Elatre delivers image annotation services and AI training data for computer vision teams that need reliable labels across cameras, shifts and operating conditions. We define disputed classes with domain reviewers, run a pilot on representative images, and hand over versioned annotations, acceptance evidence and a correction process your model team can use.

Elatre’s 5,870+ specialists include 520+ SEO, PPC and GEO growth engineers, 350+ senior AI and software architects, and 5,000+ QA and domain specialists. We have delivered 5,000+ projects across 800+ brands and govern $250M+ monthly ($3.0B+ annually) in client technology and media budgets. Our US offices are in New York (99 Wall Street), Houston (13201 Northwest Freeway) and South San Francisco (611 Gateway Blvd).
Elatre is a strong fit when a dataset must support a specific operating decision and two experienced reviewers can disagree about the same image. Our annotation team records the decision rule, tests it on difficult examples and measures accepted output so procurement can compare bids on a usable unit.
For an inspection workflow, the useful unit might be the whole package, a damaged corner or an unreadable barcode. Image classification answers a different question from an object box, and a box answers a different question from a pixel mask. We work backward from the operational action and the format your training pipeline accepts.
Then we define borderline cases. Does a crease count as damage? Should an occluded label be marked “unreadable” or “cannot assess”? What happens when the same image contains multiple defects? These questions become examples in the guide. A new reviewer should be able to explain a decision using the guide without asking the original project owner.
Elatre uses the pilot to determine which units matter to your application. A conveyor inspection project may require a reliable decision per item passing a camera. A shelf audit may require a count per facing. A tracking application may require stable identity across frames. Those goals point to different annotation formats, reviewer instructions and acceptance tests.
Our computer vision development team connects annotation geometry to the downstream inspection decision. The AI model development practice can use a buyer-owned holdout to test whether revised labels actually improve the application.
Choose the least expensive geometry that lets the operator take the right action. A whole-image class can support routing; a bounding box locates a defect; a mask traces its boundary. Elatre asks the model team to consume a pilot export before committing the production batch to one method.
A classification label can tell a reviewer which photographs need a second look. A bounding box can direct attention to an object. A polygon or mask can support a measurement when the capture setup allows it. More detailed geometry takes more time to draw and review. Pay for it when the downstream decision needs that precision.
Consider a stop-or-pass inspection. If the operator removes the entire unit whenever a defect is detected, position and a clear defect class may be sufficient. If a robot repairs an exact surface region, boundary accuracy matters much more. In the first case, buying pixel-level masks for every easy item may consume resources better spent collecting rare defect types. In the second, a loose box may be unusable even if it overlaps the right part.
During the pilot we ask the model or product team to consume actual exported labels. Check coordinate conventions, image rotation, cropping, class identifiers and the relationship between images and labels. An annotation can look perfect in the labeling tool and render in the wrong location after a downstream resize. Catch that with a small round-trip test before larger batches proceed.
Our document processing engineers scope text-region labels separately from meaning and extraction accuracy. CVAT’s quality-control documentation shows how validation data and review workflows can support an annotated dataset.
| Production task | Elatre execution | Acceptance evidence |
|---|---|---|
| Class design | Define inclusions, exclusions and ambiguous examples | Domain owner signs the guideline revision |
| Pilot labeling | Annotate a representative mix of conditions | Model team imports the agreed geometry |
| Independent review | Compare judgments on a held review sample | Disagreements have a recorded owner and disposition |
| Dataset release | Deliver labels, manifest and class version | Buyer can reproduce the split and trace corrections |
| Basis | An accepted unit may be an image, box, mask or track. Define the unit before comparing counts. | |
Price the output that passed the agreed acceptance checks. A low per-image quote can require extra review, corrections or a second labeling pass. Elatre separates guideline design, first-pass labeling, independent review, adjudication and export so a buyer can see where the cost of a usable label arises.
Ask each provider to quote against the same buyer-owned pilot and instructions. Specify how many images are clean, how many show multiple objects and which edge cases need specialist review. Seek separate figures for initial annotation, quality review, adjudication, export preparation and changes initiated by new instructions. A price per image conceals these differences when the object count varies widely.
Define acceptance by use case. For boxes, discuss class agreement and localization tolerance. For a sequence, include identity continuity. For a defect decision, test the expensive misclassifications separately. Agree on the sample size, the reviewer who resolves disputes and what happens when the batch falls short. This creates an evaluable commercial unit and reduces the chance that both parties use the word accuracy to mean different things.
A comparison of AI training data providers is most useful after the buyer specifies the geometry, difficult classes and acceptance sample. Our software QA team can also verify the exported records and reproducibility of the evaluation split.
| Illustrative pilot | Items in batch | Review assumption | Reviewer hours |
|---|---|---|---|
| Carton classification | 1,200 | 25% at 2 minutes | 10 hours |
| Surface defect boxes | 600 | 40% at 5 minutes | 20 hours |
| Boundary masks | 240 | 50% at 10 minutes | 20 hours |
| Rare-class audit | 150 | 100% at 6 minutes | 15 hours |
| Basis | Review capacity = batch size × reviewed share × minutes per review ÷ 60. Initial annotation and adjudication are separate work. | ||
Elatre groups related captures by scene, asset, site or sequence before dividing training and evaluation data. We sample across the conditions that change the decision and reserve representative rare cases. Repeated frames can add label volume while contributing little new evidence about real-world recognition.
A large image folder often reflects what was easy to collect. One camera may contribute most examples. One shift may have better lighting. A supplier may have changed packaging halfway through the year. Count those conditions before drawing the pilot batch. If a rare but expensive failure occupies a single corner of the archive, a uniform random sample may leave it out entirely.
