Datasets:
File size: 2,421 Bytes
a8c3e93 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 | ---
license: cc-by-4.0
task_categories:
- visual-question-answering
language:
- en
tags:
- vision-language
- compositionality
- action-object-binding
- benchmark
pretty_name: GameBind
size_categories:
- 1K<n<10K
---
# GameBind
Data for **GameBind**, a controlled benchmark that measures whether a vision-language
model binds an action to the correct object. It separates three abilities a single
accuracy number blends together: recognizing the object, recognizing the action, and
binding the action to the correct object, and reports the binding-gap diagnostic
`min(Acc_object, Acc_action) - Acc_binding`.
This is an anonymous release accompanying a paper under review.
## Subsets
- **gamebind/** — main 2D release: 1000 procedurally generated BabyAI- and
Crafter-style clips, 9668 questions. Includes the rendered keyframes and clips we
generated, plus `clips.json`, `qa.json`, `summary.json`.
- **gamebind_difficulty/** — distractor-count sweep (single keyframe), matched
multiple-choice probes.
- **gamebind_difficulty_video/** — the same sweep rendered as short videos, so the
action is depicted.
- **gamebind_real/** — real-photograph probe built from V-COCO human-object
interactions. **Metadata only** (`clips.json`, `qa.json`): each clip references a
COCO 2014 `image_id`. Fetch the images from COCO
(`http://images.cocodataset.org/val2014/COCO_val2014_{image_id:012d}.jpg`); they are
not redistributed here.
- **gamebind_alfred/** — ALFRED 3D keyframe pilot. **Metadata only**: images come
from the ALFRED dataset and are not redistributed here.
## Question format
Each entry in `qa.json` has a `qa_type` (`object`, `action`, `binding`, `object_mc`,
`action_mc`, `binding_mc`, `minimal_pair`, `distractor`, `perception`), the question,
the gold `answer`, and (for multiple choice) the `choices`. Clips in `clips.json`
carry the domain, action, object, distractors, and frame or video paths.
## Licensing and attribution
The GameBind annotations and the synthetic BabyAI/Crafter-style renders in this
repository are released under CC-BY-4.0. The `gamebind_real` subset contains only
annotations that reference COCO images (via V-COCO); the COCO photographs themselves
are not included and remain under their original licenses. The `gamebind_alfred`
subset contains only metadata derived from the ALFRED dataset. Please cite COCO,
V-COCO, and ALFRED if you use those subsets.
|