Annotating a point cloud does not resemble annotating an image. The annotator does not look at a view, they navigate a space; they do not trace a contour, they select sets; and they do not recognise appearances, they read forms.
This article sets out the tooling, the working methods and the organisation an aerial LiDAR production requires. It extends the article on national programmes.
What the tooling must permit
Six capabilities distinguish a tool suited to aerial LiDAR from a simple viewer. Loading large volumes by levels of detail, a batch far exceeding available memory and having to display progressively rather than entirely. A configurable vertical section, the practitioner’s principal tool and whose thickness, orientation and displacement must be adjustable from the keyboard. Volumetric selection, by polygon in plan view, by three-dimensional box or by attribute criterion, each serving distinct situations. Multiple colourisation, by class, by elevation, by intensity and by echo number, these representations revealing structures the others mask. Display of reference layers, cadastre or derived model, superimposed on the cloud. And end-to-end preservation of attributes and georeferencing, a disqualifying requirement the tools article in the previous series detailed.The selection modes
Four methods structure the annotation work. The choice between them determines productivity more than the skill of the gesture. Selection by polygon in plan view suits extended objects well separated in plan, isolated buildings or parcels. Selection in section suits vertical distinctions, separating ground from low vegetation or a bridge deck from the vegetation beneath, and it is the dominant mode in practice. Selection by volume suits compact three-dimensional objects, an isolated tree or a structure. And selection by criterion, relative height threshold, echo number or intensity range, suits mass treatment and combines usefully with a spatial delimitation. One important observation runs across those four modes. Combining a criterion with an extent is markedly more efficient than either alone, selecting for instance every point above a given height inside a polygon handling in one gesture what would otherwise take dozens of actions.The vertical section, central tool
This instrument dominates aerial LiDAR annotation. Mastering it separates a beginner from a practitioner. A section displays a thin slice of the cloud viewed from the side, which immediately reveals the vertical structure no other representation shows. Four distinct uses follow. Diagnosis, a ground classification defect appearing instantly as a discontinuity in the ground line. Precise selection, the limit between two classes being traced in section with an exactness plan view does not permit. Understanding a configuration, a tree overhanging a building becoming legible in section where plan view confuses them. And checking, a traverse by successive sections along a transect revealing a batch’s inconsistencies. One recommendation follows for training. Teaching the section before any other tool appreciably shortens the learning curve, an annotator working in perspective alone producing an inferior result whatever their diligence.Organising a batch
Five decisions structure how the work is divided. Their omission produces recurring difficulties. Batch size, a trade-off between a file’s manipulability and the number of interfaces to manage. Overlap between batches, necessary so that boundary objects are annotated in sufficient context. The rule for objects straddling limits, a building crossing a batch boundary having to be annotated once and not twice. Processing order, a traverse by homogeneous sector being better than an arbitrary geographic one, consistency being easier to maintain across similar configurations. And progress tracking, indispensable beyond a few dozen batches. One important observation follows for an aerial LiDAR project. Those five decisions belong to tooling as much as to method, and a tool managing none of them requires reconstituting them by external means, which works poorly beyond a few people.The families of tool available
Four categories are distinguishable and a project generally mobilises several. Open viewers offer performant navigation over large volumes and limited editing functions, which makes them excellent for exploration and checking and insufficient for production. Geospatial professional software integrates point cloud handling into an environment managing georeferencing and reference layers, with acceptable editing functions and an ergonomics designed for the geomatician rather than the annotator. Three-dimensional annotation platforms, developed for autonomous driving needs, offer the best production ergonomics and often partial geospatial handling. And processing libraries permit automation and mass treatment, with no interface. One important observation runs across those four categories. None covers the full needs of an aerial LiDAR production, which imposes a composite arrangement whose assembly is part of the engineering work and a reusable asset.The case library
This document conditions the consistency of an aerial LiDAR production. It differs from its equivalent in imagery. In image annotation, an edge case is illustrated by a thumbnail. In aerial LiDAR, it is illustrated by a section accompanied by a perspective view, understanding requiring both. Four elements compose a genuinely useful case sheet. The configuration described in one sentence, a tree overhanging a roof for instance. The section illustrating the real geometry. The decision adopted and its reason, permitting neighbouring cases to be settled by analogy. And the counter-example where one exists, a neighbouring configuration receiving a different decision and whose comparison sharpens the rule. One practical consequence follows. This library must be accessible during the work and not consulted at the start, a requirement the preceding series established and which weighs more here, the variety of three-dimensional configurations exceeding that of planar ones.Assistance functions
