Logistics is a natural extension of industrial defect detection, and yet it inverts several of its principles. In production, the inspected object is known, positioned and lit under controlled conditions, and defects are rare. On a sorting conveyor, the object is unknown until it arrives, it presents in any orientation, it may be resting on another, and what is being sought is not always a defect but an item of information.
That inversion changes the nature of the annotation tasks. A logistics project typically combines reading, measurement, recognition and anomaly detection within one flow, where a quality control project usually isolates a single one. That heterogeneity is the field’s main source of complexity.
This article covers the three main families of logistics tasks, parcels, pallets and sorting, and what they imply for annotation. It extends the complete guide to defect detection, whose protocol and control principles it carries over.
What separates logistics from production defect detection
Four structural differences separate the two contexts, and each has direct consequences for corpus design.
The object is unknown
A production line handles a reference whose geometry is documented. A sorting centre handles whatever arrives: cartons of every size, soft bags, envelopes, unpackaged items, heterogeneous pallets. No reference template exists, which rules out template-comparison approaches from the outset and requires methods tolerant to shape variability.
Presentation is unconstrained
In production, conveying hardware guarantees a repeatable position. In logistics, a parcel arrives in any orientation, sometimes resting on an unexpected face, possibly touching or overlapping its neighbours. Partial occlusion is not an edge case but a routine situation, and the annotation protocol must treat it as such rather than as an exception.
Throughput is extreme
Sorting centre rates run to thousands of parcels an hour, leaving very little processing time per object and often requiring several cameras covering complementary angles. That multiplicity raises a specific annotation question: several views of one parcel must be linked together, otherwise the corpus contains observations believed independent that are not.
The purpose is not conformity
Quality control produces a conformity decision. A logistics system produces routing information, a measurement, or a record of condition with contractual value. That last category introduces an evidential requirement absent from production: when damage is recorded, the image and its timestamp serve to establish liability between carrier, sender and recipient.
Reading information
This is the largest task family in logistics, and the furthest from defect detection in the strict sense.
Codes and symbols
Not every reading task requires machine learning, and it is worth saying so plainly.
Reading barcodes and two-dimensional symbols is a largely solved problem handled by dedicated algorithms that require no learning. Annotation comes in when reading fails: locating the code in the image, qualifying the cause of failure, code obscured, damaged, blurred, badly oriented, absent. That corpus of failure cases is what improves the read rate, and it is more useful than a corpus of legible codes.
Handwritten and printed text
Addresses, fragility markings, order numbers and handwritten instructions fall to text recognition. Annotation here combines two levels: locating the text zones and transcribing them. Transcription is a task of a different nature from the others, requiring linguistic attention and raising its own questions, notably on transcribing ambiguous characters and on handling struck-through or corrected entries, common on redirected labels.
Languages and formats
An international flow mixes labels in several languages, several scripts and address formats that differ by country. That diversity concerns annotation on two counts. Transcription requires annotators able to read the scripts present, which is a linguistic rather than a vision competence. And structuring, meaning identifying recipient, street and postal code, depends on national conventions an unfamiliar annotator will apply wrongly. Restricting scope to the flows and languages the site actually handles, rather than aiming at universal coverage, is the first saving available on these tasks.
The personal data question
A parcel label carries a name, an address, often a phone number. A logistics corpus therefore contains personal data in bulk, which a quality control corpus does not. That difference requires specific handling: masking name-bearing zones where the task does not concern them, a controlled working environment where transcription is necessary, and an explicit retention policy.
This is a point logistics projects often discover late, although it governs the annotation architecture. An address-reading corpus is not handled like a damage-detection corpus, and one can be shared more widely than the other.
Measurement and dimensioning
Computing parcel volume governs pricing, load optimisation and detection of declaration discrepancies. It is a measurement task, not a recognition one.
Three approaches coexist and differ in their annotation requirements. Measurement by dedicated three-dimensional sensor, which does not involve image annotation at all. Estimation by stereo vision or depth camera, which requires reference annotation obtained by other means. And monocular estimation, which presupposes a known reference in the scene and remains the least accurate.
The annotation consequence is important: a measurement task is not annotated by outlining the object, it is annotated by associating each image with a physically measured reference value. Annotation work becomes a matter of matching images to measurements, with the timestamping and pairing questions that entails.
One intermediate case deserves flagging: delineating the faces of the parcel, which then supports geometric computation. That annotation resembles conventional segmentation, but its rule differs, since the task is to delineate the visible face of a cuboid rather than the apparent outline of the object, which includes parts hidden by perspective.
Damage detection, the core of logistics defect detection
This is the family closest to conventional industrial inspection, and where the field’s principles apply most directly.
The taxonomy
The usual damage categories are relatively few and well identified: crushing or deformation, puncture or tear, opening or tape lifting, soiling or moisture staining, and unsuitable packaging. To these is added a category decisive in logistics: closure anomaly, a parcel badly resealed or repackaged, whose implications are as much about security as quality.
