Data Annotation

Why Recycling Robots Need More Than Computer Vision Walk into a modern recycling facility, and the first thing you notice is the sheer volume of it. Thousands of items move across conveyor belts every hour. Plastic bottles tangled with crushed cardboard. Aluminium cans buried under food packaging. A lithium battery hiding beneath a pile of mixed waste with just a corner poking out. Everything arrives together, dirty, damaged, and constantly moving. For people working on the line, understanding this feels almost automatic. We know a crushed can is still aluminium. We know a greasy pizza box belongs somewhere different from clean cardboard. We spot a battery in the wrong stream even when it is barely visible. A recycling robot does not have any of that instinct. It has to be taught. Every material. Every condition. Every edge case. And that teaching happens through data. # Spotting an Object Is the Easy Part Here is where most people get it wrong when they think about recycling AI. They imagine a model drawing boxes around objects — bottle, can, battery, cardboard — and assume that once the system can detect things, the problem is mostly solved. It is not. Detection tells the model what is on the belt. But a recycling robot needs to know a lot more than that before it can act: Answering these questions is where the real complexity lives. And none of it gets answered through detection alone. Modern recycling systems often combine multiple annotation layers within the same dataset. An object may first be detected with a bounding box annotation, precisely segmented using polygons or instance segmentation, assigned material-specific attributes, and then tracked across multiple frames in video to understand movement and interaction. Each annotation contributes a different piece of information, allowing the model to build a far richer understanding than detection alone ever could. # The Problem With Clean Training Data Here is something that does not get talked about enough in recycling AI — most training datasets look nothing like a real facility. Public datasets have clearly visible objects, consistent lighting, and clean backgrounds. Everything is neat and easy to annotate. The model trains on those examples, hits good accuracy numbers in testing, and then gets deployed into a facility where objects arrive crushed, wet, folded, overlapping each other, and half hidden under other waste. And then it struggles. Not because the model is bad. Because the data was not prepared for what it was going to face. A transparent plastic bottle flattened under a sheet of cardboard looks completely different from the same bottle in a clean dataset. A battery with one corner sticking out from a pile looks nothing like a neatly placed battery on a white background. A food container covered in residue is a different visual problem than a clean one. The gap between testing performance and production performance in recycling AI is almost always a data gap. The model learned from easy examples. The real world sent it hard ones. The goal of annotation is not simply to label what is common — it is to prepare the model for everything that is uncommon, unpredictable, and difficult. # Why Different Annotation Types Each Do a Different Job To build a recycling AI system that actually holds up in production, you need training data that teaches the model several different things at once. Different annotation types handle different parts of that job. – Bounding boxes are the starting point. They teach the model where an object is — its rough location on the belt. Fast to produce and easy to scale, they work well for detection tasks where you just need to know something is there. But the moment objects start overlapping — which in recycling is almost always — bounding boxes stop being enough. – Polygon annotation and instance segmentation are where things get more precise. Instead of a rough rectangle, polygon annotation traces the exact boundary of each object. Instance segmentation goes further — it separates every individual object even when they are touching or partially overlapping, so two crushed cans stacked against each other are annotated as two distinct objects, not one combined shape. For robotic picking systems, this precision has a direct operational impact. A picking arm guided by imprecise object boundaries will miss, knock items off the belt, or attempt grips that fail mid-motion. What happens at annotation time shows up at deployment time. – Material classification teaches the model that looking similar and being similar are two completely different things. Clear PET and clear HDPE bottles look almost identical. Clean cardboard and grey chipboard share the same color palette. Aluminium foil and certain plastics reflect light in very similar ways. Without explicit material labels in the training data, the model sorts by visual appearance — which fails exactly when the visual cues are most ambiguous. – Attribute annotation is the layer that brings operational context into the dataset. It answers the questions that change what should happen to an object, even when the object class stays the same: This is the difference between a model that can identify a battery and a model that knows this particular battery is a lithium cell partially buried under packaging; which means something very specific for how the system handles it. – Video annotation and object tracking become equally important in facilities where waste is constantly moving. Instead of treating every frame independently, tracking allows the model to maintain the identity of an object as it travels along a conveyor, passes beneath cameras, becomes partially occluded, or rotates into a different orientation. This temporal understanding improves sorting consistency and reduces duplicate detections. – Quality assurance and annotation consistency are often overlooked but are just as important as the annotation itself. Two annotators interpreting contamination, occlusion, or object boundaries differently introduce inconsistency into the dataset. For the model, those inconsistencies become conflicting examples – making annotation guidelines, reviewer calibration, and structured QA essential parts of the dataset-building process. #

