Dataset Card for RevisitOP (Oxford5k, Paris6k, RevisitOP1M)

Dataset Description

RevisitOP provides popular benchmark datasets for large-scale image retrieval research:

• roxford5k: Oxford 5k buildings dataset containing ~5,000 images.
• rparis6k: Paris 6k buildings dataset with ~6,000 images.
• revisitop1m: RevisitOP 1M distractor dataset with ~1 million distractor images.
• oxfordparis: Combination of Oxford 5k and Paris 6k datasets.

These datasets are widely used for evaluating image retrieval algorithms and contain real-world building photographs and distractors.

Dataset Features

Each example contains:

• image (Image): An image file (JPEG or PNG).
• filename (string): The original filename of the image.
• dataset (string): The source dataset the image belongs to (roxford5k, rparis6k, or revisitop1m).
• query_id (int32): Query ID for query images (-1 for database images).
• bbx (Sequence[float32]): Bounding box coordinates [x1, y1, x2, y2] for query images.
• easy (Sequence[int32]): Easy relevant images for queries.
• hard (Sequence[int32]): Hard relevant images for queries.
• junk (Sequence[int32]): Junk images for queries.

Dataset Splits

• qimlist: Query images with ground truth annotations (bounding boxes and relevance labels).
• imlist: Database images for retrieval.

Dataset Versions

• Version 1.0.0

Example Usage

Use the Hugging Face datasets library to load one of the configs:

import datasets
from aiohttp import ClientTimeout

dataset_name = "randall-lab/revisitop"
timeout_period = 500000  # very long timeout to prevent timeouts
storage_options = {"client_kwargs": {"timeout": ClientTimeout(total=timeout_period)}}

# These are the config names defined in the script
dataset_configs = ["roxford5k", "rparis6k", "oxfordparis"]  # "revisitop1m" is large and may take a long time to load

# Load query split for evaluation
for i, config_name in enumerate(dataset_configs, start=1):
    # Load query images
    query_dataset = datasets.load_dataset(
        path=dataset_name,
        name=config_name,
        split="qimlist",
        trust_remote_code=True,
        storage_options=storage_options,
    )

    # Load database images
    db_dataset = datasets.load_dataset(
        path=dataset_name,  
        name=config_name,
        split="imlist",
        trust_remote_code=True,
        storage_options=storage_options,
    )


    # Example query image
    query_example = query_dataset[0]

Dataset Structure

• The datasets consist of images downloaded and extracted from official URLs hosted by the Oxford Visual Geometry Group and the RevisitOP project.
• The roxford5k and rparis6k datasets come from .tgz archives with corresponding .pkl ground truth files.
• The revisitop1m dataset consists of 100 .tar.gz archives with JPEG images as distractors.
• The combined oxfordparis dataset merges the Oxford and Paris sets.
• Ground truth files contain query lists, database lists, and annotations (bounding boxes, easy/hard/junk labels).

Dataset Citation

If you use this dataset, please cite the original paper:

@inproceedings{Radenovic2018RevisitingOP,
  title={Revisiting Oxford and Paris: Large-Scale Image Retrieval Benchmarking},
  author={Filip Radenovic and Ahmet Iscen and Giorgos Tolias and Yannis Avrithis and Ondrej Chum},
  year={2018}
}

Dataset Homepage

RevisitOP project page