VisualCortex Takes Video Analytics to Enterprise with W&B

"The reason we can have such a small team is because W&B helps us coordinate everything. The platform removes a lot of the manual work for us and helps us automate processes so each member can focus on their core role and not worry about the mechanics and governance."
Mike Seddon
Head of Machine Learning and Artificial Intelligence

Video is the New Data Frontier

Of the technologies driving innovation today, AI-powered video analytics is catching on fast. This thriving market is projected to reach 100.42 billion USD by 2032—with no signs of slowing down.

Video analytics isn’t just for security anymore. It’s valuable to almost all industries and can address a variety of use cases, from enriching the shopper experience in retail to optimizing traffic flow for city planners. But to make video data truly actionable is a task often meant for those familiar with machine learning and computer vision initiatives. Well, that’s a thing of the past, as VisualCortex, a cloud-based video intelligence platform, is on a mission to make video analytics accessible and valuable for all video-rich industries and business functions.

From Vehicle Centric to People Centric

While there are many different applications for video analytics, VisualCortex currently focuses on delivering vehicle and people-centric models with a very high level of accuracy. And where can you find a situation where there is a need to analyze foot traffic and monitor parking areas? That’s right, shopping centers.

These once-popular meeting hubs are facing intense pressure to provide an enticing experience that caters to the many demographic segments of the population. VisualCortex helps modern mall operators tap into the data already available to them through existing security and surveillance systems—revealing valuable customer behavior insights and opportunities to streamline operations.

Retail of the future

Car parking issues are one of the main frustrations that make people avoid brick-and-mortar stores. Traditionally, shopping centers rely on an extensive and expensive number of devices—with very tight specifications—to manage parking lots.

What VisualCortex offers is a way for the placement of those devices to be much more flexible, and to do that means focusing on the generalization of their models. “Our automatic number plate recognition (ANPR) solution relies on heavily augmenting images of license plates—things like image distortion or simply changing colors,” said Mike Seddon, Head of Machine Learning and Artificial Intelligence at VisualCortex. “All that work is geared towards training a much more generalized model than you might traditionally get.” Training models in this way allows VisualCortex’s ANPR solution to be quickly deployed in a broad range of environments and begin generating highly accurate results from the outset.
The ANPR solution mentioned by Mike has a hierarchical set of models composed of three phases: vehicle detection (VD), license plate detection (LPD), and license plate recognition (LPR).
To start, as the name implies, VD detects if a vehicle is present. This helps with tracking a number of useful things simultaneously—like peak times and flow patterns in the lot. Then, LPD aims to identify and localize a license plate by generating a bounding box around the plate. Finally, LPR is where character segmentation and character recognition take place to read the specific letters and numbers on the plate. Each of these processing steps plays an important role to achieve high accuracy in real-time. And VisualCortex has clearly knocked it out of the park.
“For one of our clients, we’re seeing a 10% increase in accuracy with zero downtime,” said Patrick Elliott, Co-Founder and CEO at VisualCortex.
Moving to the heart of shopping centers, the models deployed by VisualCortex also help improve a range of practices for in-store retail. Layout optimization is one of those revenue-boosting opportunities. Using a similar object detection model used for parking management, VisualCortex can provide retailers with analytics that track customer movement around the store. Questions like how long is somebody standing in front of a display or which part of the store is lacking staff can easily be answered.
The best part? These video analytics are accessible via a web-based interface that empowers both technical users like software developers and data scientists, as well as business users—from marketing executives to business analytics—to quickly and easily produce actionable data streams about any aspect of their operations.
“We really try to take the science and fear out of computer vision technology and focus on the end-users of the business world,” said Patrick.

Driving efficiency and productivity

Although VisualCortex owns the end-to-end training from data collection and labeling, to training and deployment, the bulk of their ML work is made possible by a small team of three developers. To ensure they have the tools to move fast and keep productivity levels high, they lean on W&B to streamline their ML workflow.
“The reason we can have such a small team is because W&B helps us coordinate everything,” explained Patrick. “The platform removes a lot of the manual work for us and helps us automate processes so each member can focus on their core role and not worry about the mechanics and governance.”
In particular, W&B Sweeps accelerated the rate of finding the most optimal data augmentations. By programmatically building their augmentation into Sweeps, VisualCortex can efficiently sample the space of augmentation combinations to push forward the best model. An additional bonus for the team was seeing these attributes displayed through the W&B hyperparameter importance panel, helping them uncover valuable patterns and relationships.
As VisualCortex gears up to start a new round of model training, there is a greater need to effectively orchestrate a disparate set of compute nodes. To address that, the team plans on leveraging W&B Launch to enable easy access to compute resources and dramatically scale their training workloads.
“We’ve got Kubernetes running on our end, and having that ability to create queues to each cluster and activate them is exactly what we want to do,” said Mike.

The Rise of Video in the Enterprise

The potential of video analytics is growing and proving to be undeniably diverse. Yet, making video data actionable is challenging, especially at the enterprise level. As a platform designed to tap into this unprecedented data mining opportunity, VisualCortex is making it possible to efficiently generate reliable insights—solving real-world commercial challenges at scale.
To continue the vision of scaling and productionizing computer vision technology, having an end-to-end MLOps platform is key. With W&B, VisualCortex can simplify the management and execution of ML workflows, automate the search for optimal augmentations, explore interactive visualizations of model behavior, and above all, catapult their productivity to the next level