Short answer
Hazard communication labels carry a signal word, pictogram, hazard statements and precautionary statements, and they signal how a container should be handled. Image data for recognizing them comes from distributors and facilities that stock chemicals, and often needs commissioned photography of real containers. Recognition supports flagging and routing. It cannot certify safe handling or replace the safety data sheet.
Why hazard labels deserve their own data
For most objects, a recognition error is an inconvenience. For a chemical container, it can be a hazard. A container marked corrosive or oxidizing may call for different handling and storage than one of detergent, as its safety data sheet and the facility's hazard program set out. A humanoid that moves containers in a stockroom, laboratory or janitorial closet may be required to detect the label first and then apply a different rule.
OSHA's hazard communication standard describes the system the US workplace uses: chemicals are classified by manufacturers and importers, hazard information reaches workers through container labels, safety data sheets and training, and employers must keep a written program, keep SDSs accessible and train workers at initial assignment and when new hazards are introduced. The standard was aligned with the UN Globally Harmonized System in 2012, with labels carrying a signal word, a pictogram, hazard statements and precautionary statements (OSHA, 29 CFR 1910.1200). That is the label structure a recognizer has to read. This page draws no conclusion about any party's obligations under the standard.
This page is about sourcing the image data. The signage side of the problem is in workplace signage and safety labels, and the handling consequences in safety data sheets as handling constraints.
What a label recognizer has to do
A usable system needs more than a pictogram classifier:
- Find the label on a curved, partly rotated or partly hidden container.
- Detect pictograms, which may be printed small, on colored backgrounds or faded.
- Read text: signal word, product identifier, hazard statements, supplier details.
- Associate the label with the container, not with a neighbor.
- Tell apart GHS-style labels from transport placards, NFPA diamonds and supplier-specific marks.
- Handle failure: missing, damaged, handwritten or non-compliant labels.
What the data can support and what it cannot
Image data can support:
- Pictogram and signal-word detection and classification.
- Text reading from labels, combined with OCR evaluation.
- Container-to-label association and label-orientation estimation.
- Routing rules, such as "do not carry; call a person" for flagged items.
- Evaluation of the failure modes above.
It cannot support:
- Hazard classification. The label on a container is the manufacturer's statement. A recognizer reads it; it does not verify the contents. The safety data sheet and the facility's hazard program remain the controlling documents.
- Compliance. Image datasets do not demonstrate that any deployment meets workplace safety obligations. That is a matter for the employer's program and qualified safety staff. This is not legal advice.
- Safe handling. Knowing a drum is corrosive does not teach a robot to move it. Control and safety behavior need robot data and engineering controls, not images alone.
- Contents. Mislabeled, decanted and unlabeled containers look like ordinary bottles.
Sources of images
| Source | What it offers | Typical gap |
|---|---|---|
| Chemical and janitorial distributors | Many products, many label designs | Photos may be catalog packshots, not in-situ |
| Facilities with chemical storage | Real containers in real lighting | Access, safety and consent constraints |
| Manufacturers' product photography | Clean, front-on label views | Little occlusion or wear |
| Training and compliance material | Illustrative pictograms | Not photos of real containers |
| Commissioned photography | Designed viewpoints and conditions | New cost; needs site approval |
For labels, the decisive variable is condition. Faded, torn, stained, relabeled and secondary-container labels are where recognizers are likely to struggle, and they are probably rare in catalog sources.
Illustrative collection specification
The specification below is Illustrative. All values are examples a buyer would set.
# Illustrative only - not an available dataset
collection: ghs_label_images_v0
containers:
types: [bottle, jug, drum, aerosol, bag, secondary_container]
label_states: [new, faded, torn, stained, handwritten, missing]
views:
label_front: true
oblique_angles: [30, 60]
container_in_context: shelf, cabinet, cart
resolution: label text legible at capture distance; buyer sets distance
lighting: warehouse, cabinet interior, glare cases
annotations:
pictogram_boxes: with class labels
signal_word: text + box
hazard_statement_text: transcribed
container_mask: instance
label_to_container_link: id
exclusions:
- containers whose contents are unknown
- images showing people
metadata: site_id, product_family, label_standard_observed, capture_date
Images should be captured by people trained in the facility's safety rules, and containers should be photographed in place without being moved or opened.
Acceptance criteria you can set
- Pictogram coverage. Each hazard pictogram appears across container types and conditions, reported as counts.
- Condition coverage. Degraded and non-standard labels are included, not just clean ones.
- Legibility. Sample images to confirm text is readable at the stated resolution. See small object resolution requirements.
- Label accuracy. A human reviewer checks pictogram classes and transcribed text on a sample.
- Held-out split. Evaluation sets separate by product family and site.
- Metadata. Label standard observed and capture conditions are recorded.
Questions to ask suppliers
- Are the images of real stocked containers, or catalog artwork?
- How many products, brands and label designs are represented?
- Are older or non-standard labels present, such as pre-alignment formats or imports?
- Were images captured under the facility's safety procedures, and who approved capture?
- Do the images show employees, badges or screens?
- Who can license product label images: the distributor, the manufacturer or the facility?
- Is any content from safety data sheets, and who owns it?
- Can the supplier add photographs of degraded or secondary-container labels?
Permissions and provenance
Labels carry manufacturer branding and text, and photographs may be covered by photographer or brand rights. Facilities may restrict photography in chemical stores for safety or confidentiality, and images can reveal stock lists, quantities and layouts. Redaction of names or quantities lowers exposure but does not settle whether the facility may license the images. Treat site access and licensing as separate checks.
Where this fits
Hazard labels are one input to hazard properties of objects and to safety-critical steps in procedures. Recognition is a perception layer; it does not replace engineering controls or human oversight. The objects hub and the humanoid training data guide map the neighboring data types.
For printed text on packages in general, including barcodes and ingredient panels, see labels, barcodes and text on packages.
Discuss your requirements
If you need images of hazard labels on real containers, describe the label types, container conditions, annotations and permitted uses to SourceX. SourceX reviews the requirement, looks for US businesses that may hold suitable images, checks the data and the supplier's licensing permissions, and follows up about potential sources. If nothing existing fits, a commissioned collection can be discussed. A request does not guarantee a match. The catalog entry for images and inspection photos is the nearest existing category.
Discuss your humanoid training-data requirements with SourceX
Sources
- 29 CFR 1910.1200 - Hazard Communication — label elements, safety data sheets, written programs and training under the US hazard communication standard.