TRACKING / VOS / ACTION / POSE

Video annotation,tracked across every frame.

Multi-object tracking, video object segmentation, action localization, pose, re-ID. SAM 2 assisted, kappa-gated, HOTA-reported per delivery.

  • HOTA + IDF1 reporting
  • EEA-resident
Scope a Data Project See sample annotations
FORMAT MOT + COCO video

Schematic. Production runs against your raw footage, multi-camera, with SAM 2 tracker-assist.

PROCUREMENT READINESS

Compliance posture for video training data.

Article 10 + Article 5(1)(f) emotion-recognition restrictions effective 2026. GDPR Article 9 biometric category applies. Every delivery ships with the audit pack.

Per-delivery artefact pack

Annotation guideline (versioned)
Face / plate de-identification log
HOTA + IDF1 report
Track continuity + ID-switch rate
Records of processing (Article 30)
Signed DPA + sub-processor list

EU AI Act Article 10 + 5

Data governance for high-risk video. Annex III restrictions on workplace + education emotion recognition (effective 2 Feb 2026).

GDPR Articles 9, 28, 35

Biometric special category. Processor agreement. Article 35 DPIA for systematic monitoring. 30-day erasure SLA.

EEA-resident processing

Norwegian company structure. EEA contributor network. EEA infrastructure. Outside US CLOUD Act reach.

Request a Procurement Readiness Brief →

We map the evidence package to your data, risk class, and deployment environment.

PRIMITIVE TYPES

Five video annotation primitives.

Different perception stacks need different temporal annotation. The double-spend failure mode is ordering MOT when you needed VOS with mask precision.

MOT (tracking)

bbox + persistent ID

Frame-by-frame bounding box with persistent instance ID. Identity preservation across occlusion and re-entry.

USED FOR Person counting, vehicle tracking, surveillance

VOS (segmentation)

per-frame masks + ID

Per-frame pixel-precise masks with continuous instance IDs. The DAVIS / YouTube-VOS task.

USED FOR Sports, broadcast, content moderation

Action recognition

temporal localization

Start/end timestamps for actions or activities. ActivityNet / AVA protocols. mAP at tIoU thresholds.

USED FOR Sports, surveillance, surgical phase

Pose tracking

skeleton across frames

Per-frame joint keypoints (typically 17-30) tracked with persistent person IDs. COCO OKS evaluation.

USED FOR Sports analytics, medical rehab, animation

Re-identification

cross-camera ID

Same identity preserved across non-overlapping cameras. Critical for multi-site surveillance and retail.

USED FOR Multi-camera surveillance, retail dwell, sports

WHERE VIDEO IS PROCURED

Five verticals where temporal accuracy matters.

Each industry has its own ontology and regulatory frame. EEA-resident annotation, DPIA-aligned, Article 30 records per delivery.

Automotive (in-cabin + outward)

Driver monitoring, occupant detection, gaze, gesture (in-cabin). Multi-frame perception, prediction (outward). UN R157 + DMS DPIA.

DMSALKSDPIA

Surveillance + security

Object tracking, person re-identification, event detection. Article 35 DPIA mandatory. Retention limits, signage requirements.

ReIDDPIAMOT

Surgical video

Phase recognition, tool tracking, anatomy delineation. JIGSAWS / Cholec80 protocols. HIPAA + GDPR Article 9 patient consent.

PhaseToolsCholec80

Sports analytics

Player tracking, ball, action recognition, formation analysis. SoccerNet / NBA / FineGym protocols. Athlete-data agreements.

SoccerNetPoseAction

Retail + content

Queue analysis, dwell-time, theft prevention. Content moderation, shot detection. Legitimate-interest balancing tests.

DwellModerationShot
Discuss your video modality

Or describe the video type and YPAI will scope ontology, kappa target, and delivery format.

HOW WE LABEL

Every project clears the same six gates.

De-identification policy decided BEFORE annotation. HOTA + IDF1 thresholds documented. SAM 2 tracker-assist with human review.

01

Schema + de-id policy

Class taxonomy, ID continuity rules, face / plate / patient de-identification policy locked with your team.

Deliverable: Versioned schema + de-id policy

02

Annotation guideline

Class definitions, ID continuity across occlusion, partial-visibility policy, action boundary tolerance.

Deliverable: Annotation guideline document

03

Calibration round

Pilot batch on shared subset. Per-frame mIoU + track-level IDF1 between annotators. Disagreement patterns drive refinement.

Deliverable: Calibration HOTA / IDF1 report

04

IAA gate

Production starts only when calibration clears HOTA 0.65+, IDF1 0.75+ on the schema. Lower thresholds for surveillance far-field allowed.

Deliverable: Gate-pass attestation

05

Production with SAM 2 + tracker-assist

Keyframe annotation with SAM 2 promptable tracking + interpolation. Cross-frame consistency check. Single-tenant CVAT video.

