Image recognition (IR) powered retail execution and auditing are a new reality that all CPG brands and retailers are adopting at an ever-increasing pace. IR enables highly accurate retail execution KPIs ranging from on-shelf availability to planogram and price compliance as well as branding material visibility at accuracy and speed, which were unimaginable a couple of years ago or so. This power-up of retail execution practices is largely thanks to the advent of cutting-edge technological development in AI, and more precisely, in Deep Learning. Today, automated visual tasks such as product recognition, price reading, and brand material recognition is no longer a dream of the future but rather a strong reality of the present, when delivered by the right IR vendor with the right MLOps and Deep Learning know-how, expertise, and capabilities.  

We’re on the cusp of a new revolution. Thanks to breakthroughs in AI and image recognition, it is now possible for grocery retailers to achieve perfect stores with a data-driven approach, and offer wowing moments to the shoppers. If the objective is to meet the customer expectations at the best possible quality, detail, and speed, CPG brands and retailers MUST get the best possible IR service from a top-tier provider. And there are three things in common while choosing the right cutting-edge IR provider in the new reality of retail. Let’s dive deep into them. 

1) Practical Setup and Annotation Processes 

Deep Learning-based IR relies on a paradigm called "learning by examples". That means, in order to automatically recognize products and properties in a retail scene relevant to a particular FMCG or retailer, several visual examples of the entities to be recognized have to be provided to the recognition engine prior to and during the actual operation. Outdated practices involved taking pictures of products in a controlled setting, but it is no longer the case due to cross-domain limitations and additional data collection efforts. All to the good, using actual retail images to perform this must-do annotation in the early setup phase of production is the new modern reality.  

This pioneering approach should be one thing to consider when evaluating an IR vendor. It requires the annotation of real shelf scenes with the provided metadata of the CPG brand to make engines ready in a short time. In the meantime, this process also exploits the concept of “Transfer Learning”. Transfer Learning means transferring and building upon the knowledge acquired from earlier engine versions that were specifically trained for a large corpus of already known retail products and their seen examples. With this concept, the need for annotation for novel ones is substantially limited. And this whole process of setup and annotation should be transparent between the vendor and the CPG supplier or retailer. 

2) Quality Control at Its Best 

Cutting-edge IR systems need regular quality control to assess and report performance and repurpose new images to optimize recognition engines. Otherwise, performance drift is unavoidable. This is especially pronounced when considering the fact that each month an average of 3% of the CPG products seen on shelves are novel due to a newly introduced brand, promotion or packaging, which was not there before.   

From an IR engine’s perspective, this corresponds to changing the ground rules of recognition during production with new target classes to choose from. Thus, unlike generic IR tasks, like person detection, gap detection, or scene text reading; SKU-level product recognition is not a visual task to deploy and forget about. It requires a continuous and carefully leveled human intervention, at the very least to enroll novel products and their master data to the system.  

The randomly sampled human validation comes into play while relabeling the novel product discovery, which is robustly automated by IR technology, according to dynamic master data along with other quality assurance processes. This intervention feeds the IR engine training even further and paves the way for better performance in quality. 

Consequently, achieving the most superior and sustainable quality is only possible by maintaining the IR engine. Making an eye-catching demo at once is the simplest. However, the ones who provide the most accurate IR engines and ensuring regular quality controls by taking the images from the retail scene as ground truth could deliver stable and sustainable IR operations involving millions of images across several years. Therefore, it is important to take into account the approach of the IR vendor's quality control techniques when considering long-term projects with a sheer number of images. 

3) Interactive Field Force Support 

Driving perfect stores and boosting sales is achievable among optimizing field productivity by empowering field personnel to the right execution strategies. Field force plays a vital role in reaching in-store efficiency at most and creates a memorable experience for shoppers. Considering the sheer amount of performance indicators that contribute to scoring the retail execution in grocery, the field team's performance also needed to be tracked. Image recognition-based unbiased data analytics solutions could fairly measure the in-store performance of the team while quantifying the merchandising efforts. These merchandising tasks can range from fixing the shelves to achieving sales targets. 

On the other hand, managers need to monitor the measurement of field scorecards and see how well sales reps are performing. IR could reveal objective references to identify the performance gap that falls behind the expectations. The insight on the performance can be monitored by the HQ as well as the field force itself. Even though IR fairly evaluates the performance based on unbiased data coming from the images taken and analyzed for the store conditions, the right IR vendor should provide an interactive platform for sales reps to make themselves heard and convey their messages if there is something wrong with the evaluation carried out and the scoring. Instead of obliging the team to be satisfied by what they get, it is critical to provide an objections service in FMCG self-auditing settings so that errors can be reported from the field and corrected by the objection operators. When sales reps record their objections, HQ can make systematic fixes according to the correction of the IR provider to uplift sales across business and motivation across teams. When considering the motivation and execution quality, preparing the right medium for the two-way communication between the sales reps and the HQ is critical and needed to be sustained. 

Conclusion 

In Vispera, as a top-leading IR vendor in the retail execution and auditing space, we are committed to providing our customers with the most performant and versatile set of IR capabilities at the best possible quality, detail, and speed. Parallel to that, we also feel responsible for demystifying the magic behind Deep Learning-based visual AI by explaining the fundamental dynamics of retail-specific IR so that our FMCG and retailer customers can make informed decisions.   

We underline the fact that all the human involvement in image recognition operates on state-of-the-art Vispera recognition engines, which record the best-automated performance as corroborated by our customers. The mission of the core R&D team is to minimize these efforts by continually incorporating new technical developments into our pipeline. As experts of computer vision and image recognition, that's actually what excites us the most. In the meantime, as an IR provider, which uniquely understands the quality requirement of retail execution, we don't let our customers suffer from quality issues, however minor they can be.   

To learn more about Vispera’s best-of-breed retail intelligence solutions and experience the magic, talk to us now.  

Authors:

Ceyhun Burak Akgül - Co-Founder and Co-CEO of Vispera

Erdem Yörük - CTO of Vispera