You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

By requesting access, you agree not to use OpenHotels-DMLR 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.

Log in or Sign Up to review the conditions and access this dataset content.

OpenHotels-DMLR

OpenHotels-DMLR is a benchmark for retrieving matching hotels from photographs. It contains 420,682 hotel images from 37,728 hotels, with room-type labels, object bounding boxes, and public hotel-ID ground truth. The gallery provides reference images; the test split provides 10,000 query images.

Split Images Hotels TAR shards
gallery 410,682 37,728 211
test 10,000 8,962 6

Every test hotel is represented in the gallery. The gallery also contains 28,766 hotels without test queries, which serve as distractors. There is no training split. The full package is approximately 217.5 GB, including about 216 GB of uncompressed image-and-annotation shards.

Repository: imagingforgood/OpenHotels-DMLR.

Quick start: stream from Hugging Face

Requires Python 3.11 or later. Request access on the dataset page and accept the usage expectations. Access requests are approved automatically. Then install the dependencies and sign in with the account that requested access:

python -m pip install "datasets>=5.2.0,<6" Pillow
hf auth login
from datasets import load_dataset

queries = load_dataset(
    "imagingforgood/OpenHotels-DMLR", split="test", streaming=True
)
sample = next(iter(queries))
image = sample["jpg"]                  # PIL image, already upright
metadata = sample["json"]              # seven-field annotation dictionary
image_path = sample["__key__"] + ".jpg"
print(image_path, metadata["hotel_id"], metadata["room_labels"])

Use split="gallery" for the reference collection. To read a downloaded copy, replace the repository ID with its local directory. For reproducible evaluations, pass revision="<commit hash>" to load_dataset and record that revision. Authenticated access is required for Hub downloads and streaming.

The card declares a JSON feature to preserve object-label dictionaries whose keys vary between images. In Datasets 5.2.0, its JSON conversion rounds floating point values to about ten decimal places; the measured maximum coordinate change in the loader checks was about 5e-11 pixels. The archived JSON retains the original precision. Use WebDataset or the included reader for exact values.

Download all or part of the dataset

from huggingface_hub import snapshot_download

snapshot_download(
    "imagingforgood/OpenHotels-DMLR",
    repo_type="dataset",
    local_dir="OpenHotels-DMLR",
    # revision="<commit hash>",  # Pin the same revision used for evaluation.
)

To download only the test images and the files needed to inspect them, add:

allow_patterns = [
    "README.md", "load_openhotels.py", "examples/*",
    "test.json", "test_image_mapping.json", "hotel_mapping.json",
    "shard_index.json", "shard_manifest.json", "shards/test-*.tar",
]
# Pass allow_patterns=allow_patterns to snapshot_download(...).

Retrieval evaluation requires the gallery as well. Full-package validation requires every file listed in shard_manifest.json. The partial-download example is for query inspection and streaming; it is not a complete benchmark copy.

Data format

Each sample contains exactly seven metadata fields:

Field Type Meaning
hotel_id string Shared five-digit hotel label, including leading zeros
room string Submitted room identifier; distinct from a semantic room type
timestamp string Upload timestamp preserved as recorded, not verified capture time
width, height integer Dimensions of the saved upright JPEG, in pixels
room_labels list of strings Semantic room-type annotations
objects dictionary Object label to a list of [x_min, y_min, x_max, y_max] boxes

Bounding boxes use floating-point pixel coordinates in the original saved image. Images already have their orientation baked into their pixels, with EXIF Orientation removed. Do not apply another orientation transform. Resizing or cropping an image requires transforming its boxes if they are used with the transformed image. Empty object dictionaries are valid annotation results. Inherited boxes can extend up to one pixel beyond the image bounds.

Room strings preserve case, whitespace, and leading zeros. They are submitted identifiers rather than independently verified room identities. Timestamps are preserved without inferring local time zones.

Identifiers and paths

Hotel IDs are strings from 00000 through 37727, shared across both splits. Gallery image IDs start at 0000 within each hotel; test image IDs run globally from 0000 through 9999. All image IDs use four digits. Directory groups contain every 1,000 hotel IDs in the gallery and every 1,000 image IDs in the test set.

Each JPEG is immediately followed by a JSON member with the same stem:

images/gallery/00/00000/0000.jpg
images/gallery/00/00000/0000.json
...
images/test/0/0000.jpg
images/test/0/0000.json

These are paths inside the TAR archives. Extraction into one directory restores the paths referenced by the JSON files; extraction is optional. WebDataset's __key__ is the member path without its extension. __url__ identifies the shard.

Accompanying files

File Contents
gallery.json hotels[hotel_id].visits[visit_id].images[image_path]; hotel names, coordinates, and visit context remain here
test.json Flat image_path -> metadata dictionary with the seven fields above
hotel_mapping.json Release hotel ID to original TraffickCam hotel ID
gallery_image_mapping.json, test_image_mapping.json Image path to release/TCAM IDs, original source path, size, and image SHA-256
test_iamge_mapping.json Identical compatibility copy of test_image_mapping.json
shard_index.json Image path to shard, JPEG/JSON byte offsets, and payload sizes
shard_manifest.json Shard memberships, sizes, SHA-256 hashes, counts, and accompanying file checksums
load_openhotels.py, validate_shards.py Standalone reader and full-package validator
examples/ Runnable streaming, image-access, batching, and evaluation examples

Gallery image records contain the five image annotation fields; their hotel ID and room are supplied by the enclosing hotel and visit. Shard JSON records flatten that context into the same seven fields as test.json. Hotel names and coordinates are available through gallery.json and are not duplicated in test records. Test visit identifiers and date ranges are not part of the public test records. Original source paths in mappings are provenance references, not download URLs.

