1. Capture
- Images tagged with zone, aisle, bay and time
- Image quality check before processing
- Re-capture request for blurred or dark images
- Capture schedule set per zone
AI / Computer Vision Inventory Counting
Inventory Counting with Computer Vision
Cycle counts take staff hours, and the system still shows stock that is not on the shelf. Agnotic Technologies builds computer vision software that counts stock from shelf and rack images, matches each item to your item master and sends the differences to your WMS or ERP for a person to approve.
Counts from imagesSKUs matched to the item masterPeople approve adjustmentsPer-SKU accuracy checks
Trusted by global innovators
How does inventory counting with computer vision work?
Inventory counting with computer vision captures images of shelves, racks or yards, recognizes each item against the item master, counts what is visible and compares it with on-hand quantities in the WMS or ERP. Agnotic Technologies builds this software so differences become cycle count proposals that a person approves before any record changes.
The camera does not replace the inventory system. It gives the system a frequent look at the shelf, and the inventory team decides what to do with what it sees.
The general method for building vision models, from task choice to labeling, is on our computer vision development page. This page covers what is specific to counting stock.
How the count is built
What happens between the image and the inventory record?
Six stages run on every capture. Each stage keeps its own evidence, so any count can be traced back to the image it came from.
2. Location matching
- Image matched to the bin, slot or planogram position
- Location labels read where they are visible
- Bay map kept in step with the WMS location list
- Unmatched images held for review
3. Item recognition
- Products detected and matched to SKUs
- Barcode or QR reads used where the code faces the camera
- Unknown items flagged, never guessed
- Confidence stored for each item
4. Counting
- Visible facings counted per SKU
- Depth estimated only where the rules allow it
- Cases and pallets counted by their labels
- Visibility score attached to each count
5. Reconciliation
- Count compared with WMS on-hand for that location
- Open picks and receipts allowed for
- Differences proposed as cycle count tasks
- Proposals grouped by zone for the counter
6. Review and posting
- Inventory controller approves, edits or rejects
- Approved adjustments posted through the WMS or ERP API
- Rejected proposals saved as test cases
- Every posting linked to its image
Capture method decision table
Which capture method fits your stock?
The capture method decides what the software can see. Most sites end up with two methods for different zones.
| Method | Where it fits | What it sees and misses | Labeling needs | How counts reach the WMS |
|---|---|---|---|---|
| Fixed camera | Retail shelves, pick faces and high-value cages that need frequent counts | Sees the same view all day; misses anything behind the front facing or outside the frame | One label set per camera view, refreshed when the layout changes | Scheduled counts compared with on-hand, differences sent as cycle count tasks |
| Handheld phone | Small stores, back rooms and occasional audits by staff | Sees what the person points at; angle and distance change every time | Images from many angles and phone models, since every capture differs | Count shown on the phone for the person to confirm, then posted |
| Autonomous robot | Long store aisles and warehouse floors with level shelving | Sees low and mid shelves consistently; misses high racks and blocked aisles | Images from the robot's own cameras at its fixed heights | Results per aisle after each run, differences queued by zone |
| Indoor drone | Tall rack storage where pallet labels are out of reach | Sees pallet labels and positions high up; struggles with wrapped labels and tight aisles | Mostly pallet and label detection rather than single items | Location checks matched to the WMS bin list, exceptions sent for review |
Robots and drones need route, charging and safety rules agreed with your site team before the counting software is built.
SKU recognition
How does the software know which SKU it is looking at?
Recognition uses every clue available, not only the product image.
Item master data
- SKU, description and pack size
- Reference product images for each SKU
- Case and pallet configurations
- Active and discontinued flags
Location and planogram
- Expected SKU for each slot or bin
- Shelf and bay layout from the planogram
- Allowed substitutes for a slot
- Layout version and effective date
Codes on the item
- GS1 barcodes read when they face the camera
- 2D codes such as QR on cases and labels
- Pallet labels on rack storage
- Shelf edge labels for the slot
Look-alike handling
- Flavor and size variants with similar packaging
- Seasonal and promotional packs
- Private label items beside branded ones
- Low confidence routed to review
Counting methods
How do you count stacked, hidden or bulk stock?
Not everything on a shelf can be counted from one image. Each case gets a rule agreed with your inventory team.
01
Single facings, all visible
Example: Cans or boxes in one row on a store shelf.
Approach: Count each detected item directly and post the count after review.
02
Rows behind the front facing
Example: Products stacked several deep on a shelf.
Approach: Count the front and report the slot as full, partial or empty, or add depth sensing, rather than guess at the hidden rows.
03
Cases and pallets
Example: Full pallets in rack storage.
Approach: Read the pallet or case label, confirm the location and use the quantity the WMS holds for that pallet.
04
Loose or bulk items
Example: Bins of small parts or bagged goods.
Approach: Estimate the fill level as a band and flag the bin for a manual count, instead of an exact number the image cannot support.
05
Blocked views
Example: A person, cart or forklift in front of the shelf.
Approach: Mark the area as not visible and capture it again on the next pass. It is never reported as empty.
Reconciliation flow
How do vision counts reach the WMS or ERP?
The software proposes and a person approves. Only then does the inventory record change.
Compare with on-hand
Each count is compared with the WMS on-hand quantity for the location at the time of the image, allowing for open picks and receipts.
Filter the noise
Differences inside the tolerance your team sets, or from low-visibility images, wait for the next capture instead of being raised.
Propose a cycle count
The remaining differences become cycle count proposals that show the image, the expected quantity and the counted quantity side by side.
