This guide covers what data labeling services actually include, how scope and unit price get decided before anyone sends you a number, and how to verify quality on your own data before committing to a volume. It is written from the buyer’s side of the table.

What data labeling services actually cover

First, the terminology. “Data labeling” and “data annotation” describe the same work in commercial practice, and providers use one or the other largely out of habit — US machine-learning teams lean toward labeling, academic and Japanese sources lean toward annotation. There is no capability difference hiding behind the word choice, and a vendor is not narrower because their homepage says one rather than the other. Filtering your shortlist on vocabulary will only shorten it arbitrarily.

The distinction worth making is between data types. Most teams arrive with one type in mind and discover a second one halfway through, so it is worth knowing at the outset which of these can sit inside the same engagement.

Image and video

The largest category by volume: bounding boxes for object detection, polygons and pixel-level regions for segmentation, keypoints for pose, and identity-consistent tracking across frames. Video is not simply “more images” — the moment you need the same object to keep the same ID across a sequence, the work and the pricing change shape. Teams building perception, inspection, or behavior-analysis models usually start here. Our image and video annotation service covers this range.

3D point cloud and LiDAR

Cuboids in 3D space, 3D segmentation, and sensor fusion where each point-cloud object is linked to its counterpart in a synchronized camera image. This is the most specification-sensitive category we handle: occlusion handling, ground-plane convention, and minimum point thresholds all have to be agreed before anyone starts, because they are expensive to retrofit. If your program combines cameras and LiDAR, the two can be scoped as one order rather than split across providers — see 3D point cloud and LiDAR annotation.

Text, LLM, and audio

Preference data for RLHF and DPO, instruction–response pairs for SFT, multi-level model evaluation, transcription, and classical NLP tagging. The judgment here is subjective in a way that box-drawing is not, which makes the guideline and the reviewer selection matter more than throughput. Language coverage is a real constraint, and native-speaker work is not interchangeable with translated work. This sits under LLM, text, and audio annotation.

Collection and preprocessing

The step most first-time buyers forget to scope. Raw footage often needs deduplication, format conversion, frame extraction, face or plate anonymization, or simply to be collected in the first place because the scenario you need does not exist in your archive. Labeling a poorly assembled dataset is the most reliable way to spend a budget and learn nothing. Data collection and preprocessing can be quoted alongside the labeling itself.

Three delivery models, and what actually differs

Providers fall into roughly three shapes: a managed service that takes the work and returns finished data, a platform that licenses you tooling and leaves the workforce to you, and a crowdsourcing marketplace that distributes tasks to an open pool. Each is a legitimate answer to a different question, and which one fits depends less on your budget than on how much specification work you want to keep. If you are still deciding between these — and between doing it in-house at all — our guide to data annotation outsourcing covers vendor selection in depth; this section stays on the one thing that is easy to miss when comparing quotes.

That thing is responsibility. Two providers can quote the same rate for the same boxes and be selling substantially different products, because the boundary of what they are accountable for sits in a different place.

Who writes the labeling guidelines

Someone has to decide what counts as an instance, how to treat an object cut off at the frame edge, whether a reflection in a window is a car, and what to do with the case nobody anticipated. If the provider expects a finished specification from you, that labor stays on your engineering team — and it is not a small amount of labor. If the provider builds the guideline with you and takes responsibility for its internal consistency, that work has moved into the engagement. Ask directly which of these you are buying, because both are often described as “labeling services”.

Who absorbs the rework

Specifications change after the first training run. That is normal — the failure cases teach you something the planning document could not. The question is what happens commercially when they do. Some arrangements treat every revision as new billable volume; others distinguish between a defect against the agreed spec and a change to the spec itself, and absorb the first. At ANOSUPO, defects traceable to us against the agreed specification are corrected at no charge for one year after delivery; a change to the specification is a different conversation, and any honest provider will tell you the same. We operate as a managed service, which is why these two clauses sit where they do.

How scope gets defined before you get a quote

By the time a number reaches you, most of the cost has already been fixed by decisions made in the scoping conversation. Three of them do the heavy lifting.

Volume, class count, and edge-case rules

Volume is the obvious lever and the least interesting one. Class count matters more than it looks: every additional class multiplies the number of boundary decisions an annotator has to make, and the marginal class is usually the ambiguous one. Edge-case rules are where estimates break. A dataset of clean, well-lit, unoccluded objects and a dataset drawn from real operating conditions can differ by a factor of two in labor while looking identical in a spreadsheet. Bring your difficult data to the scoping call, not your representative data.

Guideline design is a deliverable, not a freebie

A labeling guideline is a piece of engineering. It has to be specific enough that two annotators reach the same answer independently, and it only earns that property by surviving contact with real examples. Expect at least one revision after the first batch, and treat a provider who produces the guideline with you — and shows you the disagreements it resolved — as delivering something, not as doing sales support. Where this work is unpriced, it is usually also unowned.

Acceptance criteria — sampling vs. full inspection

This is the single largest hidden variable in any quote. Sampled spot checks and full-volume inspection are different products at different costs, and a rate quoted under one is not comparable to a rate quoted under the other. Sampling tells you the error rate; it does not remove the errors from the data you are about to train on. We inspect every deliverable rather than a sample, with a measured 99.7% quality consistency. When you compare providers, hold the inspection standard constant first — otherwise you are comparing the price of two different things.

What moves the unit price

Our published rates start at from $0.036/label for bounding boxes, and every project we quote moves from that starting point according to the same handful of factors. Object density is usually the largest: billing is per object, so a crowded intersection scene costs many times a product photograph even though both count as one image. Class count and ambiguity come next, because slower, rule-checked decisions are slower work. Then the expected revision cycles, and finally turnaround — compressing a schedule means running more annotators in parallel against the same guideline, which raises the coordination cost of keeping them consistent.

What matters as much as the rate is what the rate contains. The comparison that actually predicts your invoice looks like this.

ANOSUPO charges no setup fee and no management fee, bills purely on volume created, and applies no minimum-order requirement. Full rate details for every data type are on our pricing page, and a quote against your actual specification is free.

How to evaluate quality before you commit

Quality claims are easy to make and hard to compare, so it is worth knowing which questions actually separate providers. The most informative one is how they handle guideline design: a provider who asks you about occlusion, class boundaries, and what to do with the ambiguous case has done this work before, while one who only asks for a volume and a deadline has not. The second is the inspection standard, asked as a number with a period attached rather than as a target.

After that, ask who does the work. For subjective tasks — evaluation, preference data, anything requiring domain reading — whether annotators are selected for the task or drawn from a general pool changes the output more than any tooling difference. Then ask what the correction flow looks like when you reject a batch: how it is reported, who reviews it, how long it takes, and whether it costs anything.

Finally, look at work the provider has actually delivered rather than at capability lists. Our case studies describe the constraints each project ran into and how the specification was built, which is more useful for calibration than a logo wall. A provider who cannot describe a project’s difficulties in specific terms has usually not been close to one.

Security and how your data is handled

For most buyers this is a gating question rather than a differentiator: either the arrangement satisfies your legal and compliance requirements or the rest of the conversation does not happen. Certification is the baseline evidence — ANOSUPO is ISO/IEC 27001 (ISMS) certified — but the operational details are what your security reviewer will ask about.

Ours are as follows. Data is not retained and not processed locally; work happens in the cloud environment agreed for the project. All staff sign NDAs, and teams are separated per project so that access does not accumulate across clients. We never repurpose client data to train our own models, and data is physically deleted on completion. If you have specific requirements — VPN connection, a named tool, data residency constraints — raise them in the first conversation rather than at contract stage, since some of them change how the work is set up. Details are on our security page.

Starting small

Almost every expensive labeling failure traces back to a specification that read as complete in a document and fell apart on real data. The defense is not a longer document. It is committing a small amount of data first, in two distinct stages that answer two different questions.

The first is a free trial: we annotate roughly 10 to 50 items of your real data with our production team, so you can assess actual quality before placing any order. This is not a demo run by a specialist — it is the workflow you would receive. It answers the question can this provider do the work.

