Short answer
When color is a label, such as a cap color that separates two products, images need a reference chart in frame, a recorded illuminant, fixed or logged white balance and exposure, and an unaltered raw file. Studios and brands can supply these. Calibrated images support color-dependent recognition; they do not describe how an object feels or behaves when handled.
When color is data, not decoration
Most vision models are trained to tolerate color shifts, and that is usually right. Cameras differ, lamps change, and a robot cannot control the light on a warehouse floor. But some tasks make color the decision: telling a red-capped reagent from a blue-capped one, sorting by color-coded bin, reading status lights, matching a paint or fabric, or checking that a printed package matches the approved art.
In those cases, a pretty image is not enough. If the pixels were auto-white-balanced, tone-mapped by a camera profile and retouched in post, the "red" in the file may not be the red on the object. This page specifies the metadata and capture practices that make color usable as a label. For general multi-view capture, see the multi-view product photography spec, and for the cluster, the product and object data hub.
What to specify
Reference and scene
Place a reference color chart in at least some frames, ideally every session, and log which chart. Use a neutral gray or white target for white-balance checks. Keep the background and ambient light controlled, and record any colored surfaces nearby that could reflect onto the object.
Illuminant and lighting
Log the lamp type, correlated color temperature as supplied by the manufacturer, whether lighting was diffuse or directional, polarization if used, and whether ambient light was blocked. A lighting diagram saved with the session is worth more than a free-text note.
Camera and exposure
Record camera model, lens, sensor mode, exposure time, aperture, gain or ISO, white-balance setting (fixed, preset or auto) and any in-camera picture profile. Prefer fixed settings per session. If auto functions were used, record the values the camera chose.
Files and processing
Deliver raw files where possible, plus the processed images, and document the pipeline: color space, profile, tone curve, sharpening and retouching. State plainly if clipping paths, shadows or backgrounds were edited. The question to resolve is which version is the evidence, and which is a convenience copy.
Illustrative capture and metadata specification
Illustrative only. All values are synthetic placeholders; the choices depend on your task, objects and sensors, and none are minimums.
# Illustrative color-critical capture spec (synthetic values)
session_id: sess_0093
reference_chart:
type: "24-patch reference chart"
in_frame: every_session_start_and_end
chart_id: chart_B2
neutral_target: gray_card_in_first_frame
illuminant:
lamp_type: LED_panel_diffuse
cct_manufacturer_k: 5000 # arbitrary example
polarization: cross_polarized_optional
ambient_blocked: true
camera:
model: cam_X
white_balance: fixed_preset # not auto
exposure_ms: logged_per_frame
picture_profile: neutral
files:
raw: delivered
processed_color_space: sRGB
retouching: none_except_listed
retouching_log: required
object_metadata: [sku_id, finish: matte|gloss, material]
Surface finish matters as much as color
A glossy cap and a matte cap of the same color photograph differently under the same lamp. Finish, translucency and metallic or pearlescent coatings change apparent color with viewing angle. Record the finish, and capture more than one angle for items where this matters. Reflective and transparent items bring further difficulty; see transparent and reflective objects.
A published capture-rig reference
Google Scanned Objects is a dataset of over one thousand 3D-scanned household items, captured with a structured-light scanning rig, released under a Creative Commons license and preprocessed for simulators (Downs et al., 2022). It is useful as an example of documenting a controlled capture rig and releasing assets with clear license terms.
It is a capture reference only. It is a household-item scan collection, and the paper does not establish a color-calibration procedure or accuracy for the textures. Do not treat its color as a calibrated ground truth for your objects. Check the license terms of any asset you plan to use commercially, as the Creative Commons overview explains differences between attribution and noncommercial terms; this is not legal advice.
Where color-managed images exist
Product photography studios often work with reference charts and a managed color pipeline because brand owners demand color-accurate catalog art. Their archives may therefore hold chart frames and raw files that no other supplier retains. See product photography studio archives for the rights questions that come with them. Brands and packaging printers hold approved color references for their own products.
Operating businesses seldom have this. Warehouse and workplace images are shot under mixed lighting with auto settings, which is realistic for a robot but poor for color labeling. If you need both, plan two sets: a calibrated set for color labels, and an uncontrolled set for deployment realism. The trade-off between them is the subject of packshot versus in-situ images.
If neither source fits, a purpose-built capture of your objects with the spec above is possible to discuss; see existing assets, new collection or specialist partners.
What calibrated images cannot support
A calibrated image tells you what color an object appears under a known light. It does not tell a robot what color it will see in a dim aisle with sodium lighting, and a model trained only on controlled images may not generalize to that. Calibrated product images also carry no handling information: nothing about weight, rigidity or surface friction. Manipulation learning needs robot or teleoperation data, not color-managed photography. And synthetic renders do not substitute for calibrated real images when color is a label, a point discussed in synthetic renders versus real images.
Acceptance tests
- Chart check. Measure chart patches in delivered images and report deviation from reference, with a tolerance you set.
- Metadata completeness. Illuminant, camera, exposure and white-balance fields present for every session.
- Raw availability. Raw files exist for the images you plan to use as color evidence.
- Retouching log. Edits are listed, and color-critical regions are untouched.
- Cross-session consistency. A repeated object measured across sessions falls within your tolerance.
- Split rule. Hold out sessions or lighting setups, not just random frames.
Brand color and packaging redesigns interact: a new print run can shift a color without changing the SKU, which is one reason to read packaging variants and SKU appearance alongside this spec.
Questions to ask suppliers
- Were reference charts or gray targets used, and are frames with them retained?
- What lamps and ambient-light controls were used, and are they documented per session?
- Were white balance and exposure fixed or automatic?
- Are raw files retained and licensable, or only processed JPEGs?
- What retouching was applied, and does it touch color-critical regions?
- Is surface finish recorded for each item?
- Is a brand-approved color reference available for the items, and who may license it?
- Do the same objects appear in multiple sessions, so consistency can be tested?
- Who owns the images, charts' data and any 3D assets?
Discuss your requirements
If color accuracy is part of your labeling task, describe the objects, color decisions and permitted uses to SourceX. SourceX looks for US businesses that hold suitable material, checks the data and the supplier's licensing permissions, and coordinates the license and delivery. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. A request does not guarantee a match; SourceX follows up about potential sources. The physical-world collection page describes the separate path for new recordings, and the humanoid training data guide shows the other routes.
Discuss your humanoid training-data requirements with SourceX
Sources
- Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items — structured-light scanned household items under a Creative Commons license; a capture-rig reference only.
- About CC Licenses — differences between Creative Commons license terms.