Autonomous Driving
LiDAR-camera fusion, 3D cuboid tracking, lane detection, traffic sign/light recognition, pedestrian tracking, and edge-case scenario annotation across ODD (Operational Design Domain) coverage matrices.
COMPUTER VISION
Image & Video Annotation
Services
Pixel-perfect annotations for object detection, segmentation, tracking, pose estimation, and 3D point cloud labeling — with measurable IoU thresholds, per-class accuracy tracking, and multi-tier QA at every step. From single-class bounding boxes to complex multi-sensor fusion annotation.
5M+
Images Annotated
IoU ≥ 0.92
Avg Quality Score
50K+
Frames/Week Capacity
L1→L2→L3
QA Pipeline
< 2px
Boundary Precision
25+
Automated QA Rules
TASK TYPES
Six Core Annotation Capabilities
Each task type includes specific accuracy benchmarks, throughput rates, and quality controls. All configurable to your project requirements.
2D axis-aligned bounding boxes, oriented (rotated) bounding boxes, and 3D cuboid annotations for object detection across all domains. Configurable overlap policies, occlusion flags, truncation percentage, and difficulty scoring per annotation.
TECHNICAL DETAILS
- Axis-aligned + oriented (rotated) bounding boxes
- Occlusion handling: visibility flags (full visible, partially occluded, heavily occluded)
- Difficulty scoring: easy/medium/hard per annotation for curriculum learning
- 3D cuboid annotation for LiDAR point clouds and stereo imagery
- Truncation percentage: 0–100% for objects at image boundaries
- Multi-class support: unlimited class hierarchy with parent-child relationships
PERFORMANCE
Accuracy
IoU ≥ 0.90 (standard), ≥ 0.95 (precision-critical)
Throughput
2K–8K boxes/annotator/day
Error Rate
< 2% false positives
Task Type 2
Semantic & Instance Segmentation
Pixel-level labeling for scene understanding and object separationPixel-perfect semantic segmentation, instance segmentation with unique object IDs, and panoptic segmentation combining both. Support for 100+ class taxonomies with polygon, brush, superpixel, and SAM-assisted annotation tools.
TECHNICAL DETAILS
- Semantic segmentation: every pixel assigned a class label
- Panoptic segmentation: combined stuff (background) + things (countable objects)
- Brush + superpixel tools for complex organic shapes (medical, agricultural)
- Instance segmentation: individual object masks with unique IDs
- Polygon annotation: sub-pixel vertex accuracy < 2px deviation
- SAM-assisted pre-annotation with human correction and validation
PERFORMANCE
Accuracy
mIoU ≥ 0.88 (standard), ≥ 0.92 (precision)
Throughput
200–500 masks/annotator/day
Boundary
< 2px boundary deviation
Task Type 3
Keypoint & Pose Estimation
Skeletal keypoints for pose estimation and landmark detectionConfigurable skeleton definitions for human pose (17–133 keypoints), hand tracking (21 points), facial landmarks (68–478 points), and custom articulated objects. Each keypoint includes visibility flags (visible, occluded, out-of-frame) and confidence indicators.
TECHNICAL DETAILS
- Human pose: COCO 17-point, MPII 16-point, or custom skeleton definitions
- Facial landmarks: 68-point (dlib), 478-point (MediaPipe), or custom configurations
- Visibility flags: visible / self-occluded / occluded by other / out-of-frame
- Hand tracking: 21 keypoints per hand with finger joint angles
- Animal pose: custom skeletons for dogs, horses, birds (AP-10K compatible)
- Inter-annotator consistency: OKS (Object Keypoint Similarity) ≥ 0.85
PERFORMANCE
Accuracy
OKS ≥ 0.85
Throughput
500–1.5K poses/annotator/day
Visibility
3-level flags per keypoint
Multi-object tracking (MOT) with persistent IDs maintained through occlusions, re-entries, and camera transitions. Keyframe annotation with linear/spline interpolation and manual correction. Support for single-object tracking (SOT), MOT, and multi-camera cross-view tracking.
