Short answer
Rack, shelf and location-label images show storage faces, slot labels and the goods in them. They come from 3PLs and distributors that photograph locations for audits, slotting reviews, cycle counts and damage claims. They can support location-label reading, slot-state recognition and navigation perception, but they do not teach handling, and client goods in frame carry rights questions.
The buyer's problem
A mobile manipulator in a warehouse spends much of its day confirming where it is and what is in front of it. Before it picks, it must find the right aisle, bay, level and slot, read a label, and decide whether the location is occupied, empty or blocked. Those are perception problems, and the images that serve them are different from task demonstrations.
Perception engineers often ask for "warehouse shelf images" without saying what they are for. The right sourcing depends on the answer: reading location labels, counting items on a shelf face, detecting slot occupancy, localizing against rack geometry, or checking pallet position in a rack opening. This guide separates those needs and tells you what to request. For demonstrations of handling, see each-picking demonstrations instead.
Which perception need are you sourcing for
| Need | What the images must show | Typical label |
|---|---|---|
| Location label reading | Label face at readable resolution, varied heights and angles | Text or barcode value, bounding box |
| Slot occupancy | Whole slot opening, front and back of the face | Empty, occupied, blocked, partial |
| Item counting and detection on shelves | Densely packed identical items | Instance boxes or masks, counts |
| Rack geometry and localization | Upright, beam and bay structure over many viewpoints | Bay and level ID, pose if available |
| Pallet position in rack | Pallet in opening, overhang, load stability | Position class, condition flag |
Densely packed identical objects are a known detection challenge. SKU-110K, introduced in a paper on detection in densely packed scenes, is an annotated dataset of retail shelves whose objects are tightly packed and often identical (Goldman et al., 2019). It is retail-shelf imagery, not warehouse racking, but it illustrates the labeling and model challenge that a pick face with identical cartons also presents.
What this data can support, and what it cannot
Can support. Training or fine-tuning detectors and readers for location labels and barcodes, building slot-state classifiers, evaluating localization against a known map, and testing perception under real lighting, shrink wrap glare and low shelf shadows. The location identity itself is a strong label: if images are tied to bay and level codes, a model gets a ground truth for free.
Cannot support. Images of a shelf do not show how an item is removed, how it shifts when its neighbor is lifted, or what a hand must do to reach into a deep slot. They contain no robot actions. They are also a snapshot of storage at one time, which may not reflect the picking state. The distinction between recognition and manipulation data is set out in object recognition data vs manipulation training data.
Where location identity comes from
Warehouse locations carry identifiers, and logistic units carry their own. The GS1 Serial Shipping Container Code is an 18-digit key that uniquely identifies a logistic unit such as a case, pallet or parcel, usually encoded in GS1-128 on the logistic label or in EPC/RFID, and compatible with the ISO/IEC 15459-1 license plate (GS1). Location labels may follow a site-specific scheme rather than GS1, so ask which scheme each site uses and whether a mapping from location code to bay and level exists. If images are paired with logistic unit labels, a pallet in a rack opening can be linked to its record.
Illustrative request package
# Illustrative only - synthetic values; a buyer adapts these fields
package: rack_location_images_v0
sites: 3 (site_009, site_031, site_052)
image_sets:
- label_close: location label at 0.5-2 m, rack levels 1-5
- bay_face: full bay front, 3-6 m
- aisle_view: down-aisle, fixed height and moving
capture_params: sensor_resolution, lens_fov, exposure_mode, lighting_notes
per_image_metadata: [site_id, aisle, bay, level, capture_time, device_id]
labels: [label_text, label_box, slot_state, occupancy_class]
exclusions: [staff_faces, client_names_on_goods]
holdout: split by site and by capture date
rights: training and evaluation use; no redistribution
Supplier types
3PLs and distributors have the widest coverage because they run many locations and often hold slotting or audit photos. Cycle-count programs may already produce location-tagged images; see cycle counting and inventory checks. Stocktake vendors and facilities teams may hold survey images. A 3PL's images usually contain clients' goods, so read client-owned inventory rights and 3PL providers as warehouse data suppliers. Facility drawings are a separate asset; see warehouse layout and facility data.
Quality checks
- Resolution at the label. Check pixel height of the label text in a random sample.
- Angle and distance variety. Ask for the distribution, not a single mounted view.
- Label condition. Worn, torn and reprinted labels belong in the set.
- Occupancy balance. Check the ratio of full, empty and blocked slots.
- Time. Images from one day hide lighting and stock changes; check coverage across shifts and seasons.
- Held-out splits. Hold out entire sites and aisles, since neighboring bays look alike and leak across a frame split.
Limitations and additional preparation
The images were taken for audits, not training. Labels are usually absent: location text and slot state need annotation unless the capture app stored them. Clutter, forklifts and staff may appear. Client goods, pallets with product names and packaging can be visible. Drawings or site maps are needed if the buyer wants localization ground truth. The data does not contain depth unless the capture device recorded it, so ask. Navigation and handling policies need robot data collected on the target hardware.
Worked example of a request
Suppose a perception team needs a location-label reader that works from a mobile base. A reasonable request is: three or more sites; label close-ups at several heights and angles; whole-bay views; per-image bay and level codes; sample sets showing damaged and reprinted labels; and a stated capture device with resolution and lens details. The team then splits by site, trains on two, tests on the third, and reports reading accuracy by label height.
Compare that with a request for "warehouse shelf images." The second request returns whatever exists, and the team discovers months later that all images are of pallet-rack faces at 1.5 m. Naming the task, the label scheme and the splits up front is cheaper than discovering gaps afterwards. For pick-face geometry that affects what a hand can reach, see pick-face types.
Questions to ask suppliers
- How were the images captured: handheld device, truck-mounted camera, fixed camera or drone?
- Is the location code stored with each image, and in what scheme?
- At what distance and height were labels photographed, and what is the label text resolution?
- How many aisles, bays and levels does each site contribute?
- Do images show full slots, or only the first row of goods?
- Are damaged, missing and reprinted labels represented?
- Is depth or lidar captured alongside RGB?
- Whose goods are visible, and has each owner agreed to licensing?
- What notice applied to staff who appear in frame?
Discuss your requirements
If you can say which locations, label types and perception tasks you need, explore sourcing and licensing options with SourceX. SourceX reviews the requirement, follows up about potential sources among US businesses, checks the data and the supplier's licensing permissions, and coordinates delivery. A request does not guarantee a match. See the humanoid training data guide and the warehouse and fulfillment hub; catalog context is on the images and inspection photos page.
Explore sourcing and licensing options with SourceX
Sources
- Precise Detection in Densely Packed Scenes — SKU-110K, dense identical-object detection.
- Serial Shipping Container Code (SSCC) — logistic unit key and label encoding.