harpreetsahota/disc21-image-link-lab

Dataset Card for DISC21 Image Link Lab subset

A 6,000-image subset of DISC21, the benchmark behind the NeurIPS 2021 Image Similarity Challenge, prepared as the demo dataset for the Image Link Lab FiftyOne plugin. Every signal the plugin needs is already computed, so you can go straight to tuning a copy-detection rule and reading the evidence behind each link.

Try it without installing anything: the Image Link Lab Space.

Installation

pip install -U fiftyone
fiftyone plugins download https://github.com/harpreetsahota204/image_link_lab
fiftyone plugins requirements @harpreetsahota/image_link_lab --install

Usage

import fiftyone as fo
from fiftyone.utils.huggingface import load_from_hub

dataset = load_from_hub("harpreetsahota/disc21-image-link-lab", persistent=True)
session = fo.launch_app(dataset)
# Add the "Image Link Lab" panel (the + next to Samples), and move the sliders

Dataset Details

Dataset Description

1,000 query images (tag final_query) and 5,000 reference images (tag reference) drawn from the Kaggle re-host of DISC21. Queries are edited copies of reference images or distractors with no source in the set; 500 of the 1,000 queries have their true original among the 5,000 references, the other 500 are distractors. The edits are DISC21's deliberately hard ones: crops, flips, rotations, blur, re-compression, text and emoji overlays, collages, filters, adversarial perturbations.

On top of the DISC21 fields, the dataset carries the Image Link Lab plugin's outputs:

• a 64-bit pHash per image, also registered as a native similarity index (brain key phash)

• a CLIP similarity index (clip-vit-base32-torch, brain key clip)

• each query's candidate originals (the 5 nearest by pHash, the 5 nearest by CLIP, and the true original) with both signals' values

• the result of the starting rule, pHash ≤ 10 bits or CLIP ≥ 0.90, as a copy_of classification

• four evaluation runs, one per rule, scored against the known originals

• Curated by: Harpreet Sahota (subset and plugin fields); DISC21 by Matthijs Douze et al. (Facebook AI, CTU Prague, TU Munich)

• License: unknown. The Kaggle listing gives the license as "Unknown". The DISC21 paper states the reference images are YFCC100M images chosen under permissive Creative Commons license types; Flickr attribution (ids, usernames, URLs) is kept on every image that has it.

Dataset Sources

• Parent dataset: Voxel51/disc21 (200,000 images, FiftyOne format), itself from the Kaggle re-host of the official release
• Paper: The 2021 Image Similarity Dataset and Challenge
• Plugin: harpreetsahota204/image_link_lab

Uses

Direct Use

Copy detection and near-duplicate detection with explainable links: for each query, decide which reference it was copied from, using a perceptual hash (pixels) and an embedding (meaning) combined by a rule you control, then score the rule against the known originals. The subset is sized for interactive work: candidate gathering takes about 4 s, re-scoring a rule about 0.5 s.

Out-of-Scope Use

This is 0.5% of DISC21's reference set, so numbers here are not comparable with published ISC2021 results. The transformations are stronger than typical production edits; recall is low for any single signal by design.

Dataset Structure

Dataset type

A flat FiftyOne image dataset (media_type="image") with 6,000 JPEG samples. Subsets are sample tags; samples are not grouped.

Tag	Samples	Description
final_query	1,000	DISC21 Phase II queries: 500 edited copies whose original is in this set, 500 distractors
reference	5,000	Unmodified reference images, including the 500 true originals

Sample fields

From DISC21:

Field	FiftyOne type	Description
disc21_id	StringField	Image id, e.g. Q50029, R003513. Unique
is_distractor	BooleanField	Queries only. True if the query has no source in DISC21's reference set
reference_id	StringField	Queries only. Id of the reference image the query was made from; None for distractors
reference_available	BooleanField	Queries only. Whether that reference image is in this subset (True for all 500 non-distractors)
matching_queries	ListField(StringField)	References only. Ids of the queries made from this image
transform_names	ListField(StringField)	Queries with a source only. Names of the edits applied, in order (30 distinct names)
transforms	ListField	The full edit records with their parameters, verbatim from DISC21 metadata
flickr_ids, flickr_usernames, flickr_urls	ListField(StringField)	Flickr attribution of the source image(s)

