AI-based counting of construction materials: where accuracy is lost
AI-based counting of construction materials starts where "by eye" measurement ends: a foreman walks around a stack and names the volume from experience, and the discrepancy surfaces later — during write-off or at the next stage.
Counting materials from a photo removes subjectivity, but accuracy rests not on shooting speed, but on how well the system distinguishes similar item names and handles partially overlapping stacks.
The first source of loss is misclassification. Rebar of different diameters lies in one bundle, pallets of bricks come with breakage, bags on a pallet differ by batch: without linkage to the nomenclature, counting yields pieces but not items.
The second is short delivery: acceptance is based on the delivery note, not on fact, so discrepancies in supply volumes settle in paperwork and do not enter construction material accounting at the site.
A separate area is unclosed write-offs. Material went to an adjacent section or to a subcontractor, no act was drawn up, and inventory of the material warehouse shows a balance that physically does not exist.
The difference accumulates over weeks: purchasing relies on an inflated balance, while the site at that time operates with a shortage of a specific item.
Then the most expensive error turns on — rework. An under-delivered volume is discovered at the moment of installation, the crew idles, the order is placed on a catch-up basis and without discounts, and part of the work has to be redone for a different size type.
Below we examine exactly where accuracy is lost and how it is closed at each step — from shooting to reconciliation with acts.
What changes after implementation: speed, accuracy, control
Before implementation, accounting relies on walking around with a tape measure and notebook: recalculation takes hours, and discrepancies surface at write-off, when there is no one left to present them to.
After implementation, the primary document becomes a photo — the model marks stacks, pallets, and trays by type, and the statement is assembled without manual recalculation.
Speed of processing site photos
The foreman photographs the storage area from two or three points, and the result arrives the same day, not by period close.
Speed matters not by itself: the shorter the path from photo to statement, the earlier it is visible that stock is being used faster than planned, and the fewer emergency purchases during peak work.
A batch of photos per shift is processed in one pass — the engineer does not recalculate items manually.
Accuracy of reconciling balances
Balances are reconciled not from one frame but from a series: the system compares views from different points, cuts off overlaps and double counting of the same stack.
Reconciliation with receipts and write-offs shows discrepancies for specific items, not "the warehouse as a whole does not reconcile." The share of discrepancies falls to isolated items, and each is closed with a document.
Control of deliveries and short loads
A photo during unloading is compared with the delivery note, so a short load is recorded at the moment of acceptance, before the truck leaves. Short-load returns no longer depend on the warehouse keeper's memory, and disputed deliveries are closed with photo evidence.
Separately, the error class "accepted according to documents, but in fact less" disappears — it no longer survives until monthly inventory.
Shooting formats and automatic recognition modes
Counting accuracy is determined not by the model but by how shooting is set up. A stack in a single frame will be assessed by the system from the visible plane, and overlaps will produce scatter.
A series of frames around the perimeter reconstructs volume, and a reference object of known size sets the scale — then the result converges on repeated runs.
How to choose angles and reference objects
We fix a protocol: 6–12 frames with at least one-third overlap, a top point for stacks and pallets, side points for wall blocks.
A reference object of known size gets into the frame — tape measure, pallet, marker — otherwise scale is reconstructed from indirect signs and scatter grows. Light should be even: automatic recognition modes read direct shadows as relief.
| Shooting option | When it fits | Limitations |
|---|---|---|
| Single frame | Batch check, accounting of visible volume | Only the visible plane, overlaps distort the result |
| Series from different angles | Stacks, pallets, truck bodies, rooms | Need a scale reference and frame overlap |
| Site flyover | Facades, roofing, large areas | Depends on light, needs binding to the plan |
| Batch archive processing | Volume audit, recalculation by periods | Requires uniform shooting and stage marking |
Repeat material recalculation
Repeat material recalculation proceeds by the difference between shoots: the system binds a new set of frames to the same zone and stage, calculates the delta, and shows where volume changed.
