• Comparison among loss functions with a ResNet-101. Dataset used for training is GLD-v1. For evaluation, GLD-v2, ROxford-5K, and RParis-6K are used. LB shows the Private LB scores.

• Comparison with dataset used for training. ArcFace and ResNet-101 are used for this experiment. By using the clean version of the GLD-v2, we can obtain better results compared to GLD-v2 without cleaning.

• Our re-ranking method is much better than other recent re-ranking methods, and can be combined with the others for even higher scores. A combination of ours and αQE achieves a competitive result to the score of 1st place with only a single model.

• https://github.com/filipradenovic/cnnimageretrieval-pytorch/tree/master/cirtorch
• https://github.com/ronghuaiyang/arcface-pytorch/blob/master/models/metrics.py
• https://github.com/4uiiurz1/pytorch-adacos/blob/master/metrics.py
• https://github.com/kevin-ssy/FishNet
• https://github.com/facebookresearch/maskrcnn-benchmark/blob/master/maskrcnn_benchmark/data/samplers/grouped_batch_sampler.py