--- title: "Semantic Segmentation Annotation for Regulated CV | YPAI" url: https://ypai.ai/annotation/semantic-segmentation-annotation-services/ description: "Pixel-accurate semantic, instance, panoptic, 3D and video segmentation for automotive, medical, satellite, agriculture and industrial CV. EEA processing." source: "src/copy/routes (route /annotation/semantic-segmentation-annotation-services/)" --- # Pixel-accurate segmentation. Your taxonomy. Your evidence pack. > Pixel-accurate semantic, instance, panoptic, 3D and video segmentation for automotive, medical, satellite, agriculture and industrial CV. EEA processing. Semantic, instance, panoptic, 3D, and video segmentation for regulated CV teams. EEA processing by default, EU AI Act Article 10 evidence, kappa-gated production. SAM-assisted on natural images, human-only on medical, aerial, 3D and microscopy. ## Compliance posture for segmentation training data. Article 10 enforcement begins 2 December 2027 for standalone Annex III systems, 2 August 2028 for AI embedded in regulated products. YPAI ships every segmentation engagement with the artefacts a regulated buyer needs in their file. - Gold set with documented injection rate - Boundary F1 / HD95 for medical work ### EU AI Act Article 10 - Data governance for high-risk CV systems. Annotation provenance, bias-examination notes, class taxonomy documentation. ### GDPR Articles 7, 28, 30 - Per-contributor consent. Processor agreement. Records of processing. 30-day erasure SLA. Sub-processor change notifications. ### EEA processing by default - Norwegian company. EEA processing by default. EEA-only sourcing on request. EU AI Act Article 10 evidence available per project. ## Five segmentation primitives. Which one you procure depends on what your model has to know. Ordering semantic masks when the model needs instance IDs means paying twice. - One class per pixel; no instance separation. Stuff classes (road, sky, vegetation) and thing classes (vehicle, pedestrian) treated identically. - Road scene parsing, organ regions, land-use classification - Each object gets a unique ID alongside its class. Counting becomes possible: 12 pedestrians, 4 cells, 3 defects per image. - Stuff classes get per-pixel labels. Thing classes get per-pixel labels plus instance IDs. One unified mask, two semantic regimes. - Full scene understanding, autonomous driving stack - Each LiDAR point or voxel labelled with a class. Used in automotive (street scenes) and volumetric medical (3D CT, MRI organs). - Per-frame masks with consistent instance IDs across time. Object tracking, surgical phase, action recognition, motion analytics. - Video annotation, surgical workflow, sports analytics ## Five verticals where pixel-class matters. Different industries, different ontologies, same delivery contract: EEA processing by default, kappa-gated production, Article 30 records on delivery. ### Automotive ADAS - Cityscapes-aligned taxonomy. Pedestrian, cyclist, vehicle, lane, traffic-sign. EU AI Act Article 10 evidence available per project. ### Medical imaging - Brain MRI tumour (BraTS protocol), kidney CT (KiTS), histopathology nuclei. DICOM SEG output, Dice + HD95 reporting. ### Satellite + aerial - Land-use classification, road extraction, building footprints, deforestation. SpaceNet schemas, GeoJSON delivery. ### Agriculture - Crop-vs-weed, fruit yield, canopy gap, disease zones. Multispectral capable. Drone + ground imagery. ### Industrial + robotics - Surface defect (scratch / dent / contamination), pick-and-place region masks, LiDAR fusion for warehouse / drone / construction. Quality gate masks. ## Every project clears the same six gates. Calibration before production. Documented mask-IoU thresholds. Human QA applied according to the agreed acceptance and sampling plan. The artefacts you need for your Article 10 file. Six gates. One trail of evidence. Every delivery. ### Schema design - Class taxonomy locked with your team. Aligned to public benchmark (Cityscapes, COCO-Stuff, BraTS, custom) or built fresh. Versioned. ### Annotation guideline - Edge cases enumerated per class. Boundary policy documented (snap-to-edge tolerance, occlusion handling, ambiguous class adjudication). ### Calibration round - Pilot batch on shared subset. Mask IoU between annotators computed. Disagreement patterns drive