18 Questions About Neural Networks for Trade Audit

The effectiveness of the application of Image Recognition strongly depends on the human factor. If the merchandiser does not rearrange the layout — the recognition accuracy will decrease — the company will not receive reliable information about the share of the shelf, representation, out-of-stock, etc. So Image Recognition is just another, albeit innovative, tool in the hands of an employee.

Our attention was attracted by the case of one of the largest FMCG companies producing alcohol goods, which introduced neural networks into trade audit 7 years ago. We bring to your attention the most interesting moments of the Q&A session about the real experience of using Image Recognition in the mentioned company.

Image Recognition Today
/
Recognition: Opportunities
/
Recognition: Limitations
/
Recognition: Price Tags
/
Recognition: Verification
/
Team: Motivation

Image Recognition Today

1. How many users are currently working with IR? What data is collected?

— We have connected more than 500 merchandisers and now we are digitizing 100% of visits, recognizing 9 sales categories, monitoring 300 competitors' SKUs. Thanks to IR, we know exactly our shelf share of whiskey, vodka, wine, sparkling wine, rum, etc.

2. Has the OSA indicator changed?

— We estimate the effect of digital merchandising on OSA as 1.5% in strong alcohol and 1.5% in wine. In general, the indicator increased by 3%.

3. How much has the budget increased since the transition from human photo recognition to automatic recognition?

— The budget has not increased, on the contrary, we have saved money. Otherwise, this is confidential information.

Recognition: Opportunities

4. What is the recognition accuracy?

— The accuracy of a category recognition is above 95%. Difficult situations occur: so, only in the "wine" category there are a lot of objects for recognition — 15-20 thousand SKUs. But even in such cases, the system performs well.

5. How are situations worked out when something is in the middle of the shelf?
Do you have to divide a shelf into several scenes?

— If the shelf is divided, then, of course, the scene is interrupted, and a new one begins after the obstacle. Overlapping errors may occur when merging photos. It all depends on what is photographed and how. If one rack was photographed 2 times with an incorrect overlap, then we will not get the correct merging. This may lead to a deterioration in accuracy, but it will not affect it critically.

6. Have you considered changing the label design to improve the recognition quality?

— Designs are developed for the consumer, not for systems. And our task is to recognize the goods correctly.

7. What is the recognition speed?

— If an employee has internet, we get feedback within 3-5 minutes. As an example, for a supermarket inside a metro with low connection. an OSA report will be generated within 5 minutes after the last photo is sent.

8. Are tasks automatically generated for the merchandiser after recognizing the shelf in order to immediately correct the situation in the store?

— The system compares the positions from the matrix and the actual availability of goods. The list of unrecognized SKUs arrives on the merchandiser's device within 5 minutes. He can manually specify the reason for the OOS, for example: position is in block, missing in the point of sale, recognition error. When the system often does not recognize some SKU that is available, we transmit the information to our IR provider.

Recognition: Limitations

9. Do you take pictures of all categories every visit?

— At first, we took pictures of the entire shelf space every visit, and then we counted the average values on the shelf for a month. But practice has shown that the shelf does not change so often, it is enough to fix it once every three months.

10. How do you technically recognize secondary placement in the photo?

— Scenes with different types of secondary placements are uploaded to the application, and the merchandiser notes in advance what he's going to photograph. The neural network determines which goods should be displayed there. As far as we know, similar projects work on the same principle: neural networks are not yet able to recognize secondary placements.

11. Does recognition work in off-line mode?

— The recognition itself does not work without the Internet, but you can take photos. The merchandiser takes pictures through the SFA application, so it is not necessary to be online at the time of the visit. The data will be sent for recognition when the employee reaches the network coverage area.

Recognition: Price Tags

12. What is the level of accuracy of price recognition?

— Over 90%. The price tags in each retail chain look different. Plus, the trend for electronic price tags is gaining popularity — there are generally different promotions, the neural network needs to be trained in such things. Accuracy of price recognition strongly depends on the quality of the merchandiser's work. In order for all price tags to be correctly recognized, they must be correct and be under their positions, and the goods must be deployed by faces.

13. How does the system work out cases when there is no product on the shelf, but there is a price tag for it?

— We do not verify the price tag by the name of the position, but only recognize the price overall and check the availability of promotions. Therefore, in our case, the system will assume that the product is in general in the matrix of this sales point, and compare the actual availability with the matrix.

Recognition: Verification

14. How is the percentage of recognition accuracy measured, is the quality of price merging and the quality of price tag recognition included in the percentage?

— We measure the recognition accuracy using our algorithm on a monthly basis.

15. How do you deal with the fact that not the whole category is photographed, which is why the share of the shelf is considered incorrect?

— We conduct a selective audit, adhering to a special algorithm for detecting violations. We can be deceived by some small number of faces. If the faces are underestimated globally, we will definitely identify the errors.

Team: Motivation

16. Did you manage to link the IR results to the bonus matrix of your field team? How was this perceived, because IR is never 100% accurate?

— When the recognition accuracy reached 93%, we tied the bonuses to the execution of the assortment. We understand that there may be exceptions: an employee came to the store, and there was no light there — he had to take pictures in the dark, and end something didn't count. Therefore, we plan goals and objectives taking into account the percentage of error. But in general, we can be sure that we are getting high—quality data: the current recognition level is 95%.

17. If the system did not recognize 10%, will the employee receive a 100% or 90% representation bonus? Won't it be demotivating?

— We cannot have 10% errors: with an accuracy of 93%, the percentage of errors does not exceed 7%. At the same time, the brand is recognized in 98-99% of cases. Even if the litre is incorrectly determined, the KPI error on the shelf share will be so negligible that the employee will not suffer.

18. If you need to take a lot of photos from top to bottom, from left to right, at right angles — this may cause the agent's dissatisfaction due to additional physical activity. Are there any complaints?

— The work of a merchandiser involves physical exercises: you need to sit down somewhere, bend over, bring something, put it out. So there is nothing new in terms of tasks. But — as many people as there are opinions, there may be a negative part in any case. We have clearly defined rules that make the work of the merchandiser clear and transparent.

Do you have any questions?

We are here to help you

Leave your contact information and we will contact you promptly.

Privacy Policy