ethics.md

SteelBench: Ethics, Privacy, and Responsible Use

Source and consent

SteelBench is constructed from CCTV footage of an operating integrated steel plant, captured by 64 fixed cameras installed and maintained by plant safety personnel for routine industrial-safety monitoring. The research team obtained the footage under a data-use agreement that permits academic publication of curated, anonymized clips for non-commercial research.

Workers depicted in the footage were aware of the monitoring infrastructure as part of their employment terms. The original capture purpose was plant safety oversight; secondary use for benchmark construction was approved under the data-use agreement. No additional consent was obtained from individual workers for benchmark publication. This is documented explicitly here so users can evaluate fitness for their own intended use.

Anonymization

Version history

• v1.0 (initial release): face blur on the 10,760 JPG frames only, via MediaPipe BlazeFace. The 50 sample MP4s were not anonymized; manifests and annotations contained one SAIL railcar identifier and the work-area name "Go Down South Side".
• v1.1 (this release, 2026-05-15): added YuNet face blur on all 50 sample MP4s; added an EasyOCR on-pixel text-blur pass on every JPG and every 5th MP4 frame; scrubbed the SAIL railcar identifier from one annotation; renamed "Go Down South Side" → "Warehouse South Side" in 6 annotations + 224 manifest rows + 1 camera-zones row.

Methods (v1.1)

(1) Face anonymization — JPGs (carried over from v1.0). MediaPipe long-range face detector (solutions.face_detection.FaceDetection(model_selection=1, min_detection_confidence=0.3)). Detected face regions are blurred with a Gaussian kernel (99×99, σ=30) plus 20% padding. Detection rate on the released JPG set: 4.81%. The low rate is structural: cameras are 7–10 m from typical worker positions, so faces occupy <20 px and fall below the detector's reliable threshold. This distance also means faces in the unblurred majority are too small to support reliable identification.

(2) Face anonymization — sample MP4s (new in v1.1). OpenCV YuNet (face_detection_yunet_2023mar.onnx, score threshold 0.5), applied to every frame of every MP4 with the same 99×99 Gaussian + 20% padding. We switched to YuNet (rather than re-using MediaPipe BlazeFace) because the MediaPipe solutions API was removed in the Python 3.12 wheel; YuNet ships as a 232 KB ONNX model loadable via OpenCV directly. Audio is stripped during re-encoding. Across all 50 MP4s (13,714 frames), 19,306 face-blur applications were made.

(3) On-pixel text anonymization (new in v1.1). EasyOCR (English + Hindi/Devanagari) detects all visible text with a confidence threshold of 0.20. Detected text is blurred with a Gaussian kernel (51×51) plus 8 px padding when it matches any of:

• Brand/company identifiers: SAIL, BSL, Steel Authority
• Indian location words: Bokaro, Jharkhand, India(n), and any Devanagari (Hindi) script
• Other Indian steel plant names: IISCO, Durgapur, Bhilai, Rourkela, Jamshedpur
• Area-board signage referring to the source plant's internal zone names
• CCTV timestamp / camera-ID overlay patterns

In addition, all detected text in the top 80 px and bottom 80 px of every frame is blurred regardless of content (these bands typically carry CCTV camera-ID and timestamp overlays).

Across the released frame set: 12,573 sensitive-pattern regions and 10,229 overlay-band regions were blurred in 10,616 of 10,760 JPGs. In the sample MP4s (OCR run every 5th frame): 3,347 sensitive + 2,363 overlay regions were blurred.

Anonymization is best-effort, not exhaustive. Specifically:

• Small faces (<20 px) may not be detected at all and therefore not blurred.
• Side / back views and partial occlusions reduce detection rates.
• Helmet-occluded faces are sometimes missed.
• Detection precision is high (manual spot-check showed no false positives in 100 sampled blurred regions), but recall is bounded by the model.

Users requiring exhaustive de-identification beyond face blurring (e.g., gait, body shape, distinctive PPE) should not assume this dataset provides it.

Restrictions on use

By accessing or using SteelBench, users agree to the following beyond the CC-BY-NC 4.0 base license:

1. No re-identification. Users will not attempt to identify individuals depicted in the dataset. This includes (but is not limited to): combining the dataset with other footage of the same site, reverse-image search, biometric matching, or contacting plant personnel to identify workers.

2. No biometric training. The dataset must not be used to train face-recognition, gait-recognition, voice (the audio track is stripped, but if recovered), or other biometric identification models.

3. No surveillance-product deployment. The dataset is for VLM evaluation in research contexts. Any deployment of derived models in production worker-monitoring systems requires independent safety validation appropriate to the deployment jurisdiction (workplace surveillance laws vary widely).

4. No use against protected categories. The dataset must not be used to predict or infer membership in protected categories (race, religion, union activity, disability, etc.) of depicted workers.

5. Honor the redistribution clause in LICENSE: redistributors must include this ethics.md unchanged.

Risks and limitations

• False-safe predictions are particularly dangerous. The paper documents that several evaluated VLMs have higher false-safe rates (predicting "no violation" when one exists) than false-alarm rates. Users building safety-critical systems on top of these models should treat the dataset's findings as a warning sign rather than a green light.

• Single-facility data. All clips originate from one integrated steel plant. Sites in other facilities — different equipment, PPE conventions, lighting, or worker demographics — may exhibit different model behavior than what this dataset reveals.

• Annotation reflects observer judgment. Action codes and safety-rule violations are observer judgments by the annotation team, trained against the schema in annotation_tool/schema_validator.py. Inter-annotator agreement and inter-expert agreement are reported in the companion paper. The released annotations are not legal or regulatory determinations.

Audit transparency

The dataset was constructed with a model-assisted annotation pipeline (VLM pre-fill followed by human review). The companion paper's Section 4 introduces an audit protocol that empirically bounds the influence of the pre-fill on the resulting annotations, including:

• Override rate (fields modified by humans relative to VLM outputs)
• Direction analysis (productive vs harmful anchoring)
• Calibration on a blind subset (annotators received no VLM pre-fill)
• Inter-expert agreement on doubly-annotated clips

The full audit code and intermediate audit data are in the companion code repository (scripts/anchoring_bias_analysis.py, scripts/compute_blind_ece.py, scripts/audit_and_repair_annotations.py, and paper_reference/anchoring_analysis/*.json).

Reporting issues

If you identify privacy concerns, anonymization failures, or unintended uses of this dataset, please file an issue on the companion code repository (link in the dataset README) or contact the authors directly (author identities withheld during double-blind review; consult the published paper version after acceptance).