Deliverable: Annotated videos + tracks

06

QA + adjudication + delivery

HOTA, IDF1, ID-switch rate, track length distribution. De-identification audit. Article 30 records, signed DPA, sub-processor list.

Deliverable: Final delivery + metrics report

Six gates. One trail of evidence. De-identification policy decided upfront.

Start the scoping process →

WHAT WE DELIVER

Hover any scene to see the annotation.

Schematic previews. Production work runs against your raw footage, multi-camera, with SAM 2 tracker-assist.

HOVER reveal annotation

Automotive in-cabin DMS

Driver, occupant, gaze, gesture

HOVER reveal annotation

Outward driving

Vehicle, pedestrian, cyclist, lane

HOVER reveal annotation

Surgical video

Tool, anatomy, phase

HOVER reveal annotation

Sports tracking

Player, ball, formation

HOVER reveal annotation

Surveillance re-ID

Person, cross-camera ID

HOVER reveal annotation

Retail dwell

Person, queue, dwell-zone

PUBLIC BENCHMARK COVERAGE

Taxonomy-aligned with the corpora your model trained on.

Most video benchmarks are research-licensed. Production work runs against your own footage under your engagement DPA.

MOT Challenge (MOT17/20)

Research only

MOT tracking

Samples
14 + 8 sequences
Schema
MOT format
License
CC BY-NC-SA
Taxonomy reference

DAVIS

Research only

VOS challenge

Samples
150 video sequences
Schema
Per-frame masks
License
Non-commercial
Evaluation reference

YouTube-VOS

Research only

Large-scale VOS

Samples
4,453 videos
Schema
COCO mask format
License
CC BY-NC-SA
Taxonomy reference

Kinetics 700

Research only

Action recognition

Samples
650K clips, 700 classes
Schema
10-sec clips
License
CC BY 4.0 annotations
Taxonomy reference

AVA

Research only

Action localization

Samples
80 atomic actions
Schema
Per-frame bbox + action
License
CC BY 4.0
Evaluation reference

Cholec80

Research only

Surgical phase

Samples
80 cholecystectomy videos
Schema
Phase + tool labels
License
Research only
Domain reference

License status reflects publicly stated terms. Verify per engagement before commercial training use.

WHAT YOU RECEIVE

Every delivery ships with the artefact pack your Article 10 file needs.

The records a regulated video buyer expects. No upgrade tier, no separate request.

Annotation guideline + de-id policy.

Versioned guideline with class definitions, ID continuity rules, occlusion policy. Face / plate / patient de-identification policy documented. Every version preserved for audit.

HOTA + IDF1 + per-class metrics.

HOTA (Higher Order Tracking Accuracy), IDF1, MOTA per delivery. Per-class breakdown. Per-frame mIoU for VOS. Temporal mAP for action localization.

Track continuity + ID-switch rate.

Per-track length distribution. ID switches per minute. Fragmentation rate. Re-identification accuracy across cameras (when applicable).

De-identification audit.

Face detection + replacement log. License plate blur audit. Patient identifier handling. Skeleton-only delivery option for sensitive surveillance / surgical.

Records of processing, DPA, and sub-processor list.

Article 30 records of processing, signed Article 28 DPA, Article 35 DPIA scope where relevant, lawful-basis documentation, 30-day erasure SLA, and full sub-processor list.

START A PROJECT

Brief us. We reply within one business day.

Short brief now, deeper scoping in the reply.

Capability lanes (NER, RLHF, etc.), languages, volume, regulatory context.

QUESTIONS BUYERS ACTUALLY ASK

Frequently asked questions

De-identification policy decided BEFORE annotation. Face detection + replacement (not blur, which is reversible). License plate blur audit. Patient identifier scrubbing for medical. Skeleton-only delivery option for sensitive content.

Yes. Cross-camera person and vehicle re-ID. Proprietary multi-camera viewer with synchronised timelines. ID continuity across camera handoff.

EEA-resident. Norwegian company, EEA contributor network, EEA infrastructure. 30-day GDPR Article 17 erasure SLA. Outside US CLOUD Act reach.

GDPR Article 28 processor obligations, EU AI Act Article 10 data governance, GDPR Article 9 biometric handling, EEA-residency, DPA-by-default, and single-tenant isolation. Security artefacts package available on request.

COCO format with video extension, MOT format, custom JSON, AV2 SDK for automotive. Per-frame masks for VOS. ActivityNet temporal format for action recognition. Customer-owned work product.

GDPR-Native EU AI Act Article 10 EEA Operations Consent Evidence

Brief us on your video project.

One business day reply. NDA on request. DPA included.

SAM 2 + tracker-assisted HOTA per delivery
Or connect on LinkedIn →

Your information is never shared. We respond with the next scoping step.