Back in 2019, Venture Beat outlined that even with AI’s prowess, only one in ten data science models make it to production. Fast forward to 2022, and Gartner revealed that only 54% of the AI models successfully navigate the pilot-to-production journey. Of course, there’s progress in this space – but it’s not as concrete as the evolution of technology itself. 

The same rings true for computer vision (CV). For sure, the technology has matured, the use cases have proliferated, and more verticals have begun to understand the value of CV and how it can help them improve in real-time. But what about its applicability at scale? It’s the actual deployment post the success of the proof-of-concept (POC) pilot that needs more airtime in strategic discussions.

And that’s precisely where a comprehensive CV-based visual analytics platform comes in. More than just offering a suite of AI-based capabilities, the platform approach to visual analytics can help integrate the CV capabilities with other enterprise systems – making it easier for an enterprise to achieve immense scalability in line with the success that the POC exhibited.

To profoundly understand the rationale for a platform approach, it’s vital for organizations to define a roadmap to bridge the POC-to-production gap. To that end, let’s understand what businesses must do after their computer vision POC succeeds. 

Addressing Training Data Discrepancies

The cornerstone to the success of any computer-vision-based solution is the training data. In essence, it’s the credibility of information that defines how a computer learns and acts on what it sees in the physical space. Now, there can be two scenarios:

• The training data isn’t enough to train a computer vision model or perhaps yield sound outcomes beyond specific use cases.
• There’s massive data obtained from numerous systems which must be filtered and made sense of to train a model and guide it towards scalability.

The latter is more likely as facilities can be equipped with several IoT devices, capturing data across various endpoints and, in the process, generating numerous streams of data. For this, it’s imperative that:

• The data goes through a process of selective filtering – ensuring that the noise or abnormalities are removed and only highly relevant data is retained to train the model.
• The process through which data is filtered shouldn’t fail with the failure of a single step. These need to be connected yet separately operational.

Bringing Everyone Onboard with the Potential Change

Let’s backtrack a moment – wasn’t it the POC that helped businesses realize the potential of computer vision? Surely, yes. And in the process, they’ve identified potential use cases that can fall in the purview of improving, say, object detection, object dimensioning, or worker safety management, among others. 

Now that the solution is set to scale, it’s crucial for the business to ensure that all departments are on board with the change. For example, an enterprise might be deploying a CV solution that’s focused on improving maintenance operations. As such, the maintenance personnel will need to be trained on the new protocol to be followed about, say, proactively acting on the predictions made by the solution about a probable machine failure.

In the same vein, other stakeholders must also be brought into the loop about:

• Why is the solution needed? 
• What are the exact outcomes anticipated?
• What value should it deliver in the short and long run?
• How are teams expected to work in tandem with the solution?
• What should be the scope of this solution based on its utility, purpose, and use case?

All these must be considered so that the organization has a realistic framework within which to deploy the solution at scale – while everyone’s onboard to support the same.

Getting the Infrastructure Right

The success of any CV endeavor demands the right infrastructure. In this context, that means setting up a network of devices, including cameras, sensors, etc., to capture the relevant data. The installations must be in line with the use case that’s being pursued. 

It also means ensuring that the data is captured in a manner that’s conducive to the target use case – making it easier for the CV solution to work. And, of course, there must be a clear plan for data storage, management, and filtering (as discussed above). 

Then, of course, the discussion must look at whether to:

• Utilize the existing infrastructure: This might constrain the scalability of the solution and require more investment to scale up going forward.
• Start from scratch: Setting up devices based on the use cases that the POC has substantiated will enable a more scalable, efficient solution. This would also allow flexibility in deciding upon the computing requirements – whether they should center on a cloud-based, on-premises, or edge-based computing environment.

With all these considerations in place, it’s time to address the real issue — that of making everything work using a platform.

The Key Role of an Intelligent Computer Vision Platform

It’s very clear that an effective, scalable solution can’t be achieved with a series of disconnected components or technologies. There must be a central body that can integrate these different components to simultaneously act on the inputs. This is where an intelligent computer vision-based automation platform like KamerAI proves viable to bridge the POC-to-production gap. It can:

• Act as the brain of the entire end-to-end visual analytics function enabled across the establishment.
• Provide a seamless experience for everyone working with the solution to maintain standardization across all the use cases.
• Bridge the gap between data capture and data interpretation.
• Enable proactive action so that organizations don’t have to be subject to post-mortem analysis.
• Automate the inspection, detection, and alerting functions.
• Integrate with existing systems seamlessly.

The platform approach aligns several computer-vision-enabled applications, including timekeeping, barcode scanning, object dimensioning, optical code recognition, hard hat detection, alignment, unattended object identification, perimeter security, and more.

Interested in learning more? Connect with us today!