The difficulty specific to cardboard
This is where logistics defect detection meets its own difficulty. A pristine package is rare. A carton that has travelled normally shows handling marks, blunted corners, carrier tape, layered labels. The boundary between normal wear and real damage is therefore blurred, and it depends on the contractual reference rather than on a physical criterion.
That situation resembles wood or textile treated elsewhere: conformity is partly commercial. The consequence is the same, the ontology must be descriptive rather than decisional, and annotation must capture severity through measurable criteria, apparent depth of the dent, length of the tear, soiled area, rather than through a global appraisal.
The evidential stake
One logistics particularity changes the whole arrangement: a damage record carries contractual weight. Three requirements follow that are absent in production. The timestamp and location of the record must be retained. The image that grounded the decision must be archived, since it will be examined in the event of a dispute. And comparison between checkpoints becomes useful, damage recorded on leaving a site but absent on arrival identifying the responsible stage.
That last requirement has a strong annotation implication: the most useful corpus is not made of isolated images but of pairs or series following one parcel across several checkpoints. Building such a corpus requires traceability infrastructure, and its absence structurally limits what the system will be able to establish.
Pallets and load units
The pallet introduces an intermediate scale, between the individual parcel and the full load, with its own tasks.
Four task families arise at this scale. Inspection of the pallet itself, whose condition governs safety, broken boards, missing blocks, protruding nails. Inspection of wrapping and strapping, whose inadequacy causes collapses in transit. Verification of stacking integrity, overhang, leaning, excessive height. And counting or identifying the parcels making up the pallet, a recognition task on a heavily occluded scene.
That last task is by far the hardest in the logistics domain. On a wrapped pallet, most parcels are only partly visible, through a reflective and distorting plastic film. Annotation must state explicitly what counts as a visible parcel, what is treated as partly occluded and what is inferred rather than observed. Without that convention, two annotators produce divergent counts with neither being wrong.
Loading and transport space
A fourth scale completes the picture of logistics defect detection tasks: the transport container, truck, trailer or shipping container.
Three distinct tasks arise there. Assessing fill rate, which governs logistics optimisation and is handled by volumetric estimation rather than object recognition. Checking load securing, which concerns safety and requires recognising dangerous configurations, unchocked loads, unstable stacking, unbalanced distribution. And the condition of the container itself, cleanliness, watertightness, wall integrity, which governs acceptance of the goods.
These tasks share an acquisition difficulty: the interior of a container is a dark, deep and narrow space where artificial lighting produces strong gradients between front and back. A defect detection project on this perimeter must address that optical constraint before any method choice, otherwise the rear of the trailer stays out of reach whatever the model’s quality.
Automated sorting
Sorting combines several of the preceding tasks into one flow, under a real-time constraint.
The typical sequence chains presence detection, separation of touching objects, reading of routing information, optional measurement and condition checking. Each stage can fail, and failure of one conditions the next: a poorly separated object produces a reading attributed to the wrong parcel.
Object separation deserves particular attention here. It is instance segmentation under unfavourable conditions: identical objects, in contact, partly overlapping, with no clear visual boundary between two cartons of the same colour. The protocol must settle contact cases, define what constitutes a distinct instance and provide an uncertainty class, preferable to an arbitrary separation.
One particularly useful practice consists of annotating the existing system’s failure cases rather than the ordinary flow. A sorting centre generally has a manual reject mechanism, and the objects that take it constitute a corpus dense in difficult cases, directly relevant to improvement.
Acquisition conditions in a logistics environment
A warehouse environment is markedly less controlled than an industrial defect detection station, and that reality must be built into the design.
Four environmental factors weigh particularly. Lighting varies with time of day, season and proximity to openings, which makes surface appearance unstable. Dust accumulates on optics faster than in a clean workshop. Conveyor vibration degrades sharpness. And the field to cover is wide, which limits available resolution per object.
These constraints bear directly on which tasks are feasible. Detecting obvious damage is robust to these conditions; finely assessing packaging condition is not. Scoping must therefore start from what the conditions actually allow to be observed, rather than from what would be desirable.
Annotation in logistics: what changes
Three practical differences separate logistics annotation from that of a production defect detection project.
The first is the heterogeneity of annotation primitives. One corpus may call for boxes to detect parcels, masks to segment touching instances, polygons for text zones, transcriptions, damage labels and measurement values. That diversity requires versatile tooling and a longer protocol than usual, along with an organisation by task rather than exhaustive annotation of every image.
The second is volume. Logistics flows produce images in considerable quantity, which makes sampling even more decisive than elsewhere. Annotating a flow as it comes is waste; selecting informative cases, failures, rejects, disputes, is the only sustainable approach.
The third is traceability. Tying each image to a parcel identifier, a checkpoint and a timestamp is not a convenience but a condition of the corpus’s value, both for multi-view tracking and for evidential use.
The seasonal rhythm of logistics projects
One organisational characteristic clearly separates this sector from manufacturing and deserves anticipating.
Logistics flows are seasonal, sometimes very strongly so. A corpus collected in a slack period reflects neither the volumes, nor the diversity, nor the damage rate of a peak period, where precisely the rate rises and handling deteriorates. A logistics defect detection project must therefore cover at least one full cycle, or explicitly document the period covered as a corpus limitation.
