By requesting access, you agree not to use OpenHotels-Updated or its associated models to identify, locate, track, profile, or target specific individuals; infer sensitive personal information; facilitate surveillance, harassment, discrimination, exploitation, or other harm; reconstruct or expose information that has been removed, obscured, or anonymized; or create derivative resources that materially increase the risk of misuse without appropriate ethical and safety review.

You also agree not to redistribute the resources or access credentials to individuals who have not separately received access. You understand that proposed uses in law enforcement, immigration enforcement, military, intelligence, or commercial surveillance contexts warrant particular scrutiny and should undergo appropriate independent ethical, legal, and institutional review before proceeding.

OpenHotels2

OpenHotels2 is a benchmark for identifying hotels from images. It contains 474,322 images from 40,974 hotels, with a gallery and one test set for single-image and multi-image retrieval. Each image includes room-type labels, object labels with bounding boxes, a timestamp, and hotel and visit membership.

Split	Images	Hotels	Visits
gallery	464,322	40,974	96,185
test	10,000	4,613	5,156

Every test hotel has images in the gallery, which also contains distractor hotels. There is no training split. The full download is approximately 254 GB in 257 TAR shards containing upright JPEG and PNG images.

Download and use

Request access on the dataset page, then install the dependencies and sign in:

pip install huggingface_hub Pillow
hf auth login

Download the release:

from huggingface_hub import snapshot_download

snapshot_download("imagingforgood/OpenHotels2", repo_type="dataset",
                  local_dir="OpenHotels2")

Run this from the downloaded OpenHotels2 directory:

from load_openhotels import OpenHotels2

dataset = OpenHotels2(".")
hotel_id, visit_id, path, annotation = next(dataset.iter_images("test"))
image = dataset.load_image(path)  # returns a Pillow image
print(hotel_id, annotation["room_labels"], annotation["objects"])

# Use dataset.iter_images("gallery") to iterate the reference images.

The loader reads images directly from the TAR shards; extraction is unnecessary. Loading the metadata and index requires several GB of RAM. To download only test images and their metadata, add allow_patterns=["*.json", "*.py", "README.md", "shards/test-*.tar"] to snapshot_download. Retrieval evaluation requires the gallery as well.

Annotations and files

dataset.json: metadata grouped as hotels[hotel_id].visits[visit_id].images[image_path], including hotel names, coordinates, and image/visit counts.
shards/: image archives; shard_index.json maps image paths to their archive locations.
load_openhotels.py: the Python loader used above (Python 3.10+).
release_summary.json: dataset counts and label distributions. Additional manifests, checksums, and validation results describe the packaged files.

Each image has timestamp, width, height, room_labels, and objects. objects maps each label to a list of [x_min, y_min, x_max, y_max] boxes in pixel coordinates. Images are already upright, and dimensions and boxes refer to those pixels. Multiple room labels are possible, including ambiguous.

Use the visit's split field for gallery/test membership; path names may contain legacy split names. Observed room types are in room_labels.

Evaluation tasks

Single-image hotel retrieval: use each test image as a query and rank gallery images. A result matches when its hotel ID equals the query's hotel ID. Report Recall@K: the fraction of queries with at least one matching image in the top K.
Multi-image hotel retrieval: combine two or more test images from the same hotel or visit into a query set, then retrieve from the gallery. Report the grouping, query-set sizes, aggregation method, and set-level Recall@K.
Performance by scene or object: break down retrieval results using room_labels and object labels to compare performance across image content.

Evaluate on the provided test split, and report the dataset revision and retrieval protocol. Bounding boxes support object-focused analysis, but annotations should not be assumed exhaustive for detection or instance-count evaluation.

License and access

Released under CC0-1.0. Access requires accepting the intended-use terms on the dataset page.

Downloads last month: -

Total file size:

255 GB