Short answer
Package label reading data pairs images of barcodes, logistic labels and printed text with the decoded value as ground truth. Buyers need real capture variety: distance, angle, glare, curvature, print quality, damage and partial occlusion. Images without the decoded identifier or transcription cannot score a reader. Good data separates three tasks: finding the label, decoding the code and reading the text.
The buyer's problem
A robot that receives, sorts or restocks goods must know which item or logistic unit it holds. The identifier is printed on the package: a barcode, a shipping label, a lot code, an expiry date or hazard text. Reading it from a moving camera at arm's length is a separate perception task from recognizing the product by appearance, and it fails in particular ways, from curved surfaces to thermal print fade to shrink-wrap glare.
This page covers sourcing image data for label and text reading on packages. For hazard-specific label sets, see hazmat and GHS labels, and for workplace signs, see workplace signage and safety labels.
Three tasks that look like one
Localization. Find where the label or code is on the package or pallet. General detection datasets help with this step. Microsoft COCO provides 328,000 images, about 2.5 million labeled instances and 91 object categories with per-instance segmentation (Lin et al., 2014). That coverage is about common object kinds in context, not about reading identifiers from labels, so it supports general detection but does not stand in for label data.
Decoding. Convert a barcode or 2D code to its value. Decoding is deterministic once the image is good enough, so your evaluation is read rate, not label accuracy. Ground truth is the encoded value.
Reading. Transcribe printed text: product names, dates, lot numbers, addresses, warnings. This is optical character recognition and depends on font, language, print method and background.
What identifiers carry
GS1 defines the Serial Shipping Container Code (SSCC) as an 18-digit key that uniquely identifies a logistic unit such as a case, pallet or parcel. It is the key to advance shipping notice data and is usually encoded in GS1-128 on the GS1 Logistic Label, or in EPC/RFID, and it is compatible with the ISO/IEC 15459-1 license plate (GS1, SSCC).
For a buyer this means two things. A pallet or case label may carry a machine-readable key that connects to a shipment record, so ground truth for a label image can in principle be checked against shipping data. And label layouts follow a standard, so a dataset that includes standard logistic labels is more portable across suppliers than one built on private formats. Item-level product barcodes follow other GS1 rules; confirm which symbologies a supplier's images actually contain.
Capture variables to specify
| Variable | Range to ask for | Why it matters |
|---|---|---|
| Distance and angle | Close, arm's length, oblique | Resolution per module and skew |
| Surface | Flat, curved, wrinkled, film-wrapped | Distortion and glare |
| Print method | Thermal transfer, direct thermal, inkjet, offset | Contrast and fade |
| Condition | New, scuffed, wet, torn, partly covered | Real operating failures |
| Lighting | Dim warehouse, glare, backlight | Exposure and specular highlights |
| Motion | Static, handheld, moving conveyor | Blur |
| Symbology | Linear, 2D, GS1-128, OCR text | Decoder choice |
| Language and script | English, bilingual, other | OCR coverage |
State your robot's camera and typical distance, and ask the supplier to report pixels across the narrowest bar or module. The small object resolution requirements page covers how resolution limits recognition.
Illustrative record
# Illustrative: synthetic values only
label_image:
image_id: lbl_30188
source_site: dc_site_009
surface: shrink_wrapped_pallet
label_type: logistic_label_gs1
regions:
- {type: barcode_gs1_128, box: [x1,y1,x2,y2], decoded_value: "(00)000000000000000000", decode_verified: true}
- {type: text_block, box: [x1,y1,x2,y2], transcription: "SHIP TO: SITE 12", lang: en}
capture: {device: handheld_scanner_camera, distance_cm: 60, angle_deg: 25}
condition: [glare, partial_wrinkle]
pii_present: false # addresses may be personal data; review required
Values like those in decoded_value should be synthetic or masked in any shared sample, since real shipping labels can carry names, addresses and customer references.
Supplier types and routes
Third-party logistics providers may hold label images from receiving, putaway and shipping, plus scan logs that give ground-truth values. They hold the data under client contracts, so ask whose permission is needed. See 3PL providers as warehouse data suppliers.
Manufacturers hold images from end-of-line label verification, which are consistent but narrow in condition range.
Retailers hold receiving-dock and shelf label images, including price labels.
Existing assets suit this task well, because scan systems already log the decoded value next to an image. The caution is that images from fixed scanners show the label under controlled conditions, which differs from a moving robot camera.
Scoring a label-reading system
Define metrics before you look at the data. For decoding, report the read rate (decoded correctly over total attempts), the misread rate (decoded to the wrong value) and the no-read rate separately, since a wrong read is usually worse than no read. For text, report character and field-level accuracy, and score high-consequence fields such as lot numbers and destinations separately. Stratify by the capture variables in the table, and report time per read if your robot has a cycle-time budget. Keep a set of label templates out of training. For wrapped or worn-out labels, check whether handling marks on packaging, such as arrows and fragile symbols, add clutter around the code or carry meaning your robot needs.
Acceptance criteria
- Ground truth verified. Decoded values come from a scanner log or a second decoder, not only from the dataset's own reader.
- Failure cases retained. Images where the scanner failed are valuable; confirm they were not dropped.
- Condition spread per the table above, with counts.
- Split by site and by label template, so test labels are unseen formats.
- Read-rate baseline. Run an off-the-shelf decoder on a sample and record the rate before purchase.
- Personal data handling. Documented treatment of names and addresses.
What this data cannot support
Reading an identifier does not tell a robot how to pick, orient or place the package. Label images contain no motion or force information. A label can be wrong, outdated or swapped, so identification by label is not proof of contents. Models trained on one language, print method or template may not generalize. Reading text also does not establish meaning: a hazard phrase needs a lookup table or procedure, not just OCR.
Permissions and provenance
Shipping labels frequently contain personal names, addresses and customer or order references, and may reveal business relationships. Ask what has been masked, whether masking affects the barcode region, and who can license training use. Rights questions also arise for the images themselves and for third-party brand marks. See product image rights. This is not legal advice.
Questions to ask suppliers
- Which symbologies and label standards appear, and in what proportion?
- Is the decoded value or transcription attached to each image, and how was it verified?
- What camera, distance and lighting were used, and what resolution per module?
- Are failed scans, damaged labels and occlusions included?
- How many sites, label templates and print methods?
- What personal or customer data appears, and how was it handled?
- Are images tied to shipment or product master records that can be checked, see product master data quality audit?
- Who owns the images and the logs, and can both be licensed?
Discuss your requirements
If you can describe the label types, camera setup and read-rate target, describe the requirement to SourceX. SourceX reviews it, follows up about potential US sources such as logistics providers and manufacturers, and checks the data and licensing permissions. A request does not guarantee a match. Start from the humanoid training data guide or the product and object data hub. The site's entry on supply chain and logistics data is a related category.
Discuss your humanoid training-data requirements with SourceX
Sources
- Serial Shipping Container Code (SSCC) — SSCC: 18-digit GS1 key identifying a logistic unit, usually encoded in GS1-128 on the GS1 Logistic Label.
- Microsoft COCO: Common Objects in Context — COCO: general object-in-context dataset used as a contrast with label-reading needs (328k images, about 2.5 million labeled instances).