EvoLogics Builds Subsea Perception Models with FiftyOne

Introduction: EvoLogics

EvoLogics designs and manufactures underwater communication systems, positioning networks, and autonomous vehicles for the world's most demanding subsea environments. Based in Germany, the company has built technology for clients across commercial, offshore, defense, and research sectors.

Operating at depths and in conditions unreachable by conventional means, EvoLogics systems locate survivors in search and rescue operations, inspect subsea pipelines and infrastructure, monitor ocean environments, and support naval mine countermeasure activities. Their systems process multimodal data, including sonar data from various types, underwater and surface-mounted cameras, LiDARs, and hydrophones, in real time, during live missions.

“FiftyOne is a vital part of our dataset and model development pipeline. Since incorporating the tool into our workflows, we have better control over what data should be used to train and evaluate our models. As a result, we can deliver dataset updates and fixes into production faster, as our model training pipelines read data directly from FiftyOne” — Dmytro Prylipko, ML Engineer, EvoLogics

Unique model development challenges for the underwater environment

Building perception models for underwater environments comes with unique constraints and challenges. There aren't good pre-training models for sonar imagery since the data is scarce and domain-specific. The same object can appear very differently across missions as the sonar scans the target from multiple passes. And annotation guidelines, such as what to annotate, what to ignore, and which class to assign to are often developed in a data-driven manner, which requires a short feedback loop between annotation and performance analysis.

Before deploying FiftyOne, EvoLogics handled its data and model workflows with custom software. This required dedicated development and maintenance.

Why FiftyOne: A fully programmatic SDK and on-prem deployment

When choosing a tool for EvoLogics' specialized use case, two requirements were non-negotiable: a full programmatic SDK so the team could manage datasets with ETL scripts and incorporate data processing into CI/CD pipelines, and an on-premise deployment to keep sensitive mission data secure and off the public cloud.

FiftyOne met both with its SDK and GUI approach to development and deployment flexibility.

"FiftyOne gives us the ability to quickly observe and explore data in different ways—finding outliers, annotation errors, and an overview of the datasets we possess. And we had the requirement to be able to do it from a script, not just a UI." — Dmytro Prylipko, ML Engineer, EvoLogics

How EvoLogics uses FiftyOne

The Evologics team collects data from various sources such as underwater and surface-mounted cameras, LiDARs, sonars, and hydrophones, for various purposes like underwater object detection, autonomous navigation, and harbor surveillance. An extract-transform-load (ETL) pipeline converts sonar-specific formats into images and matched annotations and manages them in FiftyOne.

Dataset search and query. Since a single sidescan dataset covers multiple use cases, the team uses FiftyOne dataset views and filtering capabilities to slice the right subset for each model rather than duplicating data across separate training sets. This keeps the source of truth clean while giving every project the selection it needs.

Dataset bootstrapping. Using FiftyOne Brain enabled the team to quickly turn a vast amount of automatically generated image patches into a set of unique samples clustered by visual appearance. Furthermore, using seed models pre-trained on a small amount of manually labelled samples, these samples were pre-annotated and sorted out such that minimal human effort was needed to prepare a training-ready dataset.

Visual QA and annotation error detection. Engineers inspect data distributions, run queries against metadata, and catch labeling errors before they reach training pipelines. When a model flags a false positive, the team loads predictions back into FiftyOne to root-cause whether a labeling error or a true model misprediction caused it.

Interactive model evaluation. Teams load model predictions directly into FiftyOne and use the evaluation workflow to perform class-level analysis, identify false positives, false negatives, and edge cases to improve model accuracy and bridge data coverage gaps.

Quick navigation to events of interest. FiftyOne automatically clips long video files into short segments containing only labeled events, making review faster without manually scrubbing through full mission recordings.

Results: More visibility into data helps EvoLogics build accurate perception models

Since using the tool in their complex pipeline, Evologics now has better visibility and control of their data and model workflows.

“FiftyOne is a vital part of our dataset and model development pipeline. Since incorporating the tool into our workflows, we have better control over what data should be used to train and evaluate our models. As a result, we can deliver dataset updates and fixes into production faster, as our model training pipelines read data directly from FiftyOne” — Dmytro Prylipko, ML Engineer, EvoLogics

FiftyOne has also expanded what the team can do with their data. For classification projects bootstrapped from scratch with minimal seed labels, the team uses FiftyOne's embedding extractors and visualization capabilities to build out datasets faster using semi-supervised workflows.

Multimodal workflows

Working with complex multimodal data such as images, video, sensors, lidar, and geolocation, requires a platform that can handle it all in one place. FiftyOne’s grouped datasets make it easy to visualize, query, annotate, and evaluate across every data modality.

Learn about FiftyOne’s multimodal support →

How EvoLogics Powers the World’s Most Challenging Subsea Missions with FiftyOne