Errors are part of the process. There’s no need to be afraid of them. This is a normal aspect of any computer vision technology. Even a person with perfect eyesight and twenty years of retail experience can sometimes confuse two similar yogurts under poor lighting.

The system doesn’t see the way a human does. It analyzes pixels in an image: colors, shapes, textures, and object positions. If two products look similar, it may confuse them. If the lighting is poor, details are lost. If a product is partially blocked, the system works with incomplete information.

No reason to panic. Errors will happen. The key is to focus on how to detect and fix them quickly.

Step 1: Gather Information About the Error
To fix the issue, you need to understand it precisely.
What to record:

• When: date and time (errors may occur at specific hours)
• Where: which shelf, section, and store
• What exactly: which product and what the system identified instead
• How often: one-time or recurring
• Conditions: lighting, shelf fullness, neighboring products

Take photos: capture the problematic area as a person sees it, and compare it with what the camera sees.

Step 2: Check the Lighting
In most cases, the problem lies in lighting.
What to check:

• Are all lights working?
• Have new shadows appeared on the shelves (something moved, new rack installed)?
• Are there new glares (lights replaced, additional lighting added)?
• Has natural lighting changed (seasonal changes, curtains added/removed)?

Step 3: Check Camera Positioning
Cameras may have shifted, tilted, or been obstructed.
What to check:

• Is the camera pointing in the correct direction (has it moved)?
• Is the lens clean (dust, droplets, condensation)?
• Is anything blocking the view (new ads, price tags, hanging elements)?
• Are there new obstacles in the frame (new shelf, stand)?

How to check: look at the current camera feed. Is the entire shelf visible? Is everything in focus? Is there any blur?

Step 4: Evaluate Frequency and Severity
Not all errors require immediate action.
Isolated error:

• Happens rarely (1 in 1000 cases)
• Affects non-critical products
• Occurs under specific conditions

Solution: Log it, but don’t spend resources fixing it immediately. It can be added to the dataset for future model retraining.
Systemic error:

• Occurs regularly
• Affects important products (top sellers, promo items)
• Happens under normal conditions

Solution: Fix as a priority. This directly impacts operations.

Step 5: Collect Data for Developers
If simple fixes don’t work, escalate the issue to specialists.
What to collect:

• Examples of errors (10–20 camera images)
• Photos of the correct situation for comparison
• Description: what should be there vs. what the system sees
• Error frequency (percentage, absolute numbers)
• Conditions under which the error consistently occurs

In typical retail environments, 10–15% deviation is considered normal. The key is to detect and correct errors promptly.

• Large products with bright packaging under good lighting: 90–95% accuracy.
• Small, similar products under average conditions: 80–85% accuracy.
• Systems in the first months: 70–80% accuracy, improving to 85–90% within six months.

Accuracy below 80% usually indicates issues with equipment or configuration.
Above 90% is an excellent result — but not always achievable for complex assortments.