Short answer
A multi-view product photography specification fixes how each item is shot so images are consistent and reusable for recognition, retrieval and 3D reconstruction: views on a turntable at set elevations, including top and underside, fixed lens and lighting, a color reference, calibrated cameras, per-frame pose and masks, and per-item identifiers and variants. The buyer sets the counts and tolerances; acceptance tests check sharpness, exposure, color and coverage.
When to commission photography instead of licensing a catalog
Catalog imagery is the cheapest object data to license, but it is uneven. ABO, built from a large retail catalog, has 360-degree turntable views for 8,222 of its 147,702 product listings (Collins et al., 2021): rich views exist for a small minority. Commercial catalogs show the same pattern, and their images were shot for selling, not for recognition from a robot's angles.
Commissioned photography is worth the cost when you need every target item from every side, the same lighting and lens across thousands of items, known camera poses for reconstruction, or images you own outright. The trade-offs against licensed catalogs are summarized in the product and object data hub.
What the images will be used for
The specification should follow the use, because each use stresses different parameters.
| Use | Parameters that matter most |
|---|---|
| Instance recognition and retrieval galleries | View coverage, consistent background, identifiers and variants |
| Fine-grained discrimination of look-alikes | Resolution, label legibility, color accuracy |
| 3D reconstruction or novel-view synthesis | Overlap between views, calibrated intrinsics, per-frame pose, masks |
| Pose-estimation training | Pose metadata, a mesh or reconstruction, consistent scale |
| Seeds for synthetic data | Clean masks, color accuracy, diffuse lighting without baked shadows |
Published 3D object sets show what reconstruction needs. OmniObject3D provides about 6,000 real-scanned objects with real-captured videos, foreground masks and COLMAP camera poses alongside meshes (Wu et al., 2023). Google Scanned Objects captured household items on a structured-light scanning rig and prepared them for simulation (Downs et al., 2022). If you need simulation-ready geometry rather than images, add scanning to the order and read simulation-ready object assets.
Illustrative capture specification
# Illustrative only — every value is an example for the buyer to adapt
spec_id: MVPP_v0.3
rig:
turntable: indexed, angle logged per frame
azimuth_step_deg: 15 # 24 frames per ring
elevation_rings_deg: [0, 30, 60]
extra_views: [top_down, underside_after_flip]
camera:
sensor: full-frame RGB, RAW capture
lens: fixed focal length, manual focus, aperture chosen for full-item depth of field
output: 16-bit TIFF (archive) + 8-bit PNG, sRGB
min_object_pixels: 1,500 px on the longest object side
lighting:
primary: diffuse, fixed positions, recorded in a lighting diagram
variant: one hard key light (optional, for specular cues)
white_balance: fixed per session from a gray card
background: uniform neutral gray; optional second pass on a textured mat
references_per_session: [color_chart, scale_bar, gray_card]
calibration:
intrinsics: planar target at varied orientations, per camera per session
extrinsics: camera pose per ring relative to turntable axis
per_frame_metadata: [item_id, instance_id, ring_deg, azimuth_deg, camera_id, session_id, timestamp]
masks: foreground mask per frame (PNG), human-checked on a sample
per_item_metadata:
identifiers: {gtin, sku, variant_of, packaging_version}
physical: {length_mm, width_mm, height_mm, weight_g, measured: true}
condition: new | opened | damaged
states: [closed, lid_open] # only where the item has states
instances_per_sku: 2 # different production lots where possible
naming: "{item_id}_{instance}_{state}_{ring}_{azimuth}.tif"
Field notes
Views. Include the underside and the back. Barcodes, nutrition panels and ingredient lists sit there, and a robot often sees items from above or at odd angles. If items cannot be flipped safely, say how the underside is captured.
Lighting. Diffuse lighting gives clean appearance for recognition and texture extraction. A second pass with hard light helps models learn specular cues on glossy packaging. Record positions so sessions months apart match.
Color. Put a color chart in a reference frame each session and fix white balance. Color is often a label (flavor, variant, cable rating), so drift between sessions becomes label noise; see color calibration in product images.