There are 211 gallery and 6 test shards, totaling 216,038,348,800 bytes. Each uncompressed TAR is at most 1,000,000,000 bytes. Shards preserve the ascending padded-ID order. Image/JSON pairs stay together; a hotel can span shards. No resizing or recompression occurs during packaging. Sequential streaming does not require loading gallery.json or the full index into RAM. Random access loads the index; loading the nested gallery JSON requires more memory.

Local usage examples

Run these from a downloaded release directory:

python -m pip install -r examples/requirements.txt
python -B examples/stream_huggingface.py --split test --limit 3
python -B examples/stream_webdataset.py --split gallery --limit 3
python -B examples/read_image.py --image-path images/test/0/0000.jpg

The standard-library reader also works directly:

from load_openhotels import OpenHotels, iter_samples

dataset = OpenHotels(".")
image_bytes, metadata = dataset.read("images/test/0/0000.jpg")
path, image_bytes, metadata = next(iter(iter_samples(".", "gallery")))

For image batches, install PyTorch and run:

python -m pip install torch
python -B examples/pytorch_dataloader.py --split test --workers 2 --batches 2

The example uses CUDA, Apple MPS, or CPU when available and keeps string IDs and variable-length annotations alongside the batch. Its image preprocessing is an illustration; use the preprocessing required by your embedding model. See the examples guide for details and evaluation commands.

Evaluation protocol

For single-image hotel retrieval, rank gallery images for each test query. A retrieved image is a positive when its hotel_id equals the query's hotel_id. Recall@K is the fraction of test queries with at least one positive image among the first K results. The unit of ranking is a gallery image; do not silently collapse repeated hotels into a hotel ranking.

The evaluation example accepts a JSON dictionary from each test image path to its ranked gallery image paths. It checks complete query coverage, valid gallery membership, and duplicate image results before calculating Recall@K. Report the dataset revision, model/preprocessing, ranking method, and K values. There are no baseline model scores in this release package.

Annotations can support subgroup analyses by room type or object presence. Report the query counts for each subgroup. Multi-image query experiments should state their grouping rule and aggregation method; public test records support hotel-level grouping, but do not expose source visit membership.

Source and construction

The source is hotel photographs submitted through the TraffickCam application before January 1, 2026. Hotel assignments use the resolved canonical server hotel identities from the 2026 export. Existing room and object annotations were matched through source image identity, then retained with upright image pixels. The construction did not regenerate those annotations.

Images without usable room identifiers or completed annotation records were excluded, as were images on the supplied people-detection list. Subsequent reviews removed flagged locations/non-hotel content, reviewed cross-hotel matches, and identified rental properties. Review coverage is incomplete, so these exclusions do not certify every remaining image or property.

Original source visits group the same hotel and exact room string within a seven-day window anchored at the first upload. For splitting, same-hotel visit intervals are merged transitively whenever the gap is less than 30 days, regardless of room string. Entire temporal groups are held out from the gallery when any of their images are selected as test queries. Every test hotel retains same-hotel references separated by at least 30 days in upload time.

Query selection uses prior human assessments, historical image-similarity scores, and duplicate screening, with seed 20261104; it is not uniform random sampling. The historical scores were not recomputed against the final gallery. The 52,831 unselected images from held-out temporal groups are excluded from this release and must not be added to the evaluation gallery. The split was frozen on November 4, 2026.

Inherited bounding boxes were processed with greedy nonmaximum suppression within each image and exact object label, suppressing at IoU >= 0.95. Larger boxes have priority, with original annotation order breaking ties. Remaining coordinates and annotation structures are preserved.

Intended use and limitations

This dataset supports research on visual hotel retrieval, image representations, and analyses of room and object content. Gallery/test hotel overlap is deliberate: the task is to identify which reference hotel matches a query.

  • The release is an evaluation benchmark. Temporal separation does not establish separation from external training datasets or earlier OpenHotels releases. Do not use test labels for model fitting or parameter selection when reporting held-out retrieval performance.
  • Coverage follows contributor submissions and available annotations. Geography, hotel identity, room types, and object categories are unevenly represented. In the frozen test set, 92.84% of images are assigned to U.S. hotels using hotel coordinates and Natural Earth country polygons, including nearest-land fallback.
  • Object annotations are not guaranteed exhaustive; missing boxes do not establish object absence. Rare categories support only limited subgroup evaluation.
  • Upload timestamps are not capture dates, submitted room identifiers can be noisy, and hotel information reflects the source export rather than current property status. Duplicate screening reduces reviewed leakage without proving that every near-duplicate or incorrect hotel assignment has been removed.
  • Hotel matches are dataset labels. Visual similarity alone does not establish an individual's identity, presence, or movements.

Integrity checks

For a complete local package, run:

python -B validate_shards.py --workers 4

This reads every archived image, checks its SHA-256 against the image mappings, compares paired annotations with the public JSON, verifies TAR offsets and archive checksums, and checks exact split membership and padded numbering. It also checks the accompanying files against the manifest. The release is self-contained and requires no construction inputs or local audit files. The Hub-managed .gitattributes file, local Python bytecode, and Hugging Face download caches are ignored by the inventory check.

The package was checked with Hugging Face Datasets 5.2.0 and WebDataset 1.0.2. Both loaders were exercised across every gallery shard and all 10,000 test images.

Downloads last month
4,278