Datasets:
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... |
Glance-COCO
Grounding data on COCO images: one image, many text requests, and every request is answered with one box, several
boxes, or "not present". This repository was called Glance-Data; the old URL redirects here, so
load_dataset("Vision-Decision/Glance-Data") keeps working.
It now also holds the COCO-image datasets that were published separately: the training supplements that fix
Glance-Data's known shortcuts, the trainer-format records, the evaluation sets and the build intermediates. Each one
is a config or a folder here, and its original card is kept unchanged in provenance/.
Ground truth comes from COCO annotations. Language models only write the wording. Images are COCO 2014 train and are not included.
Configs
| config | split(s) | rows | use | formerly |
|---|---|---|---|---|
default |
train, diagnostic |
24,406 / 300 images | the main data: 1,831,146 / 23,578 requests in 14 kinds | Glance-Data |
plus |
train |
57,092 rows, 317,500 requests | train-only supplements that fix shortcuts in default (see below); same schema plus a source column, so it concatenates with default |
Glance-Data-Plus-clean |
diagnostic_pos_sole |
test |
160 records, 764 requests | evaluation only: sole-instance position / superlative positives on the diagnostic images | glance-ground-v2 |
robust |
zh, para, func, vague |
6,236 items | evaluation only: paraphrased, functional, vague and Chinese requests on RefCOCO/gRefCOCO val images | glance-robust-ref-clean |
Trainer-format records (jsonl files)
records/ holds default + pos_pos_sole in the Glance trainer's record format, as plain jsonl (formerly
glance-ground-v2): full/ (51,177 records, ≤48 requests each, every target in boxes, for multi-box heads),
single/ (45,027 records, n_targets ≤ 1, for single-box heads) and diagnostic.jsonl (300 records). They are files,
not configs, because one Hub dataset can only expose one file format:
from datasets import load_dataset
full = load_dataset("json", data_files="hf://datasets/Vision-Decision/Glance-COCO/records/full/*.jsonl", split="train")
What plus fixes
problem in default |
fix | source in plus |
|---|---|---|
| position words ("the leftmost cat") appear only in negatives, so a model learns to reject them (Glance v1 rejected 86.5% of sole-instance position requests) | rule-made positives on the only instance of a category, sampled 1:1 against neg_position |
ground_v2_gap_rule |
neg_count can be recognised from the text alone 98.6% of the time |
"exactly n" positives (n = 2–5), so a count becomes a minimal pair | v1.3_count_pairs |
| almost every request names the category | function, supercategory and relation requests without the category word, with negatives of the same form; blind re-identification check plus a leak lexicon | v1.2_catfree |
| regressions seen in v1.2 | targeted negatives: absent class at a position or in a relation | v1.3_negatives |
plus/manifest/ lists, by SHA-1 only, the RefCOCO expressions that were replayed during training and the Chinese
translations that are not redistributed (licence of RefCOCO-family text is unclear).
Folders
| folder | contents |
|---|---|
data/ |
default parquet files (unchanged paths) |
plus/ |
plus parquet files, manifest/, stats.json |
records/ |
trainer-format jsonl: full/, single/, diagnostic.jsonl, stats.json |
eval/ |
the original diagnostic_pos_sole.jsonl and robust/ jsonl files (with the withheld/ reference list and stats.json); parquet/ holds row-for-row parquet copies that back the diagnostic_pos_sole and robust configs |
intermediates/ |
build intermediates: rule-made scene specs, raw and filtered category-free generations (catfree_v12/), v1.3 negatives and count positives, decontamination id and hash lists, the Glance-X data preparation script and provenance |
provenance/ |
the original cards of the five merged repositories, unchanged |
Decontamination
All training images are COCO train2014. No training image appears in the RefCOCO/+/g val or test splits, the
gRefCOCO val/testA/testB splits, COCO val2017, the diagnostic split or the robust set. The robust and
diagnostic* configs are for evaluation only: do not train on them.
Licence
CC BY-NC-SA 4.0, for non-commercial research only (inherited from gRefCOCO). Images are not included; COCO
annotations are CC BY 4.0. Per-part provenance and licences are in the original cards under provenance/.
Pipeline code: Vision-Decision/Glance-Data on GitHub.
- Downloads last month
- 6