The second is a PoC. There is no minimum-order requirement at ANOSUPO, so this is a planning decision rather than a contractual threshold; for image annotation we take on PoCs from around 50 to 100 items, which is typically enough to train a baseline and see where it fails. It answers a different question: is my specification correct. Keep the two separate in your planning — collapsing them into one batch means you learn one of those answers and assume the other. You can move from the trial into a PoC without renegotiating anything, and scale from there when the guideline has survived the difficult cases. Rates for both are the same published rates on our pricing page.

Frequently asked questions

What to take away

Quotes for data labeling services diverge because scope diverges, not because quality does. Before comparing numbers, hold three things constant across providers: who writes the guideline, whether inspection is sampled or full-volume, and who pays for rework when the specification changes. Once those are fixed, the unit rates become comparable — and usually much closer together than they first appeared. Then commit a small batch of your real data rather than a large batch of your assumptions. We have delivered 1,000+ projects, and the ones that went well almost always started that way.

Free TrialTry ANOSUPO annotation free. Test our quality and communication on a sample of your real data before you commit — images, video, 3D/LiDAR and LLM evaluation.annotation-support.com

投稿 Data Labeling Services: What You’re Actually Buying は アノサポ｜AIアノテーション・学習データ作成の伴走パートナー に最初に表示されました。

This guide covers what an image annotation service actually delivers, which annotation type your problem needs, what moves the unit price, and how to scope a first batch small enough that a wrong assumption stays cheap. The rates shown are ANOSUPO’s published USD rates — use them as one reference point, not as an industry benchmark.

Image annotation, image labeling, and image tagging are used almost interchangeably across vendors. This guide says “annotation”, but everything below applies whichever term your team uses.

At a glance: ANOSUPO image annotation & labeling services

• Bounding boxes from $0.036 / label, keypoints from $0.021 / point, segmentation from $0.152 / region
• No setup fee, no management fee, no minimum order
• Full-volume inspection with 99.7% label consistency across 2025 deliveries
• Free quote against your spec, or a free sample on your own data for companies developing AI products

Request a quote or apply for a free sample.

When you need an image annotation service

Outsourcing is not automatically the right answer. Keeping the work in-house usually makes sense in three situations:

• The dataset is small and finite. A few hundred images, labeled once, is faster to do yourself than to specify for someone else.
• The spec is still changing daily. If your class definitions shift every time you look at the data, you are still doing research, not production labeling.
• The judgment cannot be transferred. Some domains — early-stage medical research, novel defect categories — depend on knowledge that only two people in your company have.

The picture changes when volume becomes recurring. Labeling scales linearly with data, engineering time does not, and consistency across annotators degrades quietly as the team grows. The comparison that matters is not cost per label against zero — it is cost per label against the engineer hours you are currently spending, plus the cost of the rework that inconsistent labels create downstream.

A useful signal: if you have already built an internal labeling rotation and it keeps slipping, you have passed the point where outsourcing pays for itself. For the broader decision framework — in-house, crowdsourcing, or a specialist vendor — see our guide to data annotation outsourcing.

Annotation types, and what each costs you in effort

Choosing the type is a modeling decision before it is a budget decision. Picking a heavier type than your model needs is the single most common way image annotation budgets get inflated.

The ratio is the part worth internalizing: a segmentation region starts at roughly four times a bounding box. That gap is labor, not markup — a box is two clicks, an outline is thirty.

Polygon annotation sits between the two. It gives you a tighter outline than a rectangle without demanding pixel-perfect edges, and where it lands on a rate card depends entirely on how tight the outline has to be, so it is quoted against your spec rather than published as a flat rate.

Two practical notes. First, keypoint pricing is per point, not per object — a 17-point skeleton on 1,000 people is 17,000 units, not 1,000. Second, if you are unsure whether boxes are enough, they usually are: start with detection, measure, and only move to segmentation when box edges demonstrably cost you accuracy. Our semantic segmentation guide covers where that line falls, and the YOLO training data workflow covers the detection case end to end.

What actually drives your unit price

Quotes from different vendors are rarely comparable, because the same phrase covers different amounts of work. Five things move the number:

• Objects per image, not images. Per-object billing means a crowded street scene with 40 vehicles costs forty times a product shot. Estimate your average object density before you estimate your budget.
• Edge-case density. Occlusion, truncation at the frame border, ambiguous class boundaries — these are where annotator disagreement lives, and handling them properly requires written rules and slower work.
• Inspection standard. Sampled spot checks and full-volume inspection are very different products at very different costs. ANOSUPO inspects every deliverable rather than a sample, with a measured label consistency of 99.7% across 2025 deliveries.
• Model-assisted pre-labeling. A first pass from a foundation model followed by human correction reduces effort on well-defined classes. It helps least on exactly the ambiguous cases that need humans most, so treat it as a cost lever, not a quality claim.
• What the quote excludes. Setup fees, project management fees, monthly minimums, and paid rework all sit outside the headline rate at some vendors. ANOSUPO charges no setup or management fee and bills purely on volume created.

A rough illustration at published rates: 10,000 images averaging three objects each is 30,000 boxes, or about $1,080. The same 10,000 images at pixel level, averaging two regions each, is 20,000 regions — around $3,040. The image count did not change; the type and the object count did. Full rate details are on our pricing page, and a quote against your actual spec is free.

How to scope a first batch

Do not send 50,000 images as your first order. Almost every expensive annotation failure traces back to a specification that looked complete in a document and fell apart on real data. Three steps, in order:

• Trial (roughly 10–50 items). ANOSUPO annotates a small sample of your real data under the production workflow, at no cost, so you can judge output quality and communication before any commitment. This is a vendor test, not a dataset.
• PoC (roughly 50–100 images). This is a specification test. Deliberately include the hard frames — heavy occlusion, poor lighting, the classes your team argues about — because a PoC made of clean images tells you nothing you needed to know.
• Scale. Only once the guideline survives contact with the difficult cases. There is no minimum-order requirement, so scaling up is a decision about your data readiness, not about contract thresholds.

Trial (roughly 10–50 items). For companies developing AI products, ANOSUPO annotates a small sample of your real data under the production workflow, at no cost, so you can judge output quality and communication before any commitment. Each request is reviewed first. This is a vendor test, not a dataset.

What to check before you commit to a vendor

Six questions that separate providers more reliably than price does:

• Is inspection full-volume or sampled? Ask for the actual figure and the period it covers, not a target.
• Who fixes errors, and for how long? ANOSUPO corrects defects attributable to us against the agreed specification at no charge for one year after delivery. Changes to the specification itself are a different conversation, and any honest vendor will say so.
• Do the formats and tools fit your pipeline? COCO, Pascal VOC, YOLO text, and PNG mask output cover most cases; tool-specific delivery through CVAT, Labelbox, or SuperAnnotate matters if your team already lives in one of them.
• Is security certified or merely described? A certification such as ISO/IEC 27001 is externally audited. A security page is not.
• Do they ask you spec questions in the first call? A provider who asks how you want occlusion and truncation handled has annotated images before. One who only asks for image counts has not.
• Can they handle genuinely awkward work? Complexity, not volume, is where vendors fail. For a construction-site AI project, our team color-coded rebar by type and documented the ambiguous cases as they appeared — see the Pixie Dust Technologies case.

Working with a provider in another country

Once your shortlist crosses borders, three factors that rarely appear on a rate card start to matter.

Data residency. Ask where your images are stored, where they are processed, and whether anything persists after delivery. ANOSUPO works cloud-only with no local retention, separates teams per project, never reuses client data to train its own models, and physically deletes data when a project closes. If your images carry regulatory or contractual location constraints, raise them during scoping rather than after the quote — they can change the workflow.

Certification as common ground. ISO/IEC 27001 is useful precisely because it is international: it gives your security reviewer a comparable frame of reference for a vendor they cannot visit. ANOSUPO is certified, and the specifics are on our security page.

Time zone offset. A large offset is a liability for real-time collaboration and an asset for batch work — deliverables land while your team sleeps. It rewards written specifications and structured handoffs, and punishes teams that rely on quick clarifying calls. ANOSUPO is based in Fukuoka, Japan, and operated by Borderless Japan, with delivery across more than 1,000 projects. For a sense of sustained throughput at that distance, our environmental-AI client reached roughly 100,000 items per month at about half their previous cost — the Pirika case has the details.

