Vision AI · ADAS · Pillar 1 — AI Training

4 Million 3D Cuboids Delivered: Building ADAS-Ready Annotation Infrastructure for Autonomous Driving Perception

To train autonomous driving perception systems at production scale, a client needed 3D cuboid annotations across ego-vehicle camera feeds — covering vehicles, pedestrians, and more — with consistent accuracy at 100K+ images per month. Oprimes deployed 75 annotators, 20 QC inspectors, and a two-layered quality framework to deliver 4 million verified annotations over 8 months.

Schedule a Demo Explore Case Studies
[ Annotation Viewer · cuboid_preview.3d ] 3D Lock
DIM4.6m × 1.8m × 1.5m YAW38°
VEH_041 · cuboid locked
PED_017 VEH_042 PED_018 4M+ total annotated
4M+
annotations delivered
100K+
images/month throughput
[ Annotations ]
4M+
Accurate 3D cuboid bounding box annotations delivered across vehicles, pedestrians, and more
[ Throughput ]
100K+
Annotated images delivered per month, sustained consistently across an 8-month engagement
[ Team Scale ]
95
Annotators and QC inspectors deployed — 75 annotators and 20 dedicated quality inspectors
[ Duration ]
8mo
Months of continuous, uninterrupted ADAS-ready annotation delivery
The Challenge

Annotating 3D cuboids around moving objects — vehicles, pedestrians, cyclists — from ego-vehicle camera feeds at 100K+ images per month required sustained accuracy across varied object types, camera angles, and real-world scene complexity. Managing annotation quality across 4M+ instances at that throughput, while maintaining 3D spatial accuracy, was the central operational challenge.

The Approach

Deployed 75 specialized annotators and 20 QC inspectors with a two-layered quality framework: 100% QC in Layer 1 and 30% sampling in Layer 2. A one-week incubation phase aligned team skills and guidelines before full-scale execution. The client's platform was used throughout, tailored to cuboid annotation requirements.

The Outcome

Over 4 million accurate 3D cuboid annotations delivered across 8 months at sustained 100K+ image-per-month throughput. The resulting dataset — ADAS-ready, layered-QA validated — provides the training and validation foundation for autonomous driving perception systems designed to operate in real-world traffic environments.

[ The Challenge ]

Sustaining 3D Accuracy Across 4 Million Instances at 100K Images Per Month

Three-dimensional cuboid annotation is substantially more complex than 2D bounding box labeling. Where a 2D box requires two points, a 3D cuboid requires the annotator to accurately estimate object depth, orientation, and spatial extent across three axes simultaneously — from a single 2D camera image. In real traffic conditions, objects overlap, partially occlude each other, appear at oblique angles, and vary in size across a single frame. A vehicle partially blocked by another vehicle requires the annotator to infer the hidden geometry and draw a cuboid that accurately represents the full physical object, not just the visible portion.

Sustaining that accuracy at 100K+ images per month for 8 months — across vehicles, pedestrians, cyclists, and other object classes — required a team and a workflow that could maintain consistent labeling quality under production-scale volume pressure. Quality degradation under throughput pressure is the canonical failure mode of large annotation projects: the team speeds up, precision erodes, and the resulting dataset introduces systematic perception errors that only surface when the model is tested in real conditions.

Multi-object tracking added a further dimension of complexity. Camera feeds from front and side angles needed aligned annotations across sequential frames — meaning annotators had to maintain object identity and 3D spatial consistency not just within a single frame, but across the full temporal sequence of an ego-vehicle camera recording.

[ What Was at Stake ]
  • 3D cuboid inaccuracies in training data produce perception systems with miscalibrated depth and orientation estimates — a direct safety risk for ADAS systems operating in real traffic environments
  • Annotation quality degradation under throughput pressure at 100K+ images/month would require costly re-labeling cycles and delay model training timelines
  • Multi-object tracking failures across sequential frames introduce object ID inconsistency in training data — training the model to lose track of objects precisely when they overlap or occlude, the highest-risk scenario in real driving
  • Team misalignment with client platform-specific cuboid tooling before the engagement started would embed workflow errors at the foundation of an 8-month production run
[ The Approach ]

75 Annotators, 20 QC Inspectors, Two-Layered Quality Framework, 8-Month Execution

01
Use Case Scoped and Team Structured

Defined the annotation requirement — 3D cuboid bounding boxes across vehicles, pedestrians, and other moving objects from ego-vehicle camera feeds — and structured a team of 75 annotators and 20 QC inspectors capable of sustaining the required throughput without compromising 3D accuracy.

