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.

DEAR is released for non-commercial research purposes only. The images remain the property of their respective copyright holders. By requesting access, you agree to use the dataset solely for non-commercial research, not to redistribute the images, and to remove any content upon request from the authors or a rights holder.

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

DEAR: Dense Entity-Attribute Retrieval

DEAR is the benchmark introduced in SAGE: Semantic Attribute Graphs for Multi-Entity Visual Retrieval (EMNLP 2026). It evaluates retrieval and question answering over dense product images: very tall, full-page captures of product detail pages and catalogs, each packing dozens of attributes into a single image. A question names one or two products, and answering it requires finding the right image(s) and the right attribute region(s) inside them.

See the paper for how the benchmark was constructed.

Contents

File Description
images/<category>/ 502 full-page images (PNG/JPEG), about 2.1 GB
images.jsonl One record per image file
qa.jsonl 1,055 question–answer pairs

Statistics

Category Images QA pairs
Appliances 91 126
Electronics 113 140
Headphone 36 72
Laptop 49 48
Phone 44 354
TV 37 113
Video game 15 14
Automobile 117 188
Total 502 1,055
QA type Entities Attributes Count
A single single 212
B single multiple 313
C multiple single 173
D multiple multiple 357

525 questions refer to one image and 530 refer to two. 366 of the images are the target of at least one question, and the rest serve as distractors in the retrieval corpus.

Counts follow the paper. They include 41 duplicated Phone QA rows (1,014 distinct qa_id values) and 11 Automobile images stored as both .png and .jpg under the same image_id (491 distinct images).

Usage

import json, os
from huggingface_hub import snapshot_download
from PIL import Image

root = snapshot_download("All4Nothing/DEAR", repo_type="dataset")

images = [json.loads(l) for l in open(os.path.join(root, "images.jsonl"))]
qa = [json.loads(l) for l in open(os.path.join(root, "qa.jsonl"))]

img = Image.open(os.path.join(root, images[0]["file_name"]))

Fields

images.jsonl: image_id, file_name (path relative to the dataset root), category, source (bestbuy or autocatalogarchive).

qa.jsonl: qa_id, category, qa_type (A–D), question, answer, image_ids (ground-truth images for retrieval).

Source and license

The images are captures of publicly accessible pages from Best Buy (https://www.bestbuy.com) and Auto Catalog Archive (https://autocatalogarchive.com). All images, product names, logos and page content remain the property of their respective owners. They are redistributed here solely for non-commercial research. The annotations (questions and answers) are released under CC BY-NC 4.0.

If you are a rights holder and would like any content removed, contact sd081104@korea.ac.kr or open a discussion in the Community tab, and it will be taken down.

Citation

@inproceedings{kim2026sage,
  title     = {SAGE: Semantic Attribute Graphs for Multi-Entity Visual Retrieval},
  author    = {Kim, Yongjoo and Kwon, Mincheol and Choi, Seonga and Lee, Minseung and Oh, Kyeong-Jin and Lee, Hyunyoung and Choi, Yunsu and Lee, Jungbeom},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing},
  year      = {2026}
}
Downloads last month
-

Paper for All4Nothing/DEAR