Data for environmental AI half the cost, at scaleCase study: Pirika's litter-survey AI "Takanome." ANOSUPO took on annotation of ever-growing video data, cutting cost to about half and scaling throughput to 100K items a month and 1M+ delivered — a data foundation in-house work could never reach.annotation-support.com

Frequently asked questions

Where to go from here

Image annotation is a solved problem operationally and an unsolved one specifically: the difficulty is never drawing the box, it is agreeing on which boxes to draw. Choose the lightest annotation type your model actually needs, estimate cost from object counts rather than image counts, and prove the specification on a small batch that includes your hardest frames before committing to volume.

If you want a second opinion on a spec, or a cost estimate against your real images, we are glad to look at them with you. Full coverage of what we handle is on our image and video annotation page.

Free TrialTry ANOSUPO annotation free. Test our quality and communication on a sample of your real data before you commit — images, video, 3D/LiDAR and LLM evaluation.annotation-support.com

投稿 Image Annotation & Labeling Services: Types, Cost, and When to Outsource は アノサポ｜AIアノテーション・学習データ作成の伴走パートナー に最初に表示されました。

That is not a scaling problem. Video annotation adds a requirement that image work never had: consistency across frames. This guide covers how to decide a sampling rate, how object tracking drives cost, how to set quality criteria that actually apply to sequences, and where in-house pipelines tend to break.

When video annotation is different from image annotation

The most expensive assumption in a video project is that video is just a lot of images. Mechanically that is true — you extract frames, and each frame gets labeled. But the labeling task itself changes shape.

In image work, every item is independent. A frame is either correct or it isn’t, and a mistake on one image has no effect on any other. In video, three things stop being independent:

• Identity. The same object must carry the same ID from the frame it appears in to the frame it leaves. A box can be pixel-perfect and still be wrong, because it inherited the wrong ID.
• Ambiguity. A blurred or half-occluded object in a single image is simply a hard case. In a sequence, the annotator can look forward and backward — which means the correct answer depends on frames other than the one in front of them.
• The unit of review. Reviewers do not check frames. They check tracks: the full life of one object across a sequence. One bad handoff in the middle invalidates the whole track.

This is why an image-derived estimate under-reads video work. The per-frame effort may be similar, but specification, tracking and review are new line items.

ANOSUPO delivers video work as part of our image and video annotation service — that page covers what we handle and the formats we support. This article is the other half: how to design and scope the work before you request a quote.

If your project is closer to still images than to continuous footage, the type selection and pricing logic differ — see our guide to image annotation services.

Frame extraction and sampling rate: the first cost decision

There is no such thing as billing “per video.” A video is converted into frames, and those extracted frames are the items that get annotated. So the single decision that sets your total volume — and therefore your total cost — is how many frames per second you pull out.

The arithmetic is unforgiving. Ten minutes of 30 fps footage is 18,000 frames. Extract at 2 fps instead and the same ten minutes is 1,200 frames. That is a 15× difference in scope from one config value, before anyone has drawn a single box.

Sampling too densely

At high frame rates, consecutive frames are near-duplicates. You pay to label the same scene repeatedly, and the model gains very little from the redundancy — the dataset looks large but its effective diversity is low. Worse, a bloated dataset hides its own gaps: 18,000 highly correlated frames from one afternoon of shooting will not cover night, rain, or a crowded scene.

Sampling too sparsely

The opposite failure is quieter and more damaging. When objects move too far between sampled frames, tracking association becomes guesswork — annotators can no longer tell whether the object on the right is the same one that was on the left, or a different object entirely. Short events disappear completely: a pedestrian crossing the frame in half a second simply does not exist at 1 fps.

A workable rule is to set the rate from how far your objects move between frames, not from habit. Slow, wide scenes tolerate sparse sampling; fast objects close to the camera do not. Different clips in the same project can legitimately use different rates, and deciding that up front is cheaper than re-extracting after production has started.

If your project also includes still-image data, the workflow for that side is covered in our guide on creating training data for YOLO, which walks through the same steps without the temporal layer.

Object tracking and ID consistency

A track is the sequence of boxes belonging to one object across the frames it appears in. Object tracking annotation means assigning and maintaining those IDs — and it is where video projects usually lose their schedule.

Three situations create almost all of the difficulty:

• Occlusion. A car passes behind a truck for twelve frames and comes out the other side. Is that one track with a gap, or two tracks? Whatever you choose, every annotator has to choose the same way.
• Exit and re-entry. A person walks out of frame and returns forty frames later. Do they keep their original ID, or get a new one? Re-identification across a long gap is genuinely hard for humans, not only for models.
• Look-alike objects. Two similar objects cross paths and the IDs swap. Every frame is individually correct; the sequence is wrong. This defect class — the ID switch — is invisible to any frame-level check.

The practical consequence is that your spec has to answer these questions before production, in writing:

• How many frames of full occlusion before the object is treated as a new track?
• Are partially visible objects labeled, and from what visible fraction?
• Does an object that leaves and returns keep its ID, and over what time window?
• Do you need attributes that can change mid-track — occluded, truncated, an activity state?

Interpolation between keyframes reduces the drawing labor considerably. It does not reduce the judgment. Deciding identity across an occlusion is a human call every time, and that is the part that scales with sequence length rather than with frame count.

Setting quality criteria for video

If your acceptance criteria are written per frame, they will pass a dataset that is unusable. A frame-level accuracy figure treats the ID switch above as a rounding error, because only a handful of frames are affected.

Evaluate at the track level instead. In practice that means three questions per sequence:

• Does each real object correspond to exactly one track, with no ID switches?
• Is any track fragmented into pieces that should have been joined across an occlusion?
• Are the box boundaries stable, or do they jitter frame to frame around a static object?

The third one matters more than it looks. Boundary jitter across an otherwise correct track teaches a detector that the object’s extent is noisy, and it comes directly from annotators working frame-by-frame without playback review.

This is also why our inspection model is full-volume inspection rather than sampling-based spot checks. Sampling assumes defects are randomly distributed; in video they are not. They cluster exactly at the hard moments — occlusions, crossings, scene cuts — which a random sample of frames will usually miss. Across projects we hold a measured 99.7% quality consistency, and defects attributable to us against the agreed spec are corrected free of charge for one year after delivery.

Where in-house video pipelines break down

Teams that handled their image datasets internally often assume video is the same work at a larger size. Three specific points tend to break, and they break in this order.

1. Volume

Frame extraction multiplies your dataset by one or two orders of magnitude overnight. A team that comfortably labeled 3,000 images a month is not going to absorb 60,000 extracted frames by working harder. Our client Pirika hit this wall with environmental video data: after moving to us, annotation costs fell to roughly half, processing scaled to 100,000 items a month — the previous vendor handled a few thousand — and cumulative delivered data passed one million items.

Data for environmental AI half the cost, at scaleCase study: Pirika's litter-survey AI "Takanome." ANOSUPO took on annotation of ever-growing video data, cutting cost to about half and scaling throughput to 100K items a month and 1M+ delivered — a data foundation in-house work could never reach.annotation-support.com

2. Review capacity

Frame labeling parallelizes cleanly across people. Track review does not: checking identity requires watching the sequence in order, so the reviewer is a serial bottleneck no matter how many annotators you add. In most in-house setups that reviewer is the ML engineer who owns the model — the most expensive person on the project, spending their week scrubbing playback.

3. Spec drift

Halfway through, someone decides that objects under 20 pixels should be skipped, or that re-entering objects keep their ID after all. In image work you re-label the affected images. In video, a rule change about identity retroactively invalidates every track that crossed an occlusion — including ones already reviewed and signed off.

None of these means outsourcing is automatically correct; the trade-offs are laid out in our guide to data annotation outsourcing. What they do mean is that the decision should be made before extraction, not after the review queue has already backed up. If the footage itself is sensitive — recorded in public space, on a work site, or containing faces and plates — our handling of that is documented on our security page; we are certified to ISO/IEC 27001, work cloud-only with no local retention, and never reuse client data to train our own models.

