NEW: Explore the dataset visually here.

This is the second version of the Google Landmarks dataset (GLDv2), which contains images annotated with labels representing human-made and natural landmarks. The dataset can be used for landmark recognition and retrieval experiments. This version of the dataset contains approximately 5 million images, split into 3 sets of images: train, index and test. The dataset was presented in our CVPR'20 paper and Google AI blog post. A hierarchical extension of the dataset is presented in an under-submission paper to the IEEE Transactions on Pattern Analysis and Machine Intelligence. In this repository, we present download links for all dataset files, baseline models and code for metric computation.

This dataset was associated to two Kaggle challenges, on landmark recognition and landmark retrieval. Results were discussed as part of a CVPR'19 workshop. In this repository, we also provide scores for the top 10 teams in the challenges, based on the latest ground-truth version. Please visit the challenge and workshop webpages for more details on the data, tasks and technical solutions from top teams.

As a reference, the previous version of the Google Landmarks dataset (referred to as Google Landmarks dataset v1, GLDv1) was available here. It is no longer available.

If you make use of this dataset, please consider citing the following papers:

Original GLDv2 CVPR'20 paper:

"Google Landmarks Dataset v2 - A Large-Scale Benchmark for Instance-Level Recognition and Retrieval"
T. Weyand*, A. Araujo*, B. Cao, J. Sim
Proc. CVPR'20

For the hierarchical labels:

"Optimization of Rank Losses for Image Retrieval"
E. Ramzi, N. Audebert, C. Rambour, A. Araujo, X. Bitot, N. Thome
In submission to: IEEE Transactions on Pattern Analysis and Machine Intelligence.

Explore the dataset visually here.

The current dataset version is 2.1. See the release history for details, including re-scored challenge submissions based on the latest ground-truth version.

There are 4,132,914 images in the train set.

• train.csv: CSV with id,url,landmark_id fields. id is a 16-character string, url is a string, landmark_id is an integer. Available at: https://s3.amazonaws.com/google-landmark/metadata/train.csv.

• train_clean.csv: CSV with landmark_id,images fields. landmark_id is an integer, images is a space-separated list of string train image IDs. Available at: https://s3.amazonaws.com/google-landmark/metadata/train_clean.csv. Courtesy of team smlyaka (see their paper).

• train_attribution.csv: CSV with id,url,author,license,title fields. id is a 16-character string, and the other fields are strings of variable length. Available at: https://s3.amazonaws.com/google-landmark/metadata/train_attribution.csv.

• train_label_to_category.csv: CSV with landmark_id,category fields: landmark_id is an integer, category is a Wikimedia URL referring to the class definition. Available at: https://s3.amazonaws.com/google-landmark/metadata/train_label_to_category.csv.

• train_label_to_hierarchical.csv: CSV with landmark_id,category,supercategory,hierarchical_label,natural_or_human_made fields: landmark_id is an integer, category is a Wikimedia URL referring to the class definition, supercategory is a string referring to the type of landmark mined from Wikimedia, hierarchical_label is a string corresponding to the hierarchical label, natural_or_human_made is a string indicating whether the landmark is natural of human-made. Available at: https://s3.amazonaws.com/google-landmark/metadata/train_label_to_hierarchical.csv.

The train set is split into 500 TAR files (each of size ~1GB) containing JPG-encoded images. The files are located in the train/ directory, and are named images_000.tar, images_001.tar, ..., images_499.tar. To download them, access the following link:

https://s3.amazonaws.com/google-landmark/train/images_000.tar

And similarly for the other files.

mkdir train && cd train
bash ../download-dataset.sh train 499

This will automatically download, verify and extract the images to the train directory.

Note: This script downloads files in parallel. To adjust the number of parallel downloads, modify NUM_PROC in the script.

All images in the train set have CC-BY licenses without the NonDerivs (ND) restriction. To verify the license for a particular image, please refer to train_attribution.csv.

There are 761,757 images in the index set.

IMPORTANT: Note that the integer landmark id's mentioned here are different from the ones in the train set above.

• index.csv: single-column CSV with id field. id is a 16-character string. Available at: https://s3.amazonaws.com/google-landmark/metadata/index.csv.

• index_image_to_landmark.csv: CSV with id,landmark_id fields: id is a 16-character string, landmark_id is an integer. Available at: https://s3.amazonaws.com/google-landmark/metadata/index_image_to_landmark.csv.

• index_label_to_category.csv: CSV with landmark_id,category fields: landmark_id is an integer, category is a Wikimedia URL referring to the class definition. Available at: https://s3.amazonaws.com/google-landmark/metadata/index_label_to_category.csv.

• index_label_to_hierarchical.csv: CSV with landmark_id,category,supercategory,hierarchical_label,natural_or_human_made fields: landmark_id is an integer, category is a Wikimedia URL referring to the class definition, supercategory is a string referring to the type of landmark mined from Wikimedia, hierarchical_label is a string corresponding to the hierarchical label, natural_or_human_made is a string indicating whether the landmark is natural of human-made. Available at: https://s3.amazonaws.com/google-landmark/metadata/index_label_to_hierarchical.csv.

