Computer Vision for Damage Detection: The Check That Decides Who Pays
Damage is one of the few things in a facility that already has a budget line, a process and an argument attached to it. Nobody has to be convinced that damaged goods cost money. The claims file proves it every month.
What is less obvious is where the money actually goes, and where the decision that commits it gets made. In most operations it is made in about thirty seconds, by a person glancing at a pallet while a driver waits, and it leaves behind a signature rather than a record.
This article covers what damage actually costs once the whole file is counted, why the handover check carries legal weight that most operations do not treat it as carrying, what a camera adds at that specific moment, and what the published research says about how well damage detection currently works, including where it stops.
The claim is the smallest number in the file
When a shipment arrives damaged, the claim value is the number that gets recorded. It is rarely the number that gets spent.
The carrier CRST published a breakdown of one customer's annual freight damage that makes the structure visible. Against 250,000 US dollars of freight claims sat another 605,000 dollars of cost: replacement product, expedited replacement freight, claims administration, billing and cash-flow delays, customer churn, sales and account management time, and operational labour. On that account, every dollar of freight damage produced 3.42 dollars of total business cost (Inbound Logistics).
Two caveats belong on that figure. It is one customer, and it was published by a carrier selling a premium handling service, so it is an argument as well as a data point. The ratio should not be treated as an industry constant.
The structure, though, is the part worth keeping. The claim is the visible cost. Claims administration, expedited replacement and the customer relationship are the rest of it, and they are paid by departments that never see the claims report. That is why damage reduction tends to be underfunded relative to what it is worth: the saving lands in several budgets at once and shows up cleanly in none of them.
Under CMR, the handover check is a legal act
For international road freight in Europe, the condition check at handover is not an operational formality. It is the moment the burden of proof moves.
The CMR Convention governs road carriage where the place of taking over and the place of delivery are in two different countries and at least one is a party, which covers almost all cross-border road freight in Europe. Under Article 8, the carrier taking over the goods must check the number of packages, their marks, and the apparent condition of the goods and their packaging. Where the carrier has no reasonable means of checking, or finds a problem, it must enter a reasoned reservation on the consignment note.
Article 9 is the consequence. Where the consignment note carries no reservations, the goods are presumed to have been in apparent good condition when the carrier took them over (Law & More, CMR guide; UNECE, CMR Convention text).
Read those two together and the position is stark. A carrier who does not record a condition reservation at loading has accepted, by default, that the goods were fine when collected. Any damage found at destination is then presumed to have happened on their watch. The check nobody had time to do properly is the check that decided the liability.
It runs the same way at the other end. Under Article 30, visible damage should be recorded on delivery, and hidden damage reported in writing within seven days, not counting Sundays and public holidays. Miss that and the consignee is presumed to have received the goods in the condition the consignment note describes, though they may still prove otherwise.
The standard legal advice that follows from all this is worth quoting plainly, because it is also a product specification: record the damage carefully, with photographs, and where possible involve a surveyor before the goods are moved or repaired.
Photographs, taken at the handover, every time. That is already the recommended control. The reason it does not happen consistently is that it depends on a person with a phone remembering to do it while a driver waits and a dock is backing up.
This is general information about how liability is allocated in road freight, not legal advice. Your own terms, your insurer and your jurisdiction govern your position.
What a camera adds at that moment
A detection model watching the loading or unloading point does something narrow. It looks at goods crossing a fixed line and returns a class, a confidence score and a timestamp for what it sees.
What that produces is not a better eye. It is a record that exists whether or not anyone was paying attention, written the same way every time, attached to the moment the goods changed hands.
Three things follow from that, and they are worth separating because operations teams tend to care about different ones.
The check stops being sampled. A person checks carefully when the dock is quiet and quickly when it is not. A camera does not have a busy period.
The record exists by default rather than by diligence. Nobody has to remember to photograph anything. The evidence is a by-product of the goods moving past the camera, not an extra task competing with unloading the next trailer.
The record is consistent enough to argue from. Photographs taken by six different people on six different phones, in different light, from different angles, make a weak file. The same camera position, the same classes and the same timestamp format across every handover make a file that holds together.
None of this prevents damage. It changes who can prove what afterwards, and it moves the conversation with a counterparty from recollection to record. Where damage claims and disputed deliveries are numbers you already track and already argue about, that is the whole value.
What the research says, and where it stops
Damage detection is a real and active research area, and the honest summary is that it works well on the damage types you have examples of, and not at all on the ones you do not.