What drives the cost of a video annotation project

Because video is annotated as extracted frames, our published per-item rates are what apply once extraction is done. There is no separate “per second of video” rate, and any vendor quoting one is making an assumption about your sampling rate on your behalf.

Four variables move the total, and you control all of them at design time:

• Annotation type. Boxes, keypoints and segmentation differ by roughly an order of magnitude per item. Ask whether your model genuinely needs pixel-level regions across every frame, or only on a subset.
• Object density. Cost is per label, not per frame. Eight objects in frame costs eight times what one object costs, so a crowded scene is a different project from an empty road.
• Sampling rate. Covered above, and the largest single lever you have.
• Tracking. Whether you need persistent IDs at all. If your model only needs per-frame detection, dropping the identity requirement removes the most judgment-heavy part of the work.

Pricing is fully usage-based with no initial cost and no minimum order — full details are on our pricing page. Because the frame count depends entirely on your footage and sampling decisions, we quote video projects individually; send us a representative clip and the spec questions from H2-3 answered, and you can request a quote against real numbers rather than an assumed frame rate.

Starting with a PoC

Video specifications are hard to get right on paper, which is why we do not recommend committing a full dataset up front. The path we suggest has three stages.

Free trial. We annotate roughly 10 to 50 items of your real data with our production team, so you can assess actual quality before placing any order. This is not a sales demo run by a specialist — it is the workflow you would get.

PoC. There is no minimum-order requirement. For image annotation we flexibly take on PoCs from around 50 to 100 items, which for video means a short representative sequence rather than a clip count. Choose the difficult footage here: the crowded intersection, the low-light clip, the one with the occlusion you argued about internally. A PoC on clean footage tells you nothing you needed to know.

Production. Once the spec survives contact with hard data, scale up. Because billing is per item with no fixed fees, scope can move up or down between batches without renegotiating a contract.

One practical note for teams working with footage or documentation in Japanese: our annotators are Japanese-native, so specs, edge-case discussions and on-site terminology do not have to be translated into English before work can start.

Frequently asked questions

Summary

Video annotation is not image annotation at a larger volume. The added requirement is consistency across frames, and it shows up in three places: the sampling rate that sets your total scope, the identity rules that drive most of the judgment cost, and the track-level criteria that determine whether a delivered dataset is usable. Decide all three before extraction, and validate them on your hardest footage rather than your cleanest.

Free TrialTry ANOSUPO annotation free. Test our quality and communication on a sample of your real data before you commit — images, video, 3D/LiDAR and LLM evaluation.annotation-support.com

投稿 Video Annotation Services: How to Build a Tracking Dataset は アノサポ｜AIアノテーション・学習データ作成の伴走パートナー に最初に表示されました。

That’s the point where outsourcing starts to make sense. This guide covers what actually drives annotation cost, how the three sourcing models differ, seven criteria for evaluating a vendor, and what to prepare before you request a quote.

When outsourcing makes sense — and when it doesn’t

Outsourcing is a good fit when the annotation work is well-defined enough to be written down, and large enough that doing it internally costs you something you care about — usually engineering time.

Outsourcing tends to work when:

• The volume is beyond what your team can absorb alongside model work
• The task can be specified in writing, including the awkward edge cases
• You need consistent output over weeks or months, not a one-off batch
• You need to scale up and down without hiring

Outsourcing tends not to work when:

• The labeling rules are still changing daily because the problem definition isn’t settled
• The task requires domain expertise that only exists inside your team and can’t be transferred through a spec
• The dataset is small enough that the coordination overhead exceeds the labeling effort

If you’re in the second group, the honest advice is to label a few hundred items yourself first — the process is what forces the spec to become concrete. Our guides on building YOLO training data and semantic segmentation annotation walk through how to do that in-house. Once the spec stabilizes, outsourcing becomes a volume question rather than a judgment question.

In-house vs. crowdsourcing vs. specialist vendor

There are three realistic ways to get labeled data, and they fail in different ways.

Each model has a characteristic failure. In-house fails quietly: nobody logs the hours, so the cost shows up as delayed model releases rather than as a line item. Crowdsourcing fails on difficulty — as soon as a task requires judgment rather than pattern matching, inter-annotator agreement drops and you spend your savings on review. This is the exact reason Kyoto University came to us for evaluation data on a conversational AI project: the task was too subjective for a crowd platform to hold quality on, and it needed a selected, trained group of workers instead. Specialist vendors fail when the spec was never nailed down — a vendor will faithfully produce whatever you asked for, including the wrong thing.

You can see how these played out across different projects on our case studies page.

OMRON SINIC X is a concrete example of the third pattern. They had built their own annotation platform and hired annotators directly, but crowdsourced workers dropped off mid-task and required constant checking. Simple tasks were manageable in-house; complex ones needed a team they could rely on.

Complex, specialized annotation handled with custom UIs and flexibilityHow ANOSUPO walked alongside OMRON SINIC X's AI research through annotation — flexible handling of specialized tasks, custom UI design, and a sharp cut in dataset creation time.annotation-support.com

The three models also differ in who owns the labeling guidelines and who absorbs rework — the part that decides your real cost. We break that down in our guide to data labeling services.

What data types change your choice of vendor

Vendor shortlists narrow fast once you name your data type. Most annotation companies handle 2D images competently. Fewer handle the rest.

• Images and video — bounding boxes, segmentation, keypoints, object tracking across frames. This is the entry point for most teams and the broadest market. See image & video annotation.
• 3D point cloud and LiDAR — 3D cuboids, 3D segmentation, and sensor fusion where LiDAR and camera data must be labeled consistently across calibrated sensors. See 3D point cloud & LiDAR annotation, or our 3D point cloud annotation guide for what the work involves.
• Text, speech and LLM data — RLHF preference pairs, DPO, SFT instruction-response sets, multi-level LLM evaluation, transcription. See LLM, text & speech annotation.
• Data collection and preprocessing — sourcing or shooting the raw data, cleaning, formatting, anonymization. See data collection & preprocessing.

Two thresholds sharply reduce the number of viable vendors. The first is sensor fusion: labeling a 3D cuboid is one skill, keeping it consistent with the corresponding camera view across a calibrated rig is another. The second is tasks that need a custom annotation interface — when the labeling schema doesn’t fit any off-the-shelf tool. For OMRON SINIC X we built a dedicated UI for a complex, non-standard task rather than forcing the work into a generic tool. If your task falls into either category, ask about it in the first conversation, not after signing.

Video deserves a separate note here. It is often treated as an image project with more files, but the vendor requirements are different: you need someone who can maintain object identity across frames, not just draw accurate boxes. Our guide to video annotation and building a tracking dataset covers what changes and how it affects your estimate.

What actually drives annotation cost

Unit price is not one number. Six factors move it:

1. Annotation type — a polygon takes several times longer than a box, and price follows time.
2. Object density — 40 objects in a frame costs more than 3, because pricing is usually per object rather than per image.
3. Number of classes — more classes means more decisions per object and more chances to disagree.
4. Spec ambiguity — if “partially occluded” isn’t defined, someone has to decide, and inconsistency is expensive to fix later.
5. Inspection regime — sampling-based QA is cheaper per unit than full-volume inspection, and the difference shows up in your error rate.
6. Security requirements — isolated teams, restricted environments and non-retention policies all carry operational cost.

Many annotation vendors in the English-speaking market don’t publish rates at all — you enter a sales process before you can estimate anything. That’s a real cost in itself, because you can’t size a project or compare options without a number. We publish ours:

There is no setup fee and no management fee — billing is fully usage-based, so you pay for what’s annotated. Full rate details are on our pricing page.

One thing a unit price can’t tell you: whether the work will need redoing. A low rate with sampled QA and a high rate with full-volume inspection are not the same product, and the gap only becomes visible when you train on the data. That’s why we recommend running a small paid or free pilot before committing volume — it converts an unknown into a measured error rate.

7 criteria for choosing an annotation vendor

Seven things to evaluate, with the question to ask for each.