Four aids genuinely accelerate this work. Their availability is a tooling selection criterion. Prior automatic classification, which the previous article showed to be the starting point of most projects. Three-dimensional region growing, which aggregates connected points presenting similar properties from a designated point, particularly effective on isolated trees and planar surfaces. Plane fitting, which identifies points belonging to a single planar surface, useful on roofs and carriageways. And class propagation along a linear object, which follows a cable or a kerb and selects its points progressively. One caveat accompanies those four aids. All fail at the interfaces, precisely where the issue sits, which restates the observation the preceding series made about assistance tools in imagery.Quality control
Four arrangements answer the difficulties specific to aerial LiDAR. Automatic consistency checking, which searches for impossible configurations, ground points above building points, negative heights in a derived model or isolated floating points. Verification through a derived product, hillshading the terrain model revealing ground class defects in minutes as the corresponding article detailed. A traverse by sections along transects, a systematic method revealing the vertical inconsistencies no plan view shows. And targeted double reading on difficult configurations, measuring inter-annotator agreement where variability concentrates. One observation completes those four arrangements. The second has the best return and it should be systematic before any delivery, its execution costing minutes per batch.Throughput and what determines it
Five factors govern an annotator’s productivity, and they are not the ones supposed. The selection mode used, an annotator combining criterion and extent handling several times the volume of one selecting point by point. The quality of the initial classification, correction on a good starting point being of an entirely different order from full production. The clarity of the conventions, an ambiguous case not covered by the library either stopping the annotator or producing an individual decision. Display performance, a tool that struggles on a dense batch discouraging changes of viewpoint and therefore verification. And session length, following three-dimensional configurations being tiring and errors concentrating at the end of a session. One observation follows for an aerial LiDAR project. The first two factors dominate the other three by a wide margin, which shifts optimisation effort towards training in selection modes and towards prior assessment rather than towards pursuing individual speed.The workstation
Four hardware characteristics condition comfort and therefore work quality. The graphics card, whose capacity determines the number of points displayable in fluid interaction, a stronger constraint than in imagery. Memory, which bounds the batch size manipulable without recourse to disk. The display, a large surface permitting a section and a perspective to be shown simultaneously, which avoids round trips costly in time and attention. And the pointing device, a mouse with free-spinning wheel and programmable buttons appreciably reducing fatigue on intensive navigation. One observation follows. Those four cost little against an annotator’s salary, and their inadequacy degrades both throughput and quality, which makes them an investment with immediate return.Training an annotator
A four-stage progression leads to three-dimensional competence. It takes an image annotator a few weeks. Navigation first, moving through a space, changing viewpoint and mastering the section. Reading structure next, recognising an object by its geometry and relative arrangement in the absence of colour. Efficient selection third, choosing the mode suited to each situation rather than always using the same one. And diagnosis fourth, locating where an automatic classification’s defects sit rather than sweeping exhaustively. One important observation accompanies that progression. The fourth stage distinguishes a productive annotator from a conscientious one, the extent of the batches precluding exhaustive examination and making the targeting strategy decisive.Building a team on this modality
Four organisational decisions structure building a capability and they differ from those of an image annotation team. The profile sought privileges spatial understanding over dexterity, an annotator able to picture a three-dimensional configuration being worth more than one fast at tracing. The initial team size stays small, training being individual and consistency easier to establish on a small group before extending. Specialisation by use justifies itself quickly, forestry conventions and network ones differing enough that an annotator gains from mastering one. And the checker role requires a distinct profile, diagnosis through derived products being a different competence from annotation itself. One observation follows for an aerial LiDAR provider. That fourth decision is the most neglected, a checker trained in diagnosis detecting in an hour what a conscientious annotator would not see in a day.What machine learning contributes
One development modifies how this work is organised and it is worth bounding its real effect. Models operating directly on point clouds have progressed considerably and they now exceed geometric rules on varied landscapes. Three contributions are genuinely real. The starting point rises, which reduces the correction volume. Learning on a local corpus surpasses a generic classification, which makes investing in a corpus profitable. And propagating a correction partly automates, a model retrained on the corrected zones improving the rest. Three limits nonetheless persist. The training volumes needed are substantial and annotated aerial LiDAR corpora are scarce. Transfer between landscapes is limited, a model trained on one relief degrading elsewhere. And evaluation still requires a manually verified reference. One practical consequence follows. This development strengthens the value of an annotated corpus rather than reducing it, a provider holding local corpora being able to train models its competitors lack the means to match.Deliverables and their format
