README.md

metadata

license: cc-by-4.0
task_categories:
  - visual-question-answering
  - image-classification
tags:
  - ppe
  - construction-safety
  - vqa
pretty_name: PPE Detection Benchmark Eval Set

PPE Benchmark Eval Set (v1)

A held-out, human-verified benchmark for evaluating vision-language models on personal protective equipment (PPE) detection — specifically hardhat and safety-vest presence — framed as a VQA-style classification task.

What this is

96 images, balanced 24/24/24/24 across the four hardhat × vest combinations (yes/yes, yes/no, no/yes, no/no). Sourced from a forked, filtered subset of the karabuk-university PPE dataset on Roboflow Universe (CC BY 4.0), then individually visually verified — not just filtered by class label — to remove:

• Mislabeled images (e.g. life jackets tagged as safety vests)
• Images with an undetected second person in frame
• Images where PPE was held, not worn
• Off-context images (e.g. a child wearing a hardhat, watermarked stock photos)

~40% of the initial no-hardhat + vest bucket was found to be contaminated by these issues before this cleanup pass.

Labels

Each image has:

• hardhat_label: yes / no
• vest_label: yes / no
• framing: full-body / half-body
• angle: front / side / rear / low / high
• lighting: studio / indoor / daylight / shade / dim
• ambiguity_type: flags 5 genuine hard-negative cases (cap or beanie worn instead of a hardhat) for testing false-positive robustness

Framing split: 56 half-body / 40 full-body. Lighting: studio 25, indoor 42, daylight 22, shade 3, dim 4.

Known limitations

This is a small benchmark (24 images/class). It's sufficient for catching large, obvious differences between models, but:

• Confidence intervals on close model comparisons will be wide
• Per-condition slicing (e.g. "accuracy in dim lighting") has very few samples per cell and should be treated as indicative, not conclusive
• Only 5 hard-negative examples exist — not enough for a standalone hard-negative accuracy metric

v2 (planned) will expand volume using the same per-image verification methodology, particularly for the no-hardhat + no-vest bucket, which was the rarest combination in the source pool.

Intended use

Held-out evaluation only — do not use for training. Built to benchmark finetuned PPE-detection VLMs (SmolVLM, Qwen3.5, Qwen3-VL) against a common, verified ground truth.

License

Images sourced under CC BY 4.0 from Roboflow Universe (karabuk-university/ppe-oevsc). Attribution: karabuk-university.