IBAK Computer Vision Case Study

IBAK: Managing Computer Vision Data for Critical Infrastructure

Beneath cities and towns, hundreds of kilometers of sewer pipe quietly keep infrastructure functioning. When something goes wrong — a crack, a junction failure, structural deterioration — the consequences are costly. IBAK is a leading provider of sewer inspection and rehabilitation systems, helping municipalities and infrastructure operators assess and maintain underground pipe networks. IBAK's computer vision team develops AI models that automatically detect, classify, and quantify damage inside sewer pipes using specialized camera systems that operate in demanding underground environments.

IBAK, a customer of Voxel51, diagram of how its sewer inspection and rehabilitation systems work. — Image courtesy of IBAK, which uses FiftyOne to detect damage inside sewer pipes.

From Distributed Tooling to a Central Hub

By the time IBAK's computer vision unit was a few years old, it had reached a point most growing ML teams will recognize. Images and videos lived on shared drives, engineers had built up their own local scripts for the tasks they ran most often, and annotations existed in whichever format the current project needed. Each piece worked well on its own terms. What was missing was a common layer connecting them, a shared way to explore, filter, and validate data that everyone could reach.

"We had good data and good tooling, they just lived in different places. That was the main driver," says Gordon. What the team wanted was a single place where data, annotations, and tooling sat together and were accessible to everyone.

Beyond organization, there was the question of scale. IBAK works with two distinct image modalities: raw fisheye recordings from the inspection cameras, and cylindrical "scan" images — panoramic views stitched together from multiple frames to show an entire pipe segment at once. Across both modalities, the team's primary dataset had grown to over 35 million samples. Filtering that corpus for specific pipe materials, defect types, or metadata criteria, and then exporting clean subsets for training, required exactly the kind of fast, visual, multi-criteria exploration that purpose-built scripts weren't designed for.

Why FiftyOne

IBAK adopted FiftyOne as its central visual data hub, drawn by three things: the ability to handle tens of millions of samples across multiple image types, powerful metadata-driven filtering for building precise training sets, and the flexibility to extend the platform with custom plugins that fit IBAK's domain-specific workflows.

Critically, FiftyOne also integrated cleanly with the team's existing stack — MongoDB for inspection metadata, MLflow for model tracking, and their own internally built annotation tool — without requiring a wholesale infrastructure change.

How IBAK Uses FiftyOne

Filtering and training set construction at 35M scale

The core of IBAK's FiftyOne usage is deceptively straightforward: filter a massive corpus by specific criteria, build a view, and export it for training. In practice, this means combining pipe material type, defect category, image modality, and domain-specific metadata to assemble precisely targeted datasets. What once required writing new scripts for every variation now happens interactively, in minutes.

Annotation QA with a custom plugin

For classification label validation — such as verifying material type annotations across thousands of images — the team built a custom FiftyOne plugin that identifies potentially mislabeled samples, surfaces them visually for review, and pushes flagged IDs directly to their MongoDB instance for downstream correction. A QA workflow that was previously manual and time-intensive is now automated, auditable, and scalable.

Model evaluation built around real-world outcomes

Standard per-image metrics like IOU and mAP are computed in FiftyOne, and they remain essential for model development. But what ultimately counts for IBAK's customers is the finished inspection report: complete and accurate for every pipe segment. To reflect that, the team implemented its own evaluation metric inside FiftyOne, scoring model performance at the report level: did the model correctly identify every defect along the full length of the pipe? This kind of domain-specific extensibility is what lets FiftyOne fit IBAK's workflow rather than the other way around.

Validating SAM Segmentations at Scale

Gordon points to one experience that captures what FiftyOne makes possible. His team had applied SAM (Segment Anything Model) to generate segmentation masks across a large corpus of previously unseen images — a fast way to bootstrap annotations, but one that still requires human validation before any of it enters a training set.

"You usually can't do that fast," Gordon says. "Just validating all those automated segmentations to have them in your real training corpus — I wouldn't know how to do it faster than with this kind of tooling."

With FiftyOne, the team could pull up hundreds of SAM-generated masks at once, visually scan for errors, select the samples that passed review, and move on. What might have taken days of scripting and manual checking became a fast, visual, repeatable process.

Results

One platform, one version of the data. Fifteen engineers now work from the same dataset hub instead of individual local environments. Annotations are centralized, metadata is queryable, and there's a shared language for what "the training data" actually is.

Faster model iteration through better error visibility. By inspecting failed model predictions directly in FiftyOne, the team can quickly distinguish between model errors and annotation errors — a distinction that used to require significant manual investigation, and one that directly impacts how quickly the next training run can be improved.

Annotation format overhead, eliminated. During model experimentation, different frameworks require different annotation formats. FiftyOne handles conversion automatically, removing what Gordon described as "a recurring source of friction" from the development cycle.

Scalable validation for automated annotations. Whether reviewing SAM-generated segmentations or validating classification labels via their custom plugin, IBAK can now process and QA large volumes of automated annotations without writing new tooling for every use case.

"The main benefit is you can see very many things at once. Depending on the kind of data you have, you can spot irregularities pretty easily. And then to export those subsets for training or test — that's the workflow."

— Gordon Böer, Machine Learning Engineer, IBAK

Want to see how FiftyOne can help your team manage visual data at scale? Book a demo or explore the open-source platform.

How IBAK Uses FiftyOne to Manage 35M+ Sewer Inspection Images and Accelerate Model Development