Batch processing of the photo archive closes the audit: recognition runs in a queue, and the engineer checks disputed items manually.
How implementation proceeds: from site audit to handover
The engineering part of implementation is not "train the model and give access," but a sequence of steps where each relies on the previous one.
An error in the site accounting scheme later surfaces as systematic undercounting, so the order of stages is agreed with you in advance.
Accounting process audit
We start with an audit of the accounting process: how materials are currently accepted, written off, and reconciled — delivery notes, foreman's log, photo fixation.
The output is a list of items where a photo provides sufficient accuracy and those where measurement or a document is needed.
If there is no unified nomenclature, we first consolidate the reference list: otherwise "ceramic brick" and "solid brick" will become different items for the system.
- Data audit: review of nomenclature, supplies, and current discrepancies.
- Site accounting scheme: roles, photo fixation points, batch marking, reconciliation moment.
- Recognition setup for your reference list: classes, units, confidence thresholds.
- Integration with the accounting system: confirmed items are transferred, not raw recognitions.
- Pilot operation at one section with a control sample for verification.
- Calibration by discrepancies, repeat runs, fixation of reference photos.
- Handover: documentation, foreman training, procedure for handling disputed cases.
Control sample for verification
We calculate accuracy not "on average," but using a control sample: manually recounted batches from different shooting conditions — warehouse, unloading, open-air stack.
The discrepancy for high-volume items, where error is most costly, must fit within the agreed threshold; anything above goes to additional training, not into the report. We hand over the system with a verification log so acceptance can be repeated on your own data.
What you receive: configuration, schemes, regulations
The package is handed over not as a folder of files, but as a working system and a description of how to operate it without us. Below is what exactly is recorded in the solution documentation.
What is included in the delivery
- Configured recognition configuration — target material classes, model confidence thresholds, behavior for a low-quality frame: the photo goes to manual review instead of a silent error.
- Container marking rules — how to label pallets, stacks, and packs so that the same material is counted identically at all sites.
- Data flow schemes — the path from a photo on the warehouse keeper's phone to a record in the accounting system: entry points, intermediate storage, behavior during a communication outage.
- Data access policies — which employees see photos and results, how long they are stored, how deletion on request is processed.
- Instructions for warehouse employees — one or two pages with examples of a correct and erroneous frame, the procedure for a disputed frame, and whom to ask.
Regulations and support
- Material accounting regulation — frequency of reconciling recognized volumes with actual balance, who signs off on a discrepancy and how it is recorded.
- Procedure for changing classes — how to add a new material or a marking edit without stopping the process.
Solution documentation is maintained together with the system: a rule edit gets into it the same day, not after the next audit.
Warehouse case: accounting accuracy before and after implementation
Case: a wholesale supplier of construction materials StroyKomplekt — a warehouse with 4,000 items, acceptance of 20-25 trucks per week.
Discrepancies in balance reconciliation reached 7 items per 100, inventory took two days with six warehouse keepers, and disputed deliveries were sorted out by correspondence without evidence.
| Metric | Before implementation | After implementation |
|---|---|---|
| Discrepancies in balance reconciliation | 7 items per 100 | 1-2 items per 100 |
| Warehouse inventory time | 2 days, 6 people | 4 hours, 2 people |
| Disputed deliveries per month | 32-35 | 5-7 |
| Photo fixation of acceptance | selective, about 10% of batches | every batch |
The figures are measurements at acceptance, accounting system data, and the claims log: three months before launch and three after.
What it was before
They lost not on counting speed, but on the lack of evidence. The warehouse keeper recounted pallets manually, photos lay in the phone gallery and did not enter accounting, and the shortage surfaced at the next reconciliation — when the truck had already left.
What it became after
Each batch is photographed from three points right at acceptance: the model recognizes the material, counts pieces and stack volume, and the calculation along with photos goes into the accounting system.
