Dataset Viewer
Auto-converted to Parquet Duplicate
image_id
int64
9
582k
file
stringlengths
31
31
width
int32
128
640
height
int32
104
640
questions
listlengths
6
177
9
COCO_train2014_000000000009.jpg
640
480
[ { "text": "pink fruit bowl", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 311.7300109863281, 4.309999942779541, 631.010009765625, 232.99000549316406 ] ] }, { "text": "pink container ...
25
COCO_train2014_000000000025.jpg
640
426
[ { "text": "tall giraffe", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 385.5299987792969, 60.029998779296875, 600.5, 357.19000244140625 ] ] }, { "text": "tall giraffe eating from the...
49
COCO_train2014_000000000049.jpg
381
500
[ { "text": "black horse", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 162.57000732421875, 226.55999755859375, 292.9800109863281, 410.989990234375 ] ] }, { "text": "black horse rearin...
72
COCO_train2014_000000000072.jpg
427
640
[ { "text": "right giraffe", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 136.6300048828125, 129.44000244140625, 425.7099914550781, 628.489990234375 ] ] }, { "text": "right giraffe wit...
110
COCO_train2014_000000000110.jpg
640
480
[ { "text": "woman in black", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 0, 0, 130.27999877929688, 174.42999267578125 ] ] }, { "text": "woman in black shirt at left corner, back turn...
113
COCO_train2014_000000000113.jpg
416
640
[ { "text": "man on left", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 4.329999923706055, 31.420000076293945, 156.02999877929688, 486.5 ] ] }, { "text": "man in striped shirt standing...
127
COCO_train2014_000000000127.jpg
640
481
[ { "text": "wooden table", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 107.5199966430664, 147.1300048828125, 639.4500122070312, 481 ] ] }, { "text": "round wooden table with food in ...
144
COCO_train2014_000000000144.jpg
640
480
[ { "text": "left giraffe", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 48.540000915527344, 79.81999969482422, 442.25, 480 ] ] }, { "text": "tall giraffe on the left by the trees", ...
165
COCO_train2014_000000000165.jpg
640
536
[ { "text": "man on the left", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 6.019999980926514, 24.09000015258789, 360.1400146484375, 528.77001953125 ] ] }, { "text": "man in a suit hol...
201
COCO_train2014_000000000201.jpg
640
428
[ { "text": "rightmost snowboard", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 333.739990234375, 31.739999771118164, 488.5899963378906, 394.3399963378906 ] ] }, { "text": "black snowb...
247
COCO_train2014_000000000247.jpg
640
424
[ { "text": "blue and white airplane", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 2.859999895095825, 0.47999998927116394, 640, 418.75 ] ] }, { "text": "blue and white propeller plane...
260
COCO_train2014_000000000260.jpg
500
333
[ { "text": "woman in yellow", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 261, 139.8800048828125, 312.75, 309.3800048828125 ] ] }, { "text": "woman in yellow vest facing away", "...
263
COCO_train2014_000000000263.jpg
591
640
[ { "text": "front elephant", "lang": "en", "kind": "short", "src": "aug_llm", "present": true, "n_targets": 1, "boxes": [ [ 156.41000366210938, 169.3300018310547, 591, 632.8300170898438 ] ] }, { "text": "smaller elephant walking in f...
End of preview. Expand in Data Studio

Glance-Data

Grounding data for a Jev-style task. Each example is one image plus many text requests. Every request is answered in parallel with one box, several boxes, or "not present".

This dataset augments the RefCOCO-family training data. It adds the kinds of requests that the original data lacks:

  • a ladder of specificity (short / medium / long) and casual phrasing, all in English and Chinese;
  • category-only requests ("the tie") and all-instance requests ("all the zebras");
  • multi-target requests and descriptions by use ("something to sit on");
  • hard negatives: the same category is present, but a detail is wrong ("a black cow" when every cow is white-brown); count, position and near-category negatives.

Ground truth comes from COCO annotations. An LLM only writes the wording. Boxes, counts and "is this category in the image" are computed from the annotations. DeepSeek (deepseek-flash, vision) looked at each image and wrote the sentences for numbered objects. Rule filters remove sentences that leak the drawing and "attribute" negatives that changed the position instead.

Splits

split images questions images from
train 24,406 1,831,146 RefCOCO-family training images, with every benchmark image removed
diagnostic 300 23,578 refcoco_val images, never used for training

No image appears in both splits. No train image appears in any RefCOCO / RefCOCO+ / RefCOCOg / gRefCOCO evaluation set (6,549 images held out).

Fields

One row per image:

field type meaning
image_id int COCO image id
file string COCO_train2014_<id>.jpg (COCO 2014 train images; not included here)
width, height int original image size
questions list requests for this image

Each question:

field meaning
text the request
lang en / zh
kind see below
src aug_llm (LLM wording), aug_function (LLM names the categories, code picks every instance), aug_code (fully rule-made), aug_gref (human gRefCOCO multi-target sentence)
present whether the request matches something in the image
n_targets 0, 1 or more
boxes every target as [x0, y0, x1, y1] in pixels of the original image; empty when present is false
kind answer ground truth
short, medium, long, casual 1 box the object the LLM was asked to describe
group ≥2 boxes the objects the LLM was asked to describe together
function ≥1 box every instance of the COCO categories the LLM named for that use
cat_only / cat_all 1 / all boxes COCO annotations (category with one instance / all instances)
gref_multi ≥2 boxes human gRefCOCO annotation
hneg_attribute, hneg_relation not present LLM edit of one attribute or action, position words kept identical
neg_count, neg_position, neg_near not present COCO annotations (more than present / no instance on that side / absent near category)

Quality

A stratified human audit judged 8 random items per kind, 112 in total: 101/112 (≈90%) correct. Every kind scored ≥6/8 except category-only (5/8), which was then fixed by skipping slivers under 0.5% of the image. Hard negatives scored 16/16. The most common remaining error is a short description that is not unique (for example "green tie" when there are two green ties).

Usage

from datasets import load_dataset
ds = load_dataset("Vision-Decision/Glance-Data")
row = ds["train"][0]
for q in row["questions"][:5]:
    print(q["kind"], q["lang"], q["present"], q["text"], q["boxes"])

Images come from COCO 2014 train (file). A model with a single-box head can use one target per question. Using group, cat_all, function and gref_multi fully needs a multi-box head.

License and sources

  • Images: COCO (not redistributed here); COCO annotations are CC BY 4.0.
  • Referring expressions: RefCOCO / RefCOCO+ / RefCOCOg (research use) and gRefCOCO, which is CC BY-NC-SA 4.0 and released for non-commercial research purposes only. Its ShareAlike clause carries over, so this dataset is released under CC BY-NC-SA 4.0, for non-commercial research only.
  • Please cite gRefCOCO (GRES, CVPR 2023), RefCOCO / RefCOCO+ / RefCOCOg and COCO when using the data.
  • Generated text: written by DeepSeek (deepseek-flash) from the images.

Pipeline code: Vision-Decision/Glance-Data on GitHub.

Downloads last month
6

Models trained or fine-tuned on Vision-Decision/Glance-Data