Short answer. Sama is a computer-vision specialist that publishes a quality-led proposition — human-verified image, video, 3D and LiDAR annotation, with a stated 99% first-batch acceptance rate and professional services running from pilot to production. Lifewood is a broader managed provider: the same visual modalities plus text, audio, multilingual and LLM data, delivered from 40+ centres across 30+ countries in 50+ languages under a stated 95%+ accuracy SLA. Both publish quality metrics, and the two figures measure different things. If your programme is and will remain almost entirely computer vision, evaluate the specialist seriously. If visual work is one workstream among several, breadth changes the arithmetic.
The trap in this comparison is the numbers. "99% first-batch acceptance" and "95%+ accuracy SLA" look directly comparable and are not. One describes the share of delivered batches a client accepts without return; the other describes label correctness against a reference, with a rework obligation attached. A programme can have a high acceptance rate and mediocre labels if acceptance is a spot-check, and it can have excellent labels and a low acceptance rate if the client's criteria are stricter than the spec. Neither figure is dishonest. Neither is comparable without its definition.
This guide sets out what each company publishes, how to normalise the quality claims, and where the fit genuinely diverges.
What each company publishes about itself
| Buyer criterion | Lifewood (company-reported) | Sama (company-reported) |
|---|---|---|
| Computer vision | Image, video, AV and 3D point-cloud workflows | Human-verified image, video, 3D and LiDAR annotation |
| Other modalities | Text, audio, multilingual corpora, LLM datasets | Text, audio and multimodal combinations |
| Quality position | 95%+ accuracy SLA, dual-layer human review | 99% first-batch acceptance rate; automation plus expert human review |
| Delivery model | 40+ delivery centres across 30+ countries | Professional services from pilots to production; in-house teams |
| Language position | 50+ languages, region-native staffing | Not primarily positioned around language breadth |
| Typical buyer | Broad enterprise annotation programmes | Visual-AI and quality-focused programmes |
Both companies are credible in computer vision. Public positioning differs in scope rather than in competence, and nothing in the public record supports a claim that either is weak at what the other emphasises.
Normalising the quality claims
Before any comparison, resolve six questions with both providers. Do it in writing.
- What is the denominator? Objects, images, frames, batches or deliveries. A per-batch figure and a per-object figure differ by orders of magnitude on the same work.
- What counts as an error? A missed object, a loose box, a wrong class and an inconsistent track are four different failures with four different downstream costs.
- Is the measure chance-corrected? For any judgement-heavy label, raw agreement flatters. Cohen's kappa or Krippendorff's alpha is the honest form.
- Who owns the reference? A vendor's gold set measures the vendor's interpretation. A client-approved gold set measures yours. Only the second is evidence.
- What sample is it drawn from? Last quarter, on comparable work, at comparable volume — or a favourable engagement from two years ago.
- What happens below threshold? Who reworks, at whose cost, on what turnaround, and how does root cause feed back into training?
For computer vision specifically, add the geometric measures the headline percentage hides:
IoU = Area of overlap ÷ Area of union
F1 = 2 × (Precision × Recall) ÷ (Precision + Recall)
Effective = Items delivered × First-pass acceptance ÷ Cycle time
throughput
Ask for F1 and IoU-at-threshold by object class. An aggregate figure is dominated by large, easy, well-lit objects and hides exactly the small, distant, occluded cases where a perception model fails.
Where Sama is strong, on its own account
- A quality-led operating position. Sama's materials position the offering around automation plus expert human review, and report a 99% first-batch acceptance rate. For a buyer whose main risk is returned batches and schedule slippage, that is the metric that matches the risk.
- Computer-vision depth. Image, video, 3D and LiDAR at scale, with the tooling and review process built for visual work rather than generalised across modalities.
- Pilot-to-production services. The professional-services model explicitly covers pilots and production optimisation, which is where most annotation programmes actually fail.
Where Lifewood fits
- Scope headroom. A computer-vision engagement can later absorb text, speech, multilingual or LLM work without adding a provider. For multi-year programmes whose roadmap is not yet fixed, that optionality has real value.
- Language operations alongside vision. 50+ languages matters more in vision work than buyers expect: signage, on-screen text, OCR, local metadata, market-specific scene review and localised guidelines all need native speakers.
- Autonomous-driving breadth. Published AV service material covers perception annotation and 3D point-cloud workflows delivered as managed production.
- Distributed production. For very large programmes, geographically distributed capacity supports continuity and regional access requirements that a concentrated operation cannot.
When Sama is the better fit
- The programme is almost entirely computer vision and will stay that way.
- Quality metrics are the deciding commercial term, and their review model is the one you want.
- Language coverage and non-visual modalities are unlikely to matter within the contract term.
- You want a specialist whose entire operating model is tuned to visual data.
When Lifewood is the better fit
- The roadmap is likely to expand into additional modalities, regions or languages.
- Visual data carries language content — text in scene, OCR, localised categories, market-specific review.
- Production geography is a requirement, whether for residency, continuity or client mandate.
- You want one accountable operation across several AI workstreams rather than a set of specialists.
What to test in a computer-vision pilot
A vision pilot built from clean daylight footage measures nothing. Load it deliberately:
- Occlusion and truncation. Objects half behind other objects, cut by the frame edge, or visible for three frames.
- Small and distant objects. Where IoU tolerance and annotator patience both break down.
- Adverse conditions. Night, rain, glare, motion blur, low-resolution sensors.
- Class confusion pairs. The two classes your own team argues about. Include them and see whether the vendor asks or guesses.
- Temporal cases for video. Objects that leave and re-enter, split, merge, or change apparent identity.
- Cross-sensor cases for fusion work. The same object where the LiDAR and camera views disagree.
Score the pilot on per-class F1 and IoU, on escalation behaviour, and on how ambiguous items came back — as confident wrong labels, as questions, or as proposed guideline amendments.
Sources and further reading
- Sama capability statements — human-verified image, video, 3D and LiDAR annotation, a reported 99% first-batch acceptance rate, and professional services from pilot to production — are drawn from the company's published materials at sama.com.
- Lifewood delivery figures (50+ languages, 40+ delivery centres across 30+ countries, 95%+ accuracy SLA) are published on lifewood.com; AV scope on autonomous driving annotation.
- IoU, F1 and chance-corrected agreement are the standard measures for visual annotation quality; a single aggregate accuracy percentage is not comparable between vendors.