5 Things to Consider Before Deploying Image Recognition in Traditional Trade

traditional trade

There is considerable literature available on improving retail execution in modern and organized trade channels. However, we do not have enough information available on tools, solutions, and best practices, when it comes to optimizing retail execution using Image Recognition in traditional (or general) trade.

This information asymmetry creates a lot of confusion and misinformation around emerging technologies such as Image Recognition and its uses and applications in traditional trade. In this article, we will discuss the unique challenges and potential solutions to successfully roll out Image Recognition in traditional trade stores. 

Before we start the discussion, let’s take a look at some of the facts that make us qualified to write this article. At ParallelDots, we built our first Image Recognition solution for a large Indian tobacco firm to help them track their retail execution. The scale and technical complexity of the project made it one of the most challenging AI project for us despite our five years of experience in solving different AI use-cases in multiple industry verticals.

With no access to labeled data and frequent design changes in point of sale materials (POSM), our data science team created an image recognition algorithm without any training data. Furthermore, they ensured the algorithm is ready to recognize a new POSM as soon as it is launched in the market. 

Despite the aforementioned challenges, our team built and delivered an Image Recognition solution to the client that is currently operating at 99.3% accuracy, processing 50,000 images per day to identify 300 different types of POSMs. All the photos belong to the very small traditional trade stores in India, popularly known as paan-beedi shops. One such photo of a paan-beedi shop is shown below:

traditional trade
Source: Justdial

After our first project in traditional trade, we went on to deliver more projects covering all different type of POSMs and solving KPIs such as:

  1. POSM Presence
  2. Cooler Purity (“detecting the presence of competitors in coolers”)
  3. Distribution and Stockouts
  4. Merchandising Execution

There were many challenges that we overcame while solving the above KPIs so we thought that CPG companies looking to deploy Image Recognition can benefit from our learnings. We have listed the top 5 challenges that we have to overcome below:

Lack of Uniformity in Traditional Trades

Unlike modern trade, traditional trade stores have very different layouts and limited space. Hence, merchandisers have to get creative to fix branded assets in the retailer allocated spaces. Let’s discuss this challenge in detail with respect to our recent deployment for a large personal care brand here.

This poses a challenge to Image Recognition algorithms though. An image recognition algorithm trained to identify such sections as a whole can get confused if the asset’s orientation is changed to fit the limited space present in the store. An example is shown below:

traditional trade
Window Asset on left is long and less wide as compared to the asset on right

In Modern Trades, products are always placed neatly on the shelf. Unlike the former, here we need a robust algorithm to account for such changes and not penalize the merchandiser wrongly. 

Low Light Conditions in Traditional Trade Stores

traditional trade
Low light conditions in Traditional Trade make it difficult to analyze photos

One of the most common questions that our customers ask is, whether your algorithm can detect products in low light conditions?

Our reply to this question is, if a human eye can discern the products clearly in a photo, AI can also detect it. However, sometimes images are really bad even for a human eye so it is important to detect such conditions and ensure to turn on the flash for a better quality picture.

This ability to improve photo quality while capturing is very important for the successful deployment of IR in traditional trade. In one of our previous posts, we have covered in detail the unique features of ShelfWatch that help in improving the photo quality at the point of capture.

High Prevalence of Irrelevant/Spam Photos

Almost every project we did last year had close to 40% of the images that were irrelevant. The most common reasons for such a high percentage of irrelevant images are listed below:

  • Blurry Images
  • Too much reflection due to product packaging and/or cooler glass
  • Images containing something else but the asset photo
  • Images were taken too close or too far from the asset

The manual analysis of images is unreliable and expensive because of the sheer volume of images coming from different sources. 

Due to a lack of technological solutions, brands either do not cover all the stores or simply give up on the project after piloting with few vendors and/or solutions. 

We believe that before actual image processing, brands should look for potential solutions that can help them reduce the number of irrelevant or poor quality images. 

ShelfWatch, for instance, automatically detects blurry and irrelevant images using project-specific heuristics. There is also a dashboard for sales supervisors to monitor the performance of their reps based on the quality of photos captured by them.

traditional trade
Ability to detect blurry images at the point of capture can improve data quality

High Cost of Deployment

traditional trade
The high cost of deployment of Image Recognition in Traditional Trade can be a barrier to adoption

Due to the non-uniform and sometimes, chaotic nature of GT stores, there are very few (perhaps zero) standard solutions available in the market for Image Recognition. Therefore, even the most sophisticated technology vendor who has not worked on traditional trade images before would need to do heavy customization in their algorithms which may lead to longer times and higher costs for them

The most important metric for a CPG company to consider is the ability of a vendor to scale its pilot solution without significantly increasing the setup or image processing costs. 

Dense Product Stacking

Another important factor to consider is the dense stacking of products which occupy a lot of space, thus exposing a very small part of the packaging. 

Some product categories such as diapers, chips, breakfast cereals, toothpaste, etc. tend to take too much space if stacked with their front packaging visible. Many traditional trade store owners will stack them such that only the side facings are visible. This makes it very difficult for Image Recognition algorithms to identify the product. An example is shown below

traditional trade
Dense packing of SKUs in Traditional Trade
Source: Justdial

Such product facings require unique customizations to the algorithm and only an already experienced vendor in these categories can build such capability.

Traditional Trades form a significant sales channel in many emerging economies. In spite of the initial challenges, rewards are huge for the brands investing resources to make every traditional trade store a “Perfect Store”.

Leave a Reply