From the Image Link Lab plugin:

Field	FiftyOne type	Description
phash	StringField	64-bit perceptual hash as 16 hex characters
copy_of_candidates	ListField(DictField)	Queries only. Up to 11 candidate originals, each {sample_id, original, phash, clip, is_truth}: differing hash bits, CLIP cosine similarity, and whether it is the true original
copy_of	Classification	Queries only. The original the starting rule linked the query to (label = its disc21_id) or none; attributes original_id, phash, clip, confidence
copy_of_gt	Classification	Queries only. copy if the query has a known original, unique otherwise. Ground truth for the evaluations
copy_of_pred	Classification	Queries only. copy if the rule linked the query to its true original, unique otherwise. Prediction for the evaluations
copies_phash10, copies_clip90, copies_phash10_and_clip90, copies_phash10_or_clip90	StringField	Per-query outcome (TP, FP, FN, TN) of each evaluation run

Label types and why

copy_of is a Classification so the link shows up as an ordinary label in the App and can be filtered on, and so that copy_of_pred versus copy_of_gt can be scored with FiftyOne's native evaluate_classifications(method="binary"). Candidates are plain dicts because they are evidence, not labels: the panel reads them to redraw the graph as the sliders move, with no server round-trip.

Runs on the dataset

Kind	Key	What it is
Brain (similarity)	phash	pHash bits as a similarity index (euclidean, so it sorts like Hamming distance)
Brain (similarity)	clip	CLIP ViT-B/32 embeddings, cosine
Evaluation	copies_phash10	pHash ≤ 10 bits: precision 99%, recall 17%
Evaluation	copies_clip90	CLIP ≥ 0.90: precision 97%, recall 7%
Evaluation	copies_phash10_and_clip90	both must pass: precision 100%, recall 4%
Evaluation	copies_phash10_or_clip90	either passes (the starting rule): precision 98%, recall 20%
Saved view	queries	the 1,000 query images

Evaluations are per query: a copy counts as found only when linked to its true original; a distractor linked to anything is a false positive.

dataset.info contents

Key	Description
source, paper, subset_folders	Carried over from the parent dataset
image_link_lab	The plugin's settings: query scope, originals scope, ID and truth fields, output field, and the last rule applied. The panel reads these, so the dataset works in the plugin as soon as it is loaded

Parsing decisions

• The 1,000 queries are the first 1,000 Phase II queries whose true original, if any, is among the 5,000 references chosen; the 500 distractors have no original anywhere in DISC21's reference set.
• Candidate lists always include the true original when known, even when it is not among the 10 nearest, so misses can be drawn and explained.
• Signals were computed with the plugin's compute_signals operator; candidates, copy_of and the evaluations with find_copies.

Dataset Creation

Curation Rationale

A subset small enough to explore interactively on a laptop and in a free Hugging Face Space, but with enough copies and distractors to show where a copy-detection rule succeeds and fails, and to make the trade-off between precision and recall visible.

Source Data

See the parent card, Voxel51/disc21: YFCC100M images under permissive Creative Commons licenses plus DFDC frames of consenting paid actors, edited manually in GIMP or automatically with AugLy.

Annotations

Ground-truth links come from DISC21's generation process: each query was made from a known reference image. The plugin's fields are derived, not annotated.

Personal and Sensitive Information

As in the parent dataset: images were selected without identifiable humans except DFDC frames of paid actors. Attribution fields contain Flickr usernames and photo URLs of the creators.

Citation

@article{douze2021disc21,
  title={The 2021 Image Similarity Dataset and Challenge},
  author={Douze, Matthijs and Tolias, Giorgos and Pizzi, Ed and Papakipos, Zo{\"e} and Chanussot, Lowik and Radenovic, Filip and Jenicek, Tomas and Maximov, Maxim and Leal-Taix{\'e}, Laura and Elezi, Ismail and Chum, Ond{\v r}ej and Canton Ferrer, Cristian},
  journal={arXiv preprint arXiv:2106.09672},
  year={2021}
}

Dataset Card Authors

Harpreet Sahota