Four elements compose an annotation delivery and omitting the last three strongly limits use. The classified cloud in the standard exchange format, compressed for archiving, with every original attribute preserved. The conventions document, describing the classes, the thresholds adopted and the illustrated case library. The control report, indicating the method used, the zones verified and the agreement rates measured. And the problem zone log, listing what could not be reliably classified and why. That last element deserves emphasis in aerial LiDAR. A dense understorey zone where no ground point exists cannot be correctly classified, and flagging it explicitly distinguishes a known limit from an invisible defect, which changes the deliverable’s value for any downstream use. One further recommendation follows. Also delivering a hillshaded terrain model lets the client verify the work’s quality themselves in minutes, a gesture costing little and installing a confidence figures do not establish.What this tooling costs
Five components make up the cost of an aerial LiDAR annotation chain. Their distribution is surprising. Licences, a visible and generally modest line, several performant tools being open. Hardware, a real line and higher than for an imagery chain, displaying dense clouds requiring graphics cards and memory. Storage, considerable given the volumes the previous article described. Developing the complements, conversions, automatic checks and tracking tools the software does not cover. And training, the dominant line on the first project and nil on the following ones. One important observation follows for a provider. The first project bears most of these costs and the following ones benefit, which makes entry into this modality expensive and continuation economical, a configuration explaining this market’s concentration.What differs from image annotation
Five differences justify a competence of its own. They explain why the transfer requires training. Navigation replaces observation, understanding being built by movement rather than by examining a fixed view. Selection replaces tracing, precision depending on the criterion rather than on the gesture. Form replaces appearance, colour being absent from a raw aerial LiDAR cloud. Volume changes scale, a batch containing millions of points where an image contains millions of pixels but organised into objects identifiable at a glance. And verification proceeds through derived products, a defect revealing itself better in a terrain model than in the cloud itself.Starting an aerial LiDAR production
Five steps take a team from training to routine production, and their order matters. Test the tooling on a real batch before committing, a tool behaving differently on several million points than on the vendor’s example. Train two people rather than one, comparing their results revealing the convention gaps a single annotator would not detect. Build the case library on that first batch, with the sections illustrating each decision. Measure the throughput obtained, a figure replacing any estimate and serving as the basis of every subsequent quotation. And systematise the derived product check from the first batch, which installs the habit before volume makes it hard to introduce. Those five steps take a few weeks and they constitute the investment the previous section described as concentrated on the first project.What this competence is worth
Three consequences emerge for a provider. They are commercial in nature. The first is that this competence’s scarcity is real and durable. It does not follow from image annotation, it requires several weeks of training and particular tooling, and few actors make that investment. The second is that this scarcity is maintained by practice rather than by the initial barrier. A competitor can make the same investment, and what durably distinguishes an actor is the accumulated case library and the tested conventions, assets that cannot be bought. And the third is that the competence transposes between uses with modest effort, an annotator trained on forestry aerial LiDAR becoming operational on a network corridor after adapting the conventions rather than retraining. That third consequence appreciably improves the investment’s return. It means a capability built for one market serves the others, which reduces the risk of committing to a modality whose sectoral demand might fluctuate.What to fix first in an existing chain
One diagnostic sequence deserves setting out, since most teams inherit a chain rather than design one. Four checks establish in an afternoon where an existing aerial LiDAR chain loses most. Watch an annotator work for twenty minutes. If they are not using sections, or are selecting point by point, the largest available gain is training rather than tooling. Ask to see the case library. If it exists as prose without sections, the conventions are being interpreted rather than applied. Check whether a derived product is generated before delivery. Its absence means defects are reaching clients that minutes of work would have caught. And ask what the measured throughput is. If nobody knows, every quotation on this modality has been a guess. Those four checks have a common property. Each identifies a gap that is cheap to close and expensive to leave, and none requires new software or new hires. One observation completes them. Teams tend to look for improvement in tooling because tooling is purchasable, where the gains here sit almost entirely in method and habit, which are harder to buy and cheaper to acquire.Common mistakes
These failures recur often enough that naming them is usually enough to avoid them.- Working in perspective without using the vertical section.
- Always using the same selection mode whatever the situation.
- Neglecting the combination of attribute criterion and spatial extent.
- Illustrating a case library without sections.
- Making the library inaccessible during the work.
- Omitting overlap between batches.
- Not fixing the rule for objects straddling limits.
- Delivering without having generated a derived product for checking.
- Sweeping exhaustively rather than targeting defective zones.
- Using image annotators without three-dimensional training.
