Industrial & manufacturing annotation

Industrial computer vision pays for itself faster than almost any other application - a defect detector that catches one bad batch, a safety model that prevents one incident, a process monitor that finds one bottleneck. But industrial datasets have a distinctive shape: the objects are specific to your line (your part, your fixture, your packaging), the defects are rare by definition, and much of the footage is video of continuous processes rather than still images. Payana's pipeline handles all three, and the platform's origins in analyzing real factory-floor and workshop footage show in the details.

Off-the-shelf models know what a "car" is; none of them know what your conveyor bracket looks like, or that a hairline weld crack is the difference between pass and scrap. Industrial vision is therefore custom-dataset territory almost by definition - which makes annotation cost and quality the controlling factors of the whole project. AI-assisted labeling changes that equation: even though the base models have never seen your part, click-to-segment boundary finding works on any object, so annotators produce precise masks of proprietary components at the same speed as everyday objects.

What we annotate on the factory floor

Video-first annotation for continuous processes

Production lines are video, not photos. A part travels through stations; an operator performs a cycle; a bottling line fills, caps, and labels in sequence. Payana's whole-video tracking annotates these as tracks: click the part once as it enters the frame, and the platform follows it with a per-frame mask through the entire traverse - including when a mechanism briefly hides it, thanks to occlusion-aware tracking that marks hidden frames rather than guessing. We validated this workflow on real bottling-line footage: bottles tracked individually along a conveyor, each with a persistent identity from entry to exit.

Frame-level event tags turn the same footage into process data: cycle boundaries, station dwell times, anomaly moments. One annotated video becomes three datasets - a detector (where are parts?), a tracker (which part is which?), and a process log (what happened when?) - from a single labeling pass.

Rare defects and class imbalance

The defining statistical problem of quality inspection is that good parts vastly outnumber bad ones. We handle it on the annotation side three ways: exhaustive labeling of every defect instance in the footage you have (no sampling away of the rare class), severity and defect-type attributes so one pass supports fine-grained classification later, and class-distribution reporting with every delivery so you know exactly how thin the rare classes are before training. When defect examples are genuinely scarce, precise per-pixel masks matter even more - segmentation models extract more learning signal per example than box-only labels, which is why our defect work defaults to instance masks.

Deployment-format delivery

Industrial models overwhelmingly deploy on edge devices next to the line, and the export formats match: YOLO-format labels with data.yaml for the real-time detector families that dominate edge deployment, COCO JSON for segmentation-grade QC models, PNG masks for surface-inspection segmentation, and classification folder trees for grading models. Versioned splits keep line-A footage from leaking into line-B validation, and successive dataset versions let you measure whether last month's model genuinely improved or just memorized new footage.

Confidentiality

Factory footage is competitively sensitive - layouts, processes, and volumes are visible in every frame. Payana runs on infrastructure we control end-to-end: no crowdsourced annotator marketplaces, no third-party labeling farms. Footage is stored in access-controlled cloud storage, annotated by our team only, and delivered exclusively back to you.

Supervised defects versus anomaly-style datasets

Industrial QC teams choose between two data strategies, and the annotation plan differs for each. Supervised defect detection - the right choice when defect types are known and recurring - needs every defect instance labeled with type and severity attributes: our default. But when defects are rare, novel, or unbounded ("anything that doesn't look right"), anomaly-detection approaches train mostly on good parts and flag deviations - and their datasets need something subtler: exhaustively verified-good imagery (one mislabeled defect in the "good" set quietly teaches the model that the defect is normal) plus a smaller, precisely-masked defect set for threshold calibration and evaluation. We build both, and part of the pilot conversation is deciding which strategy your defect distribution actually supports - a decision that determines the whole labeling budget's shape.