guideline refinement before scale. ### IAA gate - Production starts only when calibration clears the task-specific agreement, mask, boundary, and class thresholds in the acceptance plan. - Signed record that the gate was cleared ### Production labelling - SAM-assisted where domain-shift permits (natural images). Human-only on medical, aerial, 3D point cloud, and microscopy where SAM under-performs. ### QA + adjudication + delivery - Review coverage and gold-set injection follow the agreed acceptance and sampling plan. Per-annotator Dice gate. Delivery ships with per-class IoU, boundary F1 / HD95 for medical, and Article 30 records. - Final delivery pack with metrics report ## Six scenes. Point at one to see its annotation mask. Schematic previews. Production work is delivered against your raw imagery in your chosen taxonomy. - BraTS protocol: tumour core, oedema, necrosis - Land use: building, road, vegetation, water - Crop / weed / soil / canopy gap - Stuff + things + instance IDs ## Taxonomy-aligned with the corpora your model trained on. Public datasets are useful as taxonomies, baselines, and audit references. Most are research-licensed; production work runs against your own data under your engagement DPA. Licence status is part of the procurement record. Licence status reflects publicly stated terms at the dataset source. Check per engagement before any commercial training use. ### Cityscapes ### Mapillary Vistas ### BDD100K ### A2D2 ### COCO-Stuff ### ADE20K ### BraTS ### KiTS ### LIDC-IDRI ### SA-1B ### SemanticKITTI ### DAVIS ## Every delivery includes the records your Article 10 file needs. The records a regulated buyer expects, included with every segmentation engagement. ### Annotation guideline and boundary policy. - Versioned guideline with edge cases enumerated, examples per class, glossary of domain terms, and boundary tolerance rules. Updated as adjudication surfaces new patterns; every version preserved for audit. ### Gold set with documented injection rate. - Held-out gold set injected at 5 to 10 percent during production. The rate is documented for you; annotators are not told which items are gold. Per-annotator Dice gate. Calibration before queue entry, blind re-runs on failure. Quarterly refresh. ### Per-class IoU and boundary metrics. - Per-class IoU report on every delivery. Boundary F1 and HD95 for medical, fwIoU for satellite. Confusion matrix and top failure modes by frequency. Panoptic Quality for panoptic work. ### Mask IoU inter-annotator report. - Pairwise mask IoU between annotators on the calibration subset. Target: the substantial agreement band in Landis and Koch. Per-class breakdown and per-annotator over-time trend. ### Records of processing, DPA, and sub-processor list. - Article 30 records of processing, signed Article 28 DPA, lawful-basis documentation, 30-day erasure SLA, and full sub-processor list with Article 28(2) change notifications. All included with every engagement. ## Brief us. We reply inside one EU business day. Short brief now, deeper scoping in the reply. ## Where is our data processed? Norwegian company, EEA processing by default, EEA-only sourcing on request. 30-day GDPR Article 17 erasure SLA. Norwegian jurisdiction; transfer controls and member-state region defined per engagement. ## Is our data isolated from other customers? Yes. Per-engagement self-hosted CVAT or Label Studio project, dedicated access lists, no shared annotator pools across engagements. Infrastructure isolation and separation are specified per project in the DPA and security annex. ## What does your compliance posture cover? GDPR Article 28 processor obligations, EU AI Act Article 10 data governance, EEA residency by default, DPA-by-default, and isolation controls specified per project. Security artefacts package available on request. ## Do you train on customer data? Work product is customer-owned. YPAI keeps no reuse, resale or model-training rights. Sub-processor list disclosed at engagement start with Article 28(2) change notifications. ## How do you handle annotator wellbeing on sensitive content? Opt-in only for medical pathology, defect, surveillance footage, and content with potential distress. Screening, rotation off sensitive batches, no-penalty opt-out. Specifics documented in the project brief. ## Frequently asked questions