Approve, edit or reject
An inventory controller reviews each proposal on screen, or sends a counter to check the slot in person.
Post and learn
Approved adjustments post through the WMS or ERP API with the image attached, and rejected ones become test cases for the model.
Original checklist
What must a SKU pass before its counts can post?
Accuracy is measured per SKU, not as one average. A SKU that has not passed stays in review mode, where its counts are shown but never proposed.
| Check | How it is measured | Who signs off | When it runs |
|---|---|---|---|
| Recognition accuracy | Model results compared with labeled held-out images of that SKU | Inventory lead | Before the SKU goes live and after each model update |
| Count accuracy | Vision counts compared with physical counts of the same slots at the same time | Inventory lead | During a parallel run alongside existing cycle counts |
| Confusion with look-alikes | Errors between similar SKUs reviewed as a list of pairs | Category or merchandising owner | Before go-live and when new variants arrive |
| Visibility coverage | How many of the SKU's locations the capture method actually sees | Operations lead | Monthly and after layout changes |
| Ongoing drift | Approvals and rejections from review tracked by SKU | Inventory lead | Weekly in production |
Tolerances and pass thresholds are set by your team for each SKU group. We record each one and the evidence behind every go-live decision.
Edge or cloud, and what happens to the images?
Where inference runs and how long images are kept are decided site by site.
Edge inference on site
Models run on a site gateway when bandwidth is limited or counts are needed during the shift.
Cloud inference for batch counts
Overnight or periodic runs on uploaded images, scaled with the number of sites.
People removed from images
Faces and staff in frame are blurred before images are stored or sent for labeling.
Retention by purpose
Images behind posted adjustments are kept as evidence for the period you set; the rest are deleted sooner.
Who it is for
Which teams use computer vision inventory counting?
The same counting pipeline, pointed at different stock.
Grocery and retail stores
Shelf gaps and on-hand errors found between scheduled counts.
Distribution centers
Pallet locations on tall racks checked without lifts or ladders.
Manufacturing stores and kitting
Parts bins and kits checked before they reach the line.
Spare parts warehouses
Slow-moving parts located and counted where errors can go unnoticed for months.
Yards and outdoor storage
Trailers, containers or stacked materials checked against the yard or WMS record.
Retail tech and robotics startups
Inventory counting with computer vision for startups building robots, drones or shelf cameras for retailers.
Scope drivers
What drives the cost of inventory counting with computer vision?
We size the work on a scoping call rather than quote a price. These factors move effort the most.
SKUs and look-alikes
More SKUs and similar packaging mean more reference images and more review.
Capture method
Fixed cameras need installation planning; robots and drones need route, safety and charging work.
Counting rules
Counting visible facings is simpler than depth estimation or bulk fill levels.
WMS or ERP interface
A documented inventory adjustment API is simpler than file exports or screen entry.
Sites and layouts
Each store format or warehouse layout adds its own location map and test images.
Review workflow
Mobile review for counters and approval rules by stock value add product work.
Standards
Which identification standards does the build use?
Counts must resolve to the same item keys your inventory system already uses.
GS1 General SpecificationsGS1 Digital LinkISO/IEC 18004ISO/IEC TS 4213ONNX
How we work
How does a vision counting engagement work?
Three stages, starting with one zone.
1.
Scoping call
We pick one zone and a set of SKUs, agree the capture method and check how adjustments enter your WMS.
One zone
2.
Fixed-scope first release
Capture, recognition and counting for that zone, run in parallel with your existing cycle counts so accuracy is measured per SKU.
Parallel run
3.
Scale and support
More zones, SKUs and sites, with each SKU moved from review mode to live only after it passes the checklist.
Per SKU
Facts and sources
Which standards shape vision inventory counting?
Short facts behind item identification and accuracy testing, each linked to its primary source.
The GS1 General Specifications are the foundational GS1 standard that defines how identification keys, data attributes and barcodes must be used in business applications.
Source: GS1 General Specifications, Release 26.0The GS1 Digital Link URI syntax standard, first ratified in August 2018, defines how to represent GS1 identification keys within web addresses so they can link to online information and services.
Source: GS1 Digital LinkISO/IEC 18004:2024 specifies the QR code symbology, including its data character encoding methods, symbol formats, error correction rules and a reference decoding algorithm.
Source: ISO/IEC 18004:2024ISO/IEC TS 4213:2022 specifies methodologies for measuring the classification performance of machine learning models, systems and algorithms.
Source: ISO/IEC TS 4213:2022
Frequently Asked Questions
Not entirely. It shows where counts are likely wrong, so counters go to the slots that need them instead of walking every aisle. Stock the camera cannot see, such as rows behind the front facing, still needs a rule or a manual check.
They help but are not required. Barcodes and QR codes read from images confirm the item where the code faces the camera, and RFID or scan data can check a vision count. Vision is most useful where tagging or scanning every item is not practical.
No. It creates cycle count proposals with the image attached, and an inventory controller approves, edits or rejects each one. Only approved adjustments are posted to the WMS or ERP.
Yes. We read locations and on-hand quantities and post approved adjustments through the system's API where it has one. The scoping call checks what your WMS or ERP exposes and whether it can store a reason code and an image link.
Per SKU. Vision counts run in parallel with your current counts, and each SKU group goes live only after it passes the checks your team signs off. SKUs that have not passed stay in review mode.
It depends on the number of SKUs, the capture method and your WMS interface. We fix the first release for one zone on the scoping call and plan the dates from it.
Yes. One zone, one capture method and a limited set of SKUs, run in parallel with your current counts, is a complete first release that shows whether vision counting works on your shelves.