Retail Shrinkage – Using AI to Tackle a Growing Challenge

The Challenge

Have you been to a grocery store recently where you could use self-checkout without any assistance from employees? If so, did you make sure you scanned everything properly? Have you ever accidentally mixed up product codes when identifying a specific product, like choosing a non-organic version when the actual product was a higher-priced organic? Perhaps you once put a bulk package of paper towels under your cart and accidentally forgot to scan it before you left.

If this sounds familiar, you might have (unintentionally) contributed to the problem of retail loss, aka “shrink” or “shrinkage.” This issue extends beyond grocery stores—and isn’t always accidental.

When customers are responsible for self-checkout, mistakes are inevitable – and they cost retailers billions of dollars. The total cost of retail loss in the US attributed to self-checkout shrinkage is estimated to be over $100B¹. Nearly 7% of self-checkout transactions (6.7%) had at least some partial shrinkage compared to 0.3% with checkout using cashiers.

Self-checkout (SCO) systems are not only here to stay, but they are continuing to expand despite the shrinkage issue. Evidence suggests that as much as 80% of consumer transactions in supermarkets might be processed through variants of SCO technologies².

Retail loss affects every type of store when self-checkout is available. Retail loss doesn’t happen just because customers accidentally walk out and forget a payment. It also happens when customers intentionally steal, for example, by swapping barcodes for cheaper products or outright shoplifting by not scanning items.

The same holds for employees. Trained employees can also mix up a product code or miss an item when lines are long and they are trying to speed up the process. Unfortunately, they could also be stealing as well. Employees who steal items tend to be repeat offenders, becoming more brazen the longer they go uncaught. When they collude with customers, the losses begin to pile up!

The Solution

What can a store do to prevent this from happening? Traditional camera systems are forensic in nature, only being used to investigate after the crime has already been committed. Additionally, they are only looked at when a loss is suspected to have occurred, which is too late. Thankfully, new technological advancements based on artificial intelligence (AI) can help solve this problem and reduce the financial impact of shrinkage.

Computer vision software can interpret camera feeds in real-time and identify mix-ups at the checkout as they occur, providing direct feedback to the customer. By doing this, recovery of self-checkout losses directly converts to more sales. The same technology can also be used to identify suspicious behavior, alerting staff, who can then act accordingly and take action. Research indicates that the vast majority of unscanned items are due to errors that customers are willing to address once they have been pointed out to them.

Intelligent solutions like this rely on a robust platform, using complex software and optimized hardware components. These would require a lot of time and IT resources for a retailer to develop on its own, but thankfully, the latest advanced technologies from companies like NVIDIA make the development of software a lot easier. These include NVIDIA Metropolis microservices for the NVIDIA retail loss prevention AI workflow. By using Metropolis microservices and pre-trained models, the AI workflow applies few-shot learning to rapidly scale to thousands of products, and generate intelligent, actionable alerts.

Software is just one piece of the complete solution; you also need robust and reliable hardware to run it. Supermicro’s line of enterprise edge products is purpose-built for deployments such as those in retail. These come in a myriad of form factors to cover nearly every deployment scenario. Supermicro works closely with AI software providers such as NVIDIA to offer systems that support the latest AI acceleration technology as soon as it’s available.

Reliability is critical since store staff need to focus on their day-to-day tasks. They don’t have time (nor, in most cases, the expertise) to troubleshoot hardware issues with the very systems that are supposed to be helping them. And when IT staff are stretched thin, outside vendors are often called in for technical issues. These “truck rolls” can adversely affect operational budgets and distort the financial value of loss prevention systems. Having reliable hardware on which to run your AI applications is of paramount importance.

AI Building Blocks

To make it easier for developers to quickly build and roll out AI-based applications designed to prevent retail theft, NVIDIA has created a library of AI workloads. For retail, this includes the Retail Loss Prevention AI Workflow, the Multi-Camera Tracking AI Workflow, and the Retail Store Analytics Workflow, built on its Metropolis microservices. These workflows can be used as no-code or low-code building blocks for loss prevention applications because they come pre-trained with images of the most-stolen products (such as meat or alcohol). Additionally, they come with software to plug into existing store applications for point-of-sale machines as well as object and product tracking across entire stores.

