One of the buzz words in today’s technology world is Artificial Intelligence which enables machines to think and act without any human intervention. One of the most effective way of achieving this is through Machine Learning which in simplest term is automatically learning and improving through continuous training of dataset and without explicitly being programmed. There are numerous use cases being developed for end users using ever-advancing capabilities of AI & ML and if we talk about FMCG and Beverage industry particularly, image recognition is one very important area to focus on. If we look at the current landscape of these companies in image recognition initiative, we find them in different stages: Adopted, Initiative in Progress, Not Started and the third stage hugely outnumber the other two. So, if you have not initiated this journey yet, let this article guide you how you should go about it.
But before we get into this, let us quickly talk about some of the benefits you may get by adopting image recognition:
- Get full visibility of planogram compliance of your brands
- Understand share of shelf for your products with respect to competition
- Check quantity and quality of merchandising material deployment
- Quantitatively measure performance of your merchandising team
- Get quick actionable insights on merchandising operations
- Significant cost benefits over manual audit
Let’s now discuss how should you plan the entire journey. Image recognition (and for that matter, any AI initiative) is not something which you can decide to go-live on one fine day after a successful user acceptance testing. There will be unforeseen hurdles, setbacks in terms of accuracy, multiple changes in the approach. Hence, the best way is to define the steps and perform them one by one. We have tried to explain the same at a broad level here:
Define the end objective: We discussed many benefits which can be yielded through image recognition if implemented successfully. So, you need to first decide what is your end objective such as whether you want to measure effectiveness of your merchandising operations, whether you want to reduce cost of manual operations over a period of time or may be something else.
Decide on a use case: You should no jump on to solve too many problems at one go. Pick one use case for which success of image recognition can be measured quantitatively. For example, you may want to measure share of shelf of couple of your important brands with respect to few designated competition brands.
Run a POC at a small scale: We have seen many clients asking for desired accuracy level from day 1. It’s almost impossible to predict that because of multiple factors such as type of outlets, quality of images, type of boxes / packs to be identified and many more. So, you must first run a POC with smaller number of images (let’s say 5000) to get an idea of initial accuracy level and identify the challenges.
Work on the challenges: Most of the challenges that we face are solvable – some can be done quickly and some would take time. For example, if you find that many images are taken as long shot or are angled, you must work on improving the camera of the SFA app to help your field team to take images of desired quality (i.e. not too close or too short, as much perpendicular as possible, no blur etc.).
Continuous learning / training: A machine learning algorithm is something that improves on its own with more and more data being fed into it. So, it’s important that all the machine detected images goes through a manual review process and corrections made during manual review are sent back to the image processing engine for continuous improvement.
Set an accuracy cut-off: At this point of time, you should ideally reach to a decent accuracy level where you can decide not to do manual review of all images. Set an accuracy cut-off basis results achieved so far and anything which doesn’t meet that accuracy level should only go for a manual review without feedback being passed on to the AI engine.
Set a target for continuous improvement: The whole idea of machine learning is to improve on accuracy on a continuous basis. So, you must set a target for a periodic (let’s say quarterly) improvement in accuracy level. Monitor the achievements and inaccuracies to find out the root cause and try to solve those one by one to finally reach your desired level of accuracy.
What’s important is to understand it’s a continuous process which will yield desired result in the long term if you stay on course and keep on solving the challenges.