Working with your existing cameras and vision systems

Industrial annotation rarely starts from a blank slate: there are already inspection cameras on the line, sometimes a legacy rules-based vision system, often a folder of historical images sorted by the old system's pass/fail verdict. All of that is usable. Historical pass/fail sorting becomes weak pre-labels that our annotators verify and refine - much faster than starting cold. Footage from existing line cameras defines the exact optical conditions (angle, lighting, resolution) the model will face, so we annotate that rather than idealized captures. And when a legacy system's false-reject pile exists, it is gold: those borderline images, correctly labeled, are precisely the decision boundary the new model must learn.

From labels to line metrics

Tracking annotation on process video yields more than model training data - the labels themselves are measurements. Per-item tracks with frame-accurate station entry and exit times are cycle-time data; counts per time window are throughput; the gaps and stalls visible in track timelines are the bottleneck analysis an industrial engineer would otherwise collect with a stopwatch. Several clients use the annotated pilot batch this way before any model is trained: the labeling pass doubles as a time-and-motion study of the line, which often pays for the pilot by itself and always sharpens the question the eventual model should answer.

The pilot playbook for industrial vision

A typical industrial engagement starts small and de-risks fast. Week one: you send representative footage - including known defects and known-good runs - and we return a proposed schema (classes, attributes, events) with a small annotated sample. Week two: your process engineers review the sample; the schema is corrected against reality (there is always a defect subtype nobody mentioned). Then bulk annotation proceeds against the frozen schema with QA sampling throughout, and the first dataset version ships with class-distribution and small-object reports. Subsequent versions add footage from new lines, shifts, or seasons - because lighting at the night shift and the day shift are, for a camera, two different factories.

Safety analytics in depth: PPE, zones, and events

Workplace-safety models are quietly one of the highest-volume industrial vision categories, and their datasets have a specific anatomy. Person detection and tracking form the base layer; per-person PPE attributes (helmet, high-vis vest, gloves, safety glasses, harness) ride on each track rather than existing as separate objects, so the export answers "who was non-compliant, when, for how long" directly. Zone semantics come from segmentation: walkways, exclusion zones around machinery, vehicle corridors - drawn once per fixed camera and reused across the footage. Event tags mark the moments compliance systems act on: person-enters-zone, vehicle-and-person proximity, mounting/dismounting equipment. Annotating all three layers in one project keeps them time-aligned, which is what allows a safety product to reconstruct incidents rather than merely count violations. We label this footage with the same confidentiality posture as defect work - worker imagery never leaves controlled infrastructure, and faces can be blurred in exports where the client's policy requires it.

Counting, packaging, and verification datasets

A large share of industrial requests are not defect detection at all but verification: is the right thing present, in the right quantity, in the right place? Carton-content verification (all six bottles present before sealing), pallet counting, label-placement checks, cap and seal presence, kit completeness before dispatch - these train from detection datasets with strict counting discipline, where the annotation rule is exhaustive enumeration per frame and the QA rule is that counts reconcile across consecutive frames of a track. Payana's per-frame tracking makes the reconciliation check nearly free: an item that vanishes mid-conveyor without an outside mark is a flagged annotation error, caught before export instead of after deployment. For teams searching for a manufacturing defect annotation service, PPE detection dataset labeling, or production line video annotation, the stack above - schema pilot, exhaustive rare-class labeling, tracking-derived process metrics, edge-ready exports - is the complete industrial offering, validated on real conveyor footage.

Yes. Click-to-segment boundary finding is class-agnostic: it finds the boundary of whatever you click, whether or not the object has a name in any training set. Your part's identity comes from your label schema, not the model's vocabulary.

Yes - anything exportable as standard images or video is annotatable, including thermal camera output, industrial X-ray, and line-scan captures. For unusual modalities we calibrate the workflow on a pilot batch first.

Yes: person detection and tracking plus per-person PPE attributes (helmet, vest, gloves, glasses), with frame tags for zone-entry events. This maps directly onto standard safety-monitoring model architectures.

With a pilot: send representative footage, we propose a label schema (classes, attributes, events) based on what your model needs to decide, annotate a pilot batch, and iterate on the schema before scaling to the full dataset.