Satellite Image Annotation: How AI Learns to Read the Physical World There is more intelligence packed into a single aerial image than most people realise. An overhead shot of a commercial property contains roads, parking areas, rooftops, vegetation, pathways, utility structures, shadows, boundaries, all layered on top of each other, all spatially related, all meaning something different depending on what the AI system is trying to understand. The models being built on top of satellite imagery today are genuinely impressive. They can estimate outdoor surface areas, identify vegetation density, detect parking utilisation, flag infrastructure changes, and generate property proposals — all from aerial data. But none of that happens because the model is clever. It happens because someone first sat down and taught the model what it was looking at. Feature by feature. Surface by surface. Property by property. That is satellite image annotation. And it is a lot more complex than it looks from the outside. Why Satellite Imagery Is a Different Kind of Annotation Problem Most computer vision annotation deals with objects that are familiar and visually distinct. A car is a car. A person is a person. The boundaries are usually clear and the categories are usually obvious. Satellite imagery does not work that way. From an aerial perspective: The visual complexity is high, and the margin for annotation error is low — because in geospatial AI, annotations do not just train a model to recognise objects. They train it to understand spatial relationships, measure surfaces, and reason about how a property is structured. A polygon that is slightly off does not just mislabel an object. It corrupts an area calculation. It breaks a surface estimate. It feeds wrong data into a proposal or planning workflow. This is why satellite image annotation demands a different level of precision and domain awareness than most annotation categories. What Satellite Image Annotation Actually Involves # Semantic Segmentation The foundation of most geospatial AI datasets is image semantic segmentation — classifying every visible region in the image by what it is. Grass, roads, rooftops, vegetation, pavements, parking areas, water surfaces, open land — each pixel region gets assigned to a category that tells the model how that surface type is distributed across the property. This is what allows AI systems to do: The challenge with image semantic segmentation in satellite imagery is consistency. The same surface type can look different depending on lighting, season, image resolution, and shadow coverage. Without tight ontology guidelines and disciplined annotator training, the same patch of ground gets labeled differently across different images — and that inconsistency compounds into model unreliability at scale. # Instance Segmentation Semantic segmentation tells the model what something is. Image instance segmentation tells it how many there are and where each one begins and ends. For property intelligence applications, this distinction matters enormously. A parking lot is not just a parking area — it is a collection of individual spaces, each of which can be occupied or empty, marked or unmarked, accessible or standard. A cluster of trees is not just vegetation, it is individual trees with distinct canopy boundaries that affect shading calculations and landscaping estimates. Image instance segmentation is what allows AI to: For applications like parking utilisation analysis, solar panel detection, or automated site planning, this layer of annotation is not optional — it is the core of what makes the model useful. # Polyline Annotation Not everything in a satellite image is a region. Some of the most important property features are linear — and those require a different annotation approach entirely. Roads, pathways, driveways, curbs, fence lines, utility lines, property boundaries — these are structural elements that define how a property is connected and how it is divided. Polyline annotation traces these linear features with precision, giving the model the vocabulary to understand: A model that cannot reliably trace a driveway from the road to the building entrance cannot support automated site analysis. A model that loses a property boundary mid-line cannot measure a lot accurately. # Bounding Box Annotation Not every object in a satellite dataset needs precise boundary annotation. For fast object localisation; vehicles, containers, rooftop equipment, utility poles, movable assets, 2D bounding box annotation is often the right tool. It is faster to produce, easier to scale, and sufficient for detection tasks where the exact object boundary matters less than knowing the object is there and roughly where it sits. In large-scale infrastructure monitoring or asset tracking workflows, 2D bounding box annotation forms a practical and efficient detection layer without the overhead of full segmentation. # Attribute-Based Annotation Detection and segmentation tell the model what is there and where. Attribute annotation tells it what kind of thing it is. These attribute labels are what push geospatial AI from detection into property-level intelligence. A model that knows a parking area exists is useful. A model that knows how many spaces it contains, what condition they are in, and how many are accessible is genuinely valuable for the platforms being built on top of it. Where Satellite Annotation Gets Hard — And How We Handle It Elevated structures cast shadows that hide what is underneath them. A tree canopy obscures the pathway running beneath it. A building shadow falls across a road and makes the surface boundary ambiguous. This is not just a visual problem, it is a consistency problem. If two annotators handle occluded regions differently, the model learns contradictory things about the same situation. The way we handle it is through clear occlusion protocols built into the annotation guidelines before the project starts. Every edge case, partial occlusion, full shadow coverage, seasonal variation gets a defined handling rule. Annotators are not making judgment calls on the fly. They are following a documented decision framework, and QA reviews specifically check for occlusion consistency across the dataset. This is the challenge that catches most teams off guard. From the air, concrete and asphalt are nearly indistinguishable without contextual cues. Dry vegetation and open gravel

