An industrial robot cell with machined brackets, fittings and a sensor housing moving along a conveyor belt

Anyone can label the easy frames. The hard ones decide the model.

YPAI settles the taxonomy, the edge-case rules and the review method before production, so occlusion, thin boundaries, tiny parts and dense instances are labelled the same way in every frame.

Scope an annotation project
What the guideline has to settle
  • Occlusion how much of the object must be visible to count
  • Thin boundary where the edge sits when the contrast does not
  • Tiny object the size below which an instance is out of scope
  • Dense instances when touching objects are one label or many

Drawing the shape is the easy half. Deciding what counts is the work.

Two annotators can draw the same object and still disagree. What separates them is not skill, it is whether the guideline already answered the question:

A drawn shape

  • geometry
  • a class name

A decided label

  • what counts as the object at all
  • where the boundary sits when contrast fails
  • how much occlusion still earns a label
  • whether a truncated instance is kept or dropped
  • when touching objects are one instance or many
  • the size below which an instance is out of scope
  • whether a deviation is a defect or in tolerance
  • which attributes travel with the instance
  • what an annotator does when the frame is ambiguous
  • who adjudicates, and what the record shows

YPAI settles these questions in the guideline before production, not in review afterwards.

The same holds for industrial vision: a surface mark becomes useful training data only when the guideline states the product, station, process stage, tolerance, viewpoint and inspection method that decide whether it is a defect at all.

The taxonomy is built around the decision the system must make.

That decision may be:

  • which object to manipulate
  • which instance to act on
  • whether an assembly is correct
  • whether a surface contains a defect
  • whether the system should continue, stop, retry or escalate

The ontology, capture protocol, annotation and acceptance plan follow that decision.

Start with the data bottleneck.

YPAI can run one defined stage or the complete data operation.

  1. Entry point: Write the guideline

    Start here when: Two annotators would answer the same frame differently.

    Typical output A written guideline with worked examples, a frozen class list and the rules that resolve disagreement.

  2. Entry point: Label a defined set

    Start here when: You have frames and an agreed acceptance criterion.

    Typical output Labelled frames with per-instance attributes, reviewer records, agreement results and acceptance status.

  3. Entry point: Re-label existing data

    Start here when: You have labels you cannot trust or cannot use across training and evaluation.

    Typical output A re-labelled set against the current taxonomy, with the disagreements found and how each was resolved.

  4. Entry point: Run the annotation operation

    Start here when: You need the work to keep running against a moving taxonomy.

    Typical output A controlled labelling, review and acceptance workflow connected to your model-development environment.

Edge cases

The edge is part of the task.

A model trained on clean frames meets the other kind in production. These four decide the guideline, and the guideline decides whether two annotators agree.

A forklift in a warehouse aisle, its mast and body partly hidden behind a wrapped pallet stack
Occlusion A forklift behind a pallet stack. How much of the object must remain visible before it stops being an instance.
Clear glass bottles on a production line, their outlines almost indistinguishable from the background
Thin boundary Where the edge sits when the contrast does not. Wrapping, shadow and specular surfaces move the boundary a human would draw.
A wide loading yard where the vehicles and containers are small in frame
Tiny object Below some size an instance costs more to label than it teaches. That threshold is a decision, not a preference.
A crate packed with tangled metal fittings, individual parts overlapping and touching
Dense instances Touching and stacked objects. One label or many, and whether the group is itself a class.

Each one is settled in writing before production, with a worked example, so the answer is the same in frame one and frame forty thousand.

Every class should say what it excludes.

The exact taxonomy is written for the customer's product, station, model task and acceptance criteria.

An annotation taxonomy can carry the following layers.

Representative taxonomy

Class definition
  • what the class is
  • what it excludes
  • near-miss classes
  • worked examples
  • known confusions
Geometry
  • label type per class
  • box, polygon, mask or keypoints
  • minimum instance size
  • vertex budget
  • precision tolerance
Boundary rules
  • where the edge sits
  • behaviour when contrast fails
  • shadow and reflection
  • specular and transparent surfaces
  • motion blur
Occlusion and truncation
  • visibility threshold
  • frame-edge rule
  • partial instance handling
  • occluder recorded
  • amodal or visible extent
Instance separation
  • touching objects
  • stacked objects
  • group versus instance
  • instance identity across frames
  • split and merge rules
Attributes
  • per-instance attributes
  • permitted values
  • when an attribute is required
  • null and unknown
  • attribute conflicts
Defect criteria
  • product and station
  • process stage
  • tolerance band
  • severity levels
  • in-tolerance deviations
Ambiguity and escalation
  • what the annotator does when unsure
  • flag reason
  • escalation path
  • who adjudicates
  • how the ruling propagates
Review and acceptance
  • sampling method
  • agreement measure
  • reviewer record
  • rework trigger
  • acceptance status

The taxonomy is frozen before production scale.

