To ensure the Dutch infrastructure remains safe and accessible, with an optimal through-flow of traffic, the Digital Asset Management department of BAM Infra Netherlands (BAM) uses computer vision to analyze existing road infrastructure. To create a scalable and efficient operation, BAM works together with UbiOps, the Dutch platform that enables data scientists to run and manage AI models at scale. Ruben Stam, data scientist at BAM Infra Netherlands, explains why they chose UbiOps and how the platform enables BAM to use state of the art technology with a zero upfront investment.
UbiOps is a platform to run and manage AI workloads at scale, focused explicitly on running on-demand GPU workloads. Data scientists can focus on data science and utilise the advanced underlying GPU infrastructure required for machine learning, without the need to build and maintain the complex computing system themselves.
BAM’s Digital Asset Management is a three-year-old department that focuses on creating a better decision-making environment for corrective, preventive and predictive maintenance. To achieve this, the department combines data, AI and expert domain knowledge.
Applying computer vision for digital road inspection
“Computer vision helps us analyze an infrastructure asset to define and predict its future and necessary maintenance,” says Ruben Stam. “We gather massive amounts of data on roads via a vehicle equipped with cameras and a LiDAR scanner. We process the data with our ADAPT Public Lighting model to document and analyze public lighting”.
The computer vision model registers the location, height, and misalignment of lighting fixtures in the public domain. Using artificial intelligence has some valuable advantages, as Ruben Stam explains: “our accuracy and efficiency are increased compared to traditional methods. In addition, this inspection method creates less disturbance to the public since we don’t have to close lanes. It’s also more cost-efficient; with a one-time location visit, we scan various assets. To ensure all results are reliable, an inspector is always part of the process.”
On-demand and scalable GPU workloads for computer vision
“As a data analytics team, when we develop a computer vision model, we need an underlying computing infrastructure that can process the massive amounts of image and LiDAR data. While we were testing the Public Lighting model at the 32-kilometer long Afsluitdijk on CPUs the inference time proved to be too long – 72 hours to complete one run. We, therefore, started looking for a solution that on the one hand decreases the runtime, and on the other, remains easily scalable; road inspection is a seasonal business, workloads vary per month and per day.”
“With the UbiOps platform, we have found a solution to run our workloads on on-demand GPUs. Being a data analytics team, UbiOps allows us to focus on developing new computer vision models, without the need for extensive and in-depth knowledge of IT infrastructure required to run such models at scale.”
“Moreover, UbiOps gives us the chance to scale up and down across GPUs rapidly and create applications that are ready for peak workloads, while not worrying about GPU availability. Now we can analyze our massive datasets within an adequate turnaround time; on average the runtimes of our models are accelerated by 4.5 times. UbiOps has enabled us to deliver the ADAPT Public Lighting product to our customers more quickly.”
From a successful pilot to a valuable partnership
“When we came in contact with UbiOps, there was no on-demand GPU offering at that time, Ruben Stam explains. “We sat down with the UbiOps team and explained our challenges of on-demand scaling on GPUs. UbiOps gave us the possibility to become part of the GPU on-demand pilot group, where we were able to give input during the development and discuss our requirements. For us that showed the service-oriented approach of UbiOps.”
“The on-demand feature was an extremely convenient solution for us, since we didn’t have to change our data structure and could deliver continuously. UbiOps used our real-life case to test this new feature and used our feedback to further develop this best-in-class solution. We are eager to continue further with UbiOps and have already discussed a new pilot for training AI models.
About Ruben Stam
After working as a researcher and lecturer at the Amsterdam University of Applied Sciences, Ruben Stam became a data scientist at BAM Infra Nederland. As a member of the Digital Asset Management department, he develops state-of-the-art computer vision models and artificial intelligence applications.
