OpenHotels

OpenHotels is a large-scale hotel image retrieval benchmark built from hotel-room imagery and associated hotel metadata. The dataset is designed for hotel-scale retrieval: given a query image, a system must retrieve the matching hotel from a large gallery containing both true matching classes and many distractor hotel classes.

Dataset Structure

The release contains tar-sharded image files under shards/ and four metadata files:

shards/
  gallery-00000.tar
  ...
  test_non_object-00000.tar
  ...
  test_object-00000.tar
  ...
metadata_gallery.json
metadata_test_non_object.json
metadata_test_object.json
metadata_hotels.json
croissant.json

The benchmark has three image subsets:

• gallery: searchable reference images for all hotel classes, including distractor-only hotels.
• test_non_object: held-out room-view query images.
• test_object: held-out object-centric query images.

Statistics

Subset	Hotels	Images
Gallery	41,027	253,597
Test Non-Object	15,982	62,706
Test Object	4,707	54,842

The gallery contains 140,247 non-object room-view images and 113,350 object-centric images. Across both test subsets there are 15,982 unique hotel classes; the gallery includes 25,045 distractor hotel classes that do not appear in either test subset.

Metadata

The dataset includes the following metadata fields.

Image Metadata

All image metadata rows include:

• path: image member name inside the tar file listed in shard.
• shard: relative path to the tar shard containing the image.
• hotel_id: stable anonymized hotel class identifier.
• room: room identifier associated with the image upload when available. This is not the semantic room/view label.
• timestamp: upload timestamp associated with the image.

Gallery rows additionally include:

• is_object: whether the gallery image is object-centric.
• view_type: room-level view category for non-object gallery images.
• object_type: localized object category for object-centric gallery images.

Test Non-Object rows additionally include:

• view_type: room-level view category for the query image.

Test Object rows additionally include:

• object_type: localized object category for the query image.

view_type describes the room-level view depicted in a non-object image. Possible values are: bedroom, bathroom, living area, hallway, kitchen, closet, and balcony.

object_type describes the localized hotel-room object depicted in an object-centric image. Possible values are: bed, lamp, artwork, window/curtains, toilet, nightstand, sink, seating, office desk, shower/bathtub, tv, wardrobe, door, chest, mirror, kitchen appliances, flooring, and sign.

Example gallery row:

{
  "path": "images/gallery/23/23297/0011.jpg",
  "shard": "shards/gallery-00000.tar",
  "hotel_id": "23297",
  "room": "244",
  "timestamp": "2024-12-27T04:20:21",
  "is_object": false,
  "view_type": "bedroom"
}

Example Test Non-Object row:

{
  "path": "images/test_non_object/00/000000.jpg",
  "shard": "shards/test_non_object-00000.tar",
  "hotel_id": "03875",
  "room": "405",
  "timestamp": "2016-06-25T06:13:23",
  "view_type": "bedroom"
}

Example Test Object row:

{
  "path": "images/test_object/00/000000.jpg",
  "shard": "shards/test_object-00000.tar",
  "hotel_id": "03875",
  "room": "425",
  "timestamp": "2021-07-28T09:43:57",
  "object_type": "nightstand"
}

Hotel Metadata

Each row in metadata_hotels.json describes one hotel class:

• hotel_id: stable anonymized hotel class identifier.
• name: hotel name.
• lat: hotel latitude in decimal degrees.
• lng: hotel longitude in decimal degrees.
• date_added: timestamp when the hotel record was added.
• in_gallery: whether the hotel appears in the gallery subset.
• in_test_non_object: whether the hotel appears in the Test Non-Object subset.
• in_test_object: whether the hotel appears in the Test Object subset.

{
  "hotel_id": "00000",
  "name": "Extended Stay America - Fairbanks - Old Airport Way",
  "lat": 64.83538,
  "lng": -147.8233,
  "date_added": "2015-06-25T21:34:48",
  "in_gallery": true,
  "in_test_non_object": false,
  "in_test_object": false
}

Loading Images from Shards

The same helper can load gallery images and both query subsets. The metadata row tells the loader which tar shard contains the image and which member path to read from that shard.

import json
import tarfile
from io import BytesIO
from PIL import Image

def load_image(row):
    """Load one OpenHotels image from its tar shard."""
    with tarfile.open(row["shard"], "r") as tar:
        image_file = tar.extractfile(row["path"])
        return Image.open(BytesIO(image_file.read())).convert("RGB")

def load_metadata(path):
    with open(path) as f:
        return json.load(f)

Load a gallery image. Gallery rows contain is_object; non-object gallery rows include view_type, and object-centric gallery rows include object_type.

gallery = load_metadata("metadata_gallery.json")
gallery_row = gallery[0]
gallery_image = load_image(gallery_row)

print(gallery_row["hotel_id"], gallery_row["path"])
if gallery_row["is_object"]:
    print("object type:", gallery_row["object_type"])
else:
    print("view type:", gallery_row["view_type"])

Load a Test Non-Object query image and inspect its room-level view label.

test_non_object = load_metadata("metadata_test_non_object.json")
non_object_row = test_non_object[0]
non_object_image = load_image(non_object_row)

print(non_object_row["hotel_id"], non_object_row["path"])
print("view type:", non_object_row["view_type"])

Load a Test Object query image and inspect its object label.

test_object = load_metadata("metadata_test_object.json")
object_row = test_object[0]
object_image = load_image(object_row)

print(object_row["hotel_id"], object_row["path"])
print("object type:", object_row["object_type"])

Evaluation Protocol

Use the gallery as the retrieval database. Evaluate the two query subsets separately:

• Test Non-Object evaluates retrieval from room-level query images.
• Test Object evaluates retrieval from object-centric query images.

For each query, rank gallery images or aggregate ranked images to hotel-level predictions, then score retrieval against the query hotel_id. The standard metrics are Recall@K for K in {1, 5, 10, 100}.