1. Quality assurance method. Sampling and full-volume inspection produce different error profiles. Ask: do you inspect every item, or a sample — and what percentage?
2. Security and data handling. Certification, retention policy, worker NDAs, whether your data is reused for the vendor’s own model training. Ask: where is our data stored, who can see it, and what happens to it after delivery?
3. Data type coverage. Whether the vendor can follow you from 2D images into 3D, video or LLM data as your project grows. Ask: which of these have you delivered in the last year?
4. Pricing transparency. Whether you can estimate before entering a sales cycle, and whether setup or management fees exist. Ask: what’s the unit price, and what else appears on the invoice?
5. Minimum order and pilot options. A high minimum forces you to bet before you have evidence. Ask: what’s the smallest batch you’ll take, and can we trial on our own data first?
6. Spec-building support. Most first specs are incomplete. A good vendor surfaces the edge cases before production, not after. Ask: who writes the annotation guidelines, and how are ambiguous cases escalated?
7. Post-delivery correction. What happens when delivered data doesn’t match the agreed spec. Ask: what’s covered, for how long, and at what cost?

For reference, our answers: full-volume inspection rather than sampling, with 99.7% quality consistency; ISO/IEC 27001 (ISMS) certification with a non-retention, cloud-only policy detailed on our security page; all four data categories above; published unit prices with no setup or management fee; no minimum order, with image annotation starting from PoC batches of 50–100 items; joint spec development including edge-case definition; and free correction for one year on anything that doesn’t match the agreed spec and is attributable to us.

A pre-order checklist

Settle these internally before you request quotes. Every item you can answer makes the quotes you get more accurate and more comparable.

• ☐ Annotation definition, including edge cases — occlusion, truncation at frame borders, minimum object size, ambiguous classes
• ☐ Rough volume — how many items for the first milestone, and what the full dataset looks like
• ☐ Output format — COCO, YOLO, Pascal VOC, or something custom your pipeline expects
• ☐ Evaluation metric — how you’ll judge whether the delivered data is good enough
• ☐ Deadline — including whether it’s one delivery or a recurring pipeline
• ☐ Confidentiality classification — what the data contains and what handling it requires
• ☐ Sample data — a representative subset, including the hard cases, not just clean examples
• ☐ Internal owner — who answers the vendor’s spec questions, and how fast

The last one matters more than it looks. Annotation projects stall on unanswered questions far more often than on labeling capacity.

Data residency, NDAs and time zones

If you’re outsourcing across borders, three questions come up in procurement before anything technical does.

Where the data is processed. We work cloud-only and don’t retain data locally. Data is separated by project, teams are isolated per engagement, and everything is physically deleted after completion. Your data is never used to train our own models.

The confidentiality framework. All staff work under NDA, and we hold ISO/IEC 27001 (ISMS) certification. In many Western procurement processes, certification is a pass/fail gate before capability is even discussed — so it’s worth confirming early rather than late.

Time zones. We’re based in Fukuoka, Japan. In practice this is a workflow design question rather than a drawback: asynchronous handoff means work progresses while your team is offline, provided the spec is clear enough that the annotation team isn’t blocked waiting for answers. We structure projects around a written spec and a single communication channel — for Pirika, a Slack-centered setup let us minimize meetings entirely while scaling throughput.

Frequently asked questions

Why teams choose ANOSUPO

• Published unit prices — you can estimate before talking to sales
• No minimum order, no setup or management fee — fully usage-based, so a pilot costs pilot money
• Full-volume inspection, not sampling — 99.7% quality consistency
• ISO/IEC 27001 certified, cloud-only, non-retention, per-project team isolation
• All four data categories — images and video, 3D point cloud and LiDAR, LLM/text/speech, and data collection and preprocessing
• One year of free correction against the agreed spec
• 1,000+ projects delivered for companies, universities and research institutions since 2021

ANOSUPO is operated by Borderless Japan and based in Fukuoka. Recognition includes Forbes 30 Under 30 Asia 2023, the Good Design Award 2023, and selection for Japan Innovation Campus, the Japanese government’s startup hub in the United States.

If you already know your data type and volume, request a quote and we’ll price it directly. If you’d rather see the output before deciding, start with the free trial below.

Free TrialTry ANOSUPO annotation free. Test our quality and communication on a sample of your real data before you commit — images, video, 3D/LiDAR and LLM evaluation.annotation-support.com

投稿 Data Annotation Outsourcing: How to Choose a Vendor (2026 Guide) は アノサポ｜AIアノテーション・学習データ作成の伴走パートナー に最初に表示されました。

The same problems show up whether you are building perception for autonomous vehicles, warehouse robotics, drones, construction site monitoring, or survey and mapping — anywhere a LiDAR sensor produces frames that a model has to learn from.

This guide covers what 3D point cloud annotation actually involves: which annotation types exist, the five steps of building a dataset, what drives cost per frame, where in-house point cloud annotation tends to break down, and how to evaluate an outsourcing partner. It is written for engineers and project managers in autonomous driving, ADAS, and robotics who have shipped 2D annotation before and are now sizing up the 3D version of the same problem.

When You Actually Need 3D Point Cloud Annotation

Not every perception problem needs 3D. Adding LiDAR to a project that a camera could have handled multiplies your data cost for no gain, so the first decision is whether you need it at all.

2D image annotation is usually enough when your model only needs to answer “is this object present, and where is it in the image?” Classification, counting, defect detection on a fixed camera, and most retail or inspection use cases fall here. If you are at this stage, our guide to creating training data for YOLO covers the 2D workflow end to end.

You need 3D point cloud annotation when the model has to answer “how far away, how big, which direction is it facing, and where will it be next?” Metric distance and heading cannot be recovered reliably from a single 2D box. Concretely:

• Autonomous driving and ADAS. Path planning needs the distance and orientation of every vehicle, pedestrian, and cyclist in metric space, not pixel space. Braking decisions depend on it.
• Robotics. Point cloud annotation for robotics AI supports navigation, obstacle avoidance, and grasping, where the robot needs the geometry of free space as much as the identity of objects.
• Infrastructure and site inspection. Aerial or terrestrial LiDAR scans of roads, bridges, power lines, and construction sites, where measurement is the output.
• Night, glare, and adverse weather. LiDAR does not depend on ambient light, so datasets covering these conditions often rely on it where a camera-only dataset would fail.

A useful middle case: some teams start with 2D annotation on the camera stream, confirm the model concept works, and only then extend to LiDAR point cloud annotation on the same scenes. That sequencing keeps early costs low and is worth considering before committing to a 3D dataset.

Annotation Types for LiDAR Data

“3D annotation” is not one task. The four types below have very different labor profiles, and mixing them into a single project without deciding which you need is one of the most common ways a budget gets away from a team.

3D cuboid annotation services are the entry point for most detection projects and the type teams request first. Segmentation is where the per-frame labor climbs steeply, because the unit of work moves from “one box per object” to “every point in the scene.” If you are weighing the same question in 2D, the trade-off is structurally identical and covered in our semantic segmentation annotation guide.

Tracking IDs deserve a specific warning. Annotating 200 isolated frames and annotating a 200-frame continuous sequence with persistent IDs are not the same job, even though the frame count matches. The second requires the annotator to hold object identity across occlusion events, and it costs meaningfully more. Decide which you need before you request a quote.

3D Point Cloud & LiDAR AnnotationLiDAR annotation services for autonomous driving, robotics and surveying: 3D cuboids, segmentation and sensor fusion. Full-volume QA, from $0.121 per cuboid.annotation-support.com

Building a 3D Point Cloud Dataset: Five Steps

1. Collect and sample frames

A LiDAR sensor running at 10 Hz produces 36,000 frames per hour of driving. Annotating all of them is neither affordable nor useful, because consecutive frames are highly redundant.

Sample instead. Take keyframes at a fixed interval, then deliberately oversample the situations your model will fail on: intersections, merges, dense pedestrian crossings, night, rain, and unusual objects. A dataset of 1,000 well-chosen frames will beat 10,000 frames of empty highway. Scene diversity, not frame count, is what moves accuracy.