02
One-Week Incubation Phase

Ran a structured one-week incubation before production began — aligning team skills with the client-provided annotation platform, calibrating understanding of cuboid guidelines across all annotator and QC roles, and identifying any skill gaps requiring targeted remediation before the production clock started.

03
Two-Layered QA Framework Deployed

Layer 1 applied 100% QC review of every annotated instance — every frame, every cuboid — with QC inspectors validating 3D accuracy, object class labeling, and spatial consistency before output moved to the delivery queue. Layer 2 applied 30% sampling QC as an additional quality gate, catching systematic issues that individual frame review at scale can occasionally miss.

04
Client-Platform Integration Maintained

Operated entirely within the client's proprietary cuboid annotation platform tailored to their ADAS data format requirements — ensuring output was compatible with the client's training data pipeline from day one, with no format conversion overhead or tool-switching risk across the 8-month delivery.

05
Continuous Throughput Maintained — 8 Months

Delivered 100K+ annotated images per month consistently across all 8 months of the engagement — managing team capacity, annotation workload distribution, and QC pipeline throughput to ensure uninterrupted delivery against the client's model training schedule.

[ Services Deployed ]
3D Cuboid Annotation

Large-scale 3D cuboid bounding box annotation of vehicles, pedestrians, and moving objects from ego-vehicle camera feeds across front and side angles.

Two-Layered QA Framework

100% Layer 1 QC on every annotated instance plus 30% Layer 2 sampling — ensuring consistent labeling accuracy across 4M+ annotations at production throughput.

Multi-Object Tracking Support

Aligned 3D cuboid annotations across sequential frames from front and side camera angles, maintaining object identity consistency throughout temporal sequences.

[ Team & Process Details ]
75 specialized annotators deployed for 3D cuboid labeling across all object classes
20 dedicated QC inspectors — independent of annotation team — running dual-layer quality verification
Layer 1: 100% QC review of every annotated instance before delivery acceptance
Layer 2: 30% sampling QC applied as a secondary quality gate across all delivered batches
1-week incubation phase before production — platform calibration and guideline alignment across full team
Client-provided platform used throughout — tailored for 3D cuboid annotations in ADAS data format
[ Results & Impact ]

4M+ Annotations. 100K Images/Month. 8 Months Uninterrupted.

4M+
3D Cuboid Annotations

Accurate, ADAS-ready 3D cuboid bounding boxes across vehicles, pedestrians, and moving objects — all passing two-layer QC validation.

100K+
Images Per Month

Sustained annotation throughput maintained consistently across all 8 months — matching the client's model training data intake schedule without gaps.

2
QA Layers Applied

100% Layer 1 QC plus 30% Layer 2 sampling — ensuring consistent labeling accuracy was maintained at scale across all annotation batches.

8mo
Continuous Delivery

Uninterrupted annotation and QC delivery across 8 months — providing the client a reliable, high-volume data supply for ongoing model training cycles.

The 4 million 3D cuboid annotations delivered through this engagement represent more than a volume milestone — they represent the training foundation for a perception system that will make real-time decisions in real traffic environments. The two-layered QA framework was the mechanism that made that scale achievable without quality compromise: by applying 100% review at Layer 1, no annotation uncertainty reached the delivery queue unchecked; by applying 30% sampling at Layer 2, systematic issues were caught before they compounded across batches. The combination produced consistent labeling accuracy at a throughput level that most annotation vendors struggle to sustain past month two.

The one-week incubation investment proved its value across the full 8-month run. Teams that begin production annotation without platform-specific calibration typically produce a first-month rework wave that consumes the capacity gains of starting fast. The Oprimes incubation approach front-loaded alignment cost at a scale where it was cheap — before production began — and eliminated it as a recurring cost during delivery.

-->
[ Key Takeaways ]

What This Engagement Teaches About ADAS Annotation at Production Scale

Two-Layer QA Is the Only Reliable Architecture for 4M-Instance Scale

Single-layer QC — whether 100% review or statistical sampling — misses different categories of errors. 100% review catches individual frame-level mistakes but can build in systematic biases when the same reviewers develop shared blind spots. Statistical sampling catches systematic issues but misses individual outliers. Running both layers in sequence eliminates the failure modes of each: every instance is reviewed, and systematic patterns are detected before they propagate.

Incubation Investment Pays Back Across the Full Engagement Length

A one-week incubation phase before production annotation begins costs the equivalent of one week's throughput — a small fraction of an 8-month engagement. The return is a team that starts production already calibrated to the platform, the annotation guidelines, and the client's specific cuboid requirements. Without it, calibration happens in production, where rework is expensive, throughput suffers, and quality issues from the first production batches can propagate into training data before they're caught.

