bmeivar/MetricVQA

Metric VQA

A two-split benchmark for diagnosing evidence use in vision-language models on monocular metric object-size estimation: given a single RGB image and a natural-language dimension query, predict the physical dimension in centimeters.

• In-the-wild (ITW): 331 queries over 70 cluttered scene photographs, every ground-truth dimension measured by tape measure (0.8–245 cm, ~99 distinct objects). Each query carries a human-verified instance mask.
• Objectron: 10,482 queries over 3,494 frames from ~2,600 object instances in 9 categories of Google Objectron, ground truth from the dataset's oriented 3D bounding boxes.

The benchmark is a diagnostic instrument, not a leaderboard: the companion intervention suite (object-only, object-ring, wrong-object, random-mask, zoom, lens distortion, text-only) probes which evidence channels a model's estimate actually depends on.

Paper: Ill-Posed by Design: Probing Evidence Use in VLMs (arXiv:2606.24335) · Code: https://github.com/bmeivar/MetricVQA

Layout

itw/questions.jsonl         331 records (schema below)
itw/questions.csv           same records, flat CSV
itw/images/                 70 scene photographs (PNG)
itw/masks/                  331 binary instance masks (PNG, one per query)
objectron/questions.jsonl   10,482 records
objectron/questions.csv     same records, flat CSV

Objectron frames are distributed separately (see below) to respect the source dataset's distribution terms; the extraction script reproduces them deterministically from the public Objectron release.

Record schema (JSONL, one object per line)

field	type	splits	description
query_id	str	both	stable id, itw_0001 / obj_00001
split	str	both	itw or objectron
image	str	both	path relative to the split directory
question	str	both	verbatim query; at evaluation time it is followed by the instruction "Answer with a single float number in centimeters"
answer_cm	float	both	ground-truth dimension in centimeters
axis	str	both	length / height / width / depth / diameter / size, parsed from the query
target_phrase	str	itw	referring expression grounding the query to one object
mask	str	itw	binary PNG instance mask for the target
bbox_xyxy	[int×4]	itw	tight box of the mask, pixel coordinates
mask_source	str	itw	sam3_auto_verified (SAM 3 from the phrase, passed human review) or manual_box_prompt (re-extracted from a human-drawn box)
gt_method	str	both	tape_measure or objectron_3d_box
category	str	objectron	one of the 9 Objectron categories

Annotation process (ITW)

Scenes were photographed and measured by the authors. Target masks were initialized with SAM 3 from each query's referring phrase; a human annotator reviewed all 331 masks and re-extracted those that failed review (32 objects, covering 53 queries) using manually drawn box prompts, so every mask in the release is human-verified. The mask_source field preserves provenance per query.

Evaluation protocol

Primary metric: Acc@10% — fraction of queries with |ŷ−y|/y ≤ 0.1. Responses with no parsable number are excluded from accuracy denominators (a parse failure is an instruction-following failure, not a size estimate); report parse rates alongside. Paired intervention deltas are computed on rows parsable in both conditions. Significance: scene-clustered sign-flip permutation tests with Benjamini–Hochberg correction (queries sharing a scene flip together).

Licensing

• ITW images, masks, and annotations: CC-BY-4.0.
• Objectron-derived annotations: released under the terms of the Objectron dataset license, with attribution; frames are re-derivable from the public source.
• Evaluation and intervention code: MIT.

Intended use and limitations

Intended for diagnosing evidence use and for benchmarking metric estimation under monocular ambiguity. The ITW split is evaluation-only. Ground truth is a single tape measurement per dimension; the human-reference row reported in the paper is a single annotator, not a population estimate. Queries are in English.