An Executive Guide to a Computer Vision Engagement

July 11, 2019

The use of computer vision in cutting-edge technologies like robotics and autonomous driving is well-known, but it is also rapidly becoming a potent tool to improve day-to-day operations across sectors. 

A recent Harvard Business Review report focused on Computer Vision showed that 82% of survey respondents said they will be assessing, piloting, implementing, or have adopted computer vision capabilities within two years. In the age of smartphones, surveillance cameras, and widespread digitization of records, most companies have access to an abundance of image and video data. This data if utilized smartly can be used to drive better customer experiences, increase productivity, and prevent fraud. Its potential is immense and the market is estimated to reach $25 billion by 2023. 

However the respondents of the same HBR survey also revealed that a lack of expertise is a huge challenge in the implementation of computer vision solutions, with 41% saying that their organizations lacked a visual data strategy. A clear roadmap for early adoption can make all the difference in a competitive environment, and the time is ripe for businesses to assess how they can jump on the computer vision bandwagon. 

Computer Vision Checklist

☑ What problem are we trying to solve?
☑ Do we have the visual data needed to attack this problem?
☑ What are the anticipated costs? Do we need a pilot to find out?
☑ Do we require data labeling?

Executives are becoming increasingly aware that their strategy needs to begin with the sourcing of high-quality data. Not all data is created equal, and the first and most crucial step of the process is a careful examination of useful data that a company can access. iMerit’s CEO Radha Basu in a recent article on the human element in AI says, “Evaluate where you have access to plentiful, diverse and relevant data. The better and more plentiful the data, the more likely you are to succeed at the problem you select.”

Look inwards 

An insurance company has mountains of claims reports that need to be processed before a decision can be reached, and the claims are backlogged. Artificial Intelligence can save significant time and effort in an area like this. Simple smartphone photos of damaged assets like vehicles are submitted as part of the claim. An algorithm to identify vehicular damage and its exact location in these images becomes valuable in underwriting and also predicting damage patterns. Such a model is trained with the input of millions of images, carefully labeled & classified by data annotation experts. Bounding box annotation is a technique that is widely leveraged in this process. 

Where accidents go, doctors and healthcare must follow. As more medical records are digitized, there is increased scope for computer vision to play a big part in streamlining operations within healthcare centers by extracting insights.

The long-standing joke about doctors’ illegible handwriting has given rise to an entire area of healthcare artificial intelligence — automated scanning and extraction of medical information. Blurry scans, doctors’ notes and procedure summaries need to be interpreted, marked, and fed into the engine to create a robust Optical Character Recognition (OCR) model that can make sense of doctorly scrawls. Text analysis and data extraction techniques are employed to separate the useful from the useless. This makes key medical information easily searchable, which is a huge timesaver in a fast-paced hospital environment. 

The retail race 

Retail has always been a visual medium, be it in the form of attractive storefront displays, or in the digital age, an aesthetically curated e-commerce page. It is therefore a natural use case for computer vision, and sure enough, the industry has implemented many innovative applications. Amazon’s frictionless commerce venture Amazon Go is a stunning example that immediately comes to mind, but the retail giant is far from alone in its love affair with computer vision. 

Today, most stores have CCTV cameras, and these are a valuable source of visual data. Tracking each frame for gestures and objects using bounding boxes creates training data that can be harnessed to improve product placement, and optimize customer experience. Gesture recognition in the global retail market is projected to grow by 27.54% from 2018 to 2023. 

These industries are a drop in the ocean of possibilities and applications where visual data processing can be transformative. The Harvard report estimates that the most value will be seen in business units like operations and production, marketing, risk management, supply chain, IT, and sales. Awareness about its potential is slowly but surely penetrating boardrooms across sectors. Businesses can seek inspiration for use cases from the many successful implementations we have seen, and then look inwards for a deeper understanding of their visual data and how it can be leveraged.