Throughput Consistency Matters as Much as Throughput Volume for Model Training

An annotation partner who delivers 200K images in month one and 50K in months two through eight is not a production-scale partner — they are a burst service. ADAS model training pipelines require predictable data intake rates to maintain training velocity. Consistent 100K+ images per month across 8 months is a fundamentally different capability than average throughput of the same number: it is the difference between a data supply chain and a data spike.

[ FAQ ]

Questions About This Engagement?

Common questions about 3D annotation, ADAS datasets, and high-volume AI training data delivery.

Ready to scale your annotation programme? We deliver 100K+ annotated frames per month. Talk to us

A 3D cuboid annotation is a three-dimensional bounding box drawn around an object in a camera or lidar frame — capturing not just the 2D footprint of a car, pedestrian, or cyclist, but their precise spatial extent, orientation, and position relative to the ego vehicle. ADAS perception systems use these annotations to train models that estimate distance, predict trajectory, and make real-time collision decisions. Without accurate 3D cuboids, the model cannot determine whether an object is 10 metres away or 40.

Volume and accuracy are typically in tension in annotation programmes — faster throughput degrades quality. Oprimes resolved this through a combination of structured annotator training on the client's specific object taxonomy, a two-layered QA process (peer review followed by senior QA specialist review), and real-time accuracy monitoring that flagged annotators whose inter-rater agreement dropped below threshold mid-cycle. Batches that failed QA were rerouted to a correction queue rather than delivered.

Layer one is peer review: every annotated frame is reviewed by a second annotator who checks for dimensional accuracy, correct object classification, and proper orientation tagging. Layer two is specialist review: a senior QA reviewer audits a statistically significant sample from every annotator's daily output, scoring against a precision rubric. Frames that fail either layer are rejected and re-annotated. The client receives only output that has passed both layers.

For this engagement, Oprimes delivered 4 million+ 3D cuboid annotations across 8 months — sustaining 100,000+ images per month throughout. Ramp-up (training annotators, calibrating quality thresholds, establishing the data pipeline) took 3 to 4 weeks. Full throughput was reached in month two. An ADAS perception programme of this scale typically requires 6 to 12 months of sustained annotation; Oprimes' infrastructure was designed to maintain velocity without sacrificing accuracy through that entire duration.

The primary object classes are vehicles (cars, trucks, buses, motorcycles, bicycles), pedestrians, cyclists, animals, and static obstacles (traffic cones, barriers, parked vehicles). Each class has specific annotation requirements — a pedestrian cuboid must capture the full body volume including limbs, while a vehicle cuboid must align to the outer chassis dimensions. ADAS models are typically trained separately for near-range and far-range objects, which may require different annotation precision standards.

Yes. Oprimes' annotation capacity scales by adding trained annotators to an existing workflow — the QA pipeline, tooling, and delivery mechanisms are already in place. For a programme requiring 200,000+ images per month, Oprimes can expand the annotator pool while maintaining the same two-layer QA process. The constraint is typically ramp time for new annotators (2 to 3 weeks of calibration) rather than infrastructure limits.

Building AI That Needs to See the Real World Accurately?

Oprimes delivers Vision AI training data — from 3D cuboid annotation to image classification and segmentation — with two-layered QA at production scale. If your perception model needs a reliable, high-volume data supply, we have done this before, over millions of instances.

Schedule a Demo Explore More Case Studies
From the Blog

Insights from the Oprimes team

View all posts →
The Role of AI in Enhancing User Testing
AI/ML

The Role of AI in Enhancing User Testing

In the fast-evolving landscape of app development, ensuring a seamless user experience is paramount. Traditional user testing methods, while effective,...

Read more →
The AI Revolution: A Testing Framework for the Future of Software
AI/ML

The AI Revolution: A Testing Framework for the Future of Software

What is AI? Artificial intelligence (AI) is a broad field that includes a variety of techniques and approaches for creating...

Read more →
Improve AI-ML-based facial recognition application accuracy by validation through diverse real data sets using a user testing model.
AI/ML

Improve AI-ML-based facial recognition application accuracy by validation through diverse real data sets using a user testing model.

Conducting multiple face recognition trials in different environments and backgrounds to train the AI-based app and validate how it determines...

Read more →
Get Started

Your AI was built by humans.
Let the right humans validate it.

Book a 30-minute consultation with an Oprimes AI Trust Specialist. We will map your use case, recommend the right service pillar, and give you a delivery timeline before you commit to anything.

Book an AI Trust Consultation Read Case Studies

Trusted by 80+ enterprise AI teams across 6 industries. No obligation on first consultation.