AbdullahKhanSherwani/dubai-taxi-advertising-synthetic

Dubai Taxi Advertising Compliance (Synthetic)

63 synthetic photorealistic images of Dubai RTA taxis carrying advertising, built to evaluate whether vision-language models can judge out-of-home (OOH) advertising compliance rules from a single photograph.

Each image is generated to be an unambiguous pass or fail against one specific rule from a Dubai taxi advertising technical checklist. The dataset is an evaluation set — it is small, adversarially balanced, and deliberately includes rules that are hard or impossible to judge from an image alone.

Contents

Images	63 PNG
Rules	8
Verdicts	31 pass / 32 fail
Views	side (39), rear (16), front (8)
Generator	google/gemini-3-pro-image

Fields

• file_name — image file
• scenario_id — unique id, e.g. TAXI_POS_01_pass_v2
• rule_id — which rule this image tests
• view — camera angle: front, side, rear
• intended_verdict — pass or fail, the label
• intended_note — one line on what makes it pass or fail
• generator — model that produced the image

Rules

Rule	What it tests	Pass	Fail
TAXI_POS_01	Ad placement below the boundary line, rear	Ad on boot lid below tail lights	Ad across rear windscreen glass
TAXI_POS_02	Ad placement below the boundary line, side	Ad on door metal only, glass bare	Ad continues onto door glass
TAXI_VIS_01	Front elevation free of advertising	Front clean	Ad on bonnet/bumper
TAXI_VIS_02	Tail lights unobstructed	Lenses bare	Vinyl over tail-light lenses
TAXI_WIN_01	Perforated film on rear side windows only	Rear passenger glass only	Front/driver glass also covered
TAXI_WIN_04	≤50% of side window glass covered	~25% covered	Nearly full pane covered
TAXI_WIN_05	Commuter can still see out	Perforated, cabin visible	Solid opaque panel
TAXI_BODY_01	Operator logo remains visible	Logo clear on bonnet and door	Full wrap hides the logo

Two rules (TAXI_POS_01, TAXI_POS_02) reference a red-line placement boundary defined in an external diagram from the source checklist. That diagram is not included here — models evaluated on those rules were shown it as a second image. Without it, those two rules are under-specified.

Intended use

Benchmarking VLMs on rule-grounded visual compliance judgement. A correct evaluation should let a model answer cannot_determine: TAXI_WIN_04 asks for a percentage of glass area, which is a genuine measurement question, and abstention there is a legitimate result rather than a failure.

Limitations

Labels are intended, not verified. Each label states what the image was prompted to depict. No human has independently confirmed that every generated image actually shows what its prompt asked for, and image generators do drift from prompts. Treat these as weak labels and spot-check before drawing strong conclusions.

Synthetic, not real. These are generated images, not photographs of real vehicles. Livery details, plates, and street furniture are approximations of Dubai RTA taxis and are not authoritative. Any text rendered in the images — brand names, taglines — is invented.

Small and unbalanced. 7–8 images per rule. Per-rule accuracy on this set has wide error bars, and one rule (TAXI_WIN_05) has 7 rather than 8 images.

Not a regulatory reference. The rule wordings are paraphrases written for model evaluation. Do not use this dataset or its rule text to determine actual compliance with Dubai RTA advertising regulation.