IMAGE ANNOTATION / MEASURED IN IoU

Pixel-accurate image annotation, verified by humans in the EEA

Bounding boxes through panoptic masks, reported in IoU, COCO AP, and PQ, with 100% human QA and GDPR-native EEA processing.

Scope your annotation pilot Discuss your dataset
  • 100% human QA
  • EEA-resident (Norway AS)
  • 40,000+ contributors
  • 30-day erasure SLA
Annotation overlay client data masked
bbox segmentation mask IoU, COCO AP, PQ

Built in Europe for regulated computer vision

  • EU AI Act Article 10 Data-governance dossier per dataset for high-risk computer vision.
  • GDPR Article 9 and 25 Biometric distinction handled correctly, privacy by design in-tool.
  • EEA-resident Norway AS jurisdiction, self-hosted European processing, outside US CLOUD Act reach.
  • DPA always Signed data processing agreement and sub-processor list with every engagement.

THE HARD 20%

Occlusion, tiny objects, thin boundaries, ambiguous classes

Cheap labeling reads fine on the easy 80% and collapses on the rest. We build an edge-case rulebook and decision-tree taxonomy so the hard cases stay consistent.

  • Occlusion and amodal completion

    Heavily occluded objects get their hidden geometry inferred within documented domain caps, not guessed, so detectors learn the real extent.

    Amodal rules in the guideline

  • Tiny and far-field objects

    Sub-30-pixel objects get oriented-bbox and sub-pixel placement instead of being dropped, holding small-object AP.

    AP by object size reported

  • Thin boundaries

    Hair, foliage, wire, and transparent edges get spline tracing so mask-IoU and boundary F-score survive at the contour.

    Boundary F-score gate

  • Ambiguous classes

    A decision-tree taxonomy resolves ambiguous classes the same way every time, lifting inter-annotator agreement on the cases that otherwise drift.

    Kappa lifted on ambiguous classes

Schematic previews. Production work runs against your raw data under your engagement DPA, with real annotated frames delivered to your team, not published here.

Same scene, three geometries Schematic
bounding box polygon mask

EVERY GEOMETRY

From bounding boxes to panoptic masks, the primitive your task needs

The same scene, annotated three ways. A mismatched primitive is a deployment-only failure, so we pair every geometry with the model task it feeds.

  • Bounding box and oriented box axis-aligned and rotated
    Object detection (YOLO, Faster R-CNN), counting, oriented small objects.
  • Polygon tight irregular outlines
    Irregular boundaries, holes, and nested regions.
  • Polyline and keypoint continuity and pose
    Lane lines and wires with topological continuity, pose and landmarks.
  • Semantic and instance mask per-pixel labels
    Dense prediction, drivable area, per-object masks.
  • Panoptic stuff, things, and instance IDs
    Full scene understanding, PQ-evaluated dense prediction.

QUALITY YOU CAN AUDIT

IoU, mAP, PQ, and kappa, reported per project

We agree the targets up front and report against them. A bounding box can read 0.80 IoU while the mask underneath reveals 0.45, so we measure the mask, not just the box.

Metric What it measures Threshold convention
IoU (intersection over union) Overlap over union of predicted and ground-truth regions. 0.5 acceptable, 0.75 high-precision, 0.5:0.95 the COCO AP protocol
COCO AP Average precision across IoU thresholds, reported by object size. AP at 0.5:0.95, AP50, AP75, AP-small, AP-medium, AP-large
Boundary F-score Contour quality within a 2-pixel tolerance band. Reported for thin-boundary and dense-segmentation work
Mask IoU vs box IoU The gap that reveals where cheap geometric labeling hides boundary failure. Mask IoU and Dice reported alongside box IoU
PQ (panoptic quality) Segmentation quality times recognition quality for panoptic tasks. PQ equals SQ times RQ
Kappa (Cohen and Fleiss) Inter-annotator agreement, two annotators or multi-pass. Landis-Koch 0.61 to 0.80 substantial, 0.81 to 1.00 almost-perfect
Mask IoU vs box IoU, illustrative of method Standard definitions; not a published YPAI score

Cheap geometric labeling can hold the box and lose the boundary. We measure the mask, report the gap, and agree the targets up front.

DENSE SEGMENTATION

Dense per-pixel segmentation, accelerated by models, corrected by humans

Fine pixel labeling is the most expensive primitive (Cityscapes reported about 1.5 hours per image for fine labels). Model-assist accelerates it but smooths fine boundaries, so humans refine the contour to hold mask-IoU.

  1. 01 Raw scene

    The unlabeled source image before annotation.

  2. 02 Model-assisted pre-label

    A SAM or interactive-segmentation draft mask, fast but boundary-smoothed.

  3. 03 Human-corrected delivery

    Boundary-refined per-pixel mask, measured in mask-IoU and boundary F-score.

Schematic previews. Cityscapes is a public dataset anchor for fine-label cost, not a YPAI throughput SLA. Production runs against your raw data under your engagement DPA.

100% HUMAN QA

Stable quality from the first thousand images to the ten-millionth

Quality drift from annotator churn and fatigue is the real failure mode on long projects. We run a living QA system, not a one-off pass, and the team labeling your data is a named cohort, not an anonymous crowd.

  1. 01

    Gold-set ground truth

    Held-out expert-labeled frames seeded as honeypots so every annotator is measured against known-correct ground truth continuously.