2. Define classes and write the annotation spec

This is the single highest-leverage step, and the one teams consistently under-invest in. In 3D, three questions cause almost all of the spec drift you will encounter later:

• Occlusion. A vehicle is 60% hidden behind a truck. Do you box the visible points only, or do you extrapolate the full vehicle extent from the visible portion? Both are defensible. Only one can be in your dataset.
• Sparse distant objects. At what point does an object have too few points to label? Define a minimum point count or a maximum range, and state it explicitly. “Use your judgment” guarantees inconsistency.
• The ground plane. Does a cuboid extend down to the road surface, or does it stop at the lowest returned point on the object? Wheel-level returns are noisy, and this choice shifts every box in the dataset.

Write the answers down, with annotated example frames showing the edge cases. A spec with pictures resolves disputes that a spec with prose does not.

3. Annotate

Work in a point cloud annotation tool that supports oriented cuboid fitting, multi-view panels (top-down, front, side), and frame-to-frame interpolation. Interpolation matters more in 3D than in 2D: annotating an object at frame 1 and frame 20 and letting the tool interpolate the intermediate poses removes a large share of the manual work in sequence projects. Open-source and commercial options both exist, and the right choice depends on whether you need sensor fusion views and how your team handles review. We will cover tool selection in detail in a follow-up article.

Whichever tool you use, plan the review pass at the same time as the annotation pass. In 3D point cloud labeling, review is not a formality — see the cost section below.

4. Convert and split

Export to a format your training pipeline already reads. KITTI and nuScenes are the common public conventions, and most detection frameworks ingest one or both. Confirm two things before you export at scale: the coordinate convention (which axis points forward, where the origin sits) and the heading angle convention. A 90-degree systematic rotation error is easy to introduce and hard to notice until your evaluation numbers are inexplicably poor.

Split train, validation, and test by scene or by drive, never by randomly shuffling individual frames. Consecutive frames from the same drive are near-duplicates; shuffling them puts near-identical data on both sides of the split and produces an evaluation score that flatters your model and will not survive deployment.

5. Train, evaluate, and go back to the data

Train a baseline on the smallest dataset that could plausibly work, then look at the failure cases before annotating more. In 3D, failures cluster into recognizable groups: distant objects, heavily occluded objects, and rare classes. Each of those points at a specific gap — in your sampling, in your spec, or in class balance — and fixing the gap is usually cheaper than adding volume indiscriminately. Expect to revise the spec at least once after the first training run. That is normal, not a planning failure.

How Much Does 3D LiDAR Annotation Cost per Frame?

There is no single per-frame price, because a “frame” is not a fixed unit of work. A highway frame with four vehicles and an intersection frame with forty vehicles and pedestrians differ by an order of magnitude in labor while both counting as one frame.

Pricing at ANOSUPO is per annotated object, not per frame, and billing is fully usage-based with no setup fee, no management fee, and no minimum order:

The 2D rates are shown for comparison: a 3D cuboid carries roughly three times the unit price of a 2D bounding box, which reflects the additional work of fitting depth, extent, and heading rather than a rectangle. Your effective cost per frame is therefore the object count in that frame multiplied by the unit rate, adjusted for the factors below. At the entry rate of $0.121 per cuboid, the highway frame above works out to about $0.48 and the intersection frame to about $4.84 — a tenfold difference for the same single frame, which is why frame count alone is a poor basis for a budget.

What moves the number:

• Objects per frame. The dominant factor. Scene density drives cost more than anything else on this list.
• Class count and attributes. Five classes with occlusion flags and heading precision requirements cost more per object than two classes with no attributes.
• Tracking across frames. Persistent IDs through occlusion add work that isolated-frame annotation does not have.
• Sensor fusion. Linking each 3D instance to its camera counterpart adds a second pass over the same objects.
• Inspection standard. Full-volume inspection costs more than spot-checking a sample, and produces a materially different dataset.

For current published rates across all annotation types, see our pricing page. For a number specific to your sensor setup, scene density, and spec, request a quote — we can estimate from a handful of representative frames.

PricingTransparent annotation pricing — Bbox from $0.036/label, keypoints from $0.021/point, segmentation from $0.152/region, 3D cuboids from $0.121, LLM evaluation from $0.121. $0 setup, fully usage-based, no minimum order.annotation-support.com

Three Places In-House 3D Annotation Breaks Down

Teams that successfully ran 2D annotation in-house often assume 3D is the same job with an extra axis. Three things are different, and each one has ended 3D annotation projects.

1. Occlusion rules fracture between annotators. In 2D, the outline of an object is visible and two annotators mostly agree. In 3D, a partially occluded vehicle has no observable “true” extent, so annotators infer it — and they infer it differently. The result is low inter-annotator agreement on exactly the cases your model most needs to get right. This is a specification problem, not a diligence problem, and it does not resolve by asking people to be more careful.

2. Inspection is heavy, and sampling does not work. Reviewing a 2D box is a glance. Reviewing a 3D cuboid means rotating the scene, checking the top-down view, checking the side view, and confirming the heading. Because 3D errors are systematic rather than random — a misunderstood occlusion rule affects every occluded object in the batch — inspecting a 10% sample tells you very little. You either inspect everything or you accept unknown quality. ANOSUPO uses full-volume inspection rather than sampling, with 99.7% quality consistency, and corrects any defect attributable to us against the agreed specification free of charge for one year after delivery.

3. Engineering time disappears into labeling. The engineers who understand the spec well enough to annotate correctly are the engineers who should be training models. In 3D, per-frame labor is high enough that this trade becomes expensive fast, and it is rarely visible in a budget line.

How to Choose a 3D Point Cloud Annotation Provider

Many vendors that offer image annotation list 3D as a capability without having run it at volume. Use this checklist when evaluating any 3D point cloud annotation company:

• Which 3D types can they actually deliver? Ask specifically about cuboids, point-level segmentation, cross-frame tracking IDs, and sensor fusion. These are four different capabilities. A provider handling cuboids well may have no segmentation workflow at all.
• How do they inspect? Full-volume or sampling? If sampling, at what rate, and what happens when a systematic spec error is found in the sample — is the whole batch reworked?
• How is security handled? Vehicle and site data typically contains public spaces, faces, and license plates, and is often commercially sensitive as well. Look for a certified framework rather than assurances. ANOSUPO is certified to ISO/IEC 27001 (ISMS), operates cloud-only with no local retention, has NDAs with all staff, separates teams per project, never reuses client data to train our own models, and physically deletes data at project close. Details are on our security page.
• How do they bill? Setup fees, management fees, and minimum order volumes change the economics of a first project far more than the unit rate does.
• Can you start small? A provider confident in their work will annotate a sample of your real data before you commit. If a vendor requires a large minimum commitment to begin, you cannot evaluate them on your own data — only on their marketing.
• Do they engage with your spec? The best signal in a first conversation is whether the provider asks about occlusion handling, ground-plane convention, and minimum point thresholds. Providers who ask these questions have annotated point clouds. Providers who only ask for frame counts have not.

Comparing 3D point cloud annotation companies on unit price alone is a mistake, because unit price is only meaningful once the spec and inspection standard are held constant. A lower per-cuboid rate with sampled inspection and no rework guarantee can easily cost more once you account for the model retraining that follows a bad batch.

3D point cloud work narrows your vendor options considerably. Our data annotation outsourcing guide lists seven criteria for evaluating an annotation partner.

In-House vs. Outsourcing

Neither answer is universally correct. The split usually follows project stage.

A common and effective pattern: keep the first few hundred frames in-house to discover what your spec actually needs to say, then hand the stable spec and the volume to a partner. The in-house phase is spec development, not data production.

Our 3D LiDAR annotation services cover cuboids, point cloud segmentation, cross-frame tracking, and sensor fusion — see our LiDAR annotation services page for scope. For research and robotics projects where the task does not fit a standard workflow, our OMRON SINIC X case study describes a complex, specialized annotation task handled with a purpose-built interface.

Complex, specialized annotation handled with custom UIs and flexibilityHow ANOSUPO walked alongside OMRON SINIC X's AI research through annotation — flexible handling of specialized tasks, custom UI design, and a sharp cut in dataset creation time.annotation-support.com

Frequently Asked Questions

Getting Started

3D point cloud annotation is not 2D annotation with an extra axis. The specification questions are harder, inspection is heavier, and per-frame labor is high enough that decisions made early — what to sample, how to handle occlusion, whether you need tracking or fusion — determine your budget more than any unit rate does.