How We Train Annotators on Hard Domains – and What Goes Wrong When Teams Skip the Ramp Most annotation vendors will tell you their annotators are “experienced.” What they won’t tell you is what those annotators were experienced in — and why it might have nothing to do with your data. We’ve reviewed a lot of datasets that came back from AI data annotation vendors before teams came to us.The labels are clean-looking. The turnaround was fast. The pricing was competitive. And the model trained on that data performs terribly — not because the architecture is wrong, not because the learning rate needs tuning, but because the labels themselves carry systematic errors that are almost impossible to detect without domain knowledge. This post is about how we prevent that. It’s about how annotator skill is built at Pixel Annotation — and what happens to annotation quality when teams skip the ramp. The Myth of the “Experienced Annotator” When an annotation vendor says “we have experienced annotators,” it sounds reassuring. But experience in annotation doesn’t work the way experience works in most fields. A software engineer who has spent years in backend systems can pick up a new framework faster because the underlying concepts — data structures, APIs, state management — transfer. Annotation doesn’t work like that. An annotator with 5,000 hours of e-commerce bounding box experience is a complete beginner on a radiology dataset. They don’t know what a lesion boundary should look like. They don’t know what “ambiguous margin” means. They don’t know when to flag uncertainty versus commit to a label. They will make confident mistakes — and confident mistakes are the hardest to catch. THE REAL PROBLEM: The danger isn’t that inexperienced annotators make obvious errors. It’s that they make systematic errors that look correct on the surface — wrong polygon on the right object class in image annotation, wrong boundary with the right shape in image segmentation, wrong temporal ID with the right bounding box. These errors survive QA. They ship. They break models months later. Domain knowledge isn’t a soft advantage. It’s the difference between a label that trains a model and a label that quietly misleads it. What Domain Certification Actually Means At Pixel Annotation, we don’t certify annotators per project. We certify them per attribute. That distinction matters more than it sounds. If we certify an annotator on “lesion boundary segmentation,” that certification applies to any project that uses lesion boundary as an attribute — regardless of which client, which dataset, which modality. The skill is portable. The certification travels with the annotator. Certification is tied to the spec version. When a spec is updated — even a minor version — impacted annotators go through a delta re-certification. Not a full ramp, but a targeted review of what changed and a re-test on the affected attributes. This is how annotation quality stays consistent across a 12-month engagement, not just in week one. This is how annotation quality stays consistent across a 12-month engagement, not just in week one. The Four-Stage Ramp Every annotator working on a hard domain at Pixel Annotation goes through a structured ramp. No shortcuts. No exceptions. Here is exactly what that looks like. Stage 1: Spec Mastery Before an annotator touches a single frame, they read the full annotation spec for their domain. For complex domains like medical imaging or autonomous driving, this can be a 40–60 page document covering class definitions, edge-case handling rules, ambiguity conventions, and labeling hierarchy. After reading, they take a 30-question test — auto-graded, randomized, question order and answer options shuffled each time. Minimum passing score is 95%. Fail and you re-read and retake. There’s no shortcut through this stage. This stage alone filters out annotators who skim specs. In our experience, roughly 20–25% of annotators who were confident on intake fail the first spec test on a hard-domain project. That’s not a failure of the test. That’s the test working. Stage 2: Supervised Labeling on a Graded Set This is where most vendors cut corners — and where we don’t. The annotator labels 100 pre-graded frames. These are frames we’ve labeled internally with verified ground truth. Their output is scored frame by frame. They see exactly where they diverged from ground truth and why. This stage isn’t about passing a number. It’s about building calibration. An annotator can score 93% accuracy and still be deploying a consistent systematic error — for example, drawing polygons 2–3 pixels too tight on every instance of a specific object class during image segmentation tasks. That kind of bias has to be caught at Stage 2, because by Stage 3 it becomes a review burden, and by Stage 4 it ships. They cannot proceed until they hit threshold accuracy. For most hard domains, that threshold is 90–92% IoU on image segmentation tasks, higher on classification attributes. Stage 3: Shadow Labeling Against Senior Reviewers This is the most expensive stage. It is also non-negotiable. For the first week on a live project, 100% of the annotator’s output is reviewed by a senior reviewer. Not sampled. Not spot-checked. Reviewed completely. Every disagreement between the annotator and the reviewer becomes a teaching moment. The annotator gets written feedback on each disagreement — not just “wrong” but why it’s wrong, what the spec says, and what the correct interpretation should be. Senior reviewer time is the most constrained resource in any annotation operation. This is why most vendors skip Stage 3 — it’s expensive in the short term. But an annotator who enters production without Stage 3 is carrying interpretation errors that will silently degrade your dataset for months. Stage 4: Independent Labeling with Sampled Review Once an annotator completes Stage 3 without a rising disagreement rate, they are production-certified. Review rate drops to 20–30% on a randomized sample — not the same frames each time. Their disagreement rate is tracked weekly. Not monthly. Weekly. If we see a rising trend over two consecutive weeks, that annotator goes back to shadow mode. Not retrained from zero. Not terminated. Re-shadowed — which usually catches a specific drift pattern in under three days. Why Most Operations Skip Stage