An autonomous driving technology company
100,000 fisheye images and roughly 8 million annotations delivered in five months for autonomous parking
Published 8 September 2026
At a glance
| Industry | Autonomous driving | automated parking and low-speed manoeuvring algorithms |
|---|---|---|
| Client | Autonomous driving technology company | unnamed under the confidentiality terms of the engagement |
| Programme scale | 100,000 images | fisheye camera captures across varied parking environments |
| Annotation volume | ~8 million annotations | approximately 80 annotations per image, derived from the delivered totals |
| Object categories | 60+ | pedestrians, vehicles, traffic signs, lights, lane markings |
| Duration | 5 months | from programme start to completed delivery |
| Throughput | ~20,000 images / month | derived: 100,000 images over 5 months |
| Team ramp | 40 annotators in 3 weeks | recruited, trained and productive |
| Scenario split | Indoor and outdoor | separate specialist teams rather than one pooled team |
| Method | Pre-recognition plus manual verification | proprietary platform pre-labels; every annotation is human-verified |
| Status | Delivered | full agreed scope accepted |
This is a Lifewood Data Technology case study in Autonomous driving — an engagement delivered for an autonomous driving technology company across Not disclosed. 100,000 fisheye images and roughly 8 million annotations delivered in five months for autonomous parking
Fisheye distortion breaks the geometry that ordinary annotation depends on
An autonomous driving technology company needed a training dataset for its automated parking algorithms, spanning 60+ object categories across the full range of parking environments its vehicles would encounter. Parking is where perception is hardest rather than easiest. Clearances are measured in centimetres, pedestrians appear at close range and from behind obstructions, multi-storey structures remove satellite positioning, and lighting shifts from direct sun to sodium-lit basement within a single manoeuvre. Fisheye optics are what make the annotation work genuinely different. A fisheye lens buys the wide field of view that close-quarters manoeuvring requires, and pays for it with radial distortion: straight lines bow, and a rectangle drawn in image space does not correspond to a rectangle in the world. An annotator trained on forward-facing road footage will place boxes that look correct and are geometrically wrong, and the error grows toward the frame edge — which is exactly where the obstacles that matter for parking appear. Category granularity compounds it. Sixty-plus categories across pedestrians, vehicles, traffic signs, signals and lane markings means the specification is long enough that consistency between annotators becomes the binding constraint rather than individual skill.
Pre-recognition with mandatory human verification, and separate indoor and outdoor teams
Lifewood ran the programme on a proprietary annotation platform combining automated pre-recognition with manual verification, and split the workforce by scenario rather than pooling it. The approach ran in four steps. 1. Pre-recognition as a first pass, never as the output. The platform pre-labels each frame and annotators verify and correct rather than drawing from scratch. This is what allows 8 million annotations at a viable unit cost, and the reason every annotation still passes through a human is that pre-recognition degrades precisely where fisheye distortion is worst. 2. Separate indoor and outdoor teams. Basement and multi-storey parking and open-air lots present different lighting, different obstacle profiles and different distortion behaviour. Specialist teams hold a narrower specification well rather than a broad one approximately. 3. A 40-member team recruited and trained in three weeks. Team build was treated as part of delivery rather than as a precondition, which is what kept the five-month total achievable. 4. Manual verification against the 60+ category specification. Verification covers category assignment as well as geometry, because a correctly drawn box in the wrong category is still a labelling error. Cost reduction on this programme came from the pre-recognition step rather than from reducing review coverage — the distinction matters, because reducing review is the usual way vendors hit a price point and it is the one that shows up later in model performance.
Full scope delivered in five months at roughly 80 annotations per image
The programme delivered 100,000 annotated fisheye images carrying approximately 8 million annotations across 60+ object categories, completed over five months. That averages roughly 80 annotations per image and about 20,000 images per month sustained across the run — both figures derived from the delivered totals rather than separately reported. The team ramp is the number worth noting alongside the volume. Forty trained annotators productive within three weeks is what made a five-month total possible on a dataset of this density; on a programme where team build takes two months, the same scope takes seven.
Related Lifewood services
Verified outcomes
| Metric | Value | Baseline | How measured |
|---|---|---|---|
| Images delivered | 100,000 | full agreed scope | Accepted delivery |
| Annotations delivered | ~8,000,000 | ≈80 per image, derived from delivered totals | Platform annotation count |
| Object categories | 60+ | pedestrians, vehicles, signs, signals, lane markings | Client annotation specification |
| Duration | 5 months | programme start to completed delivery | Delivery record |
| Sustained throughput | ~20,000 images / month | derived: 100,000 ÷ 5 months | Derived from delivered totals |
| Team ramp | 40 annotators in 3 weeks | team build inside the delivery window, not before it | Recruitment and training records |
| Verification coverage | Every annotation | pre-recognition output is never shipped unverified | Manual verification pass |
Method and verification. Figures on this page are Lifewood-reported. Image count, annotation count, category coverage, team size and duration are delivery actuals. The per-image annotation density and the monthly throughput figure are derived from those actuals by division and are labelled as derived rather than presented as separately measured results. Every annotation produced by platform pre-recognition passed a manual verification step; pre-recognition output was not shipped unverified at any point. Delivered accuracy is not published on this page because no verified measurement basis has been agreed. The client is not named on this page under the confidentiality terms of the engagement.
Questions about this programme
On Lifewood fisheye parking programmes the difficulty is radial distortion — a rectangle drawn in image space does not correspond to a rectangle in the world, and the error grows toward the frame edge where parking obstacles appear.
On this autonomous parking programme Lifewood used platform pre-recognition as a first pass only; every annotation then passed a manual verification step, because pre-recognition degrades exactly where fisheye distortion is worst.
Lifewood split the parking annotation workforce by scenario because basement, multi-storey and open-air environments differ in lighting, obstacle profile and distortion behaviour, and a specialist team holds a narrow specification better than a pooled team holds a broad one.
On this programme Lifewood recruited and trained a 40-member fisheye annotation team in three weeks, inside the five-month delivery window rather than ahead of it.
Run a similar program with Lifewood?
Tell us your scope and accuracy bar. We will scope a comparable engagement within one call.
Talk to our team