The index set is split into 100 TAR files (each of size ~850MB) containing JPG-encoded images. The files are located in the index/ directory, and are named images_000.tar, images_001.tar, ..., images_099.tar. To download them, access the following link:

https://s3.amazonaws.com/google-landmark/index/images_000.tar

And similarly for the other files.

mkdir index && cd index
bash ../download-dataset.sh index 99

This will automatically download, verify and extract the images to the index directory.

Note: This script downloads files in parallel. To adjust the number of parallel downloads, modify NUM_PROC in the script.

All images in the index set have CC-0 or Public Domain licenses.

There are 117,577 images in the test set.

• test.csv: single-column CSV with id field. id is a 16-character string. Available at: https://s3.amazonaws.com/google-landmark/metadata/test.csv.

• recognition_solution_v2.1.csv: CSV with three columns: id (16-character string), landmarks (space-separated list of integer landmark IDs, or empty if no landmark from the dataset is depicted), Usage (either "Public" or "Private", referring to which subset the image belongs to). Available at: https://s3.amazonaws.com/google-landmark/ground_truth/recognition_solution_v2.1.csv.

• retrieval_solution_v2.1.csv: CSV with three columns: id (16-character string), images (space-separated list of string index image IDs, or None if this image is ignored), Usage (either "Public" or "Private", referring to which subset the image belongs to). Available at: https://s3.amazonaws.com/google-landmark/ground_truth/retrieval_solution_v2.1.csv.

The test set is split into 20 TAR files (each of size ~500MB) containing JPG-encoded images. The files are located in the test/ directory, and are named images_000.tar, images_001.tar, ..., images_019.tar. To download them, access the following link:

https://s3.amazonaws.com/google-landmark/test/images_000.tar

And similarly for the other files.

mkdir test && cd test
bash ../download-dataset.sh test 19

This will automatically download, verify and extract the images to the test directory.

Note: This script downloads files in parallel. To adjust the number of parallel downloads, modify NUM_PROC in the script.

All images in the test set have CC-0 or Public Domain licenses.

We also make available md5sum files for checking the integrity of the downloaded files. Each md5sum file corresponds to one of the TAR files mentioned above; they are located in the md5sum/index/, md5sum/test/ and md5sum/train/ directories, with file names md5.images_000.txt, md5.images_001.txt, etc. For example, the md5sum file corresponding to the images_000.tar file in the index set can be found via the following link:

https://s3.amazonaws.com/google-landmark/md5sum/index/md5.images_000.txt

And similarly for the other files.

If you use the provided download-dataset.sh script, the integrity of the files is already checked right after download.

We recommend that the set of TAR files corresponding to each dataset split be extracted into a directory per split; ie, the index TARs extracted into an index directory; train TARs extracted into a train directory; test TARs extracted into a test directory. This is done automatically if you use the above download instructions/script.

The directory structure of the image data is as follows: Each image is stored in a directory ${a}/${b}/${c}/${id}.jpg, where ${a}, ${b} and ${c} are the first three letters of the image id, and ${id} is the image id found in the csv files. For example, an image with the id 0123456789abcdef would be stored in 0/1/2/0123456789abcdef.jpg.

We make available the ResNet101-ArcFace baseline model from the paper, see instructions here.

The metric computation scripts have been made available, via the DELF github repository, see the python scripts compute_recognition_metrics.py and compute_retrieval_metrics.py. These scripts accept as input the ground-truth files, along with predictions in the format submitted to Kaggle.

The annotations are licensed by Google under CC BY 4.0 license. The images listed in this dataset are publicly available on the web, and may have different licenses. Google does not own their copyright. Note: while we tried to identify images that are licensed under a Creative Commons Attribution license, we make no representations or warranties regarding the license status of each image and you should verify the license for each image yourself.

As an addition to the original dataset, we added hierarchical labels for landmarks.

Ground-truth and labelmaps released. Note that the ground-truth has been substantially updated since the end of the 2019 Kaggle challenges; it is not the one that was used for scoring in the challenge.

We have re-computed metrics for the top 10 teams in the 2019 challenges (see the Kaggle challenge webpages for precise definitions of the metrics):

Included data for test and index sets.

Initial version, including only train set.

For any questions/suggestions/comments/corrections, please open an issue in this github repository, and tag @andrefaraujo. In particular, we plan to maintain and release new versions of the ground-truth as corrections are found.

Original GLDv2 paper:

@inproceedings{weyand2020GLDv2,
  author = {Weyand, T. and Araujo, A. and Cao, B. and Sim, J.},
  title = {{Google Landmarks Dataset v2 - A Large-Scale Benchmark for Instance-Level Recognition and Retrieval}},
  year = {2020},
  booktitle = {Proc. CVPR},
}

Hierarchical extension:

@inproceedings{ramzi2023optimization,
  author = {Ramzi, E. and Audebert, N. and Rambour, C. and Araujo, A. and Bitot, X. and Thome, N.},
  title = {{Optimization of Rank Losses for Image Retrieval}},
  year = {2023},
  booktitle = {In submission to: IEEE Transactions on Pattern Analysis and Machine Intelligence},
}

The following table is necessary for this dataset to be indexed by search engines such as Google Dataset Search.