A practical constraint compounds this: peak periods are when operations are least available to support a project, and when images are most valuable. The remedy is to automate collection upstream, during the slack period, so that capturing the peaks requires no human intervention on site.
The operator’s role
One final point differentiates this sector: the place of the human in the loop is structurally different.
In production, an operator supervises a line and can examine a flagged part. In logistics, the parcel has continued its journey before verification is possible, and recalling it is expensive. The consequence is that the operating point is set differently: a flag that cannot be verified has value only if it is reliable.
Two arrangements address that constraint effectively. Output as a score rather than a decision, which allows manual verification to be reserved for the most suspicious cases while others pass with a simple record. And the deferred record, where the image is archived without triggering immediate action and is consulted only if a claim arises later. That second mode is specific to logistics and it changes the system’s economics: value comes not from real-time sorting but from an evidential capability mobilisable weeks afterwards.
Quality control of a logistics defect detection corpus
The control arrangement follows the field’s general logic, with two sector-specific inflections.
The first is stratification of control by acquisition condition. A corpus mixing images from several sites, several stations and several times of day must be controlled separately per stratum, variable lighting being the main difficulty factor.
The second concerns transcription tasks, which call for a specific measure. Agreement between annotators on text is not measured as it is on a class: it requires an edit distance and a convention on what constitutes an error, down to the character. That measure must be defined before production, otherwise the control becomes uninterpretable.
One further practice is specific to the domain: measuring performance separately on conforming and damaged parcels. In a real flow the latter are a small minority, and an aggregate metric stays excellent even where damage detection fails.
Reusing corpora across sites
A defect detection question specific to this sector arises for operators running several facilities: how far does a corpus built at one site transfer to another.
The answer differs sharply by task. Damage taxonomies transfer well, since a crushed carton looks the same everywhere. Reading tasks transfer moderately, limited by the languages and label formats present in each flow. Separation and counting tasks transfer poorly, because they depend on conveyor geometry, camera placement and the parcel mix, all of which are site-specific.
The practical consequence is a layered corpus strategy. A shared base covering the transferable tasks, built once and maintained centrally, plus site-specific increments covering what does not transfer. That structure is markedly cheaper than building one corpus per site, and markedly more reliable than assuming a single corpus serves everywhere. It requires deciding at the outset which layer each annotation belongs to, which is a metadata decision rather than a technical one.
The most common mistakes
These failures recur often enough across logistics projects that naming them is usually enough to avoid them.
- Treating partial occlusion as an edge case rather than the normal situation.
- Not linking multiple views of the same object to each other.
- Ignoring the bulk presence of personal data on labels.
- Annotating a measurement task by outlining the object rather than associating a reference value.
- Building a decisional damage ontology although the threshold is contractual.
- Not retaining timestamp and checkpoint, which deprives the record of evidential value.
- Omitting the counting convention for partly visible parcels on a pallet.
- Annotating the ordinary flow rather than the existing system’s reject cases.
- Scoping fine tasks that warehouse lighting conditions do not permit.
- Reporting aggregate performance on a flow where damaged parcels are a small minority.
Where model assistance helps
Logistics benefits from pre-trained models more than most verticals, for one simple reason: parcels, pallets and containers are ordinary objects well represented in general-purpose training data, unlike a crazing pattern on rolled steel.
The picture is uneven across tasks. The gain is immediate on parcel detection, coarse localisation and scene segmentation, where a general-purpose model produces usable proposals with no domain adaptation at all. It is partial on separating touching objects, where general models tend to merge adjacent cartons of the same colour. And it is nil on damage assessment, where the distinction between normal wear and reportable damage is contractual and appears in no general-purpose corpus.
The practical implication is a split annotation strategy. Accept model proposals for the geometric tasks and spend the annotation budget on the tasks where no prior knowledge exists, which are precisely the ones carrying the commercial value. Projects that spread their effort evenly across all tasks pay full price for work a pre-trained model would have done adequately.
What to take away
Logistics applies defect detection principles to a context that inverts its conditions: unknown object, unconstrained presentation, routine occlusion, poorly controlled environment. What transfers is the method, written protocol, explicit conventions, stratified control; what does not transfer is the assumption of a known and positioned object.
Three decisions structure a successful project. Organise the corpus by task rather than by image, since the diversity of primitives rules out uniform exhaustive annotation. Build traceability, parcel identifier, checkpoint and timestamp, from collection onwards, since it governs both multi-view tracking and evidential value. And select images on the existing system’s failure cases rather than annotating the flow, the only sustainable approach given the volume.
For approaches, ontology and annotation tasks, the complete guide to defect detection sets the frame. For another extension of the field towards equipment and structure inspection, with its access and periodicity constraints, the article on predictive maintenance by vision covers infrastructure monitoring.
To explore delivery arrangements, supported formats and applicable control mechanisms, see our dedicated page on annotation for industry. And if you are preparing a logistics corpus and want the tasks and traceability scoped before annotation starts, let us discuss your project.