A disputed delivery is closed on the day of unloading — the card contains photos, calculation, and the receiver's signature.
Inventory proceeds using the accumulated photo pool and system data: 4 hours with two people instead of two days with six, and discrepancies stay within 1-2 items per 100.
How does phone shooting differ from a drone flyover?
Both approaches produce a set of frames suitable for calculation, but they answer different questions. A phone is a close-up: a pallet, a separate stack, readable marking. A drone is the entire site, including tall stacks that cannot be approached from the ground.
The area coverage when shooting with a phone is limited to a few meters around the operator, and the upper tiers are closed by their own side faces: overlaps and perspective distortions enter the frame, and volume is counted with double counting.
A drone shoots from above but does not show side walls — inclined passes are added along the perimeter.
Lighting requirements are stricter for the drone: midday sun plunges passages between stacks into shadow, while low sun gives long shadows and glare on film and metal.
The working window is diffuse overcast light; it is also optimal for the phone, because the operator's shadow and a close flash create glare that hides the edge of a pack.
Repeat recalculation depends on shooting reproducibility. The flyover route is repeated at fixed points, and ground landmarks set the scale — two periods are compared correctly, and receipts and consumption are visible.
Phone frames from the same point are almost impossible to repeat, so a marker of known size is placed in the frame.
In practice, this is not either/or: the basis is provided by a route flyover with binding, while canopies, stacks under roofing, and narrow passages are additionally shot with a phone.
The result is continuous site coverage without blind spots, suitable both for initial measurement and for regular balance recalculation.
Technical doubts: accuracy, light, complex objects
Doubts about accuracy concern not the model, but shooting conditions: light, angle, overlapping stacks, unmarked containers. Let us address the five most frequently asked questions.
How accurate is assessment from a phone photo?
Scale is reconstructed from a reference object of known size in the frame, and marking is based on photos from real sites. For piece items, the discrepancy is a few percent; disputed ones we confirm with a second angle.
What to do in poor lighting?
We shoot in diffuse light: overexposure from a floodlight cuts off contours, while shadow helps read the stacking. The frame is checked for sharpness and overexposure — if quality is insufficient, the system asks for a retake rather than counting at random.
The stack is closed, only the top row is visible — how to count?
We count by the top row and recalculate to the stacking geometry; for pallets, by the container and known packaging. A fully closed item is treated as a forecast and confirmed at acceptance, rather than entering accounting as fact.
How to recognize unmarked containers?
The classifier relies on container geometry, packaging type, and remnants of marking, then matches against your nomenclature. Below the confidence threshold — manual review instead of a plausible guess.
Where does data go and how are site photos stored?
The result is uploaded to the accounting system via a software interface or an exchange file. Photos are stored with binding to the site and date, access is role-based, retention period is set by policy; only depersonalized results can be kept.
Let's discuss the task: what is needed to start the project
A demonstration on ideal frames proves nothing. The accuracy of photo-based counting is determined not by the model, but by input data: angles, lighting, and how materials and containers are marked at the site. That is why we start by reviewing what already exists at the sites.
To come back with an assessment and work plan, the following set of input data is needed:
- 20–30 photos from real sites: different devices, angles, lighting, including partially overlapping stacks;
- a list of materials and containers to be counted: pallets, bags, rolls, pipes, packs, sheets;
- current accounting rules: how they count now, what goes into the report, which discrepancies are critical;
- examples of past calculations — statements or acts, against which we will verify the result with a reference;
- access to the systems where the result goes: warehouse accounting, spreadsheets, or an internal portal;
- shooting constraints: devices, network availability at the site, who is responsible for photo fixation.
Then the procedure is as follows: we discuss the counting task and site conditions, review your photos and accounting rules, and come back with an assessment and work plan — with stages, timelines, and an accuracy metric by which we will accept the result.
You can order solution development immediately after the plan is approved: it is already broken down into stages, and the start does not depend on refining requirements along the way.