Calibration. Intrinsics can be estimated from a planar pattern viewed at a few orientations, with radial distortion modeled (Zhang, 2000). Deliver the calibration files and the reprojection error. Broader rules for robotics datasets are on calibration requirements.
Instances and states. Two physical instances from different lots expose printing and fill variation. Articulated or openable items get a pass per state.
Hard materials. Transparent and mirror-like items defeat both masks and depth. ClearGrasp needed dedicated depth estimation for transparent objects and built a real-world benchmark of 286 RGB-D images to test it (Sajjan et al., 2019). Agree in advance how such items are lit and masked, and do not allow matte sprays that change their appearance. Transparent and reflective objects covers what else they need.
Parameters the buyer sets
| Parameter | Typical range of choices | Drives |
|---|---|---|
| Azimuth step and elevation rings | Coarse for retrieval, dense for reconstruction | Frames per item, cost |
| Resolution floor | Set by the smallest text or feature to be read | Lens and distance |
| Lighting passes | Diffuse only, or diffuse plus hard key | Session time |
| Instances per SKU | One, or several from different lots | Sample sourcing |
| States | Closed only, or every functional state | Handling time |
| Backgrounds | Neutral only, or neutral plus contextual | Downstream domain gap |
| Masks | None, automatic, or human-checked | Annotation cost |
Acceptance tests
- Sharpness. A focus metric on the object region exceeds your threshold on every frame; failures are reshot.
- Exposure. Clipped highlight and shadow pixels on the object stay below a share you set.
- Color. Color-chart patches in each session's reference frame fall within your color-difference tolerance.
- Calibration. Reprojection error is reported per camera and session and is within your limit.
- Coverage. Every item has every specified ring, azimuth and extra view, with no gaps in the angle log.
- Masks. On a random sample, checked masks agree with automatic masks above an overlap threshold you set.
- Identity. Identifiers validate, filenames parse, and no item appears under two IDs.
Limitations
Studio images show items alone, centered and evenly lit. Shelves, totes and workbenches add clutter, occlusion and mixed light, so plan in-situ images too; packshot vs in-situ images explains the gap. Photography records appearance and geometry only: it shows nothing of weight distribution, friction, deformation or how an item behaves when handled, which comes from interaction data. Packaging also changes, so a shoot dates quickly for fast-moving categories.
Permissions
Commissioned work makes ownership simpler but not automatic. Agree in writing who owns the images and derivatives, and whether the studio may reuse them. The packaging itself still carries brands' trademarks and artwork; rights in product images covers the diligence questions.
Questions to ask photography studios
- Can you log turntable angle and camera ID per frame automatically?
- How do you calibrate cameras, and will you deliver calibration files and reprojection errors?
- How do you keep lighting and white balance consistent across sessions?
- How do you photograph undersides, transparent items and reflective packaging?
- Do you produce foreground masks, and how are they checked?
- How do you track item identity, instance and state through capture and export?
- Will images be delivered under work-for-hire or assignment terms?
- How will you handle physical samples: receipt, condition records and return or disposal?
Discuss your requirements
Share the item list, the uses and the parameters above. Describe the object imagery you need to SourceX. If no existing asset fits, SourceX can scope a commissioned collection of new recordings. SourceX's role is sourcing, rights review, licensing and delivery coordination; it can also engage specialist capture partners, scoped case by case. It reviews the request and follows up about potential sources; a request does not guarantee a match. For licensed alternatives, see licensing product catalogs and descriptions and the humanoid training data guide.
Discuss your humanoid training-data requirements with SourceX
Sources
- ABO: Dataset and Benchmarks for Real-World 3D Object Understanding — turntable views for a minority of catalog products.
- OmniObject3D: Large-Vocabulary 3D Object Dataset for Realistic Perception, Reconstruction and Generation — real-captured videos with masks and camera poses.
- Google Scanned Objects: A High-Quality Dataset of 3D Scanned Household Items — structured-light scanning prepared for simulation.
- A Flexible New Technique for Camera Calibration — planar-pattern camera calibration.
- ClearGrasp: 3D Shape Estimation of Transparent Objects for Manipulation — why transparent items need dedicated handling.