The practical starting point is small: pick a few dozen representative frames including the hard scenes, write the spec down with example images, and annotate them. You will learn more about what your dataset needs from those frames than from any amount of planning. From there, the decision of what to keep in-house and what to hand off becomes a straightforward one.

If you would like a second opinion on your spec, or a cost estimate based on your actual frames, we are happy to look at them with you.

Free TrialTry ANOSUPO annotation free. Test our quality and communication on a sample of your real data before you commit — images, video, 3D/LiDAR and LLM evaluation.annotation-support.com

投稿 3D Point Cloud Annotation: Workflow, Cost per Frame, and How to Choose a Vendor は アノサポ｜AIアノテーション・学習データ作成の伴走パートナー に最初に表示されました。

This guide covers when segmentation is actually necessary, a 5-step workflow for building a segmentation dataset, where teams typically get stuck doing it in-house, and how to start small with a PoC instead of committing to a full dataset upfront.

When You Actually Need Segmentation

Segmentation matters when the shape of an object — not just its location — determines whether your model is useful. Three cases come up repeatedly:

• Medical imaging: precisely delineating a tumor, organ, or lesion boundary
• Autonomous driving / ADAS: pixel-level drivable-area, lane, and pedestrian boundaries
• Manufacturing visual inspection: measuring the exact shape and area of a defect

If your use case only needs to know whether and roughly where an object is present — inventory counts, simple detection alerts — bounding boxes are usually sufficient and considerably cheaper to produce. Getting this distinction right up front avoids paying for precision you don’t need.

For point-level segmentation of LiDAR data rather than images, see 3D point cloud annotation.

Bounding Boxes vs. Segmentation

The higher per-unit cost of segmentation reflects the labor of tracing a precise boundary rather than dropping four points. It’s easy to over-specify — requesting pixel-perfect masks for a task that only needed coarse localization — so it’s worth confirming the precision requirement before committing to a labeling spec.

Building a Segmentation Dataset: 5 Steps

1. Collect representative data

Start by gathering images or video that reflect the conditions your model will actually see in production. Medical images vary by device and facility; driving scenes vary by weather, time of day, and region. Most downstream accuracy problems trace back to insufficient diversity at this stage, not to the model architecture.

2. Define classes and a labeling spec

Next, define exactly what gets labeled and how. This is where class-boundary ambiguity causes the most trouble: should “pedestrian” and “cyclist” be separate classes? How should overlapping or adjacent objects be handled — allowed to overlap slightly, or forced to have a gap? Writing these rules down before annotation begins prevents a costly full re-label later. Annotation guidelines here function the same way a style guide functions for text — without one, inter-annotator agreement drops fast.

3. Annotate

With the spec in place, annotators trace boundaries using polygon or brush tools. Many teams now use semi-automatic tools like Segment Anything (SAM) to generate a rough mask that a human then corrects — this speeds up throughput, but raw model output shouldn’t be accepted as-is. Boundary precision from automated pre-labeling varies enough that a review step is non-negotiable if consistency matters.

4. Convert format and split the dataset

Once labeling is done, convert annotations into the format your training pipeline expects — commonly COCO-style polygon coordinates or PNG mask files — and split the data into train, validation, and test sets.

5. Train, evaluate, and go back to the data

Train the model and evaluate using metrics like IoU (Intersection over Union). When accuracy falls short, the fix isn’t always architectural — check which classes had the most boundary ambiguity or which scenarios were underrepresented in the training set. In a data-centric AI workflow, the dataset is usually where the fastest gains are found.

Where In-House Teams Get Stuck

Three problems show up consistently in teams building segmentation datasets internally:

• Boundary drift: different annotators trace edges differently, and consistency degrades as more people are involved — a direct inter-annotator agreement problem
• Underestimated QA time: review and correction usually take longer than the initial labeling pass, and teams rarely budget for it upfront
• Engineering time diverted: hours meant for model development get absorbed by labeling logistics and spec revisions instead

None of this means segmentation should always be outsourced — but any in-house plan should account for these costs explicitly rather than discovering them mid-project.

Pixie Dust Technologies ran into exactly this. Color-coding rebar by type for their construction-site inspection AI is segmentation work, and doing it in-house alongside algorithm development was estimated at four months. A six-person annotation team completed it in two, returning roughly two months of engineer time to the development work.

Data for construction-site AI delivered at scale, fastCase study: Pixie Dust Technologies' rebar-inspection DX "TOTTARROW" recognition AI. A dedicated 6-person ANOSUPO team delivered high-quality, color-coded rebar annotation in two months — saving about two months of in-house engineer time and freeing the team for core algorithm work.annotation-support.com

Start Small: The Case for a PoC

Estimating “how many regions we’ll need” precisely, before you’ve labeled anything, is close to guesswork. A more practical approach: run a PoC on 50–100 images first, confirm both model behavior and the soundness of your labeling spec, and only then scale to a full dataset. If your raw data also needs cleanup or preprocessing before labeling can start, it’s worth scoping that alongside the PoC — see data collection and preprocessing for what that typically involves.

Data Collection & PreprocessingData collection & preprocessing outsourcing — collection/capture, cleansing, anonymization & annotation preprocessing. We collect hard-to-source data from scratch. 99.7% quality, $0 setup, fully usage-based.annotation-support.com

In-House vs. Outsourcing

Segmentation is one of the more expensive annotation types, so it’s often the first task teams outsource. See our data annotation outsourcing guide for what drives unit price and how to compare vendors.

Working with a Japan-Based Annotation Partner

ANOSUPO, operated by Borderless Japan, provides segmentation annotation as part of its image and video annotation service. Instead of sampling-based spot checks, every deliverable goes through full-volume inspection, with a measured label consistency of 99.7%. Pricing is fully usage-based — no setup fee, no account management fee, and no minimum order — with segmentation PoCs typically starting around 50–100 images. Any defects traced to our process are corrected free of charge for one year after delivery. Rate details are on the pricing page.

PricingTransparent annotation pricing — Bbox from $0.036/label, keypoints from $0.021/point, segmentation from $0.152/region, 3D cuboids from $0.121, LLM evaluation from $0.121. $0 setup, fully usage-based, no minimum order.annotation-support.com

On the security side, the company holds ISO/IEC 27001 (ISMS) certification, doesn’t retain or process data locally (cloud-only), has NDAs with all staff, isolates teams per project, and physically deletes data once a project is complete — none of it is used to train the company’s own models.

For teams working primarily with bounding boxes rather than segmentation, our companion guide on building YOLO training data covers the box-annotation workflow in the same level of detail.

How to Create Training Data for YOLO: Annotation Workflow and Starting Small with a PoCA five-step workflow for YOLO training data: class design, annotation, label format, and sizing your first PoC.annotation-support.com

Frequently asked questions

SecurityANOSUPO security. ISO/IEC 27001 (ISMS) certified, with a strict no-retention/no-local-storage policy, NDAs with all staff and per-project team isolation to keep your training data safe.annotation-support.com

Summary

Segmentation earns its cost when shape — not just location — determines whether a model is useful. The workflow itself isn’t exotic: representative data, a clear labeling spec, careful annotation, format conversion, and a feedback loop back from evaluation to the dataset. Where teams typically lose time is boundary consistency and QA, in-house or out. Starting with a small PoC, rather than committing to a full dataset upfront, is usually the fastest way to find out which approach fits.

Free TrialAnnotate a sample of your own data under our production workflow before committing to a full order.annotation-support.com

投稿 Semantic Segmentation Annotation: Pixel-Level Data and Outsourcing は アノサポ｜AIアノテーション・学習データ作成の伴走パートナー に最初に表示されました。

Tutorials work because someone already did the hard work of preparing a clean, consistently labeled dataset. In production, that preparation is the project. It is also where the “garbage in, garbage out” cliché stops being a cliché and starts costing you weeks.

This guide walks through the five steps of building YOLO training data, what the label format actually contains, and the three failure modes that quietly wreck accuracy. It ends with a more useful question than “how many images do I need?”

