When asked to list things commonly found in a park, we would likely name items like grass, benches, and trees without much difficulty. However, the process behind this seemingly simple task is complex and involves multiple interactions with the external world. Human vision goes beyond the physical process of seeing with our eyes and encompasses our understanding of concepts and experiences. Until the advent of computer vision, machines had limited capacity for independent thought. Computer vision is the technology that seeks to replicate human vision and enable computers to identify and process information in the same way humans do. It is not a recent discovery but the adoption of the technology has particularly seen a spike last few years.

What makes computer vision so special? By simulating certain complexities of the human visual system, it enables computers to recognize and analyze objects in images and videos in a manner that resembles human cognition. With its ability to detect early indicators of rising demand, computer vision can alert managers and other supply chain participants when it's time to make additional product purchases.

Importance of Computer Vision

In our current era, AI has become ubiquitous, and its applications have extended to a diverse range of industries, ranging from transportation to finance. Amid the various fields within AI, computer vision stands out as particularly powerful, captivating, and surreptitiously pervasive.

However, most proof-of-concept (PoC) experiments in computer vision have not been applied in the real world due to the PoC-to-Production gap. The challenges in computer vision include the need for massive computation for tasks such as facial recognition and autonomous driving, and difficulties in representing human experience in a computer system.

Computer vision product development typically starts with a proof of concept (PoC), which evaluates the algorithm on a sample dataset to validate its feasibility. However, it is equally important to consider how the algorithms will work on real-life data in the production environment. Unlike other AI solutions, computer vision solutions require upfront attention to factors such as camera hardware, compute, network bandwidth, and data pipelines in production. Neglecting these factors in the early stages of the project can affect the productization of the journey.

Below are the key factors influencing the productisation efforts:

• Infrastructural ecosystem: A typical computer vision solution can use various input sources like security cameras, custom camera hardware, mobile phone images, drones, or digital images. Reliable input sources meeting algorithm requirements are important, including image quality, camera positioning, field of view, and frame per second. Green field projects are easier since new cameras can be installed, while brown field projects must work with existing cameras and video sources. It's advisable to abstract underlying camera hardware through application layers for productisation.

Compute requirements and strategy for running algorithms should be part of the proof-of-concept evaluation, with a general recommendation of running the algorithm closer to the source for real-time processing with a large volume of streaming data. HW investment/cost should be validated upfront against the business returns to avoid surprises during production transition, as computer vision solutions are compute intensive.

• Management of datasets: To build a strong computer vision model, it's important to have a large and diverse dataset, which needs to be filtered and labelled in a timely manner. A computer vision data pipeline typically includes multiple critical stages, such as data ingestion, pre-processing, filtering, curation, and model training. The pipeline's processes must be segregated to allow each stage to be activated independently if one step fails, and checkpoints can help avoid repeating the entire cycle in the event of failure. Thorough documentation of the data pipeline can help maintain the system even after a member leaves the company and enable new members to redesign things as necessary.

To create a robust data pipeline, it’s essential to filter data, implement checkpoints, and thoroughly document the process. Tools like active learning, model-assisted data labelling, and dataset meta-management can simplify decision-making and ensure data quality, integrity, and governance. This investment in data analytics pipelines can lead to smarter decisions and simpler operations.

• Creating a feedback loop: The specific activities for deploying a model depend on how the business intends to use it. After deploying the solution, a feedback loop is necessary to ensure that the model performs well on real-world data. Developing tools to monitor model performance in real-time and identify accuracy issues can help take corrective action. Another option is to engage end-users to provide real-time feedback through interactive tools as part of the human-in-loop feedback. These strategies can help maintain and improve the model's effectiveness.
• Are we using it to solve the right problem?: Computer vision technology deals with images and videos, making it easy to comprehend and appreciate. This generates a lot of interest and potential in the field. However, it's crucial to translate this enthusiasm into a well-defined business problem; otherwise, we might end up creating an impressive technological solution with no business impact. We might get carried away by the potential of computer vision and try to use it to solve every problem, even when the problem can be solved effectively without it. It's important to ensure that we are solving the correct problem using the appropriate technology.

In conclusion, using computer vision involves being responsible as well as ethical. To ensure the privacy of personal identifiable information (PIA) in computer vision solutions, it is important to have strong governance policies and create awareness among developers on ethical and responsible AI. It is also essential to solve the right data problem and avoid compromising privacy when building solutions. Regular monitoring of model performance and a continuous learning loop can combat data drift and stale models. By streamlining data and model pipelines, avoiding common mistakes, and establishing automated operations, effective and scalable computer vision models can be created.

• Photo by Vishal Bansal on Unsplash