TECHNICAL DETAILS
- Persistent object IDs: maintained through occlusions up to 60+ frames
- Re-identification: same object re-entering frame receives original ID
- Temporal action boundaries: activity start/end with ±1 frame precision
- Keyframe annotation + interpolation: linear and spline with human correction
- Multi-camera tracking: cross-view identity linking with shared ID namespace
- Track-level attributes: object class, behavior state, direction, speed estimate
PERFORMANCE
Accuracy
MOTA ≥ 0.90, IDF1 ≥ 0.85
Id Switches
≤ 0.5% per sequence
Throughput
500–2K video-min/week
Task Type 5
Image Classification & Tagging
Multi-label hierarchical classification with confidence scoringSingle-label, multi-label, and hierarchical classification with configurable confidence thresholds. Support for fine-grained recognition (breed identification, species classification), quality assessment (defect grading), and content moderation across millions of images.
TECHNICAL DETAILS
- Single-label classification with top-1 and top-3 predictions
- Hierarchical taxonomy: parent-child class relationships with inheritance rules
- Quality/condition assessment: 5-point Likert scales for subjective attributes
- Multi-label tagging: unlimited tags per image with confidence scores
- Fine-grained recognition: 500+ sub-categories within a domain
- Active learning integration: model-confidence-based sample routing
PERFORMANCE
Accuracy
≥ 97% top-1 classification accuracy
Throughput
5K–20K images/annotator/day
Agreement
κ ≥ 0.90
Task Type 6
3D Point Cloud & LiDAR Annotation
3D bounding cuboids, segmentation, and lane marking for autonomous systems3D bounding cuboid annotation in LiDAR point clouds with heading angle, velocity estimation, and multi-frame tracking. Semantic segmentation of point clouds, lane/road boundary marking, and sensor fusion annotation linking LiDAR to camera imagery.
TECHNICAL DETAILS
- 3D bounding cuboids: position, dimensions, heading angle (yaw, pitch, roll)
- Lane and road boundary annotation in 3D space
- Sensor fusion: linked annotations across LiDAR, camera, radar, and IMU
- Point-level semantic segmentation: ground, vehicle, pedestrian, vegetation, etc.
- Multi-frame tracking: 3D cuboid trajectories with velocity and acceleration estimates
- Timestamp synchronization: ≤ 10ms alignment across sensor modalities
PERFORMANCE
Accuracy
3D IoU ≥ 0.70 (standard), heading error < 5°
Throughput
200–500 frames/annotator/day
Fusion
Multi-sensor linked annotations
Industries
Where Our CV Annotations Are Used
Domain-specific annotation protocols for safety-critical and high-accuracy applications across six major verticals.
Medical Imaging
DICOM annotation for CT, MRI, X-ray, histopathology WSI, and ophthalmology. Organ segmentation, lesion classification, landmark detection, and measurement by board-certified radiologists and pathologists.
Retail Intelligence
Product recognition, shelf compliance analysis, planogram verification, visual search annotation, customer behavior tracking, and inventory management for retail AI systems.
Manufacturing QC
Surface defect detection, assembly verification, weld inspection, dimensional compliance, and quality grading for industrial quality control on high-speed production lines.
Surveillance & Security
Person detection, action recognition, anomaly detection, crowd analysis, license plate recognition, and multi-camera tracking with privacy-compliant annotation workflows.
Agriculture & Earth Observation
Drone and satellite imagery annotation for crop health monitoring, pest detection, land use classification, yield estimation, and infrastructure monitoring across growing seasons.
Quality
CV-Specific Quality Controls
Computer vision annotation demands pixel-level precision. Here's how we maintain it across large teams and complex projects.
Gold Set Calibration
IoU threshold validation against expert-labeled gold sets. Annotators must achieve 0.85+ IoU on the gold set before touching production data. Gold sets refreshed 10% monthly to prevent memorization.
IoU ≥ 0.85 on gold set required for production access
Multi-Reviewer Pipeline (L1→L2→L3)
L1 annotators produce initial labels. L2 reviewers audit 100% of output (not sampling). L3 adjudicators resolve disagreements and edge cases. Complex annotations always pass through at least two sets of eyes.
100% L2 review coverage (no sampling)
Per-Class Metrics Tracking
We track accuracy, precision, recall, and IoU per class — not just aggregate metrics. Rare but critical classes (pedestrians, small objects, defects) get extra QA attention and dedicated review queues.
