README.md

metadata

license: mit
size_categories:
  - 100M<n<1B
task_categories:
  - image-text-to-text
  - object-detection
language:
  - en
tags:
  - art

AmbiBench Metadata

This dataset provides the metadata used in the AmbiBench benchmark.
Each row corresponds to an ambiguous image or video sample paired with a question–answer pair, derived from the benchmark design.

---

Columns

question

• The natural-language question posed to the model.

answer

• The reference answer to the question.
• May include:
  • A single object name ("snake"),
  • A semicolon-separated pair for bistable/multi-scene ("duck; rabbit"),
  • Or multiple objects/scenes ("bear, deer, wolf, bird").

file_name

• Cleaned filename of the associated image or video.
• Any inline bounding box tags in the original filename (e.g.
  2120_netimg_hidden_face_illusion_[1152,807,1392,1263].jpg)
  are removed, and the coordinates are stored separately in bbox.
• ⚠️ Special case: motion_video samples — the stored file names end with .png as placeholders.
  For actual usage, replace .png with .mp4 to access the corresponding video file.

bbox

• A list of bounding box coordinates [x1, y1, x2, y2] if required by the question.
• Empty list ([]) if no bounding box is provided.

type

• Combined label = Ambiguity category (from the AmbiBench taxonomy) + Task type.
• Categories include:
  Camouflaged, Bistable, Hybrid, Color, Multi-view, Geometric, Motion, Multi-scene, Mixed.
• Task types include: open-ended, localization(bbox), region.
• Example values:
  • "camouflaged_open"
  • "mix_bbox"
  • "mix_highlighted"

---

Notes

• All entries include consistent fields: question, answer, file_name, bbox, type.
• motion_video entries must be linked to the corresponding .mp4 files (replace .png extension).
• This metadata is suitable for training and evaluating vision–language models on ambiguous image understanding.

---