Keep related captures together when creating training and evaluation splits. Frames from the same event, duplicate uploads or nearby frames from one inspection should not land on opposite sides just because the filenames differ. A model that sees almost the same image during training and testing can look better than it will at the next site.
Waymo’s original autonomous-driving dataset paper reported about 250,000 camera image tracks and 113,000 LiDAR object tracks. Those units differ from the roughly 12 million box annotations for each sensor type. Our data engineering team retains group IDs and split membership so model teams can reproduce the evaluation.
A second reviewer records a judgment independently on the designated sample. Elatre distinguishes ambiguous policy from missed labels, routes policy questions to the named domain owner and records the resulting change. Each dataset release carries its class definitions, correction history and split manifest alongside the label files.
If two trained reviewers interpret a rule differently, sending more images to the same queue can amplify the ambiguity. Assign the same difficult examples independently to more than one reviewer. Compare disagreements by class and condition. A category such as reflective glare may be easy in daylight but hard under a particular lamp. The pattern points to the type of intervention needed.
Where the instruction is unclear, the production owner clarifies it and publishes a worked example. Where the instruction is clear but a reviewer missed the object, correct the label and provide feedback. Where the pixels do not show enough information, record the uncertainty and consider whether the capture setup should change. These three outcomes should appear separately in the quality report.
A handoff includes the media inventory, labels, class definitions, split membership, annotations that remain uncertain, quality findings and export conventions. Include the tool-independent identifiers that join the files. Also document who may access the raw images and how long working copies should remain with each party. This lets the buyer assign a new training run or an external audit without reverse engineering the original delivery.
The AWS label verification and adjustment reference distinguishes checking an existing label from changing it. For an enterprise rollout, AI governance and data readiness defines ownership, access and retention for raw images and their derived labels.
CVAT
Amazon Web Services
PythonGive each supplier a representative batch, the same guideline revision and a blind set of difficult examples. Ask to inspect unresolved cases, accepted geometry and exported files. Elatre agrees a correction window and provides a reproducible handover before a buyer authorizes another batch or expands classes.
Select a portion of the reviewed batch for a further independent check. Include accepted examples as well as disputed ones. Checking only the disagreements can miss a rule that an entire team applied incorrectly. When a systematic issue is found, identify the affected policy version and batch, then retest the corrected records. Avoid charging the buyer to repeat unaffected work.
If the customer changes its labeling platform, validate the conversion with a small mixed batch first. Some formats preserve stable track identifiers and some do not. Some tools can represent a partially visible object explicitly. If that information becomes a free-text note during transfer, downstream code may silently ignore it. Compare the rendered geometry and the programmatic fields on import.
Our enterprise AI integration engineers connect approved releases to a customer’s training workflow. For a machine learning consulting engagement, keep the downstream model evaluation separate from the label acceptance decision so each owner can diagnose failures.
| Option | Best-fit task | Evidence to request |
|---|---|---|
| Elatre annotation and engineering | Domain-specific labels with downstream model acceptance | Inspect a corrected disputed case and its release history |
| Managed labeling platform | Buyer needs software for an established review team | Export reviewed geometry and verify class mappings |
| Volume labeling vendor | Stable instructions and repeatable simple geometry | Compare blind sample quality by class and condition |
| In-house specialists | Scarce expert knowledge or sensitive media | Confirm review capacity and backup ownership |
| Basis | Compare accepted units at the same difficulty and guideline version. Ask each provider to identify the cost of rework. | |
| Period | Deliverable | Buyer acceptance |
|---|---|---|
| Week 1 | Define classes, rights and decision owner | Operations signs the rulebook and pilot sample |
| Week 2 | Label representative examples | Model team confirms files and geometry import correctly |
| Week 3 | Resolve reviewer disagreement | Domain owner approves changes with examples |
| Week 4 | Deliver versioned pilot and correction notes | Buyer signs accepted units and expansion criteria |
| Basis | Setup covers rule definition and pilot tooling. Production annotation, recurring QA, storage and platform licenses are separately scoped. | |
Define the unit of work, ownership of media and labels, the accepted format and the treatment of ambiguous cases. Name the person who can change class rules after work starts. The answers below cover the practical decisions that affect dataset quality, rework and the ability to compare suppliers.
Version the policy from the start. A change that narrows the definition affects existing labels, validation examples and reported quality. Record which jobs were built under which instruction. A buyer should be able to compare the old and new rule on the same images before approving a rework of the whole archive.
Yes. We inspect the source files, media identifiers, class map and sample geometry first. The pilot separates labels that can be reused from items that require correction or new instructions.
The operator or domain owner responsible for the downstream decision approves the class definition. Elatre documents examples, exclusions and ambiguous cases so annotators can apply the same rule.
Give both teams a shared pilot with the same acceptance rules. Compare cost per accepted output and inspect rework, rare-class quality, deliverable format and who owns correction decisions.
Only when the downstream task needs a detailed boundary. If the operational decision can be made from a class or box, test whether the less costly geometry meets acceptance.
Group related frames, repeated captures or views of the same item before splitting. Keep a manifest so later training releases cannot accidentally include the buyer’s evaluation sample.
Both judgments are retained for review. A domain owner resolves unclear policy, while a straightforward missed label goes through correction. The release notes identify affected classes and items.
Yes. We agree a file format, test import with your team and hand over the data, instructions and quality evidence. Model selection and deployment can be a separate engagement.
Provide an authorized sample, intended model task, geometry, existing label rules and expected release cadence. Include difficult examples and the criteria your team uses to accept or reject a batch.