The workflows mentioned above feature an advanced variation of few-shot learning, which means that an AI model only needs to process a relatively small set of images to get going. It is designed to adapt continuously with limited new product data, using object characterization and self-supervised learning algorithms. This unique method of active learning identifies and captures new products, and packaging changes scanned during checkout for future recognition.

Developers can access NVIDIA’s AI workflows via cloud-native microservices in NVIDIA AI Enterprise, enabling fast customization of the workflow. They can then index hundreds of thousands of store products for cross-camera and barcode recognition, and quickly scale deployment. For retail organizations, the availability of these workflows means that it becomes much more economically feasible to develop and adopt AI-based solutions, while improving the quality and speed of the outcome.

At the core of any AI solution is the hardware that it runs on, specifically NVIDIA accelerated computing including Graphical Processing Units (GPUs). GPUs are accelerator cards that are the true workhorses of AI. GPUs ingest data from camera feeds and run AI models governed by the Retail AI Workflows to look for anomalies.

Deployment Scenarios

It can take several cameras to cover Self-checkout (SCO) systems adequately which means there are multiple video streams to analyze. These streams must be processed simultaneously, in real-time without lag or interruptions. With that in mind, inferencing should occur locally at the retail location instead of sending the video over the internet to a cloud-based server. This is why specialized Edge AI systems, such as those from Supermicro, are ideal for these workflows.

With the most comprehensive product line of edge AI servers on the market, Supermicro is perfectly suited to round out the loss prevention solution. For optimal results and ROI, it is important that systems are “right-sized” to meet the demands of the specific deployment while keeping TCO within budget. With a broad range of form factor sizes and configurations, Supermicro systems can be equipped with the exact GPU needed for each deployment.

Unlike data center systems, edge servers are meant to be deployed in (mostly) unconditioned environments. Edge systems could be mounted directly to a wall, placed on a shelf, or installed in two-post racks that are commonly used for network equipment. Combined with wider operating temperatures, these systems provide excellent deployment versatility.

Various deployment options and the ability to operate in typical store conditions are great features, but they only speak to part of the necessary infrastructure solution. Systems also need to be capable of performing critical AI inferencing tasks. Supermicro’s edge AI portfolio is optimized to run the latest NVIDIA accelerated computing hardware—from compact GPUs like the NVIDIA L4 to powerful, double-width cards like the NVIDIA L40S. The flexibility in system configuration enables organizations to match the AI server to the required performance exactly.

Target Systems

For deployments where there is limited space and wall-mounting is the preferred method of installation, Supermicro has multiple solutions. The SYS-E403 Box PC and the SYS-E300 Mini-1U offer great performance in a small package. Sites with small networking racks can take advantage of the short-depth rackmount products, such as the SYS-112B Network Edge server featuring an Intel® Xeon® processor or the AS-1115S with an AMD EPYC™ processor.

On the higher end of the spectrum, large-format stores with several SCOs may require more than one GPU to handle the streams from dozens of cameras. For these demanding deployments, we have several 2U options, such as the Hyper-E or SYS-212B family of products come into play.

Conclusion

What does all this mean for you, the retailer? By combining the advanced AI software solutions from NVIDIA and the best-of-breed enterprise edge servers from Supermicro with NVIDIA accelerated computing, retailers finally have access to solutions that can address the ever-growing shrinkage problem. Supermicro’s software and services partners can leverage these elements to create efficient AI solutions that help ensure low TCO and fast ROI. 

Additional Resources:

If you would like to learn more about AI-driven retail loss prevention solutions, please reach out or visit the website: https://www.supermicro.com/en/solutions/retail-loss-prevention 

For more information about the NVIDIA Loss Prevention Workflow, check out the technical documentation: https://www.nvidia.com/en-us/ai-data-science/ai-workflows/retail-loss-prevention/ 

In the meantime, download our white paper to discover Supermicro’s reference model for loss prevention deployments, combining robust, enterprise-grade edge systems with NVIDIA’s library of development tools and pre-trained models: https://www.supermicro.com/en/solutions/retail-loss-prevention/whitepaper-download?destination=%2Fen%2Fsolutions%2Fretail-loss-prevention 

¹ Source: NRF

² Source: ECR Retail Loss