    Gold-set agreement report HoneypotGround truth
  2. 02

    Consensus multi-pass

    Multiple annotators on the same frames, keeping elements at a documented agreement threshold, with kappa tracked per class.

    Consensus and kappa report ConsensusKappa
  3. 03

    Audit sampling

    Statistical audit sampling with acceptance thresholds set tighter for safety-critical and ambiguous classes.

    Audit acceptance report AQLCritical classes
  4. 04

    Drift control

    Continuous sampling and dashboards across the project, with annotator retraining when class-level agreement slips, so quality holds at scale.

    Drift-monitoring dashboard DriftRetraining

Four gates, one named cohort, quality that holds from the first thousand images to the ten-millionth.

BUILT FOR YOUR DOMAIN

Per-vertical primitives and the metric your industry gates on

Different verticals measure vision differently. We map each to its dominant primitive, the metric it gates on, and the data-handling concern that decides the deal.

Vertical Dominant primitive Gates on Procurement concern
Retail and e-commerce Bounding box, attributes Precision Product detection and planogram attributes.
Agriculture and drone Polygon, semantic seg Aerial coverage Crop, weed, and disease from multispectral aerial imagery.
Manufacturing defect Polygon, mask Recall Surface defects and rare-anomaly coverage.
Geospatial and satellite Semantic, instance seg mask-IoU Building and road polygons and land-cover.
In-cabin and driver monitoring Keypoint, bounding box Gaze and pose Occupant monitoring; perception is supporting, not a primary claim.

PRIVACY BY DESIGN

Faces and plates are personal data from the first upload, and we treat them that way

An identifiable face is personal data from upload. It becomes Article 9 special-category only when processed to extract a template for unique identification, not when you simply label person. We are your processor and build to that distinction.

  • GDPR Article 9 (biometric) Labeling a person is not biometric processing; template extraction for unique identification is. We document the controller Article 6 and, where relevant, Article 9 basis.
  • GDPR Article 25 (privacy by design) In-tool face and license-plate blur before annotators see the image, pseudonymized exports, and data minimization by default.
  • Erasure and consent Consent withdrawal and a 30-day erasure SLA across systems, with a signed DPA on every engagement.
Request the DPA We map the data-handling pack to your image sources, special-category exposure, and jurisdiction.

ARTICLE 10 DOSSIER

An EU AI Act Article 10 data-governance dossier with every dataset

Article 10 cascades to the annotation supply chain for high-risk systems. We map each governance dimension to a practice and ship the dossier attachable to your technical documentation, not as an add-on.

Article 10 dimension YPAI practice Dossier artifact
Representativeness Sampling frames across geography, weather, and lighting, with distribution statistics so the dataset reflects deployment. Distribution-statistics report
Error examination Inter-annotator agreement plus per-class precision and recall plus IoU monitoring, documented per delivery. Per-delivery error report
Bias mitigation Subgroup stratification and distribution statistics so under-represented segments are examined, not buried. Bias-examination notes
Data lineage Versioned guidelines, schema, lineage logs, and workforce profiles, preserved for audit. Provenance dossier

Attachable to your technical documentation. Compliance is evidenced through named regulations and audit-ready documentation, not third-party certification badges.

EEA-RESIDENT

Your image data never leaves the EEA, with architecture your security team can review

US CLOUD Act exposure means even a US vendor with an EU subsidiary can be compelled to hand over data. We are a Norwegian company processing on self-hosted European servers, outside US CLOUD Act reach.

Data stays in the EEA.

Self-hosted European processing with no non-EEA data flows and no non-EEA support access without your explicit written instruction.

Norway AS jurisdiction.

YPAI is a Norwegian Aksjeselskap, Bronnoysund organisasjonsnummer 928 805 735, outside US CLOUD Act reach.

Reviewable architecture.

Hosting and network architecture your security team can review, so EEA residency is verifiable, not just asserted.

No accidental egress.

No global load-balancing out of the EEA and no non-EEA support access, the failure mode European buyers most often hit with global-cloud vendors.

OBJECTIONS

Formats, model-assist, residency, consent, erasure

Formats What export formats do you deliver?
COCO (including panoptic), YOLO, Pascal VOC, CVAT, PNG masks, and custom JSON or your schema. Customer-owned work product, no reuse rights retained.
Model-assist If you use model-assist, who is accountable for the label?
Model labels are suggestions only. Every label is human-verified, edge cases and safety-critical classes get double-blind human review, and every change is logged, so a human stays accountable for the delivered geometry.
Residency Where is our image data processed?
EEA-resident. Norwegian company, self-hosted European servers, EEA contributor network, outside US CLOUD Act reach. No non-EEA support access without your written instruction.
Consent How do you handle consent and special-category image data?
We are your processor and document your Article 6 and, where relevant, Article 9 basis. Faces and plates are blurred in-tool before annotators see them, exports are pseudonymized, and consent withdrawal is supported.
Erasure What is your erasure commitment?
A 30-day erasure SLA across systems for image and annotation deletion, with a signed DPA on every engagement and a sub-processor list with change notifications.

SCOPE A PILOT

Scope a metric-reported image annotation pilot

We return a sample set with the IoU, AP, or kappa report you specify, so you can verify quality before you scale.

  • Metric-reported pilot
  • Measured in IoU and COCO AP
Scope your annotation pilot Read the data residency brief

If image annotation is not the right fit for your project, we will say so directly.