Per-class IoU, precision, recall dashboards updated hourly
Automated Consistency Checks
Rule-based validation catches common errors: overlapping bounding boxes, missing labels, impossible polygon shapes, label-class mismatches, and boundary violations. Errors flagged before human review.
25+ automated validation rules per project
Drift Detection & Alerts
Statistical monitoring across batches detects quality drift before it impacts your model. Batch-over-batch IoU, accuracy, and error-type distributions are compared. Automatic alerts trigger recalibration when drift exceeds ±2%.
Drift alerts at ±2% threshold, recalibration within 4 hours
Edge Case Libraries
Growing libraries of ambiguous and edge-case examples with documented resolution decisions. Used for annotator training, guideline refinement, and quality audit. Every edge case becomes a reusable training asset.
100+ documented edge cases per mature project
COMPATIBILITY
Export Formats & Tool Integration
We deliver data in any format your training pipeline needs, and work with the tools you already use.
Export Formats
COCO JSON
Pascal VOC XML
YOLO TXT
TFRecord
Cityscapes
KITTI
MOT Challenge
NiFTI/NRRD
LabelMe JSON
Custom Schema
Tool Compatibility
CVAT
Label Studio
Labelbox
V7
Supervisely
Roboflow
VGG Image Annotator
Custom Platforms
COMPARISON
UTL CV Annotation vs. Typical Providers
| Capability | UTL Data Engine | Typical Providers |
|---|---|---|
| Gold set calibration with monthly refresh | Initial only | |
| 100% L2 review (not sampling) | 5–10% sampling | |
| Per-class IoU/precision/recall tracking | Aggregate only | |
| 3D cuboid + multi-sensor fusion | 2D only | |
| Automated consistency validation (25+ rules) | Basic checks | |
| Drift detection with auto-alerts | ||
| Edge case libraries with decision docs | ||
| SAM-assisted pre-annotation | ||
| Domain-trained annotators (20+ hr onboarding) | 2–4 hr training | |
| Sub-pixel polygon accuracy (< 2px) | 5–10px typical |
“UTL reduced our annotation rework by over 50%. Their gold set calibration and per-class IoU tracking caught quality issues that our previous vendor missed entirely. The 100% L2 review coverage is what makes the difference — no more sampling-based QA surprises.”
FAQS
Computer Vision Questions
Default targets: IoU ≥ 0.90 for bounding boxes, mIoU ≥ 0.88 for segmentation, OKS ≥ 0.85 for keypoints, 3D IoU ≥ 0.70 for cuboids. For precision-critical projects (medical, AV), we configure higher thresholds (≥ 0.95 for 2D, ≥ 0.92 for segmentation). All thresholds are agreed during scoping and validated against gold sets.
Yes. We annotate LiDAR point clouds with 3D cuboids, point-level segmentation, and trajectory tracking. Multi-sensor fusion links annotations across LiDAR, camera, radar, and IMU with ≤ 10ms timestamp synchronization. Our teams are trained on common AV sensor configurations.
Steady-state: 500–2K video-minutes/week per project with MOTA ≥ 0.90 tracking accuracy. For burst projects, we scale to 5K+ video-minutes/week with parallel annotator teams. Multi-object tracking, re-identification, and temporal action segmentation are all supported.
We build edge-case libraries during guideline creation (100+ documented cases per mature project) and maintain a structured decision taxonomy. Occluded objects receive visibility flags (fully visible, partially occluded, heavily occluded) and truncation percentages. New edge cases discovered during production are documented and added to the library with resolution decisions.
Yes. We use Segment Anything Model (SAM) for pre-annotation to accelerate segmentation tasks. Human annotators review, correct, and validate all SAM-generated masks. This hybrid approach typically delivers 2–3× throughput improvement while maintaining our quality standards.
Scoping + guideline design: 3–5 days. Team assembly + calibration: 5–7 days. Pilot (1K–5K samples): 5–10 days. First labeled batch by Day 20. Full production velocity by Day 25. We maintain pre-qualified teams across major domains for faster ramp-up.
Need Pixel-Perfect Annotations?
Let's discuss your computer vision data pipeline — from task design to quality-assured delivery. We'll scope a pilot within 48 hours.