A class list is not a taxonomy. The rules that resolve disagreement are.

A compact robot arm on a low wooden table in a domestic living room, positioned above a small block on a tray.
Applied

The taxonomy has to
survive contact with
a real frame.

Every layer above is carried per instance. A guideline that reads well and cannot be applied to this frame is not finished.

A wheeled delivery robot crossing wet pavement beside a kerb in a residential street.
Outdoor and street scenes
A warehouse aisle with racking and a worker in a high-visibility jacket walking away from the camera.
Warehouse and production lines

The output format is
a model decision.

Each label type preserves something different and costs something different. The model task decides which one is right, not annotator preference.

  1. Classification
  2. Box
  3. Polygon
  4. Mask
  5. Keypoints
Classification / one class per image, no locationBox / location and extent, fastest to reviewPolygon / an irregular boundary, ordered verticesMask / a per-pixel region, highest cost to correct

Keypoints and pose sit beyond the ladder: they preserve articulation rather than extent, and they are chosen when the model must act on the object, not only find it.

Industrial vision data should represent the decision on the line.

A defect label without process context may not be enough to train or evaluate the system that must act on it. The same surface mark is a defect at one station and in tolerance at the next.

Overhead view of a warehouse floor with stacked pallets, a worker in a high-visibility vest, and an autonomous mobile robot carrying a loaded pallet.
The same warehouse frame under strong glare.
Glare
The same warehouse frame in low light.
Low light
The same warehouse frame with motion blur.
Motion blur
The same warehouse frame partly occluded.
Occlusion

The delivered label set records not only the class and the geometry, but the capture conditions under which the inspection decision was made, so a disagreement can be traced to the frame rather than to the annotator.

Real data grounds the system.
Synthetic data expands the coverage.

Rare classes and rare conditions are the ones a real capture rarely contains. Synthetic frames arrive pre-labelled, and they pass the same guideline, review and acceptance process as the real ones.

A grid of twelve dark panels, each holding the same sculptural form photographed at a different orientation. Real baseline Synthetic coverage
  1. Baseline label set
  2. Class and condition gaps
  3. Generate
  4. Check plausibility
  5. Evaluate against real frames
  6. Close the loop

Synthetic frames are a coverage tool, not a substitute for acceptance on real ones.

A single sculptural specimen mounted on a dark panel.
Batch-level acceptance

Quality follows the complete label set.

The quality process begins with the guideline, not with review, and ends only when the label-set version is accepted.

AcceptedReworkRe-labelledAdjudicatedRejected

Where accuracy or pass rate is the appropriate metric, YPAI typically proposes a 95% to 99% target range after calibration, with the exact metric, threshold and review method defined in the SOW.

Ships with the accepted version  /  class list & taxonomy · annotation guideline with worked examples · reviewer and adjudication record · agreement report · quality report · provenance & rights record · label-set version record · checksums

Delivered in or mapped to COCO, YOLO, Pascal VOC, KITTI, Parquet, JSONL or a customer-native schema. Format support is confirmed against the actual data, loader, toolchain and delivery requirement.

Different systems need different data operations.  Robot foundation model · manipulation and humanoid · industrial inspection · sensor and device · existing-data recovery.

Send the images, the classes and the acceptance criteria.

The system, the task, the data source, the sensors, the scale, the annotation and the delivery format. YPAI returns feasibility, method, schema, QA and acceptance, delivery format and the open decisions before launch.

Scope a Physical AI data programme Send an existing dataset for review

Related

Scope a Physical AI data project

The form adapts to the work, asks only for relevant details and sends your brief to the person who can act on it.

Service required

Your selection routes the brief to the right person.

A named project lead reviews every enquiry and replies within one business day