The 2nd place solution of 2021 google landmark retrieval on kaggle.


We use cuda 11.1/python 3.7/torch 1.9.1/torchvision 0.8.1 for training and testing.

Download imagenet pretrained model ResNeXt101ibn and SEResNet101ibn from IBN-Net. ResNest101 and ResNeSt269 can be found in ResNest.

Prepare data

  1. Download GLDv2 full version from the official site.

  2. Run python tools/ This will generate clean, c2x, trainfull and all data for different stage of training.

  3. Validation annotation comes from all 1129 images in GLDv2. We expand the competition index set to index_expand. Each query could find all its GTs in the expanded index set and the validation could be more accurate.


We use 8 GPU (32GB/16GB) for training. The evaluation metric in landmark retrieval is different from person re-identification. Due to the validation scale, we skip the validation stage during training and just use the model from last epoch for evaluation.

Fast Train Script

To make quick experiments, we provide scripts for R50_256 trained for clean subset. This setting trains very fast and is helpful for debug.

python -m --standalone --nnodes=1 --nproc_per_node=8 --master_port 55555 --max_restarts 0 --config_file configs/GLDv2/R50_256.yml

Whole Train Pipeline

The whole training pipeline for SER101ibn backbone is listed below. Other backbones and input size can be modified accordingly.

python -m --standalone --nnodes=1 --nproc_per_node=8 --master_port 55555 --max_restarts 0 --config_file configs/GLDv2/SER101ibn_384.yml
python -m --standalone --nnodes=1 --nproc_per_node=8 --master_port 55555 --max_restarts 0 --config_file configs/GLDv2/SER101ibn_384_finetune.yml
python -m --standalone --nnodes=1 --nproc_per_node=8 --master_port 55555 --max_restarts 0 --config_file configs/GLDv2/SER101ibn_512_finetune.yml
python -m --standalone --nnodes=1 --nproc_per_node=8 --master_port 55555 --max_restarts 0 --config_file configs/GLDv2/SER101ibn_512_all.yml


  • With four models trained, cd to submission/code/ and modify settings in properly.

  • Then run to get submission file and evaluate on validation set offline.

General Settings

REID_EXTRACT_FLAG: Skip feature extraction when using offline code.
FEAT_DIR: Save cached features.
IMAGE_DIR: competition image dir. We make a soft link for competition data at submission/input/landmark-retrieval-2021/
RAW_IMAGE_DIR: origin GLDv2 dir
MODEL_DIR: the latest models for submission
META_DIR: saves meta files for rerank purpose
LOCAL_MATCHING and KR_FLAG disabled for our submission.

Fast Inference Script

Use R50_256 model trained from clean subset correspongding to the fast train script. Set CATEGORY_RERANK and REF_SET_EXTRACT to False. You will get about mAP=32.84% for the validation set.

Whole Inference Pipeline

  • Copy cache_all_list.pkl, cache_index_train_list.pkl and cache_full_list.pkl from cache to submission/input/meta-data-final

  • Set REF_SET_EXTRACT to True to extract features for all images of GLDv2. This will save about 4.9 million 512 dim features for each model in submission/input/meta-data-final.

  • Set REF_SET_EXTRACT to False and CATEGORY_RERANK to before_merge. This will load the precomputed features and run the proposed Landmark-Country aware rerank.

  • The notebooks on kaggle is exactly the same file as in and We also upload the same notebooks as in kaggle in kaggle.ipynb.

Kaggle and ICCV workshops

  • The challenge is held on kaggle and the leaderboard can be found here. We rank 2nd(2/263) in this challenge.

  • Kaggle Discussion post link here

  • ICCV workshop slides coming soon.


The code is motivated by AICITY2021_Track2_DMT, 2020_1st_recognition_solution, 2020_2nd_recognition_solution, 2020_1st_retrieval_solution.


If you find our work useful in your research, please consider citing:

 title={2nd Place Solution to Google Landmark Retrieval 2021},
 author={Zhang, Yuqi and Xu, Xianzhe and Chen, Weihua and Wang, Yaohua and Zhang, Fangyi},


View Github