The short version: five steps, and the one that decides your accuracy

Building a YOLO training dataset breaks down into five steps:

1. Collect the data
2. Define classes and write annotation guidelines
3. Annotate
4. Convert to YOLO format and split into train and validation sets
5. Train, evaluate, and return to the data

Step 3 consumes the time. Step 2 determines the accuracy. Most teams notice the pain of step 3 first and start looking for help there. But the datasets that underperform usually failed at step 2, months earlier, when nobody wrote down what “a valid object” actually means.

The five steps of building YOLO training data

Step 1: Collect data that matches production

Collect images under conditions as close as possible to where the model will run. Variety matters more than raw volume:

• Lighting: daylight, night, backlight, overcast
• Distance and angle: close, far, overhead, oblique
• Object state: occluded, overlapping, partially visible
• Background: different sites, seasons, levels of clutter

If a condition never appears in training, the model meets it for the first time in production. Conversely, ten thousand near-identical frames will not buy you much: a hundred genuinely different scenes usually beats a thousand near-duplicates.

If you have no imagery yet, or want the capture protocol designed before you start shooting, data collection and preprocessing can be handled as a separate stage.

Data Collection & PreprocessingNo imagery yet? We design the capture protocol, then clean, format and anonymize what comes back.annotation-support.com

Step 2: Define classes and write annotation guidelines

Annotation guidelines are the written rules for what to label, how tightly, and what to skip. Without them, the same object in the same scene gets labeled differently depending on who is working and what day it is. That inconsistency is label noise, and the model learns it faithfully.

Decide these before anyone draws a box:

• Class definitions: how many classes, and whether visually similar objects are merged or split
• Occlusion: are partially hidden objects labeled, and above what percentage hidden do you skip them
• Truncation: are objects cut off at the frame edge labeled, and do you box only the visible part
• Minimum size: below how many pixels is an object out of scope
• Box tightness: flush to the object, or with a small margin

Ambiguous cases will come up regardless. The goal is not to find the objectively correct answer — it is to make sure the same edge case always gets the same treatment. Keep a running gallery of screenshots showing resolved edge cases and append it to the guidelines. This is what keeps inter-annotator agreement from drifting once more than one person is working.

Step 3: Annotate

This is the part where someone draws boxes. Tooling ranges from lightweight local applications such as LabelImg to server-based platforms such as CVAT or Label Studio that handle multi-user workflows and progress tracking. Local tools suit small one-off batches; server-based tools suit anything ongoing or collaborative.

Time per image varies enormously with object count and difficulty. An image with one clearly separated object and an image with forty overlapping ones are different jobs, not the same job at different speeds. Estimate in total boxes, not in images — it is the only unit that tracks the actual work.

Step 4: Convert to YOLO format and split the data

YOLO uses one plain-text label file per image. The structure is minimal:

The normalization is the classic trip-up when converting from COCO JSON or Pascal VOC XML, which store absolute pixel values — and COCO stores the top-left corner rather than the center. After any conversion, render a handful of labels back onto the images and look at them. A coordinate-system mistake discovered after a full training run is an expensive way to learn this.

Then split into train and validation sets — and watch for data leakage. If consecutive video frames or shots from the same session land on both sides of the split, your validation score measures memorization, not generalization. Split by scene, session, or site, not by shuffling individual files.

Step 5: Train, evaluate, and go back to the data

Read the metrics, then look at the actual failures. Pull up the false positives and the misses and study them. When one class underperforms, or distant objects consistently drop out, the cause is usually in the dataset rather than the architecture.

Inconsistent labeling, a missing scene, a class boundary nobody agreed on — each of these sends you back to step 2. This is the core insight behind data-centric AI: iterating on the dataset generally beats iterating on the model, and treating data creation as a loop rather than a one-time task makes the whole project easier to plan.

If your project moves from camera to LiDAR, the workflow changes substantially — object extent, occlusion, and heading all become specification questions. See our guide to 3D point cloud annotation.

Three places in-house annotation breaks down

1. Guideline drift. The moment a second person joins — or the same person returns after a week away — judgments diverge. The model learns the divergence.
2. QA gets deferred. Producing labels absorbs all available capacity, so review becomes a spot check of a few files. Whatever the spot check missed surfaces later, usually during evaluation.
3. Engineering time evaporates. Researchers and engineers end up drawing boxes instead of designing experiments, which slows the entire iteration cycle.

None of these is solved by working harder. They are problems of written process, of a dedicated review stage, and of how people are allocated.

At ANOSUPO we review every delivered item rather than sampling, and maintain 99.7% quality consistency. For one year after delivery, we correct any defect attributable to us against the agreed specification at no charge.

Pirika hit this ceiling directly. In-house annotation topped out at roughly 1,000 items a month while their video data kept growing. After moving the work outside, throughput reached 100,000 items a month and cumulative delivery passed one million items, at about half the previous cost.

Data for environmental AI half the cost, at scaleCase study: Pirika's litter-survey AI "Takanome." ANOSUPO took on annotation of ever-growing video data, cutting cost to about half and scaling throughput to 100K items a month and 1M+ delivered — a data foundation in-house work could never reach.annotation-support.com

A better question than “how many images do I need?”

It is the most common question we get, and the honest answer is that there is no universal number. It depends on how visually distinct your classes are, how much your capture conditions vary, what accuracy you need, and how many false positives you can live with.

The practical move is to stop estimating and build a small batch, then train on it. A small run tells you quickly:

• Whether your class definitions hold up against real data
• Which scenes the model struggles with — that is, what to collect next
• Which edge cases your guidelines never anticipated

ANOSUPO has no minimum order quantity. For images, a PoC of roughly 50–100 items is a normal starting point. Before that, our free trial covers roughly 10–50 items of your real data, annotated by the same team that would handle the project, so you can assess quality before committing.

One caveat: everything above assumes still images. If your source material is video, the count question changes shape — you are choosing a frame sampling rate, and that one setting can move your total volume by an order of magnitude. See our guide to video annotation and tracking datasets for how to set it.

In-house or outsourced: choosing per phase

This is not an argument for outsourcing everything. It is a question of phase.

On cost: ANOSUPO bills purely by volume, with no setup fee and no management fee.

For YOLO-style bounding boxes, the first row applies. Since the estimate is simply total boxes × unit price, counting the objects in a handful of representative images is enough to size a project roughly. See our pricing page for other task types and how quotes are built.

Beyond bounding boxes, we handle segmentation, pose estimation, and video tracking — details on the image and video annotation page.

Image & Video AnnotationBounding boxes, segmentation, pose estimation and video tracking — reviewed at 100%, priced per unit.annotation-support.com

If in-house labeling has stopped scaling, our guide to data annotation outsourcing covers cost drivers and how to evaluate a vendor.

Working with a Japan-based annotation partner

ANOSUPO is operated by Borderless Japan, founded in 2021 and based in Fukuoka, Japan. Our team was selected for Forbes 30 Under 30 Asia in 2023.

Two things tend to matter to teams evaluating us from outside Japan. The first is security posture: we hold ISO/IEC 27001 (ISMS) certification, and operate on a non-retention, non-local basis — data is handled only in the cloud and never stored on local machines. Every staff member is under NDA, teams are separated per project, we never reuse client data to train our own AI, and data is physically deleted promptly on completion.

The second is language. If your project involves Japanese-language or Japan-market data — Japanese text, domestic road or retail scenes, or preference data for RLHF and DPO written by native Japanese speakers — that is work we can do natively rather than approximate. See LLM, text and audio for the language side.

Frequently asked questions

Summary

Building YOLO training data means: collect, define, annotate, convert and split, then train and return to the data. Annotation absorbs the hours, but the guidelines written beforehand decide the outcome.

And since nobody can tell you the right dataset size in advance, the reliable approach is to build a small batch, train once, find the weaknesses, and grow the dataset from there.

If you are at the stage of finding out whether this works on your data, start with a slice of the real thing. Quality and working style are both easier to judge from one small batch than from any proposal document.

Free TrialTry ANOSUPO annotation free. Test our quality and communication on a sample of your real data before you commit — images, video, 3D/LiDAR and LLM evaluation.annotation-support.com

投稿 How to Create Training Data for YOLO: Annotation Workflow and Starting Small with a PoC は アノサポ｜AIアノテーション・学習データ作成の伴走パートナー に最初に表示されました。