Pressure for Data Accuracy Makes Computer Vision Deployment Harder in Retail Stores

Deployments of shelf cameras and computer vision in retail chains go past the pilot stage in many instances, although many retailers realize an important hidden challenge when deploying the technology in stores. While computers may find discrepancies in product inventory or at checkout lines, the need for accurate data puts an increasing burden on staff who already experience shortages.

Over the years, retailers have sought ways to increase visibility about inventory levels, shrink rates and shelf availability. Computer vision is sold as a tool that allows retailers to achieve automatic store visibility without having to perform extensive manual checks. One of the challenges here is the fact that visibility relies on store conditions that stay constant enough.

Inconsistencies related to packaging designs, product displacement and layout will negatively affect the performance of the model. For example, once a camera was trained to see a product in a certain orientation, any changes in the placement of products may cause issues. Store associates have to confirm discrepancies manually to keep data accurate for replenishment teams.

This tendency grows stronger as the technology starts being deployed in different locations and not only kept in innovation laboratories or used in demonstration stores. In contrast to the latter, in multiple-store deployments, the retailers have to consider conditions that vary between facilities.

For some retailers, computer vision deployments result in a number of changes in store operations. Retail managers start paying attention to keeping shelves spaced accurately since the image recognition software performs better in consistent environments. Night-shift staff members are involved in checking inconsistencies after floor resets. Loss prevention teams get a greater number of incident reports.

All this does not imply that the technology is useless for the industry. Most retail companies face numerous challenges associated with shrinkages, distortion and out-of-stock requests. Human audits are both costly and inconsistent in most stores; thus, a solution allowing for continuous monitoring can be extremely helpful.

The difficulty is connected with the fact that automation raises expectations that are often difficult to meet in practice. Some retailers expected reductions in labor needs after deployment; yet, many of them have realized that they face an intermediate stage during which stores handle a greater volume of administrative work.

It becomes evident in procurement conversations, too. In recent procurement processes, retailers pay more attention to understanding whether the platform continues delivering accurate data when used under conditions involving store resets, product substitution or assortment changes. Questions concerning training efforts and managing exceptions become increasingly common.

Integration issues play an important role in deployment as well. Inventory systems, POS solutions, and merchandising systems often feature inconsistent information on items. While computer vision successfully identifies an inconsistency, other systems may classify the same item incorrectly, creating a situation that requires human involvement.

Budgets for retail technology investments remain active in this segment as CEOs recognize the weakness of current solutions. Still, the discussion about computer vision deployments becomes more realistic. Buyers no longer focus extensively on autonomous stores or similar concepts and pay more attention to such aspects as alert management and store compliance.

These tendencies indicate stabilization in the industry. Retailers seem to become more prepared to investing in computer vision that is aimed at solving a particular problem in the store. At the same time, there seems to be less enthusiasm towards vendors claiming to replace most front-line functions with cameras alone.