A 2025 study compared three current detection architectures on container damage specifically, training on a dataset of 278 annotated images. YOLOv11 and YOLOv12 reached 81.9 per cent mAP@50, against 77.7 per cent for RF-DETR. On less common damage types, the ranking reversed: RF-DETR detected both damaged containers and individual damage occurrences more reliably, with higher confidence (arXiv, 2025).
Two things in that study matter more than the headline numbers.
The first is the dataset size. 278 annotated images is not an industrial research budget. It is an afternoon of labelling by somebody who knows what damage looks like on a container. That is the practical point behind no-code training: the barrier to a working damage detector is not compute, it is examples of your damage, on your goods, in your light.
The second is the reversal on uncommon damage. The model that scored better on the benchmark was not the model that handled the unusual cases better. Benchmark accuracy and field reliability are not the same measurement, which is why a parallel run against your existing manual check is worth more than any published figure.
There is a structural limitation here too, and it is specific to damage rather than to detection generally. Counting is a closed problem: the item is present or it is not. Damage is open-ended. A crush, a tear, a puncture, water staining, a corner knock and a leak are different visual events, and the list is never finished. A recent survey of industrial defect detection describes the field's response to exactly this, a shift from closed-set methods, which only recognise damage classes they were trained on, towards open-set methods that reduce the need for exhaustive annotation and can flag novel anomalies (Cheng et al., arXiv, 2025).
That research direction is promising and it is not finished. The practical consequence today is to scope a damage project around the damage types you actually see and can produce examples of, rather than around the phrase "detect damage," which is not a specification.
The same check, on a production line
Everything above is written from the dock, because that is where damage has a counterparty and a claims file. The same detection task appears earlier in the chain with different vocabulary attached.
On a line, it is called defect analysis: surface damage on a part, a fault on packaging, a crushed carton before it ships. High-speed precision inspection is already solved on the lines that justified the capital, and it is a well-defended market. The work that is still done by eye is the inspection that was too small, too varied or too changeable to put an integrator on, which is a different problem with a different economics.
The detection task is the same one. The difference is who pays when it is missed: on a line it is scrap, rework and a customer return, and on a dock it is a claim and an argument.
The privacy architecture
Damage detection has a structural advantage over most video analytics, which is that it does not need to know who anyone is. The detection target is the pallet, the carton, the container or the part. Faces are not an input, and identity is not part of the data the system needs to do its job.
As with every MAKRR deployment, the model trains in the cloud and runs on an edge device inside the facility, so inference happens locally and raw footage does not have to leave the site. What leaves is the detection, the alert and the exportable record, never the video. For a camera pointed at a loading bay where staff work all day, that distinction is usually the one that decides whether the project gets approved. See our piece on edge versus cloud computer vision for the fuller architecture case.
This is general information about data architecture, not legal advice. Confirm your specific position with your data protection officer.
A realistic first project
Pick one handover point and one camera. The inbound dock or the outbound loading bay, whichever currently generates more argument.
Collect footage across normal variation, which for a dock means different light, different weather, different carriers and a Friday afternoon as well as a Tuesday morning. Label the damage types you actually see in your claims file, not the ones you can imagine. Three or four real categories beat a taxonomy of twenty you cannot produce examples for.
Then run it alongside the existing manual check for a few weeks without acting on it, and compare what it flags against what was recorded manually. That comparison is the business case, and it is also the only honest accuracy figure, because it is measured on your goods rather than on a benchmark.
A first working model can be ready in under a day of setup. The parallel run is what takes the time, and it should.
To see how this maps onto your own handover points, book a scoping call, or start a free trial and label your first sequence this week.
FAQs
-
The damage types you can show it examples of. Dents, tears, crushed corners, punctures, water staining and surface defects are all workable where you have footage of them occurring. Damage you have never captured on camera is not detectable, which is why scoping starts from your claims file rather than from a feature list.
-
No. It produces a second opinion automatically at the moment the goods are handled, and a record that exists whether or not anyone had time to make one. Where the camera and the person agree, nothing happens. Where they disagree, somebody looks again.
-
No. The platform works with the CCTV, RTSP and USB cameras already installed, which for most loading bays means the cameras are already in position. See our piece on adding computer vision to existing cameras.
-
Published work on container damage reports around 82 per cent mAP@50 on a small training set, with the caveat that benchmark scores and field reliability diverge on uncommon damage types. The figure that matters is the one from running it in parallel with your current check on your own goods.
-
Inference runs on a device inside your facility. What leaves the site is the detection, the alert and the record, not the video.
-
Asset tracking confirms location. Damage detection confirms condition. They answer different questions and neither substitutes for the other, which is